Dive Into Deep Learning: Aston Zhang, Zachary C. Lipton, Mu Li, and Alexander J. Smola

Dive into Deep Learning
Release 0.14.3
Aston Zhang, Zachary C. Lipton, Mu Li, and Alexander J. Smola
Aug 11, 2020

Contents
Preface 1
Installation 9
Notation 13
1 Introduction 17
1.1 A Motivating Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.2 The Key Components: Data, Models, and Algorithms . . . . . . . . . . . . . . . . 20
1.3 Kinds of Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.4 Roots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
1.5 The Road to Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
1.6 Success Stories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2 Preliminaries 43
2.1 Data Manipulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.1.1 Getting Started . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.1.2 Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.1.3 Broadcasting Mechanism . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.1.4 Indexing and Slicing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.1.5 Saving Memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.1.6 Conversion to Other Python Objects . . . . . . . . . . . . . . . . . . . . . 50
2.2 Data Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
2.2.1 Reading the Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
2.2.2 Handling Missing Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.2.3 Conversion to the Tensor Format . . . . . . . . . . . . . . . . . . . . . . . 53
2.3 Linear Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
2.3.1 Scalars . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
2.3.2 Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
2.3.3 Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.3.4 Tensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
2.3.5 Basic Properties of Tensor Arithmetic . . . . . . . . . . . . . . . . . . . . 58
2.3.6 Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
2.3.7 Dot Products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.3.8 Matrix-Vector Products . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.3.9 Matrix-Matrix Multiplication . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.3.10 Norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
2.3.11 More on Linear Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
2.4 Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
2.4.1 Derivatives and Differentiation . . . . . . . . . . . . . . . . . . . . . . . . 67
2.4.2 Partial Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
i
2.4.3 Gradients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
2.4.4 Chain Rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
2.5 Automatic Differentiation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
2.5.1 A Simple Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
2.5.2 Backward for Non-Scalar Variables . . . . . . . . . . . . . . . . . . . . . . 74
2.5.3 Detaching Computation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
2.5.4 Computing the Gradient of Python Control Flow . . . . . . . . . . . . . . 75
2.6 Probability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
2.6.1 Basic Probability Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
2.6.2 Dealing with Multiple Random Variables . . . . . . . . . . . . . . . . . . 81
2.6.3 Expectation and Variance . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
2.7 Documentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
2.7.1 Finding All the Functions and Classes in a Module . . . . . . . . . . . . . 85
2.7.2 Finding the Usage of Specific Functions and Classes . . . . . . . . . . . . 85
3 Linear Neural Networks 89

3.1 Linear Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
3.1.1 Basic Elements of Linear Regression . . . . . . . . . . . . . . . . . . . . . 89
3.1.2 Vectorization for Speed . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
3.1.3 The Normal Distribution and Squared Loss . . . . . . . . . . . . . . . . . 95
3.1.4 From Linear Regression to Deep Networks . . . . . . . . . . . . . . . . . 96
3.2 Linear Regression Implementation from Scratch . . . . . . . . . . . . . . . . . . 99
3.2.1 Generating the Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
3.2.3 Initializing Model Parameters . . . . . . . . . . . . . . . . . . . . . . . . 101
3.2.4 Defining the Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
3.2.5 Defining the Loss Function . . . . . . . . . . . . . . . . . . . . . . . . . . 102
3.2.6 Defining the Optimization Algorithm . . . . . . . . . . . . . . . . . . . . . 102
3.2.7 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
3.3 Concise Implementation of Linear Regression . . . . . . . . . . . . . . . . . . . . 105
3.3.1 Generating the Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
3.3.6 Defining the Optimization Algorithm . . . . . . . . . . . . . . . . . . . . . 107
3.3.7 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
3.4 Softmax Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
3.4.1 Classification Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
3.4.2 Network Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
3.4.3 Softmax Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
3.4.4 Vectorization for Minibatches . . . . . . . . . . . . . . . . . . . . . . . . 112
3.4.5 Loss Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
3.4.6 Information Theory Basics . . . . . . . . . . . . . . . . . . . . . . . . . . 114
3.4.7 Model Prediction and Evaluation . . . . . . . . . . . . . . . . . . . . . . . 115
3.5 The Image Classification Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
3.5.2 Reading a Minibatch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
3.5.3 Putting All Things Together . . . . . . . . . . . . . . . . . . . . . . . . . . 118
3.6 Implementation of Softmax Regression from Scratch . . . . . . . . . . . . . . . . 119
ii
3.6.2 Defining the Softmax Operation . . . . . . . . . . . . . . . . . . . . . . . 120
3.6.5 Classification Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
3.6.6 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
3.6.7 Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
3.7 Concise Implementation of Softmax Regression . . . . . . . . . . . . . . . . . . . 126
3.7.2 Softmax Implementation Revisited . . . . . . . . . . . . . . . . . . . . . . 127
3.7.3 Optimization Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
3.7.4 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
4 Multilayer Perceptrons 131

4.1 Multilayer Perceptrons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
4.1.1 Hidden Layers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
4.1.2 Activation Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
4.2 Implementation of Multilayer Perceptrons from Scratch . . . . . . . . . . . . . . 139
4.2.2 Activation Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
4.2.3 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
4.2.4 Loss Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
4.2.5 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
4.3 Concise Implementation of Multilayer Perceptrons . . . . . . . . . . . . . . . . . 142
4.3.1 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
4.4 Model Selection, Underfitting, and Overfitting . . . . . . . . . . . . . . . . . . . . 144
4.4.1 Training Error and Generalization Error . . . . . . . . . . . . . . . . . . . 144
4.4.2 Model Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
4.4.3 Underfitting or Overfitting? . . . . . . . . . . . . . . . . . . . . . . . . . . 148
4.4.4 Polynomial Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
4.5 Weight Decay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
4.5.1 Norms and Weight Decay . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
4.5.2 High-Dimensional Linear Regression . . . . . . . . . . . . . . . . . . . . 156
4.5.3 Implementation from Scratch . . . . . . . . . . . . . . . . . . . . . . . . 156
4.5.4 Concise Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
4.6 Dropout . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
4.6.1 Overfitting Revisited . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
4.6.2 Robustness through Perturbations . . . . . . . . . . . . . . . . . . . . . . 162
4.6.3 Dropout in Practice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
4.7 Forward Propagation, Backward Propagation, and Computational Graphs . . . . . 167
4.7.1 Forward Propagation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
4.7.2 Computational Graph of Forward Propagation . . . . . . . . . . . . . . . . 168
4.7.3 Backpropagation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
4.7.4 Training Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
4.8 Numerical Stability and Initialization . . . . . . . . . . . . . . . . . . . . . . . . . 171
4.8.1 Vanishing and Exploding Gradients . . . . . . . . . . . . . . . . . . . . . 172
4.8.2 Parameter Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
4.9 Environment and Distribution Shift . . . . . . . . . . . . . . . . . . . . . . . . . . 176
4.9.1 Types of Distribution Shift . . . . . . . . . . . . . . . . . . . . . . . . . . 177
4.9.2 Examples of Distribution Shift . . . . . . . . . . . . . . . . . . . . . . . . 179
iii
4.9.3 Correction of Distribution Shift . . . . . . . . . . . . . . . . . . . . . . . . 181
4.9.4 A Taxonomy of Learning Problems . . . . . . . . . . . . . . . . . . . . . . 184
4.9.5 Fairness, Accountability, and Transparency in Machine Learning . . . . . 186
4.10 Predicting House Prices on Kaggle . . . . . . . . . . . . . . . . . . . . . . . . . . 187
4.10.1 Downloading and Caching Datasets . . . . . . . . . . . . . . . . . . . . . 187
4.10.2 Kaggle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
4.10.3 Accessing and Reading the Dataset . . . . . . . . . . . . . . . . . . . . . . 190
4.10.4 Data Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
4.10.5 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
4.10.6 K-Fold Cross-Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
4.10.7 Model Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
4.10.8 Submitting Predictions on Kaggle . . . . . . . . . . . . . . . . . . . . . . 195
5 Deep Learning Computation 199

5.1 Layers and Blocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
5.1.1 A Custom Block . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
5.1.2 The Sequential Block . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
5.1.3 Executing Code in the Forward Propagation Function . . . . . . . . . . . . 204
5.1.4 Compilation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
5.2 Parameter Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
5.2.1 Parameter Access . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
5.2.2 Parameter Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
5.2.3 Tied Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
5.3 Deferred Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
5.3.1 Instantiating a Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
5.4 Custom Layers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
5.4.1 Layers without Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . 216
5.4.2 Layers with Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
5.5 File I/O . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
5.5.1 Loading and Saving Tensors . . . . . . . . . . . . . . . . . . . . . . . . . 218
5.5.2 Loading and Saving Model Parameters . . . . . . . . . . . . . . . . . . . . 219
5.6 GPUs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
5.6.1 Computing Devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
5.6.2 Tensors and GPUs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
5.6.3 Neural Networks and GPUs . . . . . . . . . . . . . . . . . . . . . . . . . . 225
6 Convolutional Neural Networks 227

6.1 From Fully-Connected Layers to Convolutions . . . . . . . . . . . . . . . . . . . . 228
6.1.1 Invariance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
6.1.2 Constraining the MLP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
6.1.3 Convolutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
6.1.4 “Whereʼs Waldo” Revisited . . . . . . . . . . . . . . . . . . . . . . . . . . 231
6.2 Convolutions for Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
6.2.1 The Cross-Correlation Operation . . . . . . . . . . . . . . . . . . . . . . . 233
6.2.2 Convolutional Layers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
6.2.3 Object Edge Detection in Images . . . . . . . . . . . . . . . . . . . . . . . 235
6.2.4 Learning a Kernel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
6.2.5 Cross-Correlation and Convolution . . . . . . . . . . . . . . . . . . . . . . 237
6.2.6 Feature Map and Receptive Field . . . . . . . . . . . . . . . . . . . . . . . 238
6.3 Padding and Stride . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
6.3.1 Padding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
iv
6.3.2 Stride . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
6.4 Multiple Input and Multiple Output Channels . . . . . . . . . . . . . . . . . . . . 243
6.4.1 Multiple Input Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
6.4.2 Multiple Output Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
6.4.3 1 × 1 Convolutional Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
6.5 Pooling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
6.5.1 Maximum Pooling and Average Pooling . . . . . . . . . . . . . . . . . . . 247
6.5.2 Padding and Stride . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
6.5.3 Multiple Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
6.6 Convolutional Neural Networks (LeNet) . . . . . . . . . . . . . . . . . . . . . . . 251
6.6.1 LeNet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
6.6.2 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
7 Modern Convolutional Neural Networks 257

7.1 Deep Convolutional Neural Networks (AlexNet) . . . . . . . . . . . . . . . . . . . 257
7.1.1 Learning Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
7.1.2 AlexNet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
7.1.4 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
7.2 Networks Using Blocks (VGG) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
7.2.1 VGG Blocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
7.2.2 VGG Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
7.2.3 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
7.3 Network in Network (NiN) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
7.3.1 NiN Blocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
7.3.2 NiN Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
7.3.3 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
7.4 Networks with Parallel Concatenations (GoogLeNet) . . . . . . . . . . . . . . . . 274
7.4.1 Inception Blocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
7.4.2 GoogLeNet Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
7.4.3 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278
7.5 Batch Normalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
7.5.1 Training Deep Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
7.5.2 Batch Normalization Layers . . . . . . . . . . . . . . . . . . . . . . . . . . 281
7.5.4 Applying Batch Normalization in LeNet . . . . . . . . . . . . . . . . . . . 283
7.5.6 Controversy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 286
7.6 Residual Networks (ResNet) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
7.6.1 Function Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
7.6.2 Residual Blocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
7.6.3 ResNet Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
7.6.4 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
7.7 Densely Connected Networks (DenseNet) . . . . . . . . . . . . . . . . . . . . . . 294
7.7.1 From ResNet to DenseNet . . . . . . . . . . . . . . . . . . . . . . . . . . . 294
7.7.2 Dense Blocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
7.7.3 Transition Layers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296
7.7.4 DenseNet Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
7.7.5 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
8 Recurrent Neural Networks 299
v
8.1 Sequence Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
8.1.1 Statistical Tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300
8.1.2 A Toy Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
8.1.3 Predictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
8.2 Text Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
8.2.2 Tokenization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
8.2.3 Vocabulary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
8.3 Language Models and the Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . 310
8.3.1 Estimating a Language Model . . . . . . . . . . . . . . . . . . . . . . . . . 311
8.3.2 Markov Models and n-grams . . . . . . . . . . . . . . . . . . . . . . . . . 312
8.3.3 Natural Language Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . 312
8.3.4 Training Data Preparation . . . . . . . . . . . . . . . . . . . . . . . . . . 315
8.4 Recurrent Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 318
8.4.1 Recurrent Networks Without Hidden States . . . . . . . . . . . . . . . . . 319
8.4.2 Recurrent Networks with Hidden States . . . . . . . . . . . . . . . . . . . 320
8.4.3 Steps in a Language Model . . . . . . . . . . . . . . . . . . . . . . . . . . 321
8.4.4 Perplexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
8.5 Implementation of Recurrent Neural Networks from Scratch . . . . . . . . . . . . 323
8.5.1 One-hot Encoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
8.5.2 Initializing the Model Parameters . . . . . . . . . . . . . . . . . . . . . . 324
8.5.3 RNN Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
8.5.4 Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326
8.5.5 Gradient Clipping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326
8.5.6 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327
8.6 Concise Implementation of Recurrent Neural Networks . . . . . . . . . . . . . . . 331
8.6.2 Training and Predicting . . . . . . . . . . . . . . . . . . . . . . . . . . . . 332
8.7 Backpropagation Through Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
8.7.1 A Simplified Recurrent Network . . . . . . . . . . . . . . . . . . . . . . . 334
8.7.2 The Computational Graph . . . . . . . . . . . . . . . . . . . . . . . . . . . 336
8.7.3 BPTT in Detail . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
9 Modern Recurrent Neural Networks 339

9.1 Gated Recurrent Units (GRU) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 339
9.1.1 Gating the Hidden State . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340
9.2 Long Short Term Memory (LSTM) . . . . . . . . . . . . . . . . . . . . . . . . . . . 346
9.2.1 Gated Memory Cells . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
9.3 Deep Recurrent Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
9.3.1 Functional Dependencies . . . . . . . . . . . . . . . . . . . . . . . . . . . 354
9.3.3 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355
9.4 Bidirectional Recurrent Neural Networks . . . . . . . . . . . . . . . . . . . . . . . 356
9.4.1 Dynamic Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357
9.4.2 Bidirectional Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359
9.5 Machine Translation and the Dataset . . . . . . . . . . . . . . . . . . . . . . . . . 362
vi
9.5.1 Reading and Preprocessing the Dataset . . . . . . . . . . . . . . . . . . . 363
9.5.2 Tokenization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 364
9.5.3 Vocabulary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
9.5.4 Loading the Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
9.6 Encoder-Decoder Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367
9.6.1 Encoder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367
9.6.2 Decoder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367
9.6.3 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 368
9.7 Sequence to Sequence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369
9.7.1 Encoder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370
9.7.2 Decoder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 371
9.7.3 The Loss Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372
9.7.4 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373
9.7.5 Predicting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
9.8 Beam Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 376
9.8.1 Greedy Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 376
9.8.2 Exhaustive Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 378
9.8.3 Beam Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 378
10 Attention Mechanisms 381

10.1 Attention Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381
10.1.1 Dot Product Attention . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384
10.1.2 MLP Attention . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
10.2 Sequence to Sequence with Attention Mechanisms . . . . . . . . . . . . . . . . . 386
10.2.1 Decoder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 387
10.2.2 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389
10.3 Transformer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 390
10.3.1 Multi-Head Attention . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392
10.3.2 Position-wise Feed-Forward Networks . . . . . . . . . . . . . . . . . . . . 394
10.3.3 Add and Norm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 395
10.3.4 Positional Encoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 396
10.3.5 Encoder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 398
10.3.6 Decoder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 399
10.3.7 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 401
11 Optimization Algorithms 405

11.1 Optimization and Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . 405
11.1.1 Optimization and Estimation . . . . . . . . . . . . . . . . . . . . . . . . . 406
11.1.2 Optimization Challenges in Deep Learning . . . . . . . . . . . . . . . . . 407
11.2 Convexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411
11.2.1 Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411
11.2.2 Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 414
11.2.3 Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 417
11.3 Gradient Descent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 420
11.3.1 Gradient Descent in One Dimension . . . . . . . . . . . . . . . . . . . . . 420
11.3.2 Multivariate Gradient Descent . . . . . . . . . . . . . . . . . . . . . . . . 423
11.3.3 Adaptive Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425
11.4 Stochastic Gradient Descent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429
11.4.1 Stochastic Gradient Updates . . . . . . . . . . . . . . . . . . . . . . . . . 429
11.4.2 Dynamic Learning Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . 431
vii
11.4.3 Convergence Analysis for Convex Objectives . . . . . . . . . . . . . . . . 433
11.4.4 Stochastic Gradients and Finite Samples . . . . . . . . . . . . . . . . . . . 434
11.5 Minibatch Stochastic Gradient Descent . . . . . . . . . . . . . . . . . . . . . . . . 436
11.5.1 Vectorization and Caches . . . . . . . . . . . . . . . . . . . . . . . . . . . 436
11.5.2 Minibatches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 438
11.6 Momentum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
11.6.1 Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
11.6.2 Practical Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 449
11.6.3 Theoretical Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 452
11.7 Adagrad . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 454
11.7.1 Sparse Features and Learning Rates . . . . . . . . . . . . . . . . . . . . . 455
11.7.2 Preconditioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 455
11.7.3 The Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
11.8 RMSProp . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461
11.8.1 The Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461
11.9 Adadelta . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465
11.9.1 The Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465
11.9.2 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465
11.10 Adam . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467
11.10.1 The Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 468
11.10.2 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 469
11.10.3 Yogi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 470
11.11 Learning Rate Scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 472
11.11.1 Toy Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 472
11.11.2 Schedulers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 474
11.11.3 Policies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 476
12 Computational Performance 483

12.1 Compilers and Interpreters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 483
12.1.1 Symbolic Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . 484
12.1.2 Hybrid Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 485
12.1.3 HybridSequential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 486
12.2 Asynchronous Computation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 490
12.2.1 Asynchrony via Backend . . . . . . . . . . . . . . . . . . . . . . . . . . . 490
12.2.2 Barriers and Blockers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 492
12.2.3 Improving Computation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493
12.2.4 Improving Memory Footprint . . . . . . . . . . . . . . . . . . . . . . . . . 494
12.3 Automatic Parallelism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 497
12.3.1 Parallel Computation on CPUs and GPUs . . . . . . . . . . . . . . . . . . . 497
12.3.2 Parallel Computation and Communication . . . . . . . . . . . . . . . . . 498
12.4 Hardware . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 500
12.4.1 Computers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 501
12.4.2 Memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 502
12.4.3 Storage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 503
viii
12.4.4 CPUs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 504
12.4.5 GPUs and other Accelerators . . . . . . . . . . . . . . . . . . . . . . . . . 507
12.4.6 Networks and Buses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 510
12.4.7 More Latency Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 511
12.5 Training on Multiple GPUs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513
12.5.1 Splitting the Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 514
12.5.2 Data Parallelism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 515
12.5.3 A Toy Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517
12.5.4 Data Synchronization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 518
12.5.5 Distributing Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 519
12.5.6 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 520
12.5.7 Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 521
12.6 Concise Implementation for Multiple GPUs . . . . . . . . . . . . . . . . . . . . . 522
12.6.1 A Toy Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523
12.6.2 Parameter Initialization and Logistics . . . . . . . . . . . . . . . . . . . . 523
12.6.3 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 525
12.6.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 526
12.7 Parameter Servers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 527
12.7.1 Data Parallel Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 528
12.7.2 Ring Synchronization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 530
12.7.3 Multi-Machine Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . 533
12.7.4 (key,value) Stores . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 535
13 Computer Vision 537

13.1 Image Augmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 537
13.1.1 Common Image Augmentation Method . . . . . . . . . . . . . . . . . . . 538
13.1.2 Using an Image Augmentation Training Model . . . . . . . . . . . . . . . 542
13.2 Fine-Tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 546
13.2.1 Hot Dog Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 547
13.3 Object Detection and Bounding Boxes . . . . . . . . . . . . . . . . . . . . . . . . 552
13.3.1 Bounding Box . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 553
13.4 Anchor Boxes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 554
13.4.1 Generating Multiple Anchor Boxes . . . . . . . . . . . . . . . . . . . . . . 555
13.4.2 Intersection over Union . . . . . . . . . . . . . . . . . . . . . . . . . . . . 557
13.4.3 Labeling Training Set Anchor Boxes . . . . . . . . . . . . . . . . . . . . . 557
13.4.4 Bounding Boxes for Prediction . . . . . . . . . . . . . . . . . . . . . . . . 561
13.5 Multiscale Object Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 564
13.6 The Object Detection Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 567
13.6.1 Downloading the Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . 567
13.6.3 Demonstration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 569
13.7 Single Shot Multibox Detection (SSD) . . . . . . . . . . . . . . . . . . . . . . . . . 570
13.7.1 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 570
13.7.2 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 575
13.7.3 Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 578
13.8 Region-based CNNs (R-CNNs) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 581
13.8.1 R-CNNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 581
13.8.2 Fast R-CNN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 582
13.8.3 Faster R-CNN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 585
13.8.4 Mask R-CNN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 586
13.9 Semantic Segmentation and the Dataset . . . . . . . . . . . . . . . . . . . . . . . 587
ix
13.9.1 Image Segmentation and Instance Segmentation . . . . . . . . . . . . . . 587
13.9.2 The Pascal VOC2012 Semantic Segmentation Dataset . . . . . . . . . . . . 588
13.10 Transposed Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 593
13.10.1 Basic 2D Transposed Convolution . . . . . . . . . . . . . . . . . . . . . . 593
13.10.2 Padding, Strides, and Channels . . . . . . . . . . . . . . . . . . . . . . . . 594
13.10.3 Analogy to Matrix Transposition . . . . . . . . . . . . . . . . . . . . . . . 595
13.11 Fully Convolutional Networks (FCN) . . . . . . . . . . . . . . . . . . . . . . . . . 596
13.11.1 Constructing a Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 597
13.11.2 Initializing the Transposed Convolution Layer . . . . . . . . . . . . . . . . 599
13.11.4 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 600
13.11.5 Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 601
13.12 Neural Style Transfer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 603
13.12.1 Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 604
13.12.2 Reading the Content and Style Images . . . . . . . . . . . . . . . . . . . . 605
13.12.3 Preprocessing and Postprocessing . . . . . . . . . . . . . . . . . . . . . . 606
13.12.4 Extracting Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 606
13.12.6 Creating and Initializing the Composite Image . . . . . . . . . . . . . . . 609
13.12.7 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 610
13.13 Image Classification (CIFAR-10) on Kaggle . . . . . . . . . . . . . . . . . . . . . . 613
13.13.1 Obtaining and Organizing the Dataset . . . . . . . . . . . . . . . . . . . . 614
13.13.2 Image Augmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 616
13.13.5 Defining the Training Functions . . . . . . . . . . . . . . . . . . . . . . . 619
13.13.6 Training and Validating the Model . . . . . . . . . . . . . . . . . . . . . . 620
13.13.7 Classifying the Testing Set and Submitting Results on Kaggle . . . . . . . . 620
13.14 Dog Breed Identification (ImageNet Dogs) on Kaggle . . . . . . . . . . . . . . . . 622
13.14.1 Obtaining and Organizing the Dataset . . . . . . . . . . . . . . . . . . . . 623
13.14.2 Image Augmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 624
13.14.5 Defining the Training Functions . . . . . . . . . . . . . . . . . . . . . . . 626
13.14.6 Training and Validating the Model . . . . . . . . . . . . . . . . . . . . . . 627
13.14.7 Classifying the Testing Set and Submitting Results on Kaggle . . . . . . . . 627
14 Natural Language Processing: Pretraining 631

14.1 Word Embedding (word2vec) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 632
14.1.1 Why Not Use One-hot Vectors? . . . . . . . . . . . . . . . . . . . . . . . . 632
14.1.2 The Skip-Gram Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 632
14.1.3 The Continuous Bag of Words (CBOW) Model . . . . . . . . . . . . . . . . 634
14.2 Approximate Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 636
14.2.1 Negative Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 637
14.2.2 Hierarchical Softmax . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 638
14.3 The Dataset for Pretraining Word Embedding . . . . . . . . . . . . . . . . . . . . 639
14.3.1 Reading and Preprocessing the Dataset . . . . . . . . . . . . . . . . . . . 639
14.3.2 Subsampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 640
14.3.3 Loading the Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 642
14.4 Pretraining word2vec . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 646
x
14.4.1 The Skip-Gram Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 647
14.4.2 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 648
14.4.3 Applying the Word Embedding Model . . . . . . . . . . . . . . . . . . . . 650
14.5 Word Embedding with Global Vectors (GloVe) . . . . . . . . . . . . . . . . . . . . 651
14.5.1 The GloVe Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 652
14.5.2 Understanding GloVe from Conditional Probability Ratios . . . . . . . . . 653
14.6 Subword Embedding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 654
14.6.1 fastText . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 654
14.6.2 Byte Pair Encoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 655
14.7 Finding Synonyms and Analogies . . . . . . . . . . . . . . . . . . . . . . . . . . . 658
14.7.1 Using Pretrained Word Vectors . . . . . . . . . . . . . . . . . . . . . . . . 659
14.7.2 Applying Pretrained Word Vectors . . . . . . . . . . . . . . . . . . . . . . 660
14.8 Bidirectional Encoder Representations from Transformers (BERT) . . . . . . . . . 663
14.8.1 From Context-Independent to Context-Sensitive . . . . . . . . . . . . . . 663
14.8.2 From Task-Specific to Task-Agnostic . . . . . . . . . . . . . . . . . . . . . 663
14.8.3 BERT: Combining the Best of Both Worlds . . . . . . . . . . . . . . . . . . 664
14.8.4 Input Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 665
14.8.5 Pretraining Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 667
14.9 The Dataset for Pretraining BERT . . . . . . . . . . . . . . . . . . . . . . . . . . . 671
14.9.1 Defining Helper Functions for Pretraining Tasks . . . . . . . . . . . . . . 672
14.9.2 Transforming Text into the Pretraining Dataset . . . . . . . . . . . . . . . 674
14.10 Pretraining BERT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 677
14.10.1 Pretraining BERT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 677
14.10.2 Representing Text with BERT . . . . . . . . . . . . . . . . . . . . . . . . . 679
15 Natural Language Processing: Applications 683

15.1 Sentiment Analysis and the Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . 684
15.1.1 The Sentiment Analysis Dataset . . . . . . . . . . . . . . . . . . . . . . . 684
15.2 Sentiment Analysis: Using Recurrent Neural Networks . . . . . . . . . . . . . . . 687
15.2.1 Using a Recurrent Neural Network Model . . . . . . . . . . . . . . . . . . 688
15.3 Sentiment Analysis: Using Convolutional Neural Networks . . . . . . . . . . . . . 691
15.3.1 One-Dimensional Convolutional Layer . . . . . . . . . . . . . . . . . . . . 692
15.3.2 Max-Over-Time Pooling Layer . . . . . . . . . . . . . . . . . . . . . . . . 694
15.3.3 The TextCNN Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 695
15.4 Natural Language Inference and the Dataset . . . . . . . . . . . . . . . . . . . . . 698
15.4.1 Natural Language Inference . . . . . . . . . . . . . . . . . . . . . . . . . 698
15.4.2 The Stanford Natural Language Inference (SNLI) Dataset . . . . . . . . . . 699
15.5 Natural Language Inference: Using Attention . . . . . . . . . . . . . . . . . . . . 703
15.5.1 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 703
15.5.2 Training and Evaluating the Model . . . . . . . . . . . . . . . . . . . . . . 707
15.6 Fine-Tuning BERT for Sequence-Level and Token-Level Applications . . . . . . . . 710
15.6.1 Single Text Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . 710
15.6.2 Text Pair Classification or Regression . . . . . . . . . . . . . . . . . . . . 711
15.6.3 Text Tagging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 712
15.6.4 Question Answering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 713
15.7 Natural Language Inference: Fine-Tuning BERT . . . . . . . . . . . . . . . . . . . 715
15.7.1 Loading Pretrained BERT . . . . . . . . . . . . . . . . . . . . . . . . . . . 716
15.7.2 The Dataset for Fine-Tuning BERT . . . . . . . . . . . . . . . . . . . . . . 717
15.7.3 Fine-Tuning BERT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 718
xi
16 Recommender Systems 721
16.1 Overview of Recommender Systems . . . . . . . . . . . . . . . . . . . . . . . . . 721
16.1.1 Collaborative Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 722
16.1.2 Explicit Feedback and Implicit Feedback . . . . . . . . . . . . . . . . . . 723
16.1.3 Recommendation Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 723
16.2 The MovieLens Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 724
16.2.1 Getting the Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 724
16.2.2 Statistics of the Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . 725
16.2.3 Splitting the dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 726
16.2.4 Loading the data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 727
16.3 Matrix Factorization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 728
16.3.1 The Matrix Factorization Model . . . . . . . . . . . . . . . . . . . . . . . 729
16.3.2 Model Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 730
16.3.3 Evaluation Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 730
16.4 AutoRec: Rating Prediction with Autoencoders . . . . . . . . . . . . . . . . . . . 733
16.4.1 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 733
16.4.2 Implementing the Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 734
16.4.3 Reimplementing the Evaluator . . . . . . . . . . . . . . . . . . . . . . . . 734
16.5 Personalized Ranking for Recommender Systems . . . . . . . . . . . . . . . . . . 736
16.5.1 Bayesian Personalized Ranking Loss and its Implementation . . . . . . . 737
16.5.2 Hinge Loss and its Implementation . . . . . . . . . . . . . . . . . . . . . 738
16.6 Neural Collaborative Filtering for Personalized Ranking . . . . . . . . . . . . . . 739
16.6.1 The NeuMF model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 740
16.6.3 Customized Dataset with Negative Sampling . . . . . . . . . . . . . . . . . 742
16.6.4 Evaluator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 742
16.7 Sequence-Aware Recommender Systems . . . . . . . . . . . . . . . . . . . . . . . 746
16.7.1 Model Architectures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 746
16.7.3 Sequential Dataset with Negative Sampling . . . . . . . . . . . . . . . . . 749
16.7.4 Load the MovieLens 100K dataset . . . . . . . . . . . . . . . . . . . . . . 750
16.7.5 Train the Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 751
16.8 Feature-Rich Recommender Systems . . . . . . . . . . . . . . . . . . . . . . . . . 752
16.8.1 An Online Advertising Dataset . . . . . . . . . . . . . . . . . . . . . . . . 753
16.8.2 Dataset Wrapper . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 753
16.9 Factorization Machines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 755
16.9.1 2-Way Factorization Machines . . . . . . . . . . . . . . . . . . . . . . . . 755
16.9.2 An Efficient Optimization Criterion . . . . . . . . . . . . . . . . . . . . . 756
16.9.4 Load the Advertising Dataset . . . . . . . . . . . . . . . . . . . . . . . . . 757
16.9.5 Train the Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 757
16.10 Deep Factorization Machines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 758
16.10.1 Model Architectures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 759
16.10.2 Implemenation of DeepFM . . . . . . . . . . . . . . . . . . . . . . . . . . 760
17 Generative Adversarial Networks 763

17.1 Generative Adversarial Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . 763
xii
17.1.1 Generate some “real” data . . . . . . . . . . . . . . . . . . . . . . . . . . . 765
17.1.2 Generator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 766
17.1.3 Discriminator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 766
17.1.4 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 766
17.2 Deep Convolutional Generative Adversarial Networks . . . . . . . . . . . . . . . . 769
17.2.1 The Pokemon Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 769
17.2.2 The Generator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 770
17.2.3 Discriminator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 772
17.2.4 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 773
18 Appendix: Mathematics for Deep Learning 777

18.1 Geometry and Linear Algebraic Operations . . . . . . . . . . . . . . . . . . . . . 778
18.1.1 Geometry of Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 778
18.1.2 Dot Products and Angles . . . . . . . . . . . . . . . . . . . . . . . . . . . 780
18.1.3 Hyperplanes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 782
18.1.4 Geometry of Linear Transformations . . . . . . . . . . . . . . . . . . . . 785
18.1.5 Linear Dependence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 787
18.1.6 Rank . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 787
18.1.7 Invertibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 788
18.1.8 Determinant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 789
18.1.9 Tensors and Common Linear Algebra Operations . . . . . . . . . . . . . . 790
18.2 Eigendecompositions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 794
18.2.1 Finding Eigenvalues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 794
18.2.2 Decomposing Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . 795
18.2.3 Operations on Eigendecompositions . . . . . . . . . . . . . . . . . . . . . 795
18.2.4 Eigendecompositions of Symmetric Matrices . . . . . . . . . . . . . . . . 796
18.2.5 Gershgorin Circle Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 796
18.2.6 A Useful Application: The Growth of Iterated Maps . . . . . . . . . . . . . 797
18.2.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 802
18.3 Single Variable Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 803
18.3.1 Differential Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 803
18.3.2 Rules of Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 806
18.4 Multivariable Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 813
18.4.1 Higher-Dimensional Differentiation . . . . . . . . . . . . . . . . . . . . . 814
18.4.2 Geometry of Gradients and Gradient Descent . . . . . . . . . . . . . . . . 815
18.4.3 A Note on Mathematical Optimization . . . . . . . . . . . . . . . . . . . . 816
18.4.4 Multivariate Chain Rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . 817
18.4.5 The Backpropagation Algorithm . . . . . . . . . . . . . . . . . . . . . . . 819
18.4.6 Hessians . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 822
18.4.7 A Little Matrix Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . 824
18.5 Integral Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 829
18.5.1 Geometric Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . 829
18.5.2 The Fundamental Theorem of Calculus . . . . . . . . . . . . . . . . . . . 831
18.5.3 Change of Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 833
18.5.4 A Comment on Sign Conventions . . . . . . . . . . . . . . . . . . . . . . . 834
18.5.5 Multiple Integrals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 835
18.5.6 Change of Variables in Multiple Integrals . . . . . . . . . . . . . . . . . . 837
18.6 Random Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 838
18.6.1 Continuous Random Variables . . . . . . . . . . . . . . . . . . . . . . . . 838
18.7 Maximum Likelihood . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 855
18.7.1 The Maximum Likelihood Principle . . . . . . . . . . . . . . . . . . . . . 856
xiii
18.7.2 Numerical Optimization and the Negative Log-Likelihood . . . . . . . . . 857
18.7.3 Maximum Likelihood for Continuous Variables . . . . . . . . . . . . . . . 859
18.8 Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 861
18.8.1 Bernoulli . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 861
18.8.2 Discrete Uniform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 863
18.8.3 Continuous Uniform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 864
18.8.4 Binomial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 866
18.8.5 Poisson . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 868
18.8.6 Gaussian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 871
18.8.7 Exponential Family . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 874
18.9 Naive Bayes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 875
18.9.1 Optical Character Recognition . . . . . . . . . . . . . . . . . . . . . . . . 876
18.9.2 The Probabilistic Model for Classification . . . . . . . . . . . . . . . . . . 877
18.9.3 The Naive Bayes Classifier . . . . . . . . . . . . . . . . . . . . . . . . . . 877
18.9.4 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 878
18.10 Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 882
18.10.1 Evaluating and Comparing Estimators . . . . . . . . . . . . . . . . . . . . 882
18.10.2 Conducting Hypothesis Tests . . . . . . . . . . . . . . . . . . . . . . . . . 886
18.10.3 Constructing Confidence Intervals . . . . . . . . . . . . . . . . . . . . . . 890
18.11 Information Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 893
18.11.1 Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 893
18.11.2 Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 895
18.11.3 Mutual Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 897
18.11.4 Kullback–Leibler Divergence . . . . . . . . . . . . . . . . . . . . . . . . . 901
18.11.5 Cross Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 903
19 Appendix: Tools for Deep Learning 907

19.1 Using Jupyter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 907
19.1.1 Editing and Running the Code Locally . . . . . . . . . . . . . . . . . . . . 907
19.1.2 Advanced Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 911
19.2 Using Amazon SageMaker . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 912
19.2.1 Registering and Logging In . . . . . . . . . . . . . . . . . . . . . . . . . . 912
19.2.2 Creating a SageMaker Instance . . . . . . . . . . . . . . . . . . . . . . . . 913
19.2.3 Running and Stopping an Instance . . . . . . . . . . . . . . . . . . . . . . 914
19.2.4 Updating Notebooks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 915
19.3 Using AWS EC2 Instances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 916
19.3.1 Creating and Running an EC2 Instance . . . . . . . . . . . . . . . . . . . . 916
19.3.2 Installing CUDA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 921
19.3.3 Installing MXNet and Downloading the D2L Notebooks . . . . . . . . . . . 922
19.3.4 Running Jupyter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 923
19.3.5 Closing Unused Instances . . . . . . . . . . . . . . . . . . . . . . . . . . . 924
19.4 Using Google Colab . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 924
19.5 Selecting Servers and GPUs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 925
19.5.1 Selecting Servers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 926
19.5.2 Selecting GPUs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 927
19.6 Contributing to This Book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 930
19.6.1 Minor Text Changes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 930
19.6.2 Propose a Major Change . . . . . . . . . . . . . . . . . . . . . . . . . . . 930
19.6.3 Adding a New Section or a New Framework Implementation . . . . . . . . 931
19.6.4 Submitting a Major Change . . . . . . . . . . . . . . . . . . . . . . . . . . 931
19.7 d2l API Document . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 935
xiv
Bibliography 955
Python Module Index 965
Index 967
xv
xvi
Preface
Just a few years ago, there were no legions of deep learning scientists developing intelligent prod-
ucts and services at major companies and startups. When the youngest among us (the authors)
entered the field, machine learning did not command headlines in daily newspapers. Our parents
had no idea what machine learning was, let alone why we might prefer it to a career in medicine or
law. Machine learning was a forward-looking academic discipline with a narrow set of real-world
applications. And those applications, e.g., speech recognition and computer vision, required so
much domain knowledge that they were often regarded as separate areas entirely for which ma-
chine learning was one small component. Neural networks then, the antecedents of the deep
learning models that we focus on in this book, were regarded as outmoded tools.
In just the past five years, deep learning has taken the world by surprise, driving rapid progress
in fields as diverse as computer vision, natural language processing, automatic speech recogni-
tion, reinforcement learning, and statistical modeling. With these advances in hand, we can now
build cars that drive themselves with more autonomy than ever before (and less autonomy than
some companies might have you believe), smart reply systems that automatically draft the most
mundane emails, helping people dig out from oppressively large inboxes, and software agents that
dominate the worldʼs best humans at board games like Go, a feat once thought to be decades away.
Already, these tools exert ever-wider impacts on industry and society, changing the way movies
are made, diseases are diagnosed, and playing a growing role in basic sciences—from astrophysics
to biology.
About This Book
This book represents our attempt to make deep learning approachable, teaching you the concepts,
the context, and the code.
One Medium Combining Code, Math, and HTML
For any computing technology to reach its full impact, it must be well-understood, well-
documented, and supported by mature, well-maintained tools. The key ideas should be clearly
distilled, minimizing the onboarding time needing to bring new practitioners up to date. Mature
libraries should automate common tasks, and exemplar code should make it easy for practitioners
to modify, apply, and extend common applications to suit their needs. Take dynamic web appli-
cations as an example. Despite a large number of companies, like Amazon, developing successful
database-driven web applications in the 1990s, the potential of this technology to aid creative en-
trepreneurs has been realized to a far greater degree in the past ten years, owing in part to the
development of powerful, well-documented frameworks.
1
Testing the potential of deep learning presents unique challenges because any single application
brings together various disciplines. Applying deep learning requires simultaneously understand-
ing (i) the motivations for casting a problem in a particular way; (ii) the mathematics of a given
modeling approach; (iii) the optimization algorithms for fitting the models to data; and (iv) and the
engineering required to train models efficiently, navigating the pitfalls of numerical computing
and getting the most out of available hardware. Teaching both the critical thinking skills required
to formulate problems, the mathematics to solve them, and the software tools to implement those
solutions all in one place presents formidable challenges. Our goal in this book is to present a
unified resource to bring would-be practitioners up to speed.
At the time we started this book project, there were no resources that simultaneously (i) were
up to date; (ii) covered the full breadth of modern machine learning with substantial technical
depth; and (iii) interleaved exposition of the quality one expects from an engaging textbook with
the clean runnable code that one expects to find in hands-on tutorials. We found plenty of code
examples for how to use a given deep learning framework (e.g., how to do basic numerical com-
puting with matrices in TensorFlow) or for implementing particular techniques (e.g., code snip-
pets for LeNet, AlexNet, ResNets, etc) scattered across various blog posts and GitHub repositories.
However, these examples typically focused on how to implement a given approach, but left out the
discussion of why certain algorithmic decisions are made. While some interactive resources have
popped up sporadically to address a particular topic, e.g., the engaging blog posts published on
the website Distill3 , or personal blogs, they only covered selected topics in deep learning, and
often lacked associated code. On the other hand, while several textbooks have emerged, most no-
tably (Goodfellow et al., 2016), which offers a comprehensive survey of the concepts behind deep
learning, these resources do not marry the descriptions to realizations of the concepts in code,
sometimes leaving readers clueless as to how to implement them. Moreover, too many resources
are hidden behind the paywalls of commercial course providers.
We set out to create a resource that could (i) be freely available for everyone; (ii) offer sufficient
technical depth to provide a starting point on the path to actually becoming an applied machine
learning scientist; (iii) include runnable code, showing readers how to solve problems in practice;
(iv) allow for rapid updates, both by us and also by the community at large; and (v) be comple-
mented by a forum4 for interactive discussion of technical details and to answer questions.
These goals were often in conflict. Equations, theorems, and citations are best managed and laid
out in LaTeX. Code is best described in Python. And webpages are native in HTML and JavaScript.
Furthermore, we want the content to be accessible both as executable code, as a physical book,
as a downloadable PDF, and on the Internet as a website. At present there exist no tools and no
workflow perfectly suited to these demands, so we had to assemble our own. We describe our
approach in detail in Section 19.6. We settled on GitHub to share the source and to allow for edits,
Jupyter notebooks for mixing code, equations and text, Sphinx as a rendering engine to generate
multiple outputs, and Discourse for the forum. While our system is not yet perfect, these choices
provide a good compromise among the competing concerns. We believe that this might be the
first book published using such an integrated workflow.
3
http://distill.pub
4
http://discuss.d2l.ai
2 Contents
Learning by Doing
Many textbooks teach a series of topics, each in exhaustive detail. For example, Chris Bishopʼs
excellent textbook (Bishop, 2006), teaches each topic so thoroughly, that getting to the chapter on
linear regression requires a non-trivial amount of work. While experts love this book precisely
for its thoroughness, for beginners, this property limits its usefulness as an introductory text.
In this book, we will teach most concepts just in time. In other words, you will learn concepts at the
very moment that they are needed to accomplish some practical end. While we take some time at
the outset to teach fundamental preliminaries, like linear algebra and probability, we want you to
taste the satisfaction of training your first model before worrying about more esoteric probability
distributions.
Aside from a few preliminary notebooks that provide a crash course in the basic mathematical
background, each subsequent chapter introduces both a reasonable number of new concepts and
provides single self-contained working examples—using real datasets. This presents an organi-
zational challenge. Some models might logically be grouped together in a single notebook. And
some ideas might be best taught by executing several models in succession. On the other hand,
there is a big advantage to adhering to a policy of 1 working example, 1 notebook: This makes it as
easy as possible for you to start your own research projects by leveraging our code. Just copy a
notebook and start modifying it.
We will interleave the runnable code with background material as needed. In general, we will
often err on the side of making tools available before explaining them fully (and we will follow up
by explaining the background later). For instance, we might use stochastic gradient descent before
fully explaining why it is useful or why it works. This helps to give practitioners the necessary
ammunition to solve problems quickly, at the expense of requiring the reader to trust us with
some curatorial decisions.
This book will teach deep learning concepts from scratch. Sometimes, we want to delve into fine
details about the models that would typically be hidden from the user by deep learning frame-
worksʼ advanced abstractions. This comes up especially in the basic tutorials, where we want you
to understand everything that happens in a given layer or optimizer. In these cases, we will often
present two versions of the example: one where we implement everything from scratch, relying
only on the NumPy interface and automatic differentiation, and another, more practical exam-
ple, where we write succinct code using Gluon. Once we have taught you how some component
works, we can just use the Gluon version in subsequent tutorials.
Contents 3
Content and Structure
The book can be roughly divided into three parts, which are presented by different colors in Fig.
1:
Fig. 1: Book structure
• The first part covers basics and preliminaries. Chapter 1 offers an introduction to deep learn-
ing. Then, in Chapter 2, we quickly bring you up to speed on the prerequisites required for
hands-on deep learning, such as how to store and manipulate data, and how to apply various
numerical operations based on basic concepts from linear algebra, calculus, and probabil-
ity. Chapter 3 and Chapter 4 cover the most basic concepts and techniques of deep learning,
such as linear regression, multilayer perceptrons and regularization.
• The next five chapters focus on modern deep learning techniques. Chapter 5 describes the
various key components of deep learning calculations and lays the groundwork for us to
subsequently implement more complex models. Next, in Chapter 6 and Chapter 7, we intro-
duce convolutional neural networks (CNNs), powerful tools that form the backbone of most
modern computer vision systems. Subsequently, in Chapter 8 and Chapter 9, we introduce
recurrent neural networks (RNNs), models that exploit temporal or sequential structure in
data, and are commonly used for natural language processing and time series prediction.
In Chapter 10, we introduce a new class of models that employ a technique called attention
mechanisms and they have recently begun to displace RNNs in natural language processing.
These sections will get you up to speed on the basic tools behind most modern applications
of deep learning.
• Part three discusses scalability, efficiency, and applications. First, in Chapter 11, we dis-
cuss several common optimization algorithms used to train deep learning models. The next
4 Contents
chapter, Chapter 12 examines several key factors that influence the computational perfor-
mance of your deep learning code. In Chapter 13, we illustrate major applications of deep
learning in computer vision. In Chapter 14 and Chapter 15, we show how to pretrain lan-
guage representation models and apply them to natural language processing tasks.
Code
Most sections of this book feature executable code because of our belief in the importance of an
interactive learning experience in deep learning. At present, certain intuitions can only be devel-
oped through trial and error, tweaking the code in small ways and observing the results. Ideally,
an elegant mathematical theory might tell us precisely how to tweak our code to achieve a desired
result. Unfortunately, at present, such elegant theories elude us. Despite our best attempts, for-
mal explanations for various techniques are still lacking, both because the mathematics to char-
acterize these models can be so difficult and also because serious inquiry on these topics has only
just recently kicked into high gear. We are hopeful that as the theory of deep learning progresses,
future editions of this book will be able to provide insights in places the present edition cannot.
At times, to avoid unnecessary repetition, we encapsulate the frequently-imported and referred-to
functions, classes, etc. in this book in the d2l package. For any block such as a function, a class,
or multiple imports to be saved in the package, we will mark it with #@save. We offer a detailed
overview of these functions and classes in Section 19.7. The d2l package is light-weight and only
requires the following packages and modules as dependencies:
#@save
import collections
from collections import defaultdict
from IPython import display
import math
from matplotlib import pyplot as plt
import os
import pandas as pd
import random
import re
import shutil
import sys
import tarfile
import time
import requests
import zipfile
import hashlib
d2l = sys.modules[__name__]
Most of the code in this book is based on Apache MXNet. MXNet is an open-source framework for
deep learning and the preferred choice of AWS (Amazon Web Services), as well as many colleges
and companies. All of the code in this book has passed tests under the newest MXNet version.
However, due to the rapid development of deep learning, some code in the print edition may not
work properly in future versions of MXNet. However, we plan to keep the online version up-to-
date. In case you encounter any such problems, please consult Installation (page 9) to update your
code and runtime environment.
Here is how we import modules from MXNet.
Contents 5
#@save
from mxnet import autograd, context, gluon, image, init, np, npx
from mxnet.gluon import nn, rnn
Target Audience
This book is for students (undergraduate or graduate), engineers, and researchers, who seek a
solid grasp of the practical techniques of deep learning. Because we explain every concept from
scratch, no previous background in deep learning or machine learning is required. Fully explain-
ing the methods of deep learning requires some mathematics and programming, but we will only
assume that you come in with some basics, including (the very basics of) linear algebra, calcu-
lus, probability, and Python programming. Moreover, in the Appendix, we provide a refresher
on most of the mathematics covered in this book. Most of the time, we will prioritize intuition
and ideas over mathematical rigor. There are many terrific books which can lead the interested
reader further. For instance, Linear Analysis by Bela Bollobas (Bollobas, 1999) covers linear alge-
bra and functional analysis in great depth. All of Statistics (Wasserman, 2013) is a terrific guide to
statistics. And if you have not used Python before, you may want to peruse this Python tutorial5 .
Forum
Associated with this book, we have launched a discussion forum, located at discuss.d2l.ai6 . When
you have questions on any section of the book, you can find the associated discussion page link at
the end of each chapter.
Acknowledgments
We are indebted to the hundreds of contributors for both the English and the Chinese drafts. They
helped improve the content and offered valuable feedback. Specifically, we thank every con-
tributor of this English draft for making it better for everyone. Their GitHub IDs or names are
(in no particular order): alxnorden, avinashingit, bowen0701, brettkoonce, Chaitanya Prakash
Bapat, cryptonaut, Davide Fiocco, edgarroman, gkutiel, John Mitro, Liang Pu, Rahul Agarwal,
Mohamed Ali Jamaoui, Michael (Stu) Stewart, Mike Müller, NRauschmayr, Prakhar Srivastav,
sad-, sfermigier, Sheng Zha, sundeepteki, topecongiro, tpdi, vermicelli, Vishaal Kapoor, Vish-
wesh Ravi Shrimali, YaYaB, Yuhong Chen, Evgeniy Smirnov, lgov, Simon Corston-Oliver, Igor
Dzreyev, Ha Nguyen, pmuens, Andrei Lukovenko, senorcinco, vfdev-5, dsweet, Mohammad
Mahdi Rahimi, Abhishek Gupta, uwsd, DomKM, Lisa Oakley, Bowen Li, Aarush Ahuja, Prasanth
Buddareddygari, brianhendee, mani2106, mtn, lkevinzc, caojilin, Lakshya, Fiete Lüer, Surbhi Vi-
jayvargeeya, Muhyun Kim, dennismalmgren, adursun, Anirudh Dagar, liqingnz, Pedro Larroy,
lgov, ati-ozgur, Jun Wu, Matthias Blume, Lin Yuan, geogunow, Josh Gardner, Maximilian Böther,
Rakib Islam, Leonard Lausen, Abhinav Upadhyay, rongruosong, Steve Sedlmeyer, Ruslan Bara-
tov, Rafael Schlatter, liusy182, Giannis Pappas, ati-ozgur, qbaza, dchoi77, Adam Gerson, Phuc
Le, Mark Atwood, christabella, vn09, Haibin Lin, jjangga0214, RichyChen, noelo, hansent, Giel
Dops, dvincent1337, WhiteD3vil, Peter Kulits, codypenta, joseppinilla, ahmaurya, karolszk, heyti-
tle, Peter Goetz, rigtorp, tiepvupsu, sfilip, mlxd, Kale-ab Tessera, Sanjar Adilov, MatteoFerrara,
hsneto, Katarzyna Biesialska, Gregory Bruss, duythanhvn, paulaurel, graytowne, minhduc0711,
5
http://learnpython.org/
6
https://discuss.d2l.ai/
6 Contents
sl7423, Jaedong Hwang, Yida Wang, cys4, clhm, Jean Kaddour, austinmw, trebeljahr, tbaums,
cuongvng, pavelkomarov, vzlamal, NotAnotherSystem, J-Arun-Mani, jancio, eldarkurtic, the-
great-shazbot, doctorcolossus, gducharme, cclauss, Daniel-Mietchen, hoonose, biagiom, abhi-
navsp0730, jonathanhrandall, ysraell, Nodar Okroshiashvili, UgurKap, Jiyang Kang, StevenJokes,
Tomer Kaftan, liweiwp, netyster, ypandya, NishantTharani, heiligerl, SportsTHU, nguyenhoa93,
manuel-arno-korfmann-webentwicklung, aterzis-personal, nxby, Xiaoting He, yoderj, mathre-
search.
We thank Amazon Web Services, especially Swami Sivasubramanian, Raju Gulabani, Charlie Bell,
and Andrew Jassy for their generous support in writing this book. Without the available time,
resources, discussions with colleagues, and continuous encouragement this book would not have
happened.
Summary
• Deep learning has revolutionized pattern recognition, introducing technology that now
powers a wide range of technologies, including computer vision, natural language process-
ing, automatic speech recognition.
• To successfully apply deep learning, you must understand how to cast a problem, the math-
ematics of modeling, the algorithms for fitting your models to data, and the engineering
techniques to implement it all.
• This book presents a comprehensive resource, including prose, figures, mathematics, and
code, all in one place.
• To answer questions related to this book, visit our forum at https://discuss.d2l.ai/.
• All notebooks are available for download on GitHub.
Exercises
1. Register an account on the discussion forum of this book discuss.d2l.ai7 .

2. Install Python on your computer.
3. Follow the links at the bottom of the section to the forum, where you will be able to seek out
help and discuss the book and find answers to your questions by engaging the authors and
broader community.
Discussions8
7
https://discuss.d2l.ai/
8
https://discuss.d2l.ai/t/18
Contents 7
8 Contents
Installation
In order to get you up and running for hands-on learning experience, we need to set you up with an
environment for running Python, Jupyter notebooks, the relevant libraries, and the code needed
to run the book itself.
Installing Miniconda
The simplest way to get going will be to install Miniconda9 . The Python 3.x version is required.
You can skip the following steps if conda has already been installed. Download the corresponding
Miniconda sh file from the website and then execute the installation from the command line using
sh <FILENAME> -b. For macOS users:
# The file name is subject to changes

sh Miniconda3-latest-MacOSX-x86_64.sh -b
For Linux users:
# The file name is subject to changes

sh Miniconda3-latest-Linux-x86_64.sh -b
Next, initialize the shell so we can run conda directly.
~/miniconda3/bin/conda init
Now close and re-open your current shell. You should be able to create a new environment as
following:
conda create --name d2l -y
9
https://conda.io/en/latest/miniconda.html
9
Downloading the D2L Notebooks
Next, we need to download the code of this book. You can click the “All Notebooks” tab on the top
of any HTML page to download and unzip the code. Alternatively, if you have unzip (otherwise
run sudo apt install unzip) available:
mkdir d2l-en && cd d2l-en

curl https://d2l.ai/d2l-en.zip -o d2l-en.zip
unzip d2l-en.zip && rm d2l-en.zip
Now we will want to activate the d2l environment and install pip. Enter y for the queries that
follow this command.
conda activate d2l

conda install python=3.7 pip -y
Installing the Framework and the d2l Package
Before installing the deep learning framework, please first check whether or not you have proper
GPUs on your machine (the GPUs that power the display on a standard laptop do not count for our
purposes). If you are installing on a GPU server, proceed to GPU Support (page 11) for instructions
to install a GPU-supported version.
Otherwise, you can install the CPU version. That will be more than enough horsepower to get you
through the first few chapters but you will want to access GPUs before running larger models.
pip install mxnet==1.6.0
We also install the d2l package that encapsulates frequently used functions and classes in this
book.
# -U: Upgrade all packages to the newest available version

pip install -U d2l
Once they are installed, we now open the Jupyter notebook by running:
jupyter notebook
At this point, you can open http://localhost:8888 (it usually opens automatically) in your Web
browser. Then we can run the code for each section of the book. Please always execute conda ac-
tivate d2l to activate the runtime environment before running the code of the book or updating
the deep learning framework or the d2l package. To exit the environment, run conda deactivate.
10 Contents
GPU Support
By default, the deep learning framework is installed without GPU support to ensure that it will
run on any computer (including most laptops). Part of this book requires or recommends running
with GPU. If your computer has NVIDIA graphics cards and has installed CUDA10 , then you should
install a GPU-enabled version. If you have installed the CPU-only version, you may need to remove
it first by running:
pip uninstall mxnet
Then we need to find the CUDA version you installed. You may check it through nvcc --version
or cat /usr/local/cuda/version.txt. Assume that you have installed CUDA 10.1, then you can
install with the following command:
# For Windows users

pip install mxnet-cu101==1.6.0b20190926
# For Linux and macOS users

pip install mxnet-cu101==1.6.0
You may change the last digits according to your CUDA version, e.g., cu100 for CUDA 10.0 and cu90
for CUDA 9.0.
Exercises
1. Download the code for the book and install the runtime environment.
Discussions11
10
https://developer.nvidia.com/cuda-downloads
11
Contents 11
12 Contents
Notation
The notation used throughout this book is summarized below.
Numbers
• x: A scalar
• x: A vector
• X: A matrix
• X: A tensor
• I: An identity matrix
• xi , [x]i : The ith element of vector x
• xij , xi,j ,[X]ij , [X]i,j : The element of matrix X at row i and column j
Set Theory
• X : A set
• Z: The set of integers
• R: The set of real numbers
• Rn : The set of n-dimensional vectors of real numbers
• Ra×b : The set of matrices of real numbers with a rows and b columns
• |X |: Cardinality (number of elements) of set X
• A ∪ B: Union of sets A and B
• A ∩ B: Intersection of sets A and B
• A \ B: Subtraction of set B from set A
13
Functions and Operators
• f (·): A function
• log(·): The natural logarithm
• exp(·): The exponential function
• 1X : The indicator function
• (·)⊤ : Transpose of a vector or a matrix
• X−1 : Inverse of matrix X
• ⊙: Hadamard (elementwise) product
• [·, ·]: Concatenation
• |X |: Cardinality of set X
• ∥ · ∥p : Lp norm
• ∥ · ∥: L2 norm
• ⟨x, y⟩: Dot product of vectors x and y
∑
• : Series addition
∏
• : Series multiplication
def
• = : Definition
Calculus
dy
• dx : Derivative of y with respect to x
∂y
• ∂x : Partial derivative of y with respect to x
• ∇x y: Gradient of y with respect to x
∫b
• a f (x) dx: Definite integral of f from a to b with respect to x
∫
• f (x) dx: Indefinite integral of f with respect to x
Probability and Information Theory
• P (·): Probability distribution

• z ∼ P : Random variable z has probability distribution P
• P (X | Y ): Conditional probability of X | Y
• p(x): Probability density function
• Ex [f (x)]: Expectation of f with respect to x
• X ⊥ Y : Random variables X and Y are independent
14 Contents
• X ⊥ Y | Z: Random variables X and Y are conditionally independent given random vari-
able Z
• Var(X): Variance of random variable X
• σX : Standard deviation of random variable X
• Cov(X, Y ): Covariance of random variables X and Y
• ρ(X, Y ): Correlation of random variables X and Y
• H(X): Entropy of random variable X
• DKL (P ∥Q): KL-divergence of distributions P and Q
Complexity
• O: Big O notation
Discussions12
12
Contents 15
16 Contents
1 | Introduction
Until recently, nearly every computer program that we interact with daily was coded by software
developers from first principles. Say that we wanted to write an application to manage an e-
commerce platform. After huddling around a whiteboard for a few hours to ponder the prob-
lem, we would come up with the broad strokes of a working solution that might probably look
something like this: (i) users interact with the application through an interface running in a web
browser or mobile application; (ii) our application interacts with a commercial-grade database
engine to keep track of each userʼs state and maintain records of historical transactions; and (iii)
at the heart of our application, the business logic (you might say, the brains) of our application spells
out in methodical detail the appropriate action that our program should take in every conceivable
circumstance.
To build the brains of our application, weʼd have to step through every possible corner case that
we anticipate encountering, devising appropriate rules. Each time a customer clicks to add an
item to their shopping cart, we add an entry to the shopping cart database table, associating that
userʼs ID with the requested productʼs ID. While few developers ever get it completely right the
first time (it might take some test runs to work out the kinks), for the most part, we could write
such a program from first principles and confidently launch it before ever seeing a real customer.
Our ability to design automated systems from first principles that drive functioning products and
systems, often in novel situations, is a remarkable cognitive feat. And when you are able to devise
solutions that work 100% of the time, you should not be using machine learning.
Fortunately for the growing community of machine learning (ML) scientists, many tasks that we
would like to automate do not bend so easily to human ingenuity. Imagine huddling around the
whiteboard with the smartest minds you know, but this time you are tackling one of the following
problems:
• Write a program that predicts tomorrowʼs weather given geographic information, satellite
images, and a trailing window of past weather.
• Write a program that takes in a question, expressed in free-form text, and answers it cor-
rectly.
• Write a program that given an image can identify all the people it contains, drawing outlines
around each.
• Write a program that presents users with products that they are likely to enjoy but unlikely,
in the natural course of browsing, to encounter.
In each of these cases, even elite programmers are incapable of coding up solutions from scratch.
The reasons for this can vary. Sometimes the program that we are looking for follows a pattern
that changes over time, and we need our programs to adapt. In other cases, the relationship (say
between pixels, and abstract categories) may be too complicated, requiring thousands or millions
of computations that are beyond our conscious understanding (even if our eyes manage the task
17
effortlessly). ML is the study of powerful techniques that can learn from experience. As an ML
algorithm accumulates more experience, typically in the form of observational data or interac-
tions with an environment, its performance improves. Contrast this with our deterministic e-
commerce platform, which performs according to the same business logic, no matter how much
experience accrues, until the developers themselves learn and decide that it is time to update the
software. In this book, we will teach you the fundamentals of machine learning, and focus in par-
ticular on deep learning, a powerful set of techniques driving innovations in areas as diverse as
computer vision, natural language processing, healthcare, and genomics.
1.1 A Motivating Example
Before we could begin writing, the authors of this book, like much of the work force, had to be-
come caffeinated. We hopped in the car and started driving. Using an iPhone, Alex called out
“Hey Siri”, awakening the phoneʼs voice recognition system. Then Mu commanded “directions to
Blue Bottle coffee shop”. The phone quickly displayed the transcription of his command. It also
recognized that we were asking for directions and launched the Maps application to fulfill our re-
quest. Once launched, the Maps app identified a number of routes. Next to each route, the phone
displayed a predicted transit time. While we fabricated this story for pedagogical convenience, it
demonstrates that in the span of just a few seconds, our everyday interactions with a smart phone
can engage several machine learning models.
Imagine just writing a program to respond to a wake word like “Alexa”, “Okay, Google” or “Siri”. Try
coding it up in a room by yourself with nothing but a computer and a code editor, as illustrated
in Fig. 1.1.1. How would you write such a program from first principles? Think about it… the
problem is hard. Every second, the microphone will collect roughly 44,000 samples. Each sample
is a measurement of the amplitude of the sound wave. What rule could map reliably from a snippet
of raw audio to confident predictions {yes, no} on whether the snippet contains the wake word?
If you are stuck, do not worry. We do not know how to write such a program from scratch either.
That is why we use ML.
Fig. 1.1.1: Identify an awake word.
Hereʼs the trick. Often, even when we do not know how to tell a computer explicitly how to map
from inputs to outputs, we are nonetheless capable of performing the cognitive feat ourselves. In
other words, even if you do not know how to program a computer to recognize the word “Alexa”,
you yourself are able to recognize the word “Alexa”. Armed with this ability, we can collect a huge
dataset containing examples of audio and label those that do and that do not contain the wake word.
In the ML approach, we do not attempt to design a system explicitly to recognize wake words.
Instead, we define a flexible program whose behavior is determined by a number of parameters.
Then we use the dataset to determine the best possible set of parameters, those that improve the
performance of our program with respect to some measure of performance on the task of interest.
You can think of the parameters as knobs that we can turn, manipulating the behavior of the pro-
gram. Fixing the parameters, we call the program a model. The set of all distinct programs (input-
output mappings) that we can produce just by manipulating the parameters is called a family of
18 Chapter 1. Introduction
models. And the meta-program that uses our dataset to choose the parameters is called a learning
algorithm.
Before we can go ahead and engage the learning algorithm, we have to define the problem pre-
cisely, pinning down the exact nature of the inputs and outputs, and choosing an appropriate
model family. In this case, our model receives a snippet of audio as input, and it generates a se-
lection among {yes, no} as output. If all goes according to plan the modelʼs guesses will typically
be correct as to whether (or not) the snippet contains the wake word.
If we choose the right family of models, then there should exist one setting of the knobs such
that the model fires yes every time it hears the word “Alexa”. Because the exact choice of the
wake word is arbitrary, we will probably need a model family sufficiently rich that, via another
setting of the knobs, it could fire yes only upon hearing the word “Apricot”. We expect that the
same model family should be suitable for “Alexa” recognition and “Apricot” recognition because
they seem, intuitively, to be similar tasks. However, we might need a different family of models
entirely if we want to deal with fundamentally different inputs or outputs, say if we wanted to map
from images to captions, or from English sentences to Chinese sentences.
As you might guess, if we just set all of the knobs randomly, it is not likely that our model will rec-
ognize “Alexa”, “Apricot”, or any other English word. In deep learning, the learning is the process
by which we discover the right setting of the knobs coercing the desired behavior from our model.
As shown in Fig. 1.1.2, the training process usually looks like this:
1. Start off with a randomly initialized model that cannot do anything useful.
2. Grab some of your labeled data (e.g., audio snippets and corresponding {yes, no} labels)
3. Tweak the knobs so the model sucks less with respect to those examples
4. Repeat until the model is awesome.
Fig. 1.1.2: A typical training process.
To summarize, rather than code up a wake word recognizer, we code up a program that can learn
to recognize wake words, if we present it with a large labeled dataset. You can think of this act of
determining a programʼs behavior by presenting it with a dataset as programming with data. We
can “program” a cat detector by providing our machine learning system with many examples of
cats and dogs, such as the images below:
1.1. A Motivating Example 19

cat cat dog dog
This way the detector will eventually learn to emit a very large positive number if it is a cat, a very
large negative number if it is a dog, and something closer to zero if it is not sure, and this barely
scratches the surface of what ML can do.
Deep learning is just one among many popular methods for solving machine learning problems.
Thus far, we have only talked about machine learning broadly and not deep learning. To see why
deep learning is important, we should pause for a moment to highlight a couple of crucial points.
First, the problems that we have discussed thus far—learning from the raw audio signal, the raw
pixel values of images, or mapping between sentences of arbitrary lengths and their counterparts
in foreign languages—are problems where deep learning excels and where traditional ML meth-
ods faltered. Deep models are deep in precisely the sense that they learn many layers of compu-
tation. It turns out that these many-layered (or hierarchical) models are capable of addressing
low-level perceptual data in a way that previous tools could not. In bygone days, the crucial part
of applying ML to these problems consisted of coming up with manually-engineered ways of trans-
forming the data into some form amenable to shallow models. One key advantage of deep learning
is that it replaces not only the shallow models at the end of traditional learning pipelines, but also
the labor-intensive process of feature engineering. Second, by replacing much of the domain-
specific preprocessing, deep learning has eliminated many of the boundaries that previously sep-
arated computer vision, speech recognition, natural language processing, medical informatics,
and other application areas, offering a unified set of tools for tackling diverse problems.
1.2 The Key Components: Data, Models, and Algorithms
In our wake-word example, we described a dataset consisting of audio snippets and binary labels,
and we gave a hand-wavy sense of how we might train a model to approximate a mapping from
snippets to classifications. This sort of problem, where we try to predict a designated unknown
label given known inputs, given a dataset consisting of examples, for which the labels are known
is called supervised learning, and it is just one among many kinds of machine learning problems.
In the next section, we will take a deep dive into the different ML problems. First, weʼd like to
shed more light on some core components that will follow us around, no matter what kind of ML
problem we take on:
1. The data that we can learn from.
2. A model of how to transform the data.
3. A loss function that quantifies the badness of our model.
4. An algorithm to adjust the modelʼs parameters to minimize the loss.
1.2.1 Data
It might go without saying that you cannot do data science without data. We could lose hundreds of
pages pondering what precisely constitutes data, but for now, we will err on the practical side and
focus on the key properties to be concerned with. Generally, we are concerned with a collection
of examples. In order to work with data usefully, we typically need to come up with a suitable
numerical representation. Each example typically consists of a collection of numerical attributes
called features. In the supervised learning problems above, a special feature is designated as the
prediction target, (sometimes called the label or dependent variable). The given features from which
the model must make its predictions can then simply be called the features, (or often, the inputs,
covariates, or independent variables).
If we were working with image data, each individual photograph might constitute an example,
each represented by an ordered list of numerical values corresponding to the brightness of each
pixel. A 200 × 200 color photograph would consist of 200 × 200 × 3 = 120000 numerical values,
corresponding to the brightness of the red, green, and blue channels for each spatial location.
In a more traditional task, we might try to predict whether or not a patient will survive, given a
standard set of features such as age, vital signs, diagnoses, etc.
When every example is characterized by the same number of numerical values, we say that the
data consists of fixed-length vectors and we describe the (constant) length of the vectors as the
dimensionality of the data. As you might imagine, fixed-length can be a convenient property. If we
wanted to train a model to recognize cancer in microscopy images, fixed-length inputs mean we
have one less thing to worry about.
However, not all data can easily be represented as fixed-length vectors. While we might expect
microscope images to come from standard equipment, we cannot expect images mined from the
Internet to all show up with the same resolution or shape. For images, we might consider crop-
ping them all to a standard size, but that strategy only gets us so far. We risk losing information
in the cropped out portions. Moreover, text data resists fixed-length representations even more
stubbornly. Consider the customer reviews left on e-commerce sites like Amazon, IMDB, or Tri-
pAdvisor. Some are short: “it stinks!”. Others ramble for pages. One major advantage of deep
learning over traditional methods is the comparative grace with which modern models can han-
dle varying-length data.
Generally, the more data we have, the easier our job becomes. When we have more data, we
can train more powerful models and rely less heavily on pre-conceived assumptions. The regime
change from (comparatively) small to big data is a major contributor to the success of modern
deep learning. To drive the point home, many of the most exciting models in deep learning do
not work without large datasets. Some others work in the low-data regime, but are no better than
traditional approaches.
Finally, it is not enough to have lots of data and to process it cleverly. We need the right data. If
the data is full of mistakes, or if the chosen features are not predictive of the target quantity of
interest, learning is going to fail. The situation is captured well by the cliché: garbage in, garbage
out. Moreover, poor predictive performance is not the only potential consequence. In sensitive
applications of machine learning, like predictive policing, resumé screening, and risk models
used for lending, we must be especially alert to the consequences of garbage data. One common
failure mode occurs in datasets where some groups of people are unrepresented in the training
data. Imagine applying a skin cancer recognition system in the wild that had never seen black
skin before. Failure can also occur when the data does not merely under-represent some groups
but reflects societal prejudices. For example, if past hiring decisions are used to train a predictive
model that will be used to screen resumes, then machine learning models could inadvertently
1.2. The Key Components: Data, Models, and Algorithms 21

capture and automate historical injustices. Note that this can all happen without the data scientist
actively conspiring, or even being aware.
1.2.2 Models
Most machine learning involves transforming the data in some sense. We might want to build a
system that ingests photos and predicts smiley-ness. Alternatively, we might want to ingest a set of
sensor readings and predict how normal vs anomalous the readings are. By model, we denote the
computational machinery for ingesting data of one type, and spitting out predictions of a possibly
different type. In particular, we are interested in statistical models that can be estimated from
data. While simple models are perfectly capable of addressing appropriately simple problems the
problems that we focus on in this book stretch the limits of classical methods. Deep learning is
differentiated from classical approaches principally by the set of powerful models that it focuses
on. These models consist of many successive transformations of the data that are chained together
top to bottom, thus the name deep learning. On our way to discussing deep neural networks, we
will discuss some more traditional methods.
1.2.3 Objective functions
Earlier, we introduced machine learning as “learning from experience”. By learning here, we mean
improving at some task over time. But who is to say what constitutes an improvement? You might
imagine that we could propose to update our model, and some people might disagree on whether
the proposed update constituted an improvement or a decline.
In order to develop a formal mathematical system of learning machines, we need to have formal
measures of how good (or bad) our models are. In machine learning, and optimization more
generally, we call these objective functions. By convention, we usually define objective functions
so that lower is better. This is merely a convention. You can take any function f for which higher is
better, and turn it into a new function f ′ that is qualitatively identical but for which lower is better
by setting f ′ = −f . Because lower is better, these functions are sometimes called loss functions or
cost functions.
When trying to predict numerical values, the most common objective function is squared error
(y−ŷ)2 . For classification, the most common objective is to minimize error rate, i.e., the fraction of
instances on which our predictions disagree with the ground truth. Some objectives (like squared
error) are easy to optimize. Others (like error rate) are difficult to optimize directly, owing to
non-differentiability or other complications. In these cases, it is common to optimize a surrogate
objective.
Typically, the loss function is defined with respect to the modelʼs parameters and depends upon the
dataset. The best values of our modelʼs parameters are learned by minimizing the loss incurred on
a training set consisting of some number of examples collected for training. However, doing well on
the training data does not guarantee that we will do well on (unseen) test data. So we will typically
want to split the available data into two partitions: the training data (for fitting model parameters)
and the test data (which is held out for evaluation), reporting the following two quantities:
• Training Error: The error on that data on which the model was trained. You could think of
this as being like a studentʼs scores on practice exams used to prepare for some real exam.
Even if the results are encouraging, that does not guarantee success on the final exam.
• Test Error: This is the error incurred on an unseen test set. This can deviate significantly
from the training error. When a model performs well on the training data but fails to gen-
eralize to unseen data, we say that it is overfitting. In real-life terms, this is like flunking the
real exam despite doing well on practice exams.
1.2.4 Optimization algorithms
Once we have got some data source and representation, a model, and a well-defined objective
function, we need an algorithm capable of searching for the best possible parameters for mini-
mizing the loss function. The most popular optimization algorithms for neural networks follow
an approach called gradient descent. In short, at each step, they check to see, for each parameter,
which way the training set loss would move if you perturbed that parameter just a small amount.
They then update the parameter in the direction that reduces the loss.
1.3 Kinds of Machine Learning
In the following sections, we discuss a few kinds of machine learning problems in greater detail.
We begin with a list of objectives, i.e., a list of things that we would like machine learning to do.
Note that the objectives are complemented with a set of techniques of how to accomplish them,
including types of data, models, training techniques, etc. The list below is just a sampling of the
problems ML can tackle to motivate the reader and provide us with some common language for
when we talk about more problems throughout the book.
1.3.1 Supervised learning
Supervised learning addresses the task of predicting targets given inputs. The targets, which we
often call labels, are generally denoted by y. The input data, also called the features or covariates,
are typically denoted x. Each (input, target) pair is called an example or instance. Sometimes, when
the context is clear, we may use the term examples, to refer to a collection of inputs, even when the
corresponding targets are unknown. We denote any particular instance with a subscript, typically
i, for instance (xi , yi ). A dataset is a collection of n instances {xi , yi }ni=1 . Our goal is to produce a
model fθ that maps any input xi to a prediction fθ (xi ).
To ground this description in a concrete example, if we were working in healthcare, then we might
want to predict whether or not a patient would have a heart attack. This observation, heart attack
or no heart attack, would be our label y. The input data x might be vital signs such as heart rate,
diastolic and systolic blood pressure, etc.
The supervision comes into play because for choosing the parameters θ, we (the supervisors)
provide the model with a dataset consisting of labeled examples (xi , yi ), where each example xi
is matched with the correct label.
In probabilistic terms, we typically are interested in estimating the conditional probability P (y|x).
While it is just one among several paradigms within machine learning, supervised learning ac-
counts for the majority of successful applications of machine learning in industry. Partly, that is
because many important tasks can be described crisply as estimating the probability of something
unknown given a particular set of available data:
• Predict cancer vs not cancer, given a CT image.
• Predict the correct translation in French, given a sentence in English.
• Predict the price of a stock next month based on this monthʼs financial reporting data.
1.3. Kinds of Machine Learning 23

Even with the simple description “predict targets from inputs” supervised learning can take a great
many forms and require a great many modeling decisions, depending on (among other considera-
tions) the type, size, and the number of inputs and outputs. For example, we use different models
to process sequences (like strings of text or time series data) and for processing fixed-length vec-
tor representations. We will visit many of these problems in depth throughout the first 9 parts of
this book.
Informally, the learning process looks something like this: Grab a big collection of examples for
which the covariates are known and select from them a random subset, acquiring the ground truth
labels for each. Sometimes these labels might be available data that has already been collected
(e.g., did a patient die within the following year?) and other times we might need to employ human
annotators to label the data, (e.g., assigning images to categories).
Together, these inputs and corresponding labels comprise the training set. We feed the training
dataset into a supervised learning algorithm, a function that takes as input a dataset and outputs
another function, the learned model. Finally, we can feed previously unseen inputs to the learned
model, using its outputs as predictions of the corresponding label. The full process is drawn in
Fig. 1.3.1.
Fig. 1.3.1: Supervised learning.
Regression
Perhaps the simplest supervised learning task to wrap your head around is regression. Consider,
for example, a set of data harvested from a database of home sales. We might construct a table,
where each row corresponds to a different house, and each column corresponds to some relevant
attribute, such as the square footage of a house, the number of bedrooms, the number of bath-
rooms, and the number of minutes (walking) to the center of town. In this dataset, each example
would be a specific house, and the corresponding feature vector would be one row in the table.
If you live in New York or San Francisco, and you are not the CEO of Amazon, Google, Microsoft, or
Facebook, the (sq. footage, no. of bedrooms, no. of bathrooms, walking distance) feature vector
for your home might look something like: [100, 0, .5, 60]. However, if you live in Pittsburgh, it
might look more like [3000, 4, 3, 10]. Feature vectors like this are essential for most classic machine
learning algorithms. We will continue to denote the feature vector corresponding to any example
i as xi and we can compactly refer to the full table containing all of the feature vectors as X.
What makes a problem a regression is actually the outputs. Say that you are in the market for a new
home. You might want to estimate the fair market value of a house, given some features like these.
The target value, the price of sale, is a real number. If you remember the formal definition of the
reals you might be scratching your head now. Homes probably never sell for fractions of a cent,
let alone prices expressed as irrational numbers. In cases like this, when the target is actually
discrete, but where the rounding takes place on a sufficiently fine scale, we will abuse language
just a bit and continue to describe our outputs and targets as real-valued numbers.
We denote any individual target yi (corresponding to example xi ) and the set of all targets y (cor-
responding to all examples X). When our targets take on arbitrary values in some range, we call
this a regression problem. Our goal is to produce a model whose predictions closely approximate
the actual target values. We denote the predicted target for any instance ŷi . Do not worry if the
notation is bogging you down. We will unpack it more thoroughly in the subsequent chapters.
Lots of practical problems are well-described regression problems. Predicting the rating that a
user will assign to a movie can be thought of as a regression problem and if you designed a great
algorithm to accomplish this feat in 2009, you might have won the 1-million-dollar Netflix prize13 .
Predicting the length of stay for patients in the hospital is also a regression problem. A good rule
of thumb is that any How much? or How many? problem should suggest regression.
• “How many hours will this surgery take?”: regression
• “How many dogs are in this photo?”: regression.
However, if you can easily pose your problem as “Is this a _ ?”, then it is likely, classification, a
different kind of supervised problem that we will cover next. Even if you have never worked with
machine learning before, you have probably worked through a regression problem informally.
Imagine, for example, that you had your drains repaired and that your contractor spent x1 = 3
hours removing gunk from your sewage pipes. Then he sent you a bill of y1 = $350. Now imagine
that your friend hired the same contractor for x2 = 2 hours and that he received a bill of y2 = $250.
If someone then asked you how much to expect on their upcoming gunk-removal invoice you
might make some reasonable assumptions, such as more hours worked costs more dollars. You
might also assume that there is some base charge and that the contractor then charges per hour.
If these assumptions held true, then given these two data examples, you could already identify the
contractorʼs pricing structure: $100 per hour plus $50 to show up at your house. If you followed
that much then you already understand the high-level idea behind linear regression (and you just
implicitly designed a linear model with a bias term).
In this case, we could produce the parameters that exactly matched the contractorʼs prices. Some-
times that is not possible, e.g., if some of the variance owes to some factors besides your two fea-
tures. In these cases, we will try to learn models that minimize the distance between our predic-
tions and the observed values. In most of our chapters, we will focus on one of two very common
losses, the L1 loss where
∑
l(y, y ′ ) = |yi − yi′ | (1.3.1)
i
and the least mean squares loss, or L2 loss where

∑
l(y, y ′ ) = (yi − yi′ )2 . (1.3.2)
i
As we will see later, the L2 loss corresponds to the assumption that our data was corrupted by
Gaussian noise, whereas the L1 loss corresponds to an assumption of noise from a Laplace distri-
bution.
13
https://en.wikipedia.org/wiki/Netflix_Prize

Classification
While regression models are great for addressing how many? questions, lots of problems do not
bend comfortably to this template. For example, a bank wants to add check scanning to its mobile
app. This would involve the customer snapping a photo of a check with their smart phoneʼs camera
and the machine learning model would need to be able to automatically understand text seen in
the image. It would also need to understand hand-written text to be even more robust. This kind of
system is referred to as optical character recognition (OCR), and the kind of problem it addresses
is called classification. It is treated with a different set of algorithms than those used for regression
(although many techniques will carry over).
In classification, we want our model to look at a feature vector, e.g., the pixel values in an image,
and then predict which category (formally called classes), among some (discrete) set of options, an
example belongs. For hand-written digits, we might have 10 classes, corresponding to the digits 0
through 9. The simplest form of classification is when there are only two classes, a problem which
we call binary classification. For example, our dataset X could consist of images of animals and
our labels Y might be the classes {cat, dog}. While in regression, we sought a regressor to output a
real value ŷ, in classification, we seek a classifier, whose output ŷ is the predicted class assignment.
For reasons that we will get into as the book gets more technical, it can be hard to optimize a
model that can only output a hard categorical assignment, e.g., either cat or dog. In these cases,
it is usually much easier to instead express our model in the language of probabilities. Given an
example x, our model assigns a probability ŷk to each label k. Because these are probabilities,
they need to be positive numbers and add up to 1 and thus we only need K − 1 numbers to assign
probabilities of K categories. This is easy to see for binary classification. If there is a 0.6 (60%)
probability that an unfair coin comes up heads, then there is a 0.4 (40%) probability that it comes
up tails. Returning to our animal classification example, a classifier might see an image and output
the probability that the image is a cat P (y = cat | x) = 0.9. We can interpret this number by saying
that the classifier is 90% sure that the image depicts a cat. The magnitude of the probability for
the predicted class conveys one notion of uncertainty. It is not the only notion of uncertainty and
we will discuss others in more advanced chapters.
When we have more than two possible classes, we call the problem multiclass classification. Com-
mon examples include hand-written character recognition [0, 1, 2, 3 ... 9, a, b, c, ...].
While we attacked regression problems by trying to minimize the L1 or L2 loss functions, the
common loss function for classification problems is called cross-entropy.
Note that the most likely class is not necessarily the one that you are going to use for your decision.
Assume that you find this beautiful mushroom in your backyard as shown in Fig. 1.3.2.
Fig. 1.3.2: Death cap—do not eat!
Now, assume that you built a classifier and trained it to predict if a mushroom is poisonous based
on a photograph. Say our poison-detection classifier outputs P (y = deathcap|image) = 0.2. In
other words, the classifier is 80% sure that our mushroom is not a death cap. Still, you would have
to be a fool to eat it. That is because the certain benefit of a delicious dinner is not worth a 20%
risk of dying from it. In other words, the effect of the uncertain risk outweighs the benefit by far.
We can look at this more formally. Basically, we need to compute the expected risk that we incur,
i.e., we need to multiply the probability of the outcome with the benefit (or harm) associated with
it:
L(action|x) = Ey∼p(y|x) [loss(action, y)]. (1.3.3)
Hence, the loss L incurred by eating the mushroom is L(a = eat|x) = 0.2 ∗ ∞ + 0.8 ∗ 0 = ∞,
whereas the cost of discarding it is L(a = discard|x) = 0.2 ∗ 0 + 0.8 ∗ 1 = 0.8.
Our caution was justified: as any mycologist would tell us, the above mushroom actually is a death
cap. Classification can get much more complicated than just binary, multiclass, or even multi-
label classification. For instance, there are some variants of classification for addressing hierar-
chies. Hierarchies assume that there exist some relationships among the many classes. So not all
errors are equal—if we must err, we would prefer to misclassify to a related class rather than to a
distant class. Usually, this is referred to as hierarchical classification. One early example is due to
Linnaeus14 , who organized the animals in a hierarchy.
In the case of animal classification, it might not be so bad to mistake a poodle for a schnauzer,
but our model would pay a huge penalty if it confused a poodle for a dinosaur. Which hierarchy
is relevant might depend on how you plan to use the model. For example, rattle snakes and garter
snakes might be close on the phylogenetic tree, but mistaking a rattler for a garter could be deadly.
14
https://en.wikipedia.org/wiki/Carl_Linnaeus

Tagging
Some classification problems do not fit neatly into the binary or multiclass classification setups.
For example, we could train a normal binary classifier to distinguish cats from dogs. Given the
current state of computer vision, we can do this easily, with off-the-shelf tools. Nonetheless, no
matter how accurate our model gets, we might find ourselves in trouble when the classifier en-
counters an image of the Town Musicians of Bremen.
Fig. 1.3.3: A cat, a rooster, a dog and a donkey
As you can see, there is a cat in the picture, and a rooster, a dog, a donkey, and a bird, with some
trees in the background. Depending on what we want to do with our model ultimately, treating
this as a binary classification problem might not make a lot of sense. Instead, we might want to
give the model the option of saying the image depicts a cat and a dog and a donkey and a rooster
and a bird.
The problem of learning to predict classes that are not mutually exclusive is called multi-label clas-
sification. Auto-tagging problems are typically best described as multi-label classification prob-
lems. Think of the tags people might apply to posts on a tech blog, e.g., “machine learning”, “tech-
nology”, “gadgets”, “programming languages”, “linux”, “cloud computing”, “AWS”. A typical article
might have 5-10 tags applied because these concepts are correlated. Posts about “cloud comput-
ing” are likely to mention “AWS” and posts about “machine learning” could also deal with “pro-
gramming languages”.
We also have to deal with this kind of problem when dealing with the biomedical literature, where
correctly tagging articles is important because it allows researchers to do exhaustive reviews of
the literature. At the National Library of Medicine, a number of professional annotators go over
each article that gets indexed in PubMed to associate it with the relevant terms from MeSH, a
collection of roughly 28k tags. This is a time-consuming process and the annotators typically have
a one year lag between archiving and tagging. Machine learning can be used here to provide
provisional tags until each article can have a proper manual review. Indeed, for several years, the
BioASQ organization has hosted a competition15 to do precisely this.
Search and ranking
Sometimes we do not just want to assign each example to a bucket or to a real value. In the field of
information retrieval, we want to impose a ranking on a set of items. Take web search for example,
the goal is less to determine whether a particular page is relevant for a query, but rather, which
one of the plethora of search results is most relevant for a particular user. We really care about
the ordering of the relevant search results and our learning algorithm needs to produce ordered
subsets of elements from a larger set. In other words, if we are asked to produce the first 5 letters
from the alphabet, there is a difference between returning A B C D E and C A B E D. Even if the
result set is the same, the ordering within the set matters.
One possible solution to this problem is to first assign to every element in the set a corresponding
relevance score and then to retrieve the top-rated elements. PageRank16 , the original secret sauce
behind the Google search engine was an early example of such a scoring system but it was peculiar
in that it did not depend on the actual query. Here they relied on a simple relevance filter to
identify the set of relevant items and then on PageRank to order those results that contained the
query term. Nowadays, search engines use machine learning and behavioral models to obtain
query-dependent relevance scores. There are entire academic conferences devoted to this subject.
Recommender systems
Recommender systems are another problem setting that is related to search and ranking. The
problems are similar insofar as the goal is to display a set of relevant items to the user. The main
difference is the emphasis on personalization to specific users in the context of recommender sys-
tems. For instance, for movie recommendations, the results page for a SciFi fan and the results
page for a connoisseur of Peter Sellers comedies might differ significantly. Similar problems pop
up in other recommendation settings, e.g., for retail products, music, or news recommendation.
In some cases, customers provide explicit feedback communicating how much they liked a partic-
ular product (e.g., the product ratings and reviews on Amazon, IMDB, GoodReads, etc.). In some
other cases, they provide implicit feedback, e.g., by skipping titles on a playlist, which might in-
dicate dissatisfaction but might just indicate that the song was inappropriate in context. In the
simplest formulations, these systems are trained to estimate some score yij , such as an estimated
rating or the probability of purchase, given a user ui and product pj .
Given such a model, then for any given user, we could retrieve the set of objects with the largest
scores yij , which could then be recommended to the customer. Production systems are consid-
erably more advanced and take detailed user activity and item characteristics into account when
computing such scores. Fig. 1.3.4 is an example of deep learning books recommended by Amazon
based on personalization algorithms tuned to capture the authorʼs preferences.
15
http://bioasq.org/
16
https://en.wikipedia.org/wiki/PageRank

Fig. 1.3.4: Deep learning books recommended by Amazon.
Despite their tremendous economic value, recommendation systems naively built on top of pre-
dictive models suffer some serious conceptual flaws. To start, we only observe censored feedback.
Users preferentially rate movies that they feel strongly about: you might notice that items receive
many 5 and 1 star ratings but that there are conspicuously few 3-star ratings. Moreover, current
purchase habits are often a result of the recommendation algorithm currently in place, but learn-
ing algorithms do not always take this detail into account. Thus it is possible for feedback loops
to form where a recommender system preferentially pushes an item that is then taken to be bet-
ter (due to greater purchases) and in turn is recommended even more frequently. Many of these
problems about how to deal with censoring, incentives, and feedback loops, are important open
research questions.
Sequence Learning
So far, we have looked at problems where we have some fixed number of inputs and produce a fixed
number of outputs. Before we considered predicting home prices from a fixed set of features:
square footage, number of bedrooms, number of bathrooms, walking time to downtown. We
also discussed mapping from an image (of fixed dimension) to the predicted probabilities that it
belongs to each of a fixed number of classes, or taking a user ID and a product ID, and predicting
a star rating. In these cases, once we feed our fixed-length input into the model to generate an
output, the model immediately forgets what it just saw.
This might be fine if our inputs truly all have the same dimensions and if successive inputs truly
have nothing to do with each other. But how would we deal with video snippets? In this case,
each snippet might consist of a different number of frames. And our guess of what is going on in
each frame might be much stronger if we take into account the previous or succeeding frames.
Same goes for language. One popular deep learning problem is machine translation: the task of
ingesting sentences in some source language and predicting their translation in another language.
These problems also occur in medicine. We might want a model to monitor patients in the in-
tensive care unit and to fire off alerts if their risk of death in the next 24 hours exceeds some
threshold. We definitely would not want this model to throw away everything it knows about the
patient history each hour and just make its predictions based on the most recent measurements.
These problems are among the most exciting applications of machine learning and they are in-
stances of sequence learning. They require a model to either ingest sequences of inputs or to emit
sequences of outputs (or both!). These latter problems are sometimes referred to as seq2seq prob-
lems. Language translation is a seq2seq problem. Transcribing text from the spoken speech is
also a seq2seq problem. While it is impossible to consider all types of sequence transformations,
a number of special cases are worth mentioning:
Tagging and Parsing. This involves annotating a text sequence with attributes. In other words,
the number of inputs and outputs is essentially the same. For instance, we might want to know
where the verbs and subjects are. Alternatively, we might want to know which words are the
named entities. In general, the goal is to decompose and annotate text based on structural and
grammatical assumptions to get some annotation. This sounds more complex than it actually is.
Below is a very simple example of annotating a sentence with tags indicating which words refer
to named entities.
Tom has dinner in Washington with Sally.

Ent - - - Ent - Ent
Automatic Speech Recognition. With speech recognition, the input sequence x is an audio
recording of a speaker (shown in Fig. 1.3.5), and the output y is the textual transcript of what the
speaker said. The challenge is that there are many more audio frames (sound is typically sampled
at 8kHz or 16kHz) than text, i.e., there is no 1:1 correspondence between audio and text, since
thousands of samples correspond to a single spoken word. These are seq2seq problems where
the output is much shorter than the input.
Fig. 1.3.5: -D-e-e-p- L-ea-r-ni-ng-
Text to Speech. Text-to-Speech (TTS) is the inverse of speech recognition. In other words, the
input x is text and the output y is an audio file. In this case, the output is much longer than the
input. While it is easy for humans to recognize a bad audio file, this is not quite so trivial for
computers.
Machine Translation. Unlike the case of speech recognition, where corresponding inputs and
outputs occur in the same order (after alignment), in machine translation, order inversion can be
vital. In other words, while we are still converting one sequence into another, neither the number
of inputs and outputs nor the order of corresponding data examples are assumed to be the same.

Consider the following illustrative example of the peculiar tendency of Germans to place the verbs
at the end of sentences.
German: Haben Sie sich schon dieses grossartige Lehrwerk angeschaut?

English: Did you already check out this excellent tutorial?
Wrong alignment: Did you yourself already this excellent tutorial looked-at?
Many related problems pop up in other learning tasks. For instance, determining the order in
which a user reads a Webpage is a two-dimensional layout analysis problem. Dialogue problems
exhibit all kinds of additional complications, where determining what to say next requires taking
into account real-world knowledge and the prior state of the conversation across long temporal
distances. This is an active area of research.
1.3.2 Unsupervised learning
All the examples so far were related to Supervised Learning, i.e., situations where we feed the model
a giant dataset containing both the features and corresponding target values. You could think of
the supervised learner as having an extremely specialized job and an extremely anal boss. The
boss stands over your shoulder and tells you exactly what to do in every situation until you learn
to map from situations to actions. Working for such a boss sounds pretty lame. On the other hand,
it is easy to please this boss. You just recognize the pattern as quickly as possible and imitate their
actions.
In a completely opposite way, it could be frustrating to work for a boss who has no idea what
they want you to do. However, if you plan to be a data scientist, youʼd better get used to it. The
boss might just hand you a giant dump of data and tell you to do some data science with it! This
sounds vague because it is. We call this class of problems unsupervised learning, and the type and
number of questions we could ask is limited only by our creativity. We will address a number of
unsupervised learning techniques in later chapters. To whet your appetite for now, we describe a
few of the questions you might ask:
• Can we find a small number of prototypes that accurately summarize the data? Given a set of
photos, can we group them into landscape photos, pictures of dogs, babies, cats, mountain
peaks, etc.? Likewise, given a collection of usersʼ browsing activity, can we group them into
users with similar behavior? This problem is typically known as clustering.
• Can we find a small number of parameters that accurately capture the relevant properties of
the data? The trajectories of a ball are quite well described by velocity, diameter, and mass
of the ball. Tailors have developed a small number of parameters that describe human body
shape fairly accurately for the purpose of fitting clothes. These problems are referred to
as subspace estimation problems. If the dependence is linear, it is called principal component
analysis.
• Is there a representation of (arbitrarily structured) objects in Euclidean space (i.e., the space
of vectors in Rn ) such that symbolic properties can be well matched? This is called represen-
tation learning and it is used to describe entities and their relations, such as Rome − Italy +
France = Paris.
• Is there a description of the root causes of much of the data that we observe? For instance,
if we have demographic data about house prices, pollution, crime, location, education,
salaries, etc., can we discover how they are related simply based on empirical data? The
fields concerned with causality and probabilistic graphical models address this problem.
• Another important and exciting recent development in unsupervised learning is the advent
of generative adversarial networks (GANs). These give us a procedural way to synthesize data,
even complicated structured data like images and audio. The underlying statistical mech-
anisms are tests to check whether real and fake data are the same. We will devote a few
notebooks to them.
1.3.3 Interacting with an Environment
So far, we have not discussed where data actually comes from, or what actually happens when a
machine learning model generates an output. That is because supervised learning and unsuper-
vised learning do not address these issues in a very sophisticated way. In either case, we grab a big
pile of data upfront, then set our pattern recognition machines in motion without ever interacting
with the environment again. Because all of the learning takes place after the algorithm is discon-
nected from the environment, this is sometimes called offline learning. For supervised learning,
the process looks like Fig. 1.3.6.
Fig. 1.3.6: Collect data for supervised learning from an environment.
This simplicity of offline learning has its charms. The upside is we can worry about pattern recog-
nition in isolation, without any distraction from these other problems. But the downside is that
the problem formulation is quite limiting. If you are more ambitious, or if you grew up reading
Asimovʼs Robot Series, then you might imagine artificially intelligent bots capable not only of mak-
ing predictions, but of taking actions in the world. We want to think about intelligent agents, not
just predictive models. That means we need to think about choosing actions, not just making predic-
tions. Moreover, unlike predictions, actions actually impact the environment. If we want to train
an intelligent agent, we must account for the way its actions might impact the future observations
of the agent.
Considering the interaction with an environment opens a whole set of new modeling questions.
Does the environment:
• Remember what we did previously?
• Want to help us, e.g., a user reading text into a speech recognizer?
• Want to beat us, i.e., an adversarial setting like spam filtering (against spammers) or playing
a game (vs an opponent)?
• Not care (as in many cases)?
• Have shifting dynamics (does future data always resemble the past or do the patterns change
over time, either naturally or in response to our automated tools)?

This last question raises the problem of distribution shift, (when training and test data are differ-
ent). It is a problem that most of us have experienced when taking exams written by a lecturer,
while the homeworks were composed by his TAs. We will briefly describe reinforcement learning
and adversarial learning, two settings that explicitly consider interaction with an environment.
1.3.4 Reinforcement learning
If you are interested in using machine learning to develop an agent that interacts with an environ-
ment and takes actions, then you are probably going to wind up focusing on reinforcement learning
(RL). This might include applications to robotics, to dialogue systems, and even to developing AI
for video games. Deep reinforcement learning (DRL), which applies deep neural networks to RL
problems, has surged in popularity. The breakthrough deep Q-network that beat humans at Atari
games using only the visual input17 , and the AlphaGo program that dethroned the world champion
at the board game Go18 are two prominent examples.
Reinforcement learning gives a very general statement of a problem, in which an agent interacts
with an environment over a series of timesteps. At each timestep t, the agent receives some ob-
servation ot from the environment and must choose an action at that is subsequently transmitted
back to the environment via some mechanism (sometimes called an actuator). Finally, the agent
receives a reward rt from the environment. The agent then receives a subsequent observation,
and chooses a subsequent action, and so on. The behavior of an RL agent is governed by a policy.
In short, a policy is just a function that maps from observations (of the environment) to actions.
The goal of reinforcement learning is to produce a good policy.
Fig. 1.3.7: The interaction between reinforcement learning and an environment.
It is hard to overstate the generality of the RL framework. For example, we can cast any supervised
learning problem as an RL problem. Say we had a classification problem. We could create an RL
agent with one action corresponding to each class. We could then create an environment which
gave a reward that was exactly equal to the loss function from the original supervised problem.
That being said, RL can also address many problems that supervised learning cannot. For exam-
ple, in supervised learning we always expect that the training input comes associated with the
correct label. But in RL, we do not assume that for each observation, the environment tells us the
optimal action. In general, we just get some reward. Moreover, the environment may not even
tell us which actions led to the reward.
Consider for example the game of chess. The only real reward signal comes at the end of the
game when we either win, which we might assign a reward of 1, or when we lose, which we could
17
https://www.wired.com/2015/02/google-ai-plays-atari-like-pros/
18
https://www.wired.com/2017/05/googles-alphago-trounces-humans-also-gives-boost/
assign a reward of -1. So reinforcement learners must deal with the credit assignment problem:
determining which actions to credit or blame for an outcome. The same goes for an employee
who gets a promotion on October 11. That promotion likely reflects a large number of well-chosen
actions over the previous year. Getting more promotions in the future requires figuring out what
actions along the way led to the promotion.
Reinforcement learners may also have to deal with the problem of partial observability. That is,
the current observation might not tell you everything about your current state. Say a cleaning
robot found itself trapped in one of many identical closets in a house. Inferring the precise lo-
cation (and thus state) of the robot might require considering its previous observations before
entering the closet.
Finally, at any given point, reinforcement learners might know of one good policy, but there might
be many other better policies that the agent has never tried. The reinforcement learner must
constantly choose whether to exploit the best currently-known strategy as a policy, or to explore
the space of strategies, potentially giving up some short-run reward in exchange for knowledge.
MDPs, bandits, and friends
The general reinforcement learning problem is a very general setting. Actions affect subsequent
observations. Rewards are only observed corresponding to the chosen actions. The environment
may be either fully or partially observed. Accounting for all this complexity at once may ask too
much of researchers. Moreover, not every practical problem exhibits all this complexity. As a
result, researchers have studied a number of special cases of reinforcement learning problems.
When the environment is fully observed, we call the RL problem a Markov Decision Process (MDP).
When the state does not depend on the previous actions, we call the problem a contextual bandit
problem. When there is no state, just a set of available actions with initially unknown rewards, this
problem is the classic multi-armed bandit problem.
1.4 Roots
Although many deep learning methods are recent inventions, humans have held the desire to an-
alyze data and to predict future outcomes for centuries. In fact, much of natural science has its
roots in this. For instance, the Bernoulli distribution is named after Jacob Bernoulli (1655-1705)19 ,
and the Gaussian distribution was discovered by Carl Friedrich Gauss (1777-1855)20 . He invented,
for instance, the least mean squares algorithm, which is still used today for countless problems
from insurance calculations to medical diagnostics. These tools gave rise to an experimental ap-
proach in the natural sciences—for instance, Ohmʼs law relating current and voltage in a resistor
is perfectly described by a linear model.
Even in the middle ages, mathematicians had a keen intuition of estimates. For instance, the
geometry book of Jacob Köbel (1460-1533)21 illustrates averaging the length of 16 adult menʼs feet
to obtain the average foot length.
19
https://en.wikipedia.org/wiki/Jacob_Bernoulli
20
https://en.wikipedia.org/wiki/Carl_Friedrich_Gauss
21
https://www.maa.org/press/periodicals/convergence/mathematical-treasures-jacob-kobels-geometry
1.4. Roots 35
Fig. 1.4.1: Estimating the length of a foot
Fig. 1.4.1 illustrates how this estimator works. The 16 adult men were asked to line up in a row,
when leaving church. Their aggregate length was then divided by 16 to obtain an estimate for
what now amounts to 1 foot. This “algorithm” was later improved to deal with misshapen feet—
the 2 men with the shortest and longest feet respectively were sent away, averaging only over the
remainder. This is one of the earliest examples of the trimmed mean estimate.
Statistics really took off with the collection and availability of data. One of its titans, Ronald Fisher
(1890-1962)22 , contributed significantly to its theory and also its applications in genetics. Many of
his algorithms (such as Linear Discriminant Analysis) and formula (such as the Fisher Information
Matrix) are still in frequent use today (even the Iris dataset that he released in 1936 is still used
sometimes to illustrate machine learning algorithms). Fisher was also a proponent of eugenics,
which should remind us that the morally dubious use of data science has as long and enduring a
history as its productive use in industry and the natural sciences.
A second influence for machine learning came from Information Theory (Claude Shannon, 1916-
2001)23 and the Theory of computation via Alan Turing (1912-1954)24 . Turing posed the question
“can machines think?” in his famous paper Computing machinery and intelligence25 (Mind, Oc-
tober 1950). In what he described as the Turing test, a machine can be considered intelligent if it
is difficult for a human evaluator to distinguish between the replies from a machine and a human
based on textual interactions.
Another influence can be found in neuroscience and psychology. After all, humans clearly exhibit
intelligent behavior. It is thus only reasonable to ask whether one could explain and possibly re-
22
https://en.wikipedia.org/wiki/Ronald_Fisher
23
https://en.wikipedia.org/wiki/Claude_Shannon
24
https://en.wikipedia.org/wiki/Alan_Turing
25
https://en.wikipedia.org/wiki/Computing_Machinery_and_Intelligence
verse engineer this capacity. One of the oldest algorithms inspired in this fashion was formulated
by Donald Hebb (1904-1985)26 . In his groundbreaking book The Organization of Behavior (Hebb
& Hebb, 1949), he posited that neurons learn by positive reinforcement. This became known as
the Hebbian learning rule. It is the prototype of Rosenblattʼs perceptron learning algorithm and it
laid the foundations of many stochastic gradient descent algorithms that underpin deep learning
today: reinforce desirable behavior and diminish undesirable behavior to obtain good settings of
the parameters in a neural network.
Biological inspiration is what gave neural networks their name. For over a century (dating back
to the models of Alexander Bain, 1873 and James Sherrington, 1890), researchers have tried to
assemble computational circuits that resemble networks of interacting neurons. Over time, the
interpretation of biology has become less literal but the name stuck. At its heart, lie a few key
principles that can be found in most networks today:
• The alternation of linear and nonlinear processing units, often referred to as layers.
• The use of the chain rule (also known as backpropagation) for adjusting parameters in the
entire network at once.
After initial rapid progress, research in neural networks languished from around 1995 until 2005.
This was due to a number of reasons. Training a network is computationally very expensive.
While RAM was plentiful at the end of the past century, computational power was scarce. Sec-
ond, datasets were relatively small. In fact, Fisherʼs Iris dataset from 1932 was a popular tool for
testing the efficacy of algorithms. MNIST with its 60,000 handwritten digits was considered huge.
Given the scarcity of data and computation, strong statistical tools such as Kernel Methods, Deci-
sion Trees and Graphical Models proved empirically superior. Unlike neural networks, they did
not require weeks to train and provided predictable results with strong theoretical guarantees.
1.5 The Road to Deep Learning
Much of this changed with the ready availability of large amounts of data, due to the World Wide
Web, the advent of companies serving hundreds of millions of users online, a dissemination of
cheap, high-quality sensors, cheap data storage (Kryderʼs law), and cheap computation (Mooreʼs
law), in particular in the form of GPUs, originally engineered for computer gaming. Suddenly
algorithms and models that seemed computationally infeasible became relevant (and vice versa).
This is best illustrated in Table 1.5.1.
Table 1.5.1: Dataset vs. computer memory and computa-

tional power
Decade Dataset Mem- Floating Point Calculations per Second
ory
1970 100 (Iris) 1 KB 100 KF (Intel 8080)
1980 1 K (House prices in Boston) 100 KB 1 MF (Intel 80186)
1990 10 K (optical character recognition) 10 MB 10 MF (Intel 80486)
2000 10 M (web pages) 100 MB 1 GF (Intel Core)
2010 10 G (advertising) 1 GB 1 TF (Nvidia C2050)
2020 1 T (social network) 100 GB 1 PF (Nvidia DGX-2)
It is evident that RAM has not kept pace with the growth in data. At the same time, the increase
26
https://en.wikipedia.org/wiki/Donald_O._Hebb
1.5. The Road to Deep Learning 37

in computational power has outpaced that of the data available. This means that statistical mod-
els needed to become more memory efficient (this is typically achieved by adding nonlineari-
ties) while simultaneously being able to spend more time on optimizing these parameters, due
to an increased compute budget. Consequently, the sweet spot in machine learning and statis-
tics moved from (generalized) linear models and kernel methods to deep networks. This is also
one of the reasons why many of the mainstays of deep learning, such as multilayer perceptrons
(McCulloch & Pitts, 1943), convolutional neural networks (LeCun et al., 1998), Long Short-Term
Memory (Hochreiter & Schmidhuber, 1997), and Q-Learning (Watkins & Dayan, 1992), were es-
sentially “rediscovered” in the past decade, after laying comparatively dormant for considerable
time.
The recent progress in statistical models, applications, and algorithms, has sometimes been
likened to the Cambrian Explosion: a moment of rapid progress in the evolution of species. In-
deed, the state of the art is not just a mere consequence of available resources, applied to decades
old algorithms. Note that the list below barely scratches the surface of the ideas that have helped
researchers achieve tremendous progress over the past decade.
• Novel methods for capacity control, such as Dropout (Srivastava et al., 2014) have helped
to mitigate the danger of overfitting. This was achieved by applying noise injection (Bishop,
1995) throughout the network, replacing weights by random variables for training purposes.
• Attention mechanisms solved a second problem that had plagued statistics for over a cen-
tury: how to increase the memory and complexity of a system without increasing the num-
ber of learnable parameters. (Bahdanau et al., 2014) found an elegant solution by using what
can only be viewed as a learnable pointer structure. Rather than having to remember an en-
tire sentence, e.g., for machine translation in a fixed-dimensional representation, all that
needed to be stored was a pointer to the intermediate state of the translation process. This
allowed for significantly increased accuracy for long sentences, since the model no longer
needed to remember the entire sentence before commencing the generation of a new sen-
tence.
• Multi-stage designs, e.g., via the Memory Networks (MemNets) (Sukhbaatar et al., 2015) and
the Neural Programmer-Interpreter (Reed & DeFreitas, 2015) allowed statistical modelers
to describe iterative approaches to reasoning. These tools allow for an internal state of the
deep network to be modified repeatedly, thus carrying out subsequent steps in a chain of
reasoning, similar to how a processor can modify memory for a computation.
• Another key development was the invention of GANs (Goodfellow et al., 2014). Traditionally,
statistical methods for density estimation and generative models focused on finding proper
probability distributions and (often approximate) algorithms for sampling from them. As
a result, these algorithms were largely limited by the lack of flexibility inherent in the sta-
tistical models. The crucial innovation in GANs was to replace the sampler by an arbitrary
algorithm with differentiable parameters. These are then adjusted in such a way that the dis-
criminator (effectively a two-sample test) cannot distinguish fake from real data. Through
the ability to use arbitrary algorithms to generate data, it opened up density estimation to
a wide variety of techniques. Examples of galloping Zebras (Zhu et al., 2017) and of fake
celebrity faces (Karras et al., 2017) are both testimony to this progress. Even amateur doo-
dlers can produce photorealistic images based on just sketches that describe how the layout
of a scene looks like (Park et al., 2019).
• In many cases, a single GPU is insufficient to process the large amounts of data available for
training. Over the past decade the ability to build parallel distributed training algorithms
has improved significantly. One of the key challenges in designing scalable algorithms is
that the workhorse of deep learning optimization, stochastic gradient descent, relies on rel-
atively small minibatches of data to be processed. At the same time, small batches limit the
efficiency of GPUs. Hence, training on 1024 GPUs with a minibatch size of, say 32 images
per batch amounts to an aggregate minibatch of 32k images. Recent work, first by Li (Li,
2017), and subsequently by (You et al., 2017) and (Jia et al., 2018) pushed the size up to 64k
observations, reducing training time for ResNet50 on ImageNet to less than 7 minutes. For
comparison—initially training times were measured in the order of days.
• The ability to parallelize computation has also contributed quite crucially to progress in re-
inforcement learning, at least whenever simulation is an option. This has led to significant
progress in computers achieving superhuman performance in Go, Atari games, Starcraft,
and in physics simulations (e.g., using MuJoCo). See e.g., (Silver et al., 2016) for a descrip-
tion of how to achieve this in AlphaGo. In a nutshell, reinforcement learning works best if
plenty of (state, action, reward) triples are available, i.e., whenever it is possible to try out
lots of things to learn how they relate to each other. Simulation provides such an avenue.
• Deep Learning frameworks have played a crucial role in disseminating ideas. The first
generation of frameworks allowing for easy modeling encompassed Caffe27 , Torch28 , and
Theano29 . Many seminal papers were written using these tools. By now, they have been su-
perseded by TensorFlow30 , often used via its high level API Keras31 , CNTK32 , Caffe 233 , and
Apache MxNet34 . The third generation of tools, namely imperative tools for deep learning,
was arguably spearheaded by Chainer35 , which used a syntax similar to Python NumPy to
describe models. This idea was adopted by both PyTorch36 , the Gluon API37 of MXNet, and
Jax38 . It is the latter group that this course uses to teach deep learning.
The division of labor between systems researchers building better tools and statistical modelers
building better networks has greatly simplified things. For instance, training a linear logistic re-
gression model used to be a nontrivial homework problem, worthy to give to new machine learn-
ing PhD students at Carnegie Mellon University in 2014. By now, this task can be accomplished
with less than 10 lines of code, putting it firmly into the grasp of programmers.
1.6 Success Stories
Artificial Intelligence has a long history of delivering results that would be difficult to accomplish
otherwise. For instance, mail is sorted using optical character recognition. These systems have
been deployed since the 90s (this is, after all, the source of the famous MNIST and USPS sets of
handwritten digits). The same applies to reading checks for bank deposits and scoring creditwor-
thiness of applicants. Financial transactions are checked for fraud automatically. This forms the
backbone of many e-commerce payment systems, such as PayPal, Stripe, AliPay, WeChat, Apple,
Visa, MasterCard. Computer programs for chess have been competitive for decades. Machine
learning feeds search, recommendation, personalization and ranking on the Internet. In other
words, artificial intelligence and machine learning are pervasive, albeit often hidden from sight.
27
https://github.com/BVLC/caffe
28
https://github.com/torch
29
https://github.com/Theano/Theano
30
https://github.com/tensorflow/tensorflow
31
https://github.com/keras-team/keras
32
https://github.com/Microsoft/CNTK
33
https://github.com/caffe2/caffe2
34
https://github.com/apache/incubator-mxnet
35
https://github.com/chainer/chainer
36
https://github.com/pytorch/pytorch
37
https://github.com/apache/incubator-mxnet
38
https://github.com/google/jax
1.6. Success Stories 39

It is only recently that AI has been in the limelight, mostly due to solutions to problems that were
considered intractable previously.
• Intelligent assistants, such as Appleʼs Siri, Amazonʼs Alexa, or Googleʼs assistant are able to
answer spoken questions with a reasonable degree of accuracy. This includes menial tasks
such as turning on light switches (a boon to the disabled) up to making barberʼs appoint-
ments and offering phone support dialog. This is likely the most noticeable sign that AI is
affecting our lives.
• A key ingredient in digital assistants is the ability to recognize speech accurately. Gradually
the accuracy of such systems has increased to the point where they reach human parity
(Xiong et al., 2018) for certain applications.
• Object recognition likewise has come a long way. Estimating the object in a picture was a
fairly challenging task in 2010. On the ImageNet benchmark (Lin et al., 2010) achieved a
top-5 error rate of 28%. By 2017, (Hu et al., 2018) reduced this error rate to 2.25%. Similarly,
stunning results have been achieved for identifying birds, or diagnosing skin cancer.
• Games used to be a bastion of human intelligence. Starting from TDGammon [23], a pro-
gram for playing Backgammon using temporal difference (TD) reinforcement learning, al-
gorithmic and computational progress has led to algorithms for a wide range of applications.
Unlike Backgammon, chess has a much more complex state space and set of actions. Deep-
Blue beat Garry Kasparov, Campbell et al. (Campbell et al., 2002), using massive parallelism,
special purpose hardware and efficient search through the game tree. Go is more difficult
still, due to its huge state space. AlphaGo reached human parity in 2015, (Silver et al., 2016)
using Deep Learning combined with Monte Carlo tree sampling. The challenge in Poker
was that the state space is large and it is not fully observed (we do not know the opponentsʼ
cards). Libratus exceeded human performance in Poker using efficiently structured strate-
gies (Brown & Sandholm, 2017). This illustrates the impressive progress in games and the
fact that advanced algorithms played a crucial part in them.
• Another indication of progress in AI is the advent of self-driving cars and trucks. While
full autonomy is not quite within reach yet, excellent progress has been made in this direc-
tion, with companies such as Tesla, NVIDIA, and Waymo shipping products that enable at
least partial autonomy. What makes full autonomy so challenging is that proper driving re-
quires the ability to perceive, to reason and to incorporate rules into a system. At present,
deep learning is used primarily in the computer vision aspect of these problems. The rest is
heavily tuned by engineers.
Again, the above list barely scratches the surface of where machine learning has impacted prac-
tical applications. For instance, robotics, logistics, computational biology, particle physics, and
astronomy owe some of their most impressive recent advances at least in parts to machine learn-
ing. ML is thus becoming a ubiquitous tool for engineers and scientists.
Frequently, the question of the AI apocalypse, or the AI singularity has been raised in non-
technical articles on AI. The fear is that somehow machine learning systems will become sentient
and decide independently from their programmers (and masters) about things that directly af-
fect the livelihood of humans. To some extent, AI already affects the livelihood of humans in an
immediate way—creditworthiness is assessed automatically, autopilots mostly navigate vehicles,
decisions about whether to grant bail use statistical data as input. More frivolously, we can ask
Alexa to switch on the coffee machine.
Fortunately, we are far from a sentient AI system that is ready to manipulate its human creators
(or burn their coffee). First, AI systems are engineered, trained and deployed in a specific, goal-
oriented manner. While their behavior might give the illusion of general intelligence, it is a com-
bination of rules, heuristics and statistical models that underlie the design. Second, at present
tools for artificial general intelligence simply do not exist that are able to improve themselves, rea-
son about themselves, and that are able to modify, extend and improve their own architecture
while trying to solve general tasks.
A much more pressing concern is how AI is being used in our daily lives. It is likely that many me-
nial tasks fulfilled by truck drivers and shop assistants can and will be automated. Farm robots will
likely reduce the cost for organic farming but they will also automate harvesting operations. This
phase of the industrial revolution may have profound consequences on large swaths of society
(truck drivers and shop assistants are some of the most common jobs in many states). Further-
more, statistical models, when applied without care can lead to racial, gender or age bias and raise
reasonable concerns about procedural fairness if automated to drive consequential decisions. It
is important to ensure that these algorithms are used with care. With what we know today, this
strikes us a much more pressing concern than the potential of malevolent superintelligence to
destroy humanity.
Summary
• Machine learning studies how computer systems can leverage experience (often data) to im-
prove performance at specific tasks. It combines ideas from statistics, data mining, artificial
intelligence, and optimization. Often, it is used as a means of implementing artificially-
intelligent solutions.
• As a class of machine learning, representational learning focuses on how to automatically
find the appropriate way to represent data. This is often accomplished by a progression of
learned transformations.
• Much of the recent progress in deep learning has been triggered by an abundance of data
arising from cheap sensors and Internet-scale applications, and by significant progress in
computation, mostly through GPUs.
• Whole system optimization is a key component in obtaining good performance. The avail-
ability of efficient deep learning frameworks has made design and implementation of this
significantly easier.
Exercises
1. Which parts of code that you are currently writing could be “learned”, i.e., improved by
learning and automatically determining design choices that are made in your code? Does
your code include heuristic design choices?
2. Which problems that you encounter have many examples for how to solve them, yet no spe-
cific way to automate them? These may be prime candidates for using deep learning.
3. Viewing the development of artificial intelligence as a new industrial revolution, what is the
relationship between algorithms and data? Is it similar to steam engines and coal (what is
the fundamental difference)?
4. Where else can you apply the end-to-end training approach (such as in Fig. 1.1.2)? Physics?
Engineering? Econometrics?
1.6. Success Stories 41

Discussions39
39
2 | Preliminaries
To get started with deep learning, we will need to develop a few basic skills. All machine learning
is concerned with extracting information from data. So we will begin by learning the practical
skills for storing, manipulating, and preprocessing data.
Moreover, machine learning typically requires working with large datasets, which we can think
of as tables, where the rows correspond to examples and the columns correspond to attributes.
Linear algebra gives us a powerful set of techniques for working with tabular data. We will not go
too far into the weeds but rather focus on the basic of matrix operations and their implementation.
Additionally, deep learning is all about optimization. We have a model with some parameters and
we want to find those that fit our data the best. Determining which way to move each parameter at
each step of an algorithm requires a little bit of calculus, which will be briefly introduced. Fortu-
nately, the autograd package automatically computes differentiation for us, and we will cover it
next.
Next, machine learning is concerned with making predictions: what is the likely value of some un-
known attribute, given the information that we observe? To reason rigorously under uncertainty
we will need to invoke the language of probability.
In the end, the official documentation provides plenty of descriptions and examples that are be-
yond this book. To conclude the chapter, we will show you how to look up documentation for the
needed information.
This book has kept the mathematical content to the minimum necessary to get a proper under-
standing of deep learning. However, it does not mean that this book is mathematics free. Thus,
this chapter provides a rapid introduction to basic and frequently-used mathematics to allow any-
one to understand at least most of the mathematical content of the book. If you wish to understand
all of the mathematical content, further reviewing the online appendix on mathematics40 should
be sufficient.
2.1 Data Manipulation
In order to get anything done, we need some way to store and manipulate data. Generally, there
are two important things we need to do with data: (i) acquire them; and (ii) process them once they
are inside the computer. There is no point in acquiring data without some way to store it, so let us
get our hands dirty first by playing with synthetic data. To start, we introduce the n-dimensional
array, which is also called the tensor.
If you have worked with NumPy, the most widely-used scientific computing package in Python,
then you will find this section familiar. No matter which framework you use, its tensor class
40
https://d2l.ai/chapter_appendix-mathematics-for-deep-learning/index.html
43
(ndarray in MXNet, Tensor in both PyTorch and TensorFlow) is similar to NumPyʼs ndarray with
a few killer features. First, GPU is well-supported to accelerate the computation whereas NumPy
only supports CPU computation. Second, the tensor class supports automatic differentiation.
These properties make the tensor class suitable for deep learning. Throughout the book, when
we say tensors, we are referring to instances of the tensor class unless otherwise stated.
2.1.1 Getting Started
In this section, we aim to get you up and running, equipping you with the basic math and numer-
ical computing tools that you will build on as you progress through the book. Do not worry if you
struggle to grok some of the mathematical concepts or library functions. The following sections
will revisit this material in the context of practical examples and it will sink. On the other hand,
if you already have some background and want to go deeper into the mathematical content, just
skip this section.
To start, we import the np (numpy) and npx (numpy_extension) modules from MXNet. Here, the np
module includes functions supported by NumPy, while the npx module contains a set of extensions
developed to empower deep learning within a NumPy-like environment. When using tensors, we
almost always invoke the set_np function: this is for compatibility of tensor processing by other
components of MXNet.
from mxnet import np, npx

npx.set_np()
A tensor represents a (possibly multi-dimensional) array of numerical values. With one axis, a
tensor corresponds (in math) to a vector. With two axes, a tensor corresponds to a matrix. Tensors
with more than two axes do not have special mathematical names.
To start, we can use arange to create a row vector x containing the first 12 integers starting with 0,
though they are created as floats by default. Each of the values in a tensor is called an element of
the tensor. For instance, there are 12 elements in the tensor x. Unless otherwise specified, a new
tensor will be stored in main memory and designated for CPU-based computation.
x = np.arange(12)
x
array([ 0., 1., 2., 3., 4., 5., 6., 7., 8., 9., 10., 11.])
We can access a tensorʼs shape (the length along each axis) by inspecting its shape property.
x.shape
(12,)
If we just want to know the total number of elements in a tensor, i.e., the product of all of the shape
elements, we can inspect its size. Because we are dealing with a vector here, the single element
of its shape is identical to its size.
x.size
44 Chapter 2. Preliminaries
12
To change the shape of a tensor without altering either the number of elements or their values, we
can invoke the reshape function. For example, we can transform our tensor, x, from a row vector
with shape (12,) to a matrix with shape (3, 4). This new tensor contains the exact same values, but
views them as a matrix organized as 3 rows and 4 columns. To reiterate, although the shape has
changed, the elements have not. Note that the size is unaltered by reshaping.
X = x.reshape(3, 4)
X
array([[ 0., 1., 2., 3.],

[ 4., 5., 6., 7.],
[ 8., 9., 10., 11.]])
Reshaping by manually specifying every dimension is unnecessary. If our target shape is a ma-
trix with shape (height, width), then after we know the width, the height is given implicitly. Why
should we have to perform the division ourselves? In the example above, to get a matrix with 3
rows, we specified both that it should have 3 rows and 4 columns. Fortunately, tensors can au-
tomatically work out one dimension given the rest. We invoke this capability by placing -1 for
the dimension that we would like tensors to automatically infer. In our case, instead of calling
x.reshape(3, 4), we could have equivalently called x.reshape(-1, 4) or x.reshape(3, -1).
Typically, we will want our matrices initialized either with zeros, ones, some other constants, or
numbers randomly sampled from a specific distribution. We can create a tensor representing a
tensor with all elements set to 0 and a shape of (2, 3, 4) as follows:
np.zeros((2, 3, 4))
array([[[0., 0., 0., 0.],

[0., 0., 0., 0.],
[0., 0., 0., 0.]],
[[0., 0., 0., 0.],

[0., 0., 0., 0.],
[0., 0., 0., 0.]]])
Similarly, we can create tensors with each element set to 1 as follows:
np.ones((2, 3, 4))
array([[[1., 1., 1., 1.],

[1., 1., 1., 1.],
[1., 1., 1., 1.]],
[[1., 1., 1., 1.],

[1., 1., 1., 1.],
[1., 1., 1., 1.]]])
Often, we want to randomly sample the values for each element in a tensor from some probability
distribution. For example, when we construct arrays to serve as parameters in a neural network,
2.1. Data Manipulation 45

we will typically initialize their values randomly. The following snippet creates a tensor with shape
(3, 4). Each of its elements is randomly sampled from a standard Gaussian (normal) distribution
with a mean of 0 and a standard deviation of 1.
np.random.normal(0, 1, size=(3, 4))
array([[ 2.2122064 , 1.1630787 , 0.7740038 , 0.4838046 ],

[ 1.0434405 , 0.29956347, 1.1839255 , 0.15302546],
[ 1.8917114 , -1.1688148 , -1.2347414 , 1.5580711 ]])
We can also specify the exact values for each element in the desired tensor by supplying a Python
list (or list of lists) containing the numerical values. Here, the outermost list corresponds to axis
0, and the inner list to axis 1.
np.array([[2, 1, 4, 3], [1, 2, 3, 4], [4, 3, 2, 1]])
array([[2., 1., 4., 3.],

[1., 2., 3., 4.],
[4., 3., 2., 1.]])
2.1.2 Operations
This book is not about software engineering. Our interests are not limited to simply reading and
writing data from/to arrays. We want to perform mathematical operations on those arrays. Some
of the simplest and most useful operations are the elementwise operations. These apply a stan-
dard scalar operation to each element of an array. For functions that take two arrays as inputs,
elementwise operations apply some standard binary operator on each pair of corresponding ele-
ments from the two arrays. We can create an elementwise function from any function that maps
from a scalar to a scalar.
In mathematical notation, we would denote such a unary scalar operator (taking one input) by the
signature f : R → R. This just means that the function is mapping from any real number (R) onto
another. Likewise, we denote a binary scalar operator (taking two real inputs, and yielding one
output) by the signature f : R, R → R. Given any two vectors u and v of the same shape, and a binary
operator f , we can produce a vector c = F (u, v) by setting ci ← f (ui , vi ) for all i, where ci , ui , and
vi are the ith elements of vectors c, u, and v. Here, we produced the vector-valued F : Rd , Rd → Rd
by lifting the scalar function to an elementwise vector operation.
The common standard arithmetic operators (+, -, *, /, and **) have all been lifted to element-
wise operations for any identically-shaped tensors of arbitrary shape. We can call elementwise
operations on any two tensors of the same shape. In the following example, we use commas to
formulate a 5-element tuple, where each element is the result of an elementwise operation.
x = np.array([1, 2, 4, 8])
y = np.array([2, 2, 2, 2])
x + y, x - y, x * y, x / y, x ** y # The ** operator is exponentiation
(array([ 3., 4., 6., 10.]),

array([-1., 0., 2., 6.]),
(continues on next page)
(continued from previous page)
array([ 2., 4., 8., 16.]),
array([0.5, 1. , 2. , 4. ]),
array([ 1., 4., 16., 64.]))
Many more operations can be applied elementwise, including unary operators like exponentia-
tion.
np.exp(x)
array([2.7182817e+00, 7.3890562e+00, 5.4598148e+01, 2.9809580e+03])
In addition to elementwise computations, we can also perform linear algebra operations, includ-
ing vector dot products and matrix multiplication. We will explain the crucial bits of linear algebra
(with no assumed prior knowledge) in Section 2.3.
We can also concatenate multiple tensors together, stacking them end-to-end to form a larger ten-
sor. We just need to provide a list of tensors and tell the system along which axis to concatenate.
The example below shows what happens when we concatenate two matrices along rows (axis 0,
the first element of the shape) vs. columns (axis 1, the second element of the shape). We can see
that the first output tensorʼs axis-0 length (6) is the sum of the two input tensorsʼ axis-0 lengths
(3 + 3); while the second output tensorʼs axis-1 length (8) is the sum of the two input tensorsʼ axis-1
lengths (4 + 4).
X = np.arange(12).reshape(3, 4)
Y = np.array([[2, 1, 4, 3], [1, 2, 3, 4], [4, 3, 2, 1]])
np.concatenate([X, Y], axis=0), np.concatenate([X, Y], axis=1)
(array([[ 0., 1., 2., 3.],

[ 4., 5., 6., 7.],
[ 8., 9., 10., 11.],
[ 2., 1., 4., 3.],
[ 1., 2., 3., 4.],
[ 4., 3., 2., 1.]]),
array([[ 0., 1., 2., 3., 2., 1., 4., 3.],
[ 4., 5., 6., 7., 1., 2., 3., 4.],
[ 8., 9., 10., 11., 4., 3., 2., 1.]]))
Sometimes, we want to construct a binary tensor via logical statements. Take X == Y as an example.
For each position, if X and Y are equal at that position, the corresponding entry in the new tensor
takes a value of 1, meaning that the logical statement X == Y is true at that position; otherwise that
position takes 0.
X == Y
array([[False, True, False, True],

[False, False, False, False],
[False, False, False, False]])
Summing all the elements in the tensor yields a tensor with only one element.

X.sum()
array(66.)
2.1.3 Broadcasting Mechanism
In the above section, we saw how to perform elementwise operations on two tensors of the same
shape. Under certain conditions, even when shapes differ, we can still perform elementwise op-
erations by invoking the broadcasting mechanism. This mechanism works in the following way:
First, expand one or both arrays by copying elements appropriately so that after this transforma-
tion, the two tensors have the same shape. Second, carry out the elementwise operations on the
resulting arrays.
In most cases, we broadcast along an axis where an array initially only has length 1, such as in the
following example:
a = np.arange(3).reshape(3, 1)
b = np.arange(2).reshape(1, 2)
a, b
(array([[0.],
[1.],
[2.]]),
array([[0., 1.]]))
Since a and b are 3 × 1 and 1 × 2 matrices respectively, their shapes do not match up if we want
to add them. We broadcast the entries of both matrices into a larger 3 × 2 matrix as follows: for
matrix a it replicates the columns and for matrix b it replicates the rows before adding up both
elementwise.
a + b
array([[0., 1.],
[1., 2.],
[2., 3.]])
2.1.4 Indexing and Slicing
Just as in any other Python array, elements in a tensor can be accessed by index. As in any Python
array, the first element has index 0 and ranges are specified to include the first but before the last
element. As in standard Python lists, we can access elements according to their relative position
to the end of the list by using negative indices.
Thus, [-1] selects the last element and [1:3] selects the second and the third elements as follows:
X[-1], X[1:3]
(array([ 8., 9., 10., 11.]),
array([[ 4., 5., 6., 7.],
[ 8., 9., 10., 11.]]))
Beyond reading, we can also write elements of a matrix by specifying indices.
X[1, 2] = 9
X
array([[ 0., 1., 2., 3.],

[ 4., 5., 9., 7.],
[ 8., 9., 10., 11.]])
If we want to assign multiple elements the same value, we simply index all of them and then assign
them the value. For instance, [0:2, :] accesses the first and second rows, where : takes all the
elements along axis 1 (column). While we discussed indexing for matrices, this obviously also
works for vectors and for tensors of more than 2 dimensions.
X[0:2, :] = 12
X
array([[12., 12., 12., 12.],

[12., 12., 12., 12.],
[ 8., 9., 10., 11.]])
2.1.5 Saving Memory
Running operations can cause new memory to be allocated to host results. For example, if we
write Y = X + Y, we will dereference the tensor that Y used to point to and instead point Y at
the newly allocated memory. In the following example, we demonstrate this with Pythonʼs id()
function, which gives us the exact address of the referenced object in memory. After running Y =
Y + X, we will find that id(Y) points to a different location. That is because Python first evaluates Y
+ X, allocating new memory for the result and then makes Y point to this new location in memory.
before = id(Y)
Y = Y + X
id(Y) == before
False
This might be undesirable for two reasons. First, we do not want to run around allocating mem-
ory unnecessarily all the time. In machine learning, we might have hundreds of megabytes of
parameters and update all of them multiple times per second. Typically, we will want to perform
these updates in place. Second, we might point at the same parameters from multiple variables.
If we do not update in place, other references will still point to the old memory location, making
it possible for parts of our code to inadvertently reference stale parameters.
Fortunately, performing in-place operations is easy. We can assign the result of an operation to
a previously allocated array with slice notation, e.g., Y[:] = <expression>. To illustrate this

concept, we first create a new matrix Z with the same shape as another Y, using zeros_like to
allocate a block of 0 entries.
Z = np.zeros_like(Y)
print('id(Z):', id(Z))
Z[:] = X + Y
print('id(Z):', id(Z))
id(Z): 140228072655200
id(Z): 140228072655200
If the value of X is not reused in subsequent computations, we can also use X[:] = X + Y or X +=
Y to reduce the memory overhead of the operation.
before = id(X)
X += Y
id(X) == before
True
2.1.6 Conversion to Other Python Objects
Converting to a NumPy tensor, or vice versa, is easy. The converted result does not share memory.
This minor inconvenience is actually quite important: when you perform operations on the CPU
or on GPUs, you do not want to halt computation, waiting to see whether the NumPy package of
Python might want to be doing something else with the same chunk of memory.
A = X.asnumpy()
B = np.array(A)
type(A), type(B)
(numpy.ndarray, mxnet.numpy.ndarray)
To convert a size-1 tensor to a Python scalar, we can invoke the item function or Pythonʼs built-in
functions.
a = np.array([3.5])
a, a.item(), float(a), int(a)
(array([3.5]), 3.5, 3.5, 3)
Summary
• The main interface to store and manipulate data for deep learning is the tensor (n-
dimensional array). It provides a variety of functionalities including basic mathematics op-
erations, broadcasting, indexing, slicing, memory saving, and conversion to other Python
objects.
Exercises
1. Run the code in this section. Change the conditional statement X == Y in this section to X <
Y or X > Y, and then see what kind of tensor you can get.
2. Replace the two tensors that operate by element in the broadcasting mechanism with other
shapes, e.g., 3-dimensional tensors. Is the result the same as expected?
Discussions41
2.2 Data Preprocessing
So far we have introduced a variety of techniques for manipulating data that are already stored in
tensors. To apply deep learning to solving real-world problems, we often begin with preprocess-
ing raw data, rather than those nicely prepared data in the tensor format. Among popular data
analytic tools in Python, the pandas package is commonly used. Like many other extension pack-
ages in the vast ecosystem of Python, pandas can work together with tensors. So, we will briefly
walk through steps for preprocessing raw data with pandas and converting them into the tensor
format. We will cover more data preprocessing techniques in later chapters.
2.2.1 Reading the Dataset
As an example, we begin by creating an artificial dataset that is stored in a csv (comma-separated

values) file ../data/house_tiny.csv. Data stored in other formats may be processed in sim-
ilar ways. The following mkdir_if_not_exist function ensures that the directory ../data ex-
ists. Note that the comment #@save is a special mark where the following function, class,
or statements are saved in the d2l package so later they can be directly invoked (e.g., d2l.
mkdir_if_not_exist(path)) without being redefined.
import os
def mkdir_if_not_exist(path): #@save

"""Make a directory if it does not exist."""
if not isinstance(path, str):
path = os.path.join(*path)
if not os.path.exists(path):
os.makedirs(path)
Below we write the dataset row by row into a csv file.

41
2.2. Data Preprocessing 51

data_file = '../data/house_tiny.csv'
mkdir_if_not_exist('../data')
with open(data_file, 'w') as f:
f.write('NumRooms,Alley,Price\n') # Column names
f.write('NA,Pave,127500\n') # Each row represents a data example
f.write('2,NA,106000\n')
f.write('4,NA,178100\n')
f.write('NA,NA,140000\n')
To load the raw dataset from the created csv file, we import the pandas package and invoke the
read_csv function. This dataset has four rows and three columns, where each row describes the
number of rooms (“NumRooms”), the alley type (“Alley”), and the price (“Price”) of a house.
# If pandas is not installed, just uncomment the following line:

# !pip install pandas
import pandas as pd
data = pd.read_csv(data_file)
print(data)
NumRooms Alley Price

0 NaN Pave 127500
1 2.0 NaN 106000
2 4.0 NaN 178100
3 NaN NaN 140000
2.2.2 Handling Missing Data
Note that “NaN” entries are missing values. To handle missing data, typical methods include im-
putation and deletion, where imputation replaces missing values with substituted ones, while dele-
tion ignores missing values. Here we will consider imputation.
By integer-location based indexing (iloc), we split data into inputs and outputs, where the former
takes the first two columns while the latter only keeps the last column. For numerical values in
inputs that are missing, we replace the “NaN” entries with the mean value of the same column.
inputs, outputs = data.iloc[:, 0:2], data.iloc[:, 2]

inputs = inputs.fillna(inputs.mean())
print(inputs)
NumRooms Alley
0 3.0 Pave
1 2.0 NaN
2 4.0 NaN
3 3.0 NaN
For categorical or discrete values in inputs, we consider “NaN” as a category. Since the “Alley”
column only takes two types of categorical values “Pave” and “NaN”, pandas can automatically
convert this column to two columns “Alley_Pave” and “Alley_nan”. A row whose alley type is “Pave”
will set values of “Alley_Pave” and “Alley_nan” to 1 and 0. A row with a missing alley type will set
their values to 0 and 1.
inputs = pd.get_dummies(inputs, dummy_na=True)
print(inputs)
NumRooms Alley_Pave Alley_nan

0 3.0 1 0
1 2.0 0 1
2 4.0 0 1
3 3.0 0 1
2.2.3 Conversion to the Tensor Format
Now that all the entries in inputs and outputs are numerical, they can be converted to the tensor
format. Once data are in this format, they can be further manipulated with those tensor function-
alities that we have introduced in Section 2.1.
from mxnet import np
X, y = np.array(inputs.values), np.array(outputs.values)
X, y
(array([[3., 1., 0.],

[2., 0., 1.],
[4., 0., 1.],
[3., 0., 1.]], dtype=float64),
array([127500, 106000, 178100, 140000], dtype=int64))
Summary
• Like many other extension packages in the vast ecosystem of Python, pandas can work to-
gether with tensors.
• Imputation and deletion can be used to handle missing data.
Exercises
Create a raw dataset with more rows and columns.

1. Delete the column with the most missing values.
2. Convert the preprocessed dataset to the tensor format.
Discussions42
42
2.2. Data Preprocessing 53

2.3 Linear Algebra
Now that you can store and manipulate data, let us briefly review the subset of basic linear algebra
that you will need to understand and implement most of models covered in this book. Below, we
introduce the basic mathematical objects, arithmetic, and operations in linear algebra, expressing
each of them through mathematical notation and the corresponding implementation in code.
2.3.1 Scalars
If you never studied linear algebra or machine learning, then your past experience with math
probably consisted of thinking about one number at a time. And, if you ever balanced a check-
book or even paid for dinner at a restaurant then you already know how to do basic things like
adding and multiplying pairs of numbers. For example, the temperature in Palo Alto is 52 de-
grees Fahrenheit. Formally, we call values consisting of just one numerical quantity scalars. If
you wanted to convert this value to Celsius (the metric systemʼs more sensible temperature scale),
you would evaluate the expression c = 59 (f − 32), setting f to 52. In this equation, each of the
terms—5, 9, and 32—are scalar values. The placeholders c and f are called variables and they rep-
resent unknown scalar values.
In this book, we adopt the mathematical notation where scalar variables are denoted by ordinary
lower-cased letters (e.g., x, y, and z). We denote the space of all (continuous) real-valued scalars
by R. For expedience, we will punt on rigorous definitions of what precisely space is, but just
remember for now that the expression x ∈ R is a formal way to say that x is a real-valued scalar.
The symbol ∈ can be pronounced “in” and simply denotes membership in a set. Analogously, we
could write x, y ∈ {0, 1} to state that x and y are numbers whose value can only be 0 or 1.
A scalar is represented by a tensor with just one element. In the next snippet, we instantiate two
scalars and perform some familiar arithmetic operations with them, namely addition, multipli-
cation, division, and exponentiation.

npx.set_np()
x = np.array(3.0)
y = np.array(2.0)
x + y, x * y, x / y, x ** y
(array(5.), array(6.), array(1.5), array(9.))
2.3.2 Vectors
You can think of a vector as simply a list of scalar values. We call these values the elements (entries
or components) of the vector. When our vectors represent examples from our dataset, their values
hold some real-world significance. For example, if we were training a model to predict the risk that
a loan defaults, we might associate each applicant with a vector whose components correspond
to their income, length of employment, number of previous defaults, and other factors. If we
were studying the risk of heart attacks hospital patients potentially face, we might represent each
patient by a vector whose components capture their most recent vital signs, cholesterol levels,
minutes of exercise per day, etc. In math notation, we will usually denote vectors as bold-faced,
lower-cased letters (e.g., x, y, and z).
We work with vectors via one-dimensional tensors. In general tensors can have arbitrary lengths,
subject to the memory limits of your machine.
x = np.arange(4)
x
array([0., 1., 2., 3.])
We can refer to any element of a vector by using a subscript. For example, we can refer to the ith
element of x by xi . Note that the element xi is a scalar, so we do not bold-face the font when refer-
ring to it. Extensive literature considers column vectors to be the default orientation of vectors,
so does this book. In math, a vector x can be written as
 
x1
 x2 
 
x =  . , (2.3.1)
 .. 
xn
where x1 , . . . , xn are elements of the vector. In code, we access any element by indexing into the
tensor.
x[3]
array(3.)
Length, Dimensionality, and Shape
Let us revisit some concepts from Section 2.1. A vector is just an array of numbers. And just as
every array has a length, so does every vector. In math notation, if we want to say that a vector x
consists of n real-valued scalars, we can express this as x ∈ Rn . The length of a vector is commonly
called the dimension of the vector.
As with an ordinary Python array, we can access the length of a tensor by calling Pythonʼs built-in
len() function.
len(x)
When a tensor represents a vector (with precisely one axis), we can also access its length via the
.shape attribute. The shape is a tuple that lists the length (dimensionality) along each axis of the
tensor. For tensors with just one axis, the shape has just one element.
x.shape
2.3. Linear Algebra 55

(4,)
Note that the word “dimension” tends to get overloaded in these contexts and this tends to confuse
people. To clarify, we use the dimensionality of a vector or an axis to refer to its length, i.e., the
number of elements of a vector or an axis. However, we use the dimensionality of a tensor to refer
to the number of axes that a tensor has. In this sense, the dimensionality of some axis of a tensor
will be the length of that axis.
2.3.3 Matrices
Just as vectors generalize scalars from order zero to order one, matrices generalize vectors from
order one to order two. Matrices, which we will typically denote with bold-faced, capital letters
(e.g., X, Y, and Z), are represented in code as tensors with two axes.
In math notation, we use A ∈ Rm×n to express that the matrix A consists of m rows and n columns
of real-valued scalars. Visually, we can illustrate any matrix A ∈ Rm×n as a table, where each
element aij belongs to the ith row and j th column:
 
a11 a12 · · · a1n
 a21 a22 · · · a2n 
 
A= . .. .. ..  . (2.3.2)
 .. . . . 
am1 am2 · · · amn
For any A ∈ Rm×n , the shape of A is (m, n) or m × n. Specifically, when a matrix has the same
number of rows and columns, its shape becomes a square; thus, it is called a square matrix.
We can create an m × n matrix by specifying a shape with two components m and n when calling
any of our favorite functions for instantiating a tensor.
A = np.arange(20).reshape(5, 4)
A
array([[ 0., 1., 2., 3.],

[ 4., 5., 6., 7.],
[ 8., 9., 10., 11.],
[12., 13., 14., 15.],
[16., 17., 18., 19.]])
We can access the scalar element aij of a matrix A in :eqref:eq_matrix_def by specifying the in-
dices for the row (i) and column (j), such as [A]ij . When the scalar elements of a matrix A, such
as in :eqref:eq_matrix_def, are not given, we may simply use the lower-case letter of the matrix
A with the index subscript, aij , to refer to [A]ij . To keep notation simple, commas are inserted to
separate indices only when necessary, such as a2,3j and [A]2i−1,3 .
Sometimes, we want to flip the axes. When we exchange a matrixʼs rows and columns, the result is
called the transpose of the matrix. Formally, we signify a matrix Aʼs transpose by A⊤ and if B = A⊤ ,
then bij = aji for any i and j. Thus, the transpose of A in :eqref:eq_matrix_def is a n × m matrix:
 
a11 a21 . . . am1
 a12 a22 . . . am2 
 
A⊤ =  . .. . ..  . (2.3.3)
 .. . . . . 
a1n a2n . . . amn
Now we access a matrixʼs transpose in code.
A.T
array([[ 0., 4., 8., 12., 16.],

[ 1., 5., 9., 13., 17.],
[ 2., 6., 10., 14., 18.],
[ 3., 7., 11., 15., 19.]])
As a special type of the square matrix, a symmetric matrix A is equal to its transpose: A = A⊤ . Here
we define a symmetric matrix B.
B = np.array([[1, 2, 3], [2, 0, 4], [3, 4, 5]])

B
array([[1., 2., 3.],

[2., 0., 4.],
[3., 4., 5.]])
Now we compare B with its transpose.
B == B.T
array([[ True, True, True],

[ True, True, True],
[ True, True, True]])
Matrices are useful data structures: they allow us to organize data that have different modalities
of variation. For example, rows in our matrix might correspond to different houses (data exam-
ples), while columns might correspond to different attributes. This should sound familiar if you
have ever used spreadsheet software or have read Section 2.2. Thus, although the default orienta-
tion of a single vector is a column vector, in a matrix that represents a tabular dataset, it is more
conventional to treat each data example as a row vector in the matrix. And, as we will see in later
chapters, this convention will enable common deep learning practices. For example, along the
outermost axis of a tensor, we can access or enumerate minibatches of data examples, or just data
examples if no minibatch exists.
2.3.4 Tensors
Just as vectors generalize scalars, and matrices generalize vectors, we can build data structures
with even more axes. Tensors (“tensors” in this subsection refer to algebraic objects) give us a
generic way of describing n-dimensional arrays with an arbitrary number of axes. Vectors, for
example, are first-order tensors, and matrices are second-order tensors. Tensors are denoted with
capital letters of a special font face (e.g., X, Y, and Z) and their indexing mechanism (e.g., xijk and
[X]1,2i−1,3 ) is similar to that of matrices.
Tensors will become more important when we start working with images, which arrive as n-
dimensional arrays with 3 axes corresponding to the height, width, and a channel axis for stacking
the color channels (red, green, and blue). For now, we will skip over higher order tensors and
focus on the basics.

X = np.arange(24).reshape(2, 3, 4)
X
array([[[ 0., 1., 2., 3.],

[ 4., 5., 6., 7.],
[ 8., 9., 10., 11.]],
[[12., 13., 14., 15.],

[16., 17., 18., 19.],
[20., 21., 22., 23.]]])
2.3.5 Basic Properties of Tensor Arithmetic
Scalars, vectors, matrices, and tensors (“tensors” in this subsection refer to algebraic objects) of an
arbitrary number of axes have some nice properties that often come in handy. For example, you
might have noticed from the definition of an elementwise operation that any elementwise unary
operation does not change the shape of its operand. Similarly, given any two tensors with the
same shape, the result of any binary elementwise operation will be a tensor of that same shape.
For example, adding two matrices of the same shape performs elementwise addition over these
two matrices.
A = np.arange(20).reshape(5, 4)
B = A.copy() # Assign a copy of À` to `B` by allocating new memory
A, A + B
(array([[ 0., 1., 2., 3.],

[ 4., 5., 6., 7.],
[ 8., 9., 10., 11.],
[12., 13., 14., 15.],
[16., 17., 18., 19.]]),
array([[ 0., 2., 4., 6.],
[ 8., 10., 12., 14.],
[16., 18., 20., 22.],
[24., 26., 28., 30.],
[32., 34., 36., 38.]]))
Specifically, elementwise multiplication of two matrices is called their Hadamard product (math
notation ⊙). Consider matrix B ∈ Rm×n whose element of row i and column j is bij . The Hadamard
product of matrices A (defined in :eqref:eq_matrix_def) and B
 
a11 b11 a12 b12 . . . a1n b1n
 a21 b21 a22 b22 . . . a2n b2n 
 
A⊙B= . .. .. .. . (2.3.4)
 .. . . . 
am1 bm1 am2 bm2 . . . amn bmn
A * B
array([[ 0., 1., 4., 9.],
[ 16., 25., 36., 49.],
[ 64., 81., 100., 121.],
[144., 169., 196., 225.],
[256., 289., 324., 361.]])
Multiplying or adding a tensor by a scalar also does not change the shape of the tensor, where each
element of the operand tensor will be added or multiplied by the scalar.
a = 2
X = np.arange(24).reshape(2, 3, 4)
a + X, (a * X).shape
(array([[[ 2., 3., 4., 5.],

[ 6., 7., 8., 9.],
[10., 11., 12., 13.]],
[[14., 15., 16., 17.],

[18., 19., 20., 21.],
[22., 23., 24., 25.]]]),
(2, 3, 4))
2.3.6 Reduction
One useful operation that we can perform with arbitrary tensors ∑ is to calculate the sum of their
elements. In mathematical notation, we express sums∑ using the symbol. To express the sum
of the elements in a vector x of length d, we write di=1 xi . In code, we can just call the function
for calculating the sum.
x = np.arange(4)
x, x.sum()
(array([0., 1., 2., 3.]), array(6.))
We can express sums over the elements of tensors

∑of arbitrary
∑n shape. For example, the sum of the
elements of an m × n matrix A could be written m i=1 j=1 a ij .
A.shape, A.sum()
((5, 4), array(190.))
By default, invoking the function for calculating the sum reduces a tensor along all its axes to a
scalar. We can also specify the axes along which the tensor is reduced via summation. Take ma-
trices as an example. To reduce the row dimension (axis 0) by summing up elements of all the
rows, we specify axis=0 when invoking the function. Since the input matrix reduces along axis 0
to generate the output vector, the dimension of axis 0 of the input is lost in the output shape.
A_sum_axis0 = A.sum(axis=0)
A_sum_axis0, A_sum_axis0.shape

(array([40., 45., 50., 55.]), (4,))
Specifying axis=1 will reduce the column dimension (axis 1) by summing up elements of all the
columns. Thus, the dimension of axis 1 of the input is lost in the output shape.
A_sum_axis1 = A.sum(axis=1)
A_sum_axis1, A_sum_axis1.shape
(array([ 6., 22., 38., 54., 70.]), (5,))
Reducing a matrix along both rows and columns via summation is equivalent to summing up all
the elements of the matrix.
A.sum(axis=[0, 1]) # Same as À.sum()`
array(190.)
A related quantity is the mean, which is also called the average. We calculate the mean by dividing
the sum by the total number of elements. In code, we could just call the function for calculating
the mean on tensors of arbitrary shape.
A.mean(), A.sum() / A.size
(array(9.5), array(9.5))
Likewise, the function for calculating the mean can also reduce a tensor along the specified axes.
A.mean(axis=0), A.sum(axis=0) / A.shape[0]
(array([ 8., 9., 10., 11.]), array([ 8., 9., 10., 11.]))
Non-Reduction Sum
However, sometimes it can be useful to keep the number of axes unchanged when invoking the
function for calculating the sum or mean.
sum_A = A.sum(axis=1, keepdims=True)

sum_A
array([[ 6.],
[22.],
[38.],
[54.],
[70.]])
For instance, since sum_A still keeps its two axes after summing each row, we can divide A by sum_A
with broadcasting.
A / sum_A
array([[0. , 0.16666667, 0.33333334, 0.5 ],

[0.18181819, 0.22727273, 0.27272728, 0.3181818 ],
[0.21052632, 0.23684211, 0.2631579 , 0.28947368],
[0.22222222, 0.24074075, 0.25925925, 0.2777778 ],
[0.22857143, 0.24285714, 0.25714287, 0.27142859]])
If we want to calculate the cumulative sum of elements of A along some axis, say axis=0 (row by
row), we can call the cumsum function. This function will not reduce the input tensor along any
axis.
A.cumsum(axis=0)
array([[ 0., 1., 2., 3.],

[ 4., 6., 8., 10.],
[12., 15., 18., 21.],
[24., 28., 32., 36.],
[40., 45., 50., 55.]])
2.3.7 Dot Products
So far, we have only performed elementwise operations, sums, and averages. And if this was all
we could do, linear algebra probably would not deserve its own section. However, one of the most
dot product x⊤ y (or
fundamental operations is the dot product. Given two vectors x, y ∈ Rd , their ∑
⊤
⟨x, y⟩) is a sum over the products of the elements at the same position: x y = di=1 xi yi .
y = np.ones(4)
x, y, np.dot(x, y)
(array([0., 1., 2., 3.]), array([1., 1., 1., 1.]), array(6.))
Note that we can express the dot product of two vectors equivalently by performing an element-
wise multiplication and then a sum:
np.sum(x * y)
array(6.)
Dot products are useful in a wide range of contexts. For example, given some set of values, denoted
by a vector x ∈ Rd and a set of weights denoted by w ∈ Rd , the weighted sum of the values in x
according to the weights w could be (∑expressed ) as the dot product x⊤ w. When the weights are
d
non-negative and sum to one (i.e., i=1 wi = 1 ), the dot product expresses a weighted average.
After normalizing two vectors to have the unit length, the dot products express the cosine of the
angle between them. We will formally introduce this notion of length later in this section.

2.3.8 Matrix-Vector Products
Now that we know how to calculate dot products, we can begin to understand matrix-vector prod-
ucts. Recall the matrix A ∈ Rm×n and the vector x ∈ Rn defined and visualized in (2.3.2) and (2.3.1)
respectively. Let us start off by visualizing the matrix A in terms of its row vectors
 ⊤
a1
 a⊤ 
 2
A =  . , (2.3.5)
 .. 
a⊤
m
where each a⊤i ∈ R is a row vector representing the i row of the matrix A. The matrix-vector
n th
product Ax is simply a column vector of length m, whose ith element is the dot product a⊤
i x:
 ⊤  ⊤ 
a1 a1 x
 a⊤   a⊤ x 
 2  2 
Ax =  .  x =  .  . (2.3.6)
 . 
.  .. 
a⊤m a⊤mx
We can think of multiplication by a matrix A ∈ Rm×n as a transformation that projects vectors

from Rn to Rm . These transformations turn out to be remarkably useful. For example, we can
represent rotations as multiplications by a square matrix. As we will see in subsequent chapters,
we can also use matrix-vector products to describe the most intensive calculations required when
computing each layer in a neural network given the values of the previous layer.
Expressing matrix-vector products in code with tensors, we use the same dot function as for dot
products. When we call np.dot(A, x) with a matrix A and a vector x, the matrix-vector product is
performed. Note that the column dimension of A (its length along axis 1) must be the same as the
dimension of x (its length).
A.shape, x.shape, np.dot(A, x)
((5, 4), (4,), array([ 14., 38., 62., 86., 110.]))
2.3.9 Matrix-Matrix Multiplication
If you have gotten the hang of dot products and matrix-vector products, then matrix-matrix multi-
plication should be straightforward.
Say that we have two matrices A ∈ Rn×k and B ∈ Rk×m :
   
a11 a12 · · · a1k b11 b12 · · · b1m
 a21 a22 · · · a2k  b21 b22 · · · b2m 
   
A= . . .. .  , B =  .. .. . . ..  . (2.3.7)
 .. .. . ..   . . . . 
an1 an2 · · · ank bk1 bk2 · · · bkm
Denote by a⊤
i ∈ R the row vector representing the i row of the matrix A, and let bj ∈ R be the
k th k
column vector from the j th column of the matrix B. To produce the matrix product C = AB, it is
easiest to think of A in terms of its row vectors and B in terms of its column vectors:
 ⊤
a1
a⊤  [ ]
 2
A =  .  , B = b1 b2 · · · bm . (2.3.8)
 . 
.
a⊤
n
Then the matrix product C ∈ Rn×m is produced as we simply compute each element cij as the dot
product a⊤
i bj :
 ⊤  ⊤ 
a1 a1 b1 a⊤ ⊤
1 b2 · · · a1 bm
a⊤  [ ]  ⊤ ⊤ ⊤ 
 2 a2 b1 a2 b2 · · · a2 bm 
C = AB =  .  b1 b2 · · · bm =  .
.. 
. .. . . (2.3.9)
 ..   .. .. . 
a⊤n a⊤ ⊤ ⊤
n b1 an b2 · · · an bm
We can think of the matrix-matrix multiplication AB as simply performing m matrix-vector prod-

ucts and stitching the results together to form an n × m matrix. In the following snippet, we
perform matrix multiplication on A and B. Here, A is a matrix with 5 rows and 4 columns, and B is
a matrix with 4 rows and 3 columns. After multiplication, we obtain a matrix with 5 rows and 3
columns.
B = np.ones(shape=(4, 3))
np.dot(A, B)
array([[ 6., 6., 6.],

[22., 22., 22.],
[38., 38., 38.],
[54., 54., 54.],
[70., 70., 70.]])
Matrix-matrix multiplication can be simply called matrix multiplication, and should not be con-
fused with the Hadamard product.
2.3.10 Norms
Some of the most useful operators in linear algebra are norms. Informally, the norm of a vector
tells us how big a vector is. The notion of size under consideration here concerns not dimension-
ality but rather the magnitude of the components.
In linear algebra, a vector norm is a function f that maps a vector to a scalar, satisfying a handful
of properties. Given any vector x, the first property says that if we scale all the elements of a vector
by a constant factor α, its norm also scales by the absolute value of the same constant factor:
f (αx) = |α|f (x). (2.3.10)
The second property is the familiar triangle inequality:
f (x + y) ≤ f (x) + f (y). (2.3.11)
The third property simply says that the norm must be non-negative:
f (x) ≥ 0. (2.3.12)

That makes sense, as in most contexts the smallest size for anything is 0. The final property re-
quires that the smallest norm is achieved and only achieved by a vector consisting of all zeros.
∀i, [x]i = 0 ⇔ f (x) = 0. (2.3.13)
You might notice that norms sound a lot like measures of distance. And if you remember Euclidean
distances (think Pythagorasʼ theorem) from grade school, then the concepts of non-negativity and
the triangle inequality might ring a bell. In fact, the Euclidean distance is a norm: specifically it
is the L2 norm. Suppose that the elements in the n-dimensional vector x are x1 , . . . , xn . The L2
norm of x is the square root of the sum of the squares of the vector elements:
v
u n
u∑
∥x∥2 = t x2i , (2.3.14)
i=1
where the subscript 2 is often omitted in L2 norms, i.e., ∥x∥ is equivalent to ∥x∥2 . In code, we can
calculate the L2 norm of a vector as follows.
u = np.array([3, -4])
np.linalg.norm(u)
array(5.)
In deep learning, we work more often with the squared L2 norm. You will also frequently en-
counter the L1 norm, which is expressed as the sum of the absolute values of the vector elements:
∑
n
∥x∥1 = |xi | . (2.3.15)
i=1
As compared with the L2 norm, it is less influenced by outliers. To calculate the L1 norm, we
compose the absolute value function with a sum over the elements.
np.abs(u).sum()
array(7.)
Both the L2 norm and the L1 norm are special cases of the more general Lp norm:
( n )1/p
∑ p
∥x∥p = |xi | . (2.3.16)
i=1
Analogous to L2 norms of vectors, the Frobenius norm of a matrix X ∈ Rm×n is the square root of
the sum of the squares of the matrix elements:
v
u∑
um ∑ n
∥X∥F = t x2ij . (2.3.17)
i=1 j=1
The Frobenius norm satisfies all the properties of vector norms. It behaves as if it were an L2 norm
of a matrix-shaped vector. Invoking the following function will calculate the Frobenius norm of a
matrix.
np.linalg.norm(np.ones((4, 9)))
array(6.)
Norms and Objectives
While we do not want to get too far ahead of ourselves, we can plant some intuition already about
why these concepts are useful. In deep learning, we are often trying to solve optimization prob-
lems: maximize the probability assigned to observed data; minimize the distance between pre-
dictions and the ground-truth observations. Assign vector representations to items (like words,
products, or news articles) such that the distance between similar items is minimized, and the
distance between dissimilar items is maximized. Oftentimes, the objectives, perhaps the most
important components of deep learning algorithms (besides the data), are expressed as norms.
2.3.11 More on Linear Algebra
In just this section, we have taught you all the linear algebra that you will need to understand a
remarkable chunk of modern deep learning. There is a lot more to linear algebra and a lot of
that mathematics is useful for machine learning. For example, matrices can be decomposed into
factors, and these decompositions can reveal low-dimensional structure in real-world datasets.
There are entire subfields of machine learning that focus on using matrix decompositions and
their generalizations to high-order tensors to discover structure in datasets and solve prediction
problems. But this book focuses on deep learning. And we believe you will be much more inclined
to learn more mathematics once you have gotten your hands dirty deploying useful machine learn-
ing models on real datasets. So while we reserve the right to introduce more mathematics much
later on, we will wrap up this section here.
If you are eager to learn more about linear algebra, you may refer to either the online appendix
on linear algebraic operations43 or other excellent resources (Strang, 1993; Kolter, 2008; Petersen
et al., 2008).
Summary
• Scalars, vectors, matrices, and tensors are basic mathematical objects in linear algebra.
• Vectors generalize scalars, and matrices generalize vectors.
• Scalars, vectors, matrices, and tensors have zero, one, two, and an arbitrary number of axes,
respectively.
• A tensor can be reduced along the specified axes by sum and mean.
• Elementwise multiplication of two matrices is called their Hadamard product. It is different
from matrix multiplication.
• In deep learning, we often work with norms such as the L1 norm, the L2 norm, and the
Frobenius norm.
• We can perform a variety of operations over scalars, vectors, matrices, and tensors.
43
https://d2l.ai/chapter_appendix-mathematics-for-deep-learning/geometry-linear-algebraic-ops.html

Exercises
1. Prove that the transpose of a matrix Aʼs transpose is A: (A⊤ )⊤ = A.

2. Given two matrices A and B, show that the sum of transposes is equal to the transpose of a
sum: A⊤ + B⊤ = (A + B)⊤ .
3. Given any square matrix A, is A + A⊤ always symmetric? Why?
4. We defined the tensor X of shape (2, 3, 4) in this section. What is the output of len(X)?
5. For a tensor X of arbitrary shape, does len(X) always correspond to the length of a certain
axis of X? What is that axis?
6. Run A / A.sum(axis=1) and see what happens. Can you analyze the reason?
7. When traveling between two points in Manhattan, what is the distance that you need to cover
in terms of the coordinates, i.e., in terms of avenues and streets? Can you travel diagonally?
8. Consider a tensor with shape (2, 3, 4). What are the shapes of the summation outputs along
axis 0, 1, and 2?
9. Feed a tensor with 3 or more axes to the linalg.norm function and observe its output. What
does this function compute for tensors of arbitrary shape?
Discussions44
2.4 Calculus
Finding the area of a polygon had remained mysterious until at least 2,500 years ago, when ancient
Greeks divided a polygon into triangles and summed their areas. To find the area of curved shapes,
such as a circle, ancient Greeks inscribed polygons in such shapes. As shown in Fig. 2.4.1, an
inscribed polygon with more sides of equal length better approximates the circle. This process is
also known as the method of exhaustion.
Fig. 2.4.1: Find the area of a circle with the method of exhaustion.
In fact, the method of exhaustion is where integral calculus (will be described in Section 18.5) orig-
inates from. More than 2,000 years later, the other branch of calculus, differential calculus, was
invented. Among the most critical applications of differential calculus, optimization problems
consider how to do something the best. As discussed in Section 2.3.10, such problems are ubiqui-
tous in deep learning.
In deep learning, we train models, updating them successively so that they get better and better
as they see more and more data. Usually, getting better means minimizing a loss function, a score
44
that answers the question “how bad is our model?” This question is more subtle than it appears.
Ultimately, what we really care about is producing a model that performs well on data that we have
never seen before. But we can only fit the model to data that we can actually see. Thus we can
decompose the task of fitting models into two key concerns: i) optimization: the process of fitting
our models to observed data; ii) generalization: the mathematical principles and practitionersʼ
wisdom that guide as to how to produce models whose validity extends beyond the exact set of
data examples used to train them.
To help you understand optimization problems and methods in later chapters, here we give a very
brief primer on differential calculus that is commonly used in deep learning.
2.4.1 Derivatives and Differentiation
We begin by addressing the calculation of derivatives, a crucial step in nearly all deep learning
optimization algorithms. In deep learning, we typically choose loss functions that are differen-
tiable with respect to our modelʼs parameters. Put simply, this means that for each parameter,
we can determine how rapidly the loss would increase or decrease, were we to increase or decrease
that parameter by an infinitesimally small amount.
Suppose that we have a function f : R → R, whose input and output are both scalars. The derivative
of f is defined as
f (x + h) − f (x)
f ′ (x) = lim , (2.4.1)
h→0 h
if this limit exists. If f ′ (a) exists, f is said to be differentiable at a. If f is differentiable at every
number of an interval, then this function is differentiable on this interval. We can interpret the
derivative f ′ (x) in (2.4.1) as the instantaneous rate of change of f (x) with respect to x. The so-called
instantaneous rate of change is based on the variation h in x, which approaches 0.
To illustrate derivatives, let us experiment with an example. Define u = f (x) = 3x2 − 4x.
%matplotlib inline
from d2l import mxnet as d2l
npx.set_np()
def f(x):
return 3 * x ** 2 - 4 * x
By setting x = 1 and letting h approach 0, the numerical result of f (x+h)−f

h
(x)
in (2.4.1) approaches
2. Though this experiment is not a mathematical proof, we will see later that the derivative u′ is 2
when x = 1.
def numerical_lim(f, x, h):

return (f(x + h) - f(x)) / h
h = 0.1
for i in range(5):
print(f'h={h:.5f}, numerical limit={numerical_lim(f, 1, h):.5f}')
h *= 0.1
2.4. Calculus 67
h=0.10000, numerical limit=2.30000
Let us familiarize ourselves with a few equivalent notations for derivatives. Given y = f (x), where
x and y are the independent variable and the dependent variable of the function f , respectively.
The following expressions are equivalent:
dy df d
f ′ (x) = y ′ = = = f (x) = Df (x) = Dx f (x), (2.4.2)
dx dx dx
d
where symbols dx and D are differentiation operators that indicate operation of differentiation. We
can use the following rules to differentiate common functions:
• DC = 0 (C is a constant),
• Dxn = nxn−1 (the power rule, n is any real number),
• Dex = ex ,
• D ln(x) = 1/x.
To differentiate a function that is formed from a few simpler functions such as the above com-
mon functions, the following rules can be handy for us. Suppose that functions f and g are both
differentiable and C is a constant, we have the constant multiple rule
d d
[Cf (x)] = C f (x), (2.4.3)
dx dx
the sum rule
d d d
[f (x) + g(x)] = f (x) + g(x), (2.4.4)
dx dx dx
the product rule
d d d
[f (x)g(x)] = f (x) [g(x)] + g(x) [f (x)], (2.4.5)
dx dx dx
and the quotient rule
[ ]
d f (x) d
g(x) dx [f (x)] − f (x) dx
d
[g(x)]
= 2
. (2.4.6)
dx g(x) [g(x)]
Now we can apply a few of the above rules to find u′ = f ′ (x) = 3 dx x − 4 dx

d 2 d
x = 6x − 4. Thus, by
′
setting x = 1, we have u = 2: this is supported by our earlier experiment in this section where
the numerical result approaches 2. This derivative is also the slope of the tangent line to the curve
u = f (x) when x = 1.
To visualize such an interpretation of derivatives, we will use matplotlib, a popular plotting li-
brary in Python. To configure properties of the figures produced by matplotlib, we need to define
a few functions. In the following, the use_svg_display function specifies the matplotlib package
to output the svg figures for sharper images.
def use_svg_display(): #@save
"""Use the svg format to display a plot in Jupyter."""
display.set_matplotlib_formats('svg')
We define the set_figsize function to specify the figure sizes. Note that here we directly use d2l.
plt since the import statement from matplotlib import pyplot as plt has been marked for
being saved in the d2l package in the preface.
def set_figsize(figsize=(3.5, 2.5)): #@save

"""Set the figure size for matplotlib."""
use_svg_display()
d2l.plt.rcParams['figure.figsize'] = figsize
The following set_axes function sets properties of axes of figures produced by matplotlib.
#@save
def set_axes(axes, xlabel, ylabel, xlim, ylim, xscale, yscale, legend):
"""Set the axes for matplotlib."""
axes.set_xlabel(xlabel)
axes.set_ylabel(ylabel)
axes.set_xscale(xscale)
axes.set_yscale(yscale)
axes.set_xlim(xlim)
axes.set_ylim(ylim)
if legend:
axes.legend(legend)
axes.grid()
With these three functions for figure configurations, we define the plot function to plot multiple
curves succinctly since we will need to visualize many curves throughout the book.
#@save
def plot(X, Y=None, xlabel=None, ylabel=None, legend=None, xlim=None,
ylim=None, xscale='linear', yscale='linear',
fmts=('-', 'm--', 'g-.', 'r:'), figsize=(3.5, 2.5), axes=None):
"""Plot data points."""
if legend is None:
legend = []
set_figsize(figsize)
axes = axes if axes else d2l.plt.gca()
# Return True if `X` (tensor or list) has 1 axis

def has_one_axis(X):
return (hasattr(X, "ndim") and X.ndim == 1 or isinstance(X, list)
and not hasattr(X[0], "__len__"))
if has_one_axis(X):
X = [X]
if Y is None:
X, Y = [[]] * len(X), X
elif has_one_axis(Y):
Y = [Y]
if len(X) != len(Y):
2.4. Calculus 69
X = X * len(Y)
axes.cla()
for x, y, fmt in zip(X, Y, fmts):
if len(x):
axes.plot(x, y, fmt)
else:
axes.plot(y, fmt)
set_axes(axes, xlabel, ylabel, xlim, ylim, xscale, yscale, legend)
Now we can plot the function u = f (x) and its tangent line y = 2x−3 at x = 1, where the coefficient
2 is the slope of the tangent line.
x = np.arange(0, 3, 0.1)
plot(x, [f(x), 2 * x - 3], 'x', 'f(x)', legend=['f(x)', 'Tangent line (x=1)'])
2.4.2 Partial Derivatives
So far we have dealt with the differentiation of functions of just one variable. In deep learning,
functions often depend on many variables. Thus, we need to extend the ideas of differentiation to
these multivariate functions.
Let y = f (x1 , x2 , . . . , xn ) be a function with n variables. The partial derivative of y with respect to
its ith parameter xi is
∂y f (x1 , . . . , xi−1 , xi + h, xi+1 , . . . , xn ) − f (x1 , . . . , xi , . . . , xn )

= lim . (2.4.7)
∂xi h→0 h
∂y
To calculate ∂x i
, we can simply treat x1 , . . . , xi−1 , xi+1 , . . . , xn as constants and calculate the deriva-
tive of y with respect to xi . For notation of partial derivatives, the following are equivalent:
∂y ∂f
= = fxi = fi = Di f = Dxi f. (2.4.8)
∂xi ∂xi
2.4.3 Gradients
We can concatenate partial derivatives of a multivariate function with respect to all its variables
to obtain the gradient vector of the function. Suppose that the input of function f : Rn → R is an
n-dimensional vector x = [x1 , x2 , . . . , xn ]⊤ and the output is a scalar. The gradient of the function
f (x) with respect to x is a vector of n partial derivatives:
[ ]⊤
∂f (x) ∂f (x) ∂f (x)
∇x f (x) = , ,..., , (2.4.9)
∂x1 ∂x2 ∂xn
where ∇x f (x) is often replaced by ∇f (x) when there is no ambiguity.

Let x be an n-dimensional vector, the following rules are often used when differentiating multi-
variate functions:
• For all A ∈ Rm×n , ∇x Ax = A⊤ ,
• For all A ∈ Rn×m , ∇x x⊤ A = A,
• For all A ∈ Rn×n , ∇x x⊤ Ax = (A + A⊤ )x,
• ∇x ∥x∥2 = ∇x x⊤ x = 2x.
Similarly, for any matrix X, we have ∇X ∥X∥2F = 2X. As we will see later, gradients are useful for
designing optimization algorithms in deep learning.
2.4.4 Chain Rule
However, such gradients can be hard to find. This is because multivariate functions in deep learn-
ing are often composite, so we may not apply any of the aforementioned rules to differentiate these
functions. Fortunately, the chain rule enables us to differentiate composite functions.
Let us first consider functions of a single variable. Suppose that functions y = f (u) and u = g(x)
are both differentiable, then the chain rule states that
dy dy du
= . (2.4.10)
dx du dx
Now let us turn our attention to a more general scenario where functions have an arbitrary
number of variables. Suppose that the differentiable function y has variables u1 , u2 , . . . , um ,
where each differentiable function ui has variables x1 , x2 , . . . , xn . Note that y is a function of
x1 , x2 , . . . , xn . Then the chain rule gives
dy dy du1 dy du2 dy dum

= + + ··· + (2.4.11)
dxi du1 dxi du2 dxi dum dxi
for any i = 1, 2, . . . , n.
Summary
• Differential calculus and integral calculus are two branches of calculus, where the former
can be applied to the ubiquitous optimization problems in deep learning.
• A derivative can be interpreted as the instantaneous rate of change of a function with respect
to its variable. It is also the slope of the tangent line to the curve of the function.
2.4. Calculus 71
• A gradient is a vector whose components are the partial derivatives of a multivariate function
with respect to all its variables.
• The chain rule enables us to differentiate composite functions.
Exercises
1. Plot the function y = f (x) = x3 − 1

x and its tangent line when x = 1.
2. Find the gradient of the function f (x) = 3x21 + 5ex2 .
3. What is the gradient of the function f (x) = ∥x∥2 ?
4. Can you write out the chain rule for the case where u = f (x, y, z) and x = x(a, b), y = y(a, b),
and z = z(a, b)?
Discussions45
2.5 Automatic Differentiation
As we have explained in Section 2.4, differentiation is a crucial step in nearly all deep learning
optimization algorithms. While the calculations for taking these derivatives are straightforward,
requiring only some basic calculus, for complex models, working out the updates by hand can be
a pain (and often error-prone).
Deep learning frameworks expedite this work by automatically calculating derivatives, i.e., auto-
matic differentiation. In practice, based on our designed model the system builds a computational
graph, tracking which data combined through which operations to produce the output. Automatic
differentiation enables the system to subsequently backpropagate gradients. Here, backpropagate
simply means to trace through the computational graph, filling in the partial derivatives with re-
spect to each parameter.
from mxnet import autograd, np, npx

npx.set_np()
2.5.1 A Simple Example
As a toy example, say that we are interested in differentiating the function y = 2x⊤ x with respect
to the column vector x. To start, let us create the variable x and assign it an initial value.
x = np.arange(4.0)
x
array([0., 1., 2., 3.])
Before we even calculate the gradient of y with respect to x, we will need a place to store it. It is
important that we do not allocate new memory every time we take a derivative with respect to a
parameter because we will often update the same parameters thousands or millions of times and
45
could quickly run out of memory. Note that a gradient of a scalar-valued function with respect to
a vector x is itself vector-valued and has the same shape as x.
# We allocate memory for a tensor's gradient by invoking àttach_grad`

x.attach_grad()
# After we calculate a gradient taken with respect to `x`, we will be able to
# access it via the `grad` attribute, whose values are initialized with 0s
x.grad
array([0., 0., 0., 0.])
Now let us calculate y.
# Place our code inside an àutograd.record` scope to build the computational

# graph
with autograd.record():
y = 2 * np.dot(x, x)
y
array(28.)
Since x is a vector of length 4, an inner product of x and x is performed, yielding the scalar output
that we assign to y. Next, we can automatically calculate the gradient of y with respect to each
component of x by calling the function for backpropagation and printing the gradient.
y.backward()
x.grad
array([ 0., 4., 8., 12.])
The gradient of the function y = 2x⊤ x with respect to x should be 4x. Let us quickly verify that
our desired gradient was calculated correctly.
x.grad == 4 * x
array([ True, True, True, True])
Now let us calculate another function of x.
y = x.sum()
y.backward()
x.grad # Overwritten by the newly calculated gradient
array([1., 1., 1., 1.])
2.5. Automatic Differentiation 73

2.5.2 Backward for Non-Scalar Variables
Technically, when y is not a scalar, the most natural interpretation of the differentiation of a vector
y with respect to a vector x is a matrix. For higher-order and higher-dimensional y and x, the
differentiation result could be a high-order tensor.
However, while these more exotic objects do show up in advanced machine learning (including in
deep learning), more often when we are calling backward on a vector, we are trying to calculate
the derivatives of the loss functions for each constituent of a batch of training examples. Here, our
intent is not to calculate the differentiation matrix but rather the sum of the partial derivatives
computed individually for each example in the batch.
# When we invoke `backward` on a vector-valued variable `y` (function of `x`),

# a new scalar variable is created by summing the elements in `y`. Then the
# gradient of that scalar variable with respect to `x` is computed
y = x * x # `y` is a vector
y.backward()
x.grad # Equals to y = sum(x * x)
array([0., 2., 4., 6.])
2.5.3 Detaching Computation
Sometimes, we wish to move some calculations outside of the recorded computational graph. For
example, say that y was calculated as a function of x, and that subsequently z was calculated as a
function of both y and x. Now, imagine that we wanted to calculate the gradient of z with respect
to x, but wanted for some reason to treat y as a constant, and only take into account the role that
x played after y was calculated.
Here, we can detach y to return a new variable u that has the same value as y but discards any
information about how y was computed in the computational graph. In other words, the gradient
will not flow backwards through u to x. Thus, the following backpropagation function computes
the partial derivative of z = u * x with respect to x while treating u as a constant, instead of the
partial derivative of z = x * x * x with respect to x.
y = x * x
u = y.detach()
z = u * x
z.backward()
x.grad == u
Since the computation of y was recorded, we can subsequently invoke backpropagation on y to

get the derivative of y = x * x with respect to x, which is 2 * x.
y.backward()
x.grad == 2 * x
2.5.4 Computing the Gradient of Python Control Flow
One benefit of using automatic differentiation is that even if building the computational graph
of a function required passing through a maze of Python control flow (e.g., conditionals, loops,
and arbitrary function calls), we can still calculate the gradient of the resulting variable. In the
following snippet, note that the number of iterations of the while loop and the evaluation of the
if statement both depend on the value of the input a.
def f(a):
b = a * 2
while np.linalg.norm(b) < 1000:
b = b * 2
if b.sum() > 0:
c = b
else:
c = 100 * b
return c
Let us compute the gradient.
a = np.random.normal()
a.attach_grad()
d = f(a)
d.backward()
We can now analyze the f function defined above. Note that it is piecewise linear in its input a. In
other words, for any a there exists some constant scalar k such that f(a) = k * a, where the value
of k depends on the input a. Consequently d / a allows us to verify that the gradient is correct.
a.grad == d / a
array(True)
Summary
• Deep learning frameworks can automate the calculation of derivatives. To use it, we first
attach gradients to those variables with respect to which we desire partial derivatives. We
then record the computation of our target value, execute its function for backpropagation,
and access the resulting gradient.
2.5. Automatic Differentiation 75

Exercises
1. Why is the second derivative much more expensive to compute than the first derivative?
2. After running the function for backpropagation, immediately run it again and see what hap-
pens.
3. In the control flow example where we calculate the derivative of d with respect to a, what
would happen if we changed the variable a to a random vector or matrix. At this point, the
result of the calculation f(a) is no longer a scalar. What happens to the result? How do we
analyze this?
4. Redesign an example of finding the gradient of the control flow. Run and analyze the result.
df (x)
5. Let f (x) = sin(x). Plot f (x) and dx , where the latter is computed without exploiting that
f ′ (x) = cos(x).
Discussions46
2.6 Probability
In some form or another, machine learning is all about making predictions. We might want to
predict the probability of a patient suffering a heart attack in the next year, given their clinical his-
tory. In anomaly detection, we might want to assess how likely a set of readings from an airplaneʼs
jet engine would be, were it operating normally. In reinforcement learning, we want an agent to
act intelligently in an environment. This means we need to think about the probability of getting
a high reward under each of the available actions. And when we build recommender systems we
also need to think about probability. For example, say hypothetically that we worked for a large
online bookseller. We might want to estimate the probability that a particular user would buy
a particular book. For this we need to use the language of probability. Entire courses, majors,
theses, careers, and even departments, are devoted to probability. So naturally, our goal in this
section is not to teach the whole subject. Instead we hope to get you off the ground, to teach you
just enough that you can start building your first deep learning models, and to give you enough of
a flavor for the subject that you can begin to explore it on your own if you wish.
We have already invoked probabilities in previous sections without articulating what precisely
they are or giving a concrete example. Let us get more serious now by considering the first case:
distinguishing cats and dogs based on photographs. This might sound simple but it is actually a
formidable challenge. To start with, the difficulty of the problem may depend on the resolution
of the image.
46
Fig. 2.6.1: Images of varying resolutions (10 × 10, 20 × 20, 40 × 40, 80 × 80, and 160 × 160 pixels).
As shown in Fig. 2.6.1, while it is easy for humans to recognize cats and dogs at the resolution of
160 × 160 pixels, it becomes challenging at 40 × 40 pixels and next to impossible at 10 × 10 pixels.
In other words, our ability to tell cats and dogs apart at a large distance (and thus low resolution)
might approach uninformed guessing. Probability gives us a formal way of reasoning about our
level of certainty. If we are completely sure that the image depicts a cat, we say that the probability
that the corresponding label y is “cat”, denoted P (y = “cat”) equals 1. If we had no evidence to
suggest that y = “cat” or that y = “dog”, then we might say that the two possibilities were equally
likely expressing this as P (y = “cat”) = P (y = “dog”) = 0.5. If we were reasonably confident, but
not sure that the image depicted a cat, we might assign a probability 0.5 < P (y = “cat”) < 1.
Now consider the second case: given some weather monitoring data, we want to predict the proba-
bility that it will rain in Taipei tomorrow. If it is summertime, the rain might come with probability
0.5.
In both cases, we have some value of interest. And in both cases we are uncertain about the out-
come. But there is a key difference between the two cases. In this first case, the image is in fact
either a dog or a cat, and we just do not know which. In the second case, the outcome may actu-
ally be a random event, if you believe in such things (and most physicists do). So probability is a
flexible language for reasoning about our level of certainty, and it can be applied effectively in a
broad set of contexts.
2.6.1 Basic Probability Theory
Say that we cast a die and want to know what the chance is of seeing a 1 rather than another digit.
If the die is fair, all the six outcomes {1, . . . , 6} are equally likely to occur, and thus we would see
a 1 in one out of six cases. Formally we state that 1 occurs with probability 16 .
For a real die that we receive from a factory, we might not know those proportions and we would
need to check whether it is tainted. The only way to investigate the die is by casting it many times
and recording the outcomes. For each cast of the die, we will observe a value in {1, . . . , 6}. Given
these outcomes, we want to investigate the probability of observing each outcome.
2.6. Probability 77
One natural approach for each value is to take the individual count for that value and to divide it
by the total number of tosses. This gives us an estimate of the probability of a given event. The law
of large numbers tell us that as the number of tosses grows this estimate will draw closer and closer
to the true underlying probability. Before going into the details of what is going here, let us try it
out.
To start, let us import the necessary packages.
%matplotlib inline
import random
npx.set_np()
Next, we will want to be able to cast the die. In statistics we call this process of drawing examples
from probability distributions sampling. The distribution that assigns probabilities to a number
of discrete choices is called the multinomial distribution. We will give a more formal definition of
distribution later, but at a high level, think of it as just an assignment of probabilities to events.
To draw a single sample, we simply pass in a vector of probabilities. The output is another vector
of the same length: its value at index i is the number of times the sampling outcome corresponds
to i.
fair_probs = [1.0 / 6] * 6
np.random.multinomial(1, fair_probs)
array([0, 0, 0, 1, 0, 0], dtype=int64)
If you run the sampler a bunch of times, you will find that you get out random values each time.
As with estimating the fairness of a die, we often want to generate many samples from the same
distribution. It would be unbearably slow to do this with a Python for loop, so the function we are
using supports drawing multiple samples at once, returning an array of independent samples in
any shape we might desire.
np.random.multinomial(10, fair_probs)
array([1, 1, 5, 1, 1, 1], dtype=int64)
Now that we know how to sample rolls of a die, we can simulate 1000 rolls. We can then go through
and count, after each of the 1000 rolls, how many times each number was rolled. Specifically, we
calculate the relative frequency as the estimate of the true probability.
counts = np.random.multinomial(1000, fair_probs).astype(np.float32)

counts / 1000
array([0.162, 0.149, 0.178, 0.17 , 0.166, 0.175])
Because we generated the data from a fair die, we know that each outcome has true probability 16 ,
roughly 0.167, so the above output estimates look good.
We can also visualize how these probabilities converge over time towards the true probability. Let
us conduct 500 groups of experiments where each group draws 10 samples.
counts = np.random.multinomial(10, fair_probs, size=500)
cum_counts = counts.astype(np.float32).cumsum(axis=0)
estimates = cum_counts / cum_counts.sum(axis=1, keepdims=True)
d2l.set_figsize((6, 4.5))
for i in range(6):
d2l.plt.plot(estimates[:, i].asnumpy(),
label=("P(die=" + str(i + 1) + ")"))
d2l.plt.axhline(y=0.167, color='black', linestyle='dashed')
d2l.plt.gca().set_xlabel('Groups of experiments')
d2l.plt.gca().set_ylabel('Estimated probability')
d2l.plt.legend();
Each solid curve corresponds to one of the six values of the die and gives our estimated probability
that the die turns up that value as assessed after each group of experiments. The dashed black line
gives the true underlying probability. As we get more data by conducting more experiments, the
6 solid curves converge towards the true probability.
Axioms of Probability Theory
When dealing with the rolls of a die, we call the set S = {1, 2, 3, 4, 5, 6} the sample space or outcome
space, where each element is an outcome. An event is a set of outcomes from a given sample space.
For instance, “seeing a 5” ({5}) and “seeing an odd number” ({1, 3, 5}) are both valid events of
rolling a die. Note that if the outcome of a random experiment is in event A, then event A has
occurred. That is to say, if 3 dots faced up after rolling a die, since 3 ∈ {1, 3, 5}, we can say that the
event “seeing an odd number” has occurred.
Formally, probability can be thought of a function that maps a set to a real value. The probability
2.6. Probability 79
of an event A in the given sample space S, denoted as P (A), satisfies the following properties:
• For any event A, its probability is never negative, i.e., P (A) ≥ 0;
• Probability of the entire sample space is 1, i.e., P (S) = 1;
• For any countable sequence of events A1 , A2 , . . . that are mutually exclusive (Ai ∩Aj = ∅ for all
i ̸= j), the
∪∞probability
∑∞that any happens is equal to the sum of their individual probabilities,
i.e., P ( i=1 Ai ) = i=1 P (Ai ).
These are also the axioms of probability theory, proposed by Kolmogorov in 1933. Thanks to this
axiom system, we can avoid any philosophical dispute on randomness; instead, we can reason
rigorously with a mathematical language. For instance, by letting event A1 be the entire sample
space and Ai = ∅ for all i > 1, we can prove that P (∅) = 0, i.e., the probability of an impossible
event is 0.
Random Variables
In our random experiment of casting a die, we introduced the notion of a random variable. A ran-
dom variable can be pretty much any quantity and is not deterministic. It could take one value
among a set of possibilities in a random experiment. Consider a random variable X whose value
is in the sample space S = {1, 2, 3, 4, 5, 6} of rolling a die. We can denote the event “seeing a 5”
as {X = 5} or X = 5, and its probability as P ({X = 5}) or P (X = 5). By P (X = a), we make a
distinction between the random variable X and the values (e.g., a) that X can take. However, such
pedantry results in a cumbersome notation. For a compact notation, on one hand, we can just de-
note P (X) as the distribution over the random variable X: the distribution tells us the probability
that X takes any value. On the other hand, we can simply write P (a) to denote the probability that
a random variable takes the value a. Since an event in probability theory is a set of outcomes from
the sample space, we can specify a range of values for a random variable to take. For example,
P (1 ≤ X ≤ 3) denotes the probability of the event {1 ≤ X ≤ 3}, which means {X = 1, 2, or, 3}.
Equivalently, P (1 ≤ X ≤ 3) represents the probability that the random variable X can take a
value from {1, 2, 3}.
Note that there is a subtle difference between discrete random variables, like the sides of a die,
and continuous ones, like the weight and the height of a person. There is little point in ask-
ing whether two people have exactly the same height. If we take precise enough measure-
ments you will find that no two people on the planet have the exact same height. In fact, if
we take a fine enough measurement, you will not have the same height when you wake up and
when you go to sleep. So there is no purpose in asking about the probability that someone is
1.80139278291028719210196740527486202 meters tall. Given the world population of humans the
probability is virtually 0. It makes more sense in this case to ask whether someoneʼs height falls
into a given interval, say between 1.79 and 1.81 meters. In these cases we quantify the likelihood
that we see a value as a density. The height of exactly 1.80 meters has no probability, but nonzero
density. In the interval between any two different heights we have nonzero probability. In the rest
of this section, we consider probability in discrete space. For probability over continuous random
variables, you may refer to Section 18.6.
2.6.2 Dealing with Multiple Random Variables
Very often, we will want to consider more than one random variable at a time. For instance, we
may want to model the relationship between diseases and symptoms. Given a disease and a symp-
tom, say “flu” and “cough”, either may or may not occur in a patient with some probability. While
we hope that the probability of both would be close to zero, we may want to estimate these prob-
abilities and their relationships to each other so that we may apply our inferences to effect better
medical care.
As a more complicated example, images contain millions of pixels, thus millions of random vari-
ables. And in many cases images will come with a label, identifying objects in the image. We can
also think of the label as a random variable. We can even think of all the metadata as random
variables such as location, time, aperture, focal length, ISO, focus distance, and camera type. All
of these are random variables that occur jointly. When we deal with multiple random variables,
there are several quantities of interest.
Joint Probability
The first is called the joint probability P (A = a, B = b). Given any values a and b, the joint proba-
bility lets us answer, what is the probability that A = a and B = b simultaneously? Note that for
any values a and b, P (A = a, B = b) ≤ P (A = a). This has to be the case, since for A = a and
B = b to happen, A = a has to happen and B = b also has to happen (and vice versa). Thus, A = a
and B = b cannot be more likely than A = a or B = b individually.
Conditional Probability
This brings us to an interesting ratio: 0 ≤ P (A=a,B=b)

P (A=a) ≤ 1. We call this ratio a conditional probability
and denote it by P (B = b | A = a): it is the probability of B = b, provided that A = a has occurred.
Bayes’ theorem
Using the definition of conditional probabilities, we can derive one of the most useful and cel-
ebrated equations in statistics: Bayes’ theorem. It goes as follows. By construction, we have the
multiplication rule that P (A, B) = P (B | A)P (A). By symmetry, this also holds for P (A, B) =
P (A | B)P (B). Assume that P (B) > 0. Solving for one of the conditional variables we get
P (B | A)P (A)
P (A | B) = . (2.6.1)
P (B)
Note that here we use the more compact notation where P (A, B) is a joint distribution and P (A | B)
is a conditional distribution. Such distributions can be evaluated for particular values A = a, B = b.
2.6. Probability 81
Marginalization
Bayesʼ theorem is very useful if we want to infer one thing from the other, say cause and effect,
but we only know the properties in the reverse direction, as we will see later in this section. One
important operation that we need, to make this work, is marginalization. It is the operation of
determining P (B) from P (A, B). We can see that the probability of B amounts to accounting for
all possible choices of A and aggregating the joint probabilities over all of them:
∑
P (B) = P (A, B), (2.6.2)
A
which is also known as the sum rule. The probability or distribution as a result of marginalization
is called a marginal probability or a marginal distribution.
Independence
Another useful property to check for is dependence vs. independence. Two random variables A and
B being independent means that the occurrence of one event of A does not reveal any information
about the occurrence of an event of B. In this case P (B | A) = P (B). Statisticians typically
express this as A ⊥ B. From Bayesʼ theorem, it follows immediately that also P (A | B) = P (A).
In all the other cases we call A and B dependent. For instance, two successive rolls of a die are
independent. In contrast, the position of a light switch and the brightness in the room are not
(they are not perfectly deterministic, though, since we could always have a broken light bulb,
power failure, or a broken switch).
Since P (A | B) = PP(A,B)
(B) = P (A) is equivalent to P (A, B) = P (A)P (B), two random variables are
independent if and only if their joint distribution is the product of their individual distributions.
Likewise, two random variables A and B are conditionally independent given another random vari-
able C if and only if P (A, B | C) = P (A | C)P (B | C). This is expressed as A ⊥ B | C.
Application
Let us put our skills to the test. Assume that a doctor administers an AIDS test to a patient. This
test is fairly accurate and it fails only with 1% probability if the patient is healthy but reporting
him as diseased. Moreover, it never fails to detect HIV if the patient actually has it. We use D1 to
indicate the diagnosis (1 if positive and 0 if negative) and H to denote the HIV status (1 if positive
and 0 if negative). Table 2.6.1 lists such conditional probabilities.
Table 2.6.1: Conditional probability of P (D1 | H).

Conditional probability H = 1 H = 0
P (D1 = 1 | H) 1 0.01
P (D1 = 0 | H) 0 0.99
Note that the column sums are all 1 (but the row sums are not), since the conditional probabil-
ity needs to sum up to 1, just like the probability. Let us work out the probability of the patient
having AIDS if the test comes back positive, i.e., P (H = 1 | D1 = 1). Obviously this is going to
depend on how common the disease is, since it affects the number of false alarms. Assume that
the population is quite healthy, e.g., P (H = 1) = 0.0015. To apply Bayesʼ theorem, we need to
apply marginalization and the multiplication rule to determine
P (D1 = 1)
=P (D1 = 1, H = 0) + P (D1 = 1, H = 1)
(2.6.3)
=P (D1 = 1 | H = 0)P (H = 0) + P (D1 = 1 | H = 1)P (H = 1)
=0.011485.
Thus, we get
P (H = 1 | D1 = 1)
P (D1 = 1 | H = 1)P (H = 1)
= . (2.6.4)
P (D1 = 1)
=0.1306
In other words, there is only a 13.06% chance that the patient actually has AIDS, despite using a
very accurate test. As we can see, probability can be counterintuitive.
What should a patient do upon receiving such terrifying news? Likely, the patient would ask the
physician to administer another test to get clarity. The second test has different characteristics
and it is not as good as the first one, as shown in Table 2.6.2.
Table 2.6.2: Conditional probability of P (D2 | H).

Conditional probability H = 1 H = 0
P (D2 = 1 | H) 0.98 0.03
P (D2 = 0 | H) 0.02 0.97
Unfortunately, the second test comes back positive, too. Let us work out the requisite probabilities
to invoke Bayesʼ theorem by assuming the conditional independence:
P (D1 = 1, D2 = 1 | H = 0)
=P (D1 = 1 | H = 0)P (D2 = 1 | H = 0) (2.6.5)
=0.0003,
P (D1 = 1, D2 = 1 | H = 1)
=P (D1 = 1 | H = 1)P (D2 = 1 | H = 1) (2.6.6)
=0.98.
Now we can apply marginalization and the multiplication rule:
P (D1 = 1, D2 = 1)
=P (D1 = 1, D2 = 1, H = 0) + P (D1 = 1, D2 = 1, H = 1)
(2.6.7)
=P (D1 = 1, D2 = 1 | H = 0)P (H = 0) + P (D1 = 1, D2 = 1 | H = 1)P (H = 1)
=0.00176955.
In the end, the probability of the patient having AIDS given both positive tests is
P (H = 1 | D1 = 1, D2 = 1)
P (D1 = 1, D2 = 1 | H = 1)P (H = 1)
= (2.6.8)
P (D1 = 1, D2 = 1)
=0.8307.
That is, the second test allowed us to gain much higher confidence that not all is well. Despite the
second test being considerably less accurate than the first one, it still significantly improved our
estimate.
2.6. Probability 83
2.6.3 Expectation and Variance
To summarize key characteristics of probability distributions, we need some measures. The ex-
pectation (or average) of the random variable X is denoted as
∑
E[X] = xP (X = x). (2.6.9)
x
When the input of a function f (x) is a random variable drawn from the distribution P with differ-
ent values x, the expectation of f (x) is computed as
∑
Ex∼P [f (x)] = f (x)P (x). (2.6.10)
x
In many cases we want to measure by how much the random variable X deviates from its expec-
tation. This can be quantified by the variance
[ ]
Var[X] = E (X − E[X])2 = E[X 2 ] − E[X]2 . (2.6.11)
Its square root is called the standard deviation. The variance of a function of a random variable
measures by how much the function deviates from the expectation of the function, as different
values x of the random variable are sampled from its distribution:
[ ]
Var[f (x)] = E (f (x) − E[f (x)])2 . (2.6.12)
Summary
• We can sample from probability distributions.

• We can analyze multiple random variables using joint distribution, conditional distribution,
Bayesʼ theorem, marginalization, and independence assumptions.
• Expectation and variance offer useful measures to summarize key characteristics of proba-
bility distributions.
Exercises
1. We conducted m = 500 groups of experiments where each group draws n = 10 samples.

Vary m and n. Observe and analyze the experimental results.
2. Given two events with probability P (A) and P (B), compute upper and lower bounds on
P (A ∪ B) and P (A ∩ B). (Hint: display the situation using a Venn Diagram47 .)
3. Assume that we have a sequence of random variables, say A, B, and C, where B only de-
pends on A, and C only depends on B, can you simplify the joint probability P (A, B, C)?
(Hint: this is a Markov Chain48 .)
4. In Section 2.6.2, the first test is more accurate. Why not just run the first test a second time?
Discussions49
47
https://en.wikipedia.org/wiki/Venn_diagram
48
https://en.wikipedia.org/wiki/Markov_chain
49
2.7 Documentation
Due to constraints on the length of this book, we cannot possibly introduce every single MXNet
function and class (and you probably would not want us to). The API documentation and addi-
tional tutorials and examples provide plenty of documentation beyond the book. In this section
we provide you with some guidance to exploring the MXNet API.
2.7.1 Finding All the Functions and Classes in a Module
In order to know which functions and classes can be called in a module, we invoke the dir func-
tion. For instance, we can query all properties in the module for generating random numbers:

print(dir(np.random))
['__all__', '__builtins__', '__cached__', '__doc__', '__file__', '__loader__', '__name__', '_

,→_package__', '__spec__', '_mx_nd_np', 'absolute_import', 'choice', 'multinomial', 'normal',
,→ 'rand', 'randint', 'shuffle', 'uniform']
Generally, we can ignore functions that start and end with __ (special objects in Python) or func-
tions that start with a single _(usually internal functions). Based on the remaining function or
attribute names, we might hazard a guess that this module offers various methods for generating
random numbers, including sampling from the uniform distribution (uniform), normal distribu-
tion (normal), and multinomial distribution (multinomial).
2.7.2 Finding the Usage of Specific Functions and Classes
For more specific instructions on how to use a given function or class, we can invoke the help
function. As an example, let us explore the usage instructions for tensorsʼ ones function.
help(np.ones)
Help on function ones in module mxnet.numpy:
ones(shape, dtype=<class 'numpy.float32'>, order='C', ctx=None)

Return a new array of given shape and type, filled with ones.
This function currently only supports storing multi-dimensional data
in row-major (C-style).
Parameters
----------
shape : int or tuple of int
The shape of the empty array.
dtype : str or numpy.dtype, optional
An optional value type. Default is numpy.float32. Note that this
behavior is different from NumPy's ones function where float64
is the default value, because float32 is considered as the default
data type in deep learning.
order : {'C'}, optional, default: 'C'
2.7. Documentation 85
How to store multi-dimensional data in memory, currently only row-major
(C-style) is supported.
ctx : Context, optional
An optional device context (default is the current default context).
Returns
-------
out : ndarray
Array of ones with the given shape, dtype, and ctx.
Examples
--------
>>> np.ones(5)
array([1., 1., 1., 1., 1.])
>>> np.ones((5,), dtype=int)

array([1, 1, 1, 1, 1], dtype=int64)
>>> np.ones((2, 1))

array([[1.],
[1.]])
>>> s = (2,2)
>>> np.ones(s)
array([[1., 1.],
[1., 1.]])
From the documentation, we can see that the ones function creates a new tensor with the specified
shape and sets all the elements to the value of 1. Whenever possible, you should run a quick test
to confirm your interpretation:
np.ones(4)
array([1., 1., 1., 1.])
In the Jupyter notebook, we can use ? to display the document in another window. For example,
list? will create content that is almost identical to help(list), displaying it in a new browser
window. In addition, if we use two question marks, such as list??, the Python code implementing
the function will also be displayed.
Summary
• The official documentation provides plenty of descriptions and examples that are beyond
this book.
• We can look up documentation for the usage of an API by calling the dir and help functions,
or ? and ?? in Jupyter notebooks.
Exercises
1. Look up the documentation for any function or class in the deep learning framework. Can
you also find the documentation on the official website of the framework?
Discussions50
50
2.7. Documentation 87
3 | Linear Neural Networks
Before we get into the details of deep neural networks, we need to cover the basics of neural net-
work training. In this chapter, we will cover the entire training process, including defining simple
neural network architectures, handling data, specifying a loss function, and training the model.
In order to make things easier to grasp, we begin with the simplest concepts. Fortunately, classic
statistical learning techniques such as linear and softmax regression can be cast as linear neural
networks. Starting from these classic algorithms, we will introduce you to the basics, providing
the basis for more complex techniques in the rest of the book.
3.1 Linear Regression
Regression refers to a set of methods for modeling the relationship between one or more indepen-
dent variables and a dependent variable. In the natural sciences and social sciences, the purpose
of regression is most often to characterize the relationship between the inputs and outputs. Ma-
chine learning, on the other hand, is most often concerned with prediction.
Regression problems pop up whenever we want to predict a numerical value. Common exam-
ples include predicting prices (of homes, stocks, etc.), predicting length of stay (for patients in
the hospital), demand forecasting (for retail sales), among countless others. Not every prediction
problem is a classic regression problem. In subsequent sections, we will introduce classification
problems, where the goal is to predict membership among a set of categories.
3.1.1 Basic Elements of Linear Regression
Linear regression may be both the simplest and most popular among the standard tools to regres-
sion. Dating back to the dawn of the 19th century, linear regression flows from a few simple
assumptions. First, we assume that the relationship between the independent variables x and the
dependent variable y is linear, i.e., that y can be expressed as a weighted sum of the elements
in x, given some noise on the observations. Second, we assume that any noise is well-behaved
(following a Gaussian distribution).
To motivate the approach, let us start with a running example. Suppose that we wish to estimate
the prices of houses (in dollars) based on their area (in square feet) and age (in years). To actually
develop a model for predicting house prices, we would need to get our hands on a dataset consist-
ing of sales for which we know the sale price, area, and age for each home. In the terminology of
machine learning, the dataset is called a training dataset or training set, and each row (here the data
corresponding to one sale) is called an example (or data point, data instance, sample). The thing we
are trying to predict (price) is called a label (or target). The independent variables (age and area)
upon which the predictions are based are called features (or covariates).
89
Typically, we will use n to denote the number of examples in our dataset. We index the data ex-
amples by i, denoting each input as x(i) = [x1 , x2 ]⊤ and the corresponding label as y (i) .
(i) (i)
Linear Model
The linearity assumption just says that the target (price) can be expressed as a weighted sum of
the features (area and age):
price = warea · area + wage · age + b. (3.1.1)
In (3.1.1), warea and wage are called weights, and b is called a bias (also called an offset or intercept).
The weights determine the influence of each feature on our prediction and the bias just says what
value the predicted price should take when all of the features take value 0. Even if we will never
see any homes with zero area, or that are precisely zero years old, we still need the bias or else we
will limit the expressivity of our model. Strictly speaking, (3.1.1) is an affine transformation of input
features, which is characterized by a linear transformation of features via weighted sum, combined
with a translation via the added bias.
Given a dataset, our goal is to choose the weights w and the bias b such that on average, the pre-
dictions made according to our model best fit the true prices observed in the data. Models whose
output prediction is determined by the affine transformation of input features are linear models,
where the affine transformation is specified by the chosen weights and bias.
In disciplines where it is common to focus on datasets with just a few features, explicitly express-
ing models long-form like this is common. In machine learning, we usually work with high-
dimensional datasets, so it is more convenient to employ linear algebra notation. When our inputs
consist of d features, we express our prediction ŷ (in general the “hat” symbol denotes estimates)
as
ŷ = w1 x1 + ... + wd xd + b. (3.1.2)
Collecting all features into a vector x ∈ Rd and all weights into a vector w ∈ Rd , we can express
our model compactly using a dot product:
ŷ = w⊤ x + b. (3.1.3)
In (3.1.3), the vector x corresponds to features of a single data example. We will often find it
convenient to refer to features of our entire dataset of n examples via the design matrix X ∈ Rn×d .
Here, X contains one row for every example and one column for every feature.
For a collection of features X, the predictions ŷ ∈ Rn can be expressed via the matrix-vector prod-
uct:
ŷ = Xw + b, (3.1.4)
where broadcasting (see Section 2.1.3) is applied during the summation. Given features of a train-
ing dataset X and corresponding (known) labels y, the goal of linear regression is to find the weight
vector w and the bias term b that given features of a new data example sampled from the same
distribution as X, the new exampleʼs label will (in expectation) be predicted with the lowest error.
Even if we believe that the best model for predicting y given x is linear, we would not expect to
find a real-world dataset of n examples where y (i) exactly equals w⊤ x(i) + b for all 1 ≤ i ≤ n. For
example, whatever instruments we use to observe the features X and labels y might suffer small
90 Chapter 3. Linear Neural Networks

amount of measurement error. Thus, even when we are confident that the underlying relationship
is linear, we will incorporate a noise term to account for such errors.
Before we can go about searching for the best parameters (or model parameters) w and b, we will need
two more things: (i) a quality measure for some given model; and (ii) a procedure for updating the
model to improve its quality.
Loss Function
Before we start thinking about how to fit data with our model, we need to determine a measure of
fitness. The loss function quantifies the distance between the real and predicted value of the target.
The loss will usually be a non-negative number where smaller values are better and perfect pre-
dictions incur a loss of 0. The most popular loss function in regression problems is the squared
error. When our prediction for an example i is ŷ (i) and the corresponding true label is y (i) , the
squared error is given by:
1 ( (i) )2
l(i) (w, b) = ŷ − y (i) . (3.1.5)
2
The constant 21 makes no real difference but will prove notationally convenient, canceling out
when we take the derivative of the loss. Since the training dataset is given to us, and thus out of
our control, the empirical error is only a function of the model parameters. To make things more
concrete, consider the example below where we plot a regression problem for a one-dimensional
case as shown in Fig. 3.1.1.
Fig. 3.1.1: Fit data with a linear model.
Note that large differences between estimates ŷ (i) and observations y (i) lead to even larger contri-
butions to the loss, due to the quadratic dependence. To measure the quality of a model on the
entire dataset of n examples, we simply average (or equivalently, sum) the losses on the training
set.
1 ∑ (i) 1 ∑ 1 ( ⊤ (i) )2
n n
L(w, b) = l (w, b) = w x + b − y (i) . (3.1.6)
n n 2
i=1 i=1
When training the model, we want to find parameters (w∗ , b∗ ) that minimize the total loss across
all training examples:
w∗ , b∗ = argmin L(w, b). (3.1.7)

w,b
3.1. Linear Regression 91

Analytic Solution
Linear regression happens to be an unusually simple optimization problem. Unlike most other
models that we will encounter in this book, linear regression can be solved analytically by applying
a simple formula. To start, we can subsume the bias b into the parameter w by appending a column
to the design matrix consisting of all ones. Then our prediction problem is to minimize ∥y−Xw∥2 .
There is just one critical point on the loss surface and it corresponds to the minimum of the loss
over the entire domain. Taking the derivative of the loss with respect to w and setting it equal to
zero yields the analytic (closed-form) solution:
w∗ = (X⊤ X)−1 X⊤ y. (3.1.8)
While simple problems like linear regression may admit analytic solutions, you should not get
used to such good fortune. Although analytic solutions allow for nice mathematical analysis, the
requirement of an analytic solution is so restrictive that it would exclude all of deep learning.
Minibatch Stochastic Gradient Descent
Even in cases where we cannot solve the models analytically, it turns out that we can still train
models effectively in practice. Moreover, for many tasks, those difficult-to-optimize models turn
out to be so much better that figuring out how to train them ends up being well worth the trouble.
The key technique for optimizing nearly any deep learning model, and which we will call upon
throughout this book, consists of iteratively reducing the error by updating the parameters in the
direction that incrementally lowers the loss function. This algorithm is called gradient descent.
The most naive application of gradient descent consists of taking the derivative of the loss func-
tion, which is an average of the losses computed on every single example in the dataset. In prac-
tice, this can be extremely slow: we must pass over the entire dataset before making a single
update. Thus, we will often settle for sampling a random minibatch of examples every time we
need to compute the update, a variant called minibatch stochastic gradient descent.
In each iteration, we first randomly sample a minibatch B consisting of a fixed number of training
examples. We then compute the derivative (gradient) of the average loss on the minibatch with
regard to the model parameters. Finally, we multiply the gradient by a predetermined positive
value η and subtract the resulting term from the current parameter values.
We can express the update mathematically as follows (∂ denotes the partial derivative):
η ∑
(w, b) ← (w, b) − ∂(w,b) l(i) (w, b). (3.1.9)
|B|
i∈B
To summarize, steps of the algorithm are the following: (i) we initialize the values of the model
parameters, typically at random; (ii) we iteratively sample random minibatches from the data,
updating the parameters in the direction of the negative gradient. For quadratic losses and affine
transformations, we can write this out explicitly as follows:
η ∑ η ∑ (i) ( ⊤ (i) )
w←w− ∂w l(i) (w, b) = w − x w x + b − y (i) ,
|B| |B|
i∈B i∈B
η ∑ η ∑ ( ⊤ (i) ) (3.1.10)
b←b− ∂b l (w, b) = b −
(i)
w x + b − y (i) .
|B| |B|
i∈B i∈B
Note that w and x are vectors in (3.1.10). Here, the more elegant vector notation makes the math
much more readable than expressing things in terms of coefficients, say w1 , w2 , . . . , wd . The set

cardinality |B| represents the number of examples in each minibatch (the batch size) and η denotes
the learning rate. We emphasize that the values of the batch size and learning rate are manually
pre-specified and not typically learned through model training. These parameters that are tun-
able but not updated in the training loop are called hyperparameters. Hyperparameter tuning is the
process by which hyperparameters are chosen, and typically requires that we adjust them based
on the results of the training loop as assessed on a separate validation dataset (or validation set).
After training for some predetermined number of iterations (or until some other stopping criteria
are met), we record the estimated model parameters, denoted ŵ, b̂. Note that even if our function
is truly linear and noiseless, these parameters will not be the exact minimizers of the loss because,
although the algorithm converges slowly towards the minimizers it cannot achieve it exactly in a
finite number of steps.
Linear regression happens to be a learning problem where there is only one minimum over the
entire domain. However, for more complicated models, like deep networks, the loss surfaces
contain many minima. Fortunately, for reasons that are not yet fully understood, deep learning
practitioners seldom struggle to find parameters that minimize the loss on training sets. The more
formidable task is to find parameters that will achieve low loss on data that we have not seen
before, a challenge called generalization. We return to these topics throughout the book.
Making Predictions with the Learned Model
Given the learned linear regression model ŵ⊤ x + b̂, we can now estimate the price of a new house
(not contained in the training data) given its area x1 and age x2 . Estimating targets given features
is commonly called prediction or inference.
We will try to stick with prediction because calling this step inference, despite emerging as standard
jargon in deep learning, is somewhat of a misnomer. In statistics, inference more often denotes
estimating parameters based on a dataset. This misuse of terminology is a common source of
confusion when deep learning practitioners talk to statisticians.
3.1.2 Vectorization for Speed
When training our models, we typically want to process whole minibatches of examples simulta-
neously. Doing this efficiently requires that we vectorize the calculations and leverage fast linear
algebra libraries rather than writing costly for-loops in Python.
%matplotlib inline
import math
import time
To illustrate why this matters so much, we can consider two methods for adding vectors. To start
we instantiate two 10000-dimensional vectors containing all ones. In one method we will loop
over the vectors with a Python for-loop. In the other method we will rely on a single call to +.
n = 10000
a = np.ones(n)
b = np.ones(n)
Since we will benchmark the running time frequently in this book, let us define a timer.

class Timer: #@save
"""Record multiple running times."""
def __init__(self):
self.times = []
self.start()
def start(self):
"""Start the timer."""
self.tik = time.time()
def stop(self):
"""Stop the timer and record the time in a list."""
self.times.append(time.time() - self.tik)
return self.times[-1]
def avg(self):
"""Return the average time."""
return sum(self.times) / len(self.times)
def sum(self):
"""Return the sum of time."""
return sum(self.times)
def cumsum(self):
"""Return the accumulated time."""
return np.array(self.times).cumsum().tolist()
Now we can benchmark the workloads. First, we add them, one coordinate at a time, using a
for-loop.
c = np.zeros(n)
timer = Timer()
for i in range(n):
c[i] = a[i] + b[i]
f'{timer.stop():.5f} sec'
'4.22384 sec'
Alternatively, we rely on the reloaded + operator to compute the elementwise sum.
timer.start()
d = a + b
'0.00023 sec'
You probably noticed that the second method is dramatically faster than the first. Vectorizing
code often yields order-of-magnitude speedups. Moreover, we push more of the mathematics to
the library and need not write as many calculations ourselves, reducing the potential for errors.

3.1.3 The Normal Distribution and Squared Loss
While you can already get your hands dirty using only the information above, in the following we
can more formally motivate the square loss objective via assumptions about the distribution of
noise.
Linear regression was invented by Gauss in 1795, who also discovered the normal distribution
(also called the Gaussian). It turns out that the connection between the normal distribution and
linear regression runs deeper than common parentage. To refresh your memory, the probability
density of a normal distribution with mean µ and variance σ 2 (standard deviation σ) is given as
( )
1 1
p(x) = √ exp − 2 (x − µ) .2
(3.1.11)
2πσ 2 2σ
Below we define a Python function to compute the normal distribution.
def normal(x, mu, sigma):

p = 1 / math.sqrt(2 * math.pi * sigma**2)
return p * np.exp(-0.5 / sigma**2 * (x - mu)**2)
We can now visualize the normal distributions.
# Use numpy again for visualization

x = np.arange(-7, 7, 0.01)
# Mean and standard deviation pairs

params = [(0, 1), (0, 2), (3, 1)]
d2l.plot(x, [normal(x, mu, sigma) for mu, sigma in params], xlabel='x',
ylabel='p(x)', figsize=(4.5, 2.5),
legend=[f'mean {mu}, std {sigma}' for mu, sigma in params])
As we can see, changing the mean corresponds to a shift along the x-axis, and increasing the
variance spreads the distribution out, lowering its peak.
One way to motivate linear regression with the mean squared error loss function (or simply square
loss) is to formally assume that observations arise from noisy observations, where the noise is
normally distributed as follows:
y = w⊤ x + b + ϵ where ϵ ∼ N (0, σ 2 ). (3.1.12)

Thus, we can now write out the likelihood of seeing a particular y for a given x via
( )
1 1 ⊤
P (y | x) = √ exp − 2 (y − w x − b) . 2
(3.1.13)
2πσ 2 2σ
Now, according to the principle of maximum likelihood, the best values of parameters w and b are
those that maximize the likelihood of the entire dataset:
∏
n
P (y | X) = p(y (i) |x(i) ). (3.1.14)
i=1
Estimators chosen according to the principle of maximum likelihood are called maximum likeli-
hood estimators. While, maximizing the product of many exponential functions, might look diffi-
cult, we can simplify things significantly, without changing the objective, by maximizing the log
of the likelihood instead. For historical reasons, optimizations are more often expressed as mini-
mization rather than maximization. So, without changing anything we can minimize the negative
log-likelihood − log P (y | X). Working out the mathematics gives us:
∑ 1 ( (i) )2
n
1 ⊤ (i)
− log P (y | X) = log(2πσ 2 ) + y − w x − b . (3.1.15)
2 2σ 2
i=1
Now we just need one more assumption that σ is some fixed constant. Thus we can ignore the first
term because it does not depend on w or b. Now the second term is identical to the squared error
loss introduced earlier, except for the multiplicative constant σ12 . Fortunately, the solution does
not depend on σ. It follows that minimizing the mean squared error is equivalent to maximum
likelihood estimation of a linear model under the assumption of additive Gaussian noise.
3.1.4 From Linear Regression to Deep Networks
So far we only talked about linear models. While neural networks cover a much richer family of
models, we can begin thinking of the linear model as a neural network by expressing it in the
language of neural networks. To begin, let us start by rewriting things in a “layer” notation.
Neural Network Diagram
Deep learning practitioners like to draw diagrams to visualize what is happening in their models.
In Fig. 3.1.2, we depict our linear regression model as a neural network. Note that these diagrams
highlight the connectivity pattern such as how each input is connected to the output, but not the
values taken by the weights or biases.
Fig. 3.1.2: Linear regression is a single-layer neural network.
For the neural network shown in Fig. 3.1.2, the inputs are x1 , . . . , xd , so the number of inputs (or
feature dimensionality) in the input layer is d. The output of the network in Fig. 3.1.2 is o1 , so the

number of outputs in the output layer is 1. Note that the input values are all given and there is just
a single computed neuron. Focusing on where computation takes place, conventionally we do not
consider the input layer when counting layers. That is to say, the number of layers for the neural
network in Fig. 3.1.2 is 1. We can think of linear regression models as neural networks consisting
of just a single artificial neuron, or as single-layer neural networks.
Since for linear regression, every input is connected to every output (in this case there is only one
output), we can regard this transformation (the output layer in Fig. 3.1.2) as a fully-connected layer
or dense layer. We will talk a lot more about networks composed of such layers in the next chapter.
Biology
Since linear regression (invented in 1795) predates computational neuroscience, it might seem
anachronistic to describe linear regression as a neural network. To see why linear models were a
natural place to begin when the cyberneticists/neurophysiologists Warren McCulloch and Walter
Pitts began to develop models of artificial neurons, consider the cartoonish picture of a biological
neuron in Fig. 3.1.3, consisting of dendrites (input terminals), the nucleus (CPU), the axon (out-
put wire), and the axon terminals (output terminals), enabling connections to other neurons via
synapses.
Dendrite
Axon Terminal
Node of
Ranvier
Cell body
Axon Schwann cell

Myelin sheath
Nucleus
Fig. 3.1.3: The real neuron.
Information xi arriving from other neurons (or environmental sensors such as the retina) is re-
ceived in the dendrites. In particular, that information is weighted by synaptic weights wi determin-
ing the effect of the inputs (e.g., activation or inhibition via the product xi wi ). The weighted
∑ inputs
arriving from multiple sources are aggregated in the nucleus as a weighted sum y = i xi wi + b,
and this information is then sent for further processing in the axon y, typically after some nonlin-
ear processing via σ(y). From there it either reaches its destination (e.g., a muscle) or is fed into
another neuron via its dendrites.
Certainly, the high-level idea that many such units could be cobbled together with the right con-
nectivity and right learning algorithm, to produce far more interesting and complex behavior than
any one neuron alone could express owes to our study of real biological neural systems.
At the same time, most research in deep learning today draws little direct inspiration in neuro-
science. We invoke Stuart Russell and Peter Norvig who, in their classic AI text book Artificial In-
telligence: A Modern Approach (Russell & Norvig, 2016), pointed out that although airplanes might
have been inspired by birds, ornithology has not been the primary driver of aeronautics innovation

for some centuries. Likewise, inspiration in deep learning these days comes in equal or greater
measure from mathematics, statistics, and computer science.
Summary
• Key ingredients in a machine learning model are training data, a loss function, an optimiza-
tion algorithm, and quite obviously, the model itself.
• Vectorizing makes everything better (mostly math) and faster (mostly code).
• Minimizing an objective function and performing maximum likelihood estimation can
mean the same thing.
• Linear regression models are neural networks, too.
Exercises
1. Assume
∑ that we have some data x1 , . . . , xn ∈ R. Our goal is to find a constant b such that
i (xi − b) is minimized.
2
1. Find a analytic solution for the optimal value of b.

2. How does this problem and its solution relate to the normal distribution?
2. Derive the analytic solution to the optimization problem for linear regression with squared
error. To keep things simple, you can omit the bias b from the problem (we can do this in
principled fashion by adding one column to X consisting of all ones).
1. Write out the optimization problem in matrix and vector notation (treat all the data as
a single matrix, and all the target values as a single vector).
2. Compute the gradient of the loss with respect to w.
3. Find the analytic solution by setting the gradient equal to zero and solving the matrix
equation.
4. When might this be better than using stochastic gradient descent? When might this
method break?
3. Assume that the noise model governing the additive noise ϵ is the exponential distribution.
That is, p(ϵ) = 12 exp(−|ϵ|).
1. Write out the negative log-likelihood of the data under the model − log P (y | X).
2. Can you find a closed form solution?
3. Suggest a stochastic gradient descent algorithm to solve this problem. What could pos-
sibly go wrong (hint: what happens near the stationary point as we keep on updating
the parameters)? Can you fix this?
Discussions51
51

3.2 Linear Regression Implementation from Scratch
Now that you understand the key ideas behind linear regression, we can begin to work through
a hands-on implementation in code. In this section, we will implement the entire method from
scratch, including the data pipeline, the model, the loss function, and the minibatch stochastic
gradient descent optimizer. While modern deep learning frameworks can automate nearly all of
this work, implementing things from scratch is the only way to make sure that you really know
what you are doing. Moreover, when it comes time to customize models, defining our own layers
or loss functions, understanding how things work under the hood will prove handy. In this section,
we will rely only on tensors and auto differentiation. Afterwards, we will introduce a more concise
implementation, taking advantage of bells and whistles of deep learning frameworks.
%matplotlib inline
import random
npx.set_np()
3.2.1 Generating the Dataset
To keep things simple, we will construct an artificial dataset according to a linear model with
additive noise. Our task will be to recover this modelʼs parameters using the finite set of examples
contained in our dataset. We will keep the data low-dimensional so we can visualize it easily. In
the following code snippet, we generate a dataset containing 1000 examples, each consisting of 2
features sampled from a standard normal distribution. Thus our synthetic dataset will be a matrix
X ∈ R1000×2 .
The true parameters generating our dataset will be w = [2, −3.4]⊤ and b = 4.2, and our synthetic
labels will be assigned according to the following linear model with the noise term ϵ:
y = Xw + b + ϵ. (3.2.1)
You could think of ϵ as capturing potential measurement errors on the features and labels. We
will assume that the standard assumptions hold and thus that ϵ obeys a normal distribution with
mean of 0. To make our problem easy, we will set its standard deviation to 0.01. The following
code generates our synthetic dataset.
def synthetic_data(w, b, num_examples): #@save

"""Generate y = Xw + b + noise."""
X = np.random.normal(0, 1, (num_examples, len(w)))
y = np.dot(X, w) + b
y += np.random.normal(0, 0.01, y.shape)
return X, y.reshape((-1, 1))
true_w = np.array([2, -3.4])

true_b = 4.2
features, labels = synthetic_data(true_w, true_b, 1000)
Note that each row in features consists of a 2-dimensional data example and that each row in
labels consists of a 1-dimensional label value (a scalar).
3.2. Linear Regression Implementation from Scratch 99

print('features:', features[0],'\nlabel:', labels[0])
features: [2.2122064 1.1630787]

label: [4.662078]
By generating a scatter plot using the second feature features[:, 1] and labels, we can clearly
observe the linear correlation between the two.
d2l.set_figsize()
# The semicolon is for displaying the plot only
d2l.plt.scatter(d2l.numpy(features[:, 1]), d2l.numpy(labels), 1);
Recall that training models consists of making multiple passes over the dataset, grabbing one
minibatch of examples at a time, and using them to update our model. Since this process is so
fundamental to training machine learning algorithms, it is worth defining a utility function to
shuffle the dataset and access it in minibatches.
In the following code, we define the data_iter function to demonstrate one possible implemen-
tation of this functionality. The function takes a batch size, a matrix of features, and a vector of
labels, yielding minibatches of the size batch_size. Each minibatch consists of a tuple of features
and labels.
def data_iter(batch_size, features, labels):

num_examples = len(features)
indices = list(range(num_examples))
# The examples are read at random, in no particular order
random.shuffle(indices)
for i in range(0, num_examples, batch_size):
batch_indices = np.array(
indices[i: min(i + batch_size, num_examples)])
yield features[batch_indices], labels[batch_indices]
In general, note that we want to use reasonably sized minibatches to take advantage of the GPU
hardware, which excels at parallelizing operations. Because each example can be fed through our
models in parallel and the gradient of the loss function for each example can also be taken in

parallel, GPUs allow us to process hundreds of examples in scarcely more time than it might take
to process just a single example.
To build some intuition, let us read and print the first small batch of data examples. The shape of
the features in each minibatch tells us both the minibatch size and the number of input features.
Likewise, our minibatch of labels will have a shape given by batch_size.
batch_size = 10
for X, y in data_iter(batch_size, features, labels):

print(X, '\n', y)
break
[[-0.26001498 1.8115263 ]
[ 0.34593368 -0.7567092 ]
[-1.4900824 -1.0667098 ]
[-0.49231794 -1.9087739 ]
[-0.9669159 -0.8711447 ]
[-0.00572143 -0.31106037]
[-0.33182386 1.6597842 ]
[-0.18256646 0.8374752 ]
[-0.47508195 -0.24207895]
[ 0.77013063 -1.1259842 ]]
[[-2.4844196]
[ 7.478742 ]
[ 4.851502 ]
[ 9.706129 ]
[ 5.2282877]
[ 5.2416353]
[-2.114907 ]
[ 0.9842439]
[ 4.045882 ]
[ 9.558747 ]]
As we run the iteration, we obtain distinct minibatches successively until the entire dataset has
been exhausted (try this). While the iteration implemented above is good for didactic purposes,
it is inefficient in ways that might get us in trouble on real problems. For example, it requires that
we load all the data in memory and that we perform lots of random memory access. The built-in
iterators implemented in a deep learning framework are considerably more efficient and they can
deal with both data stored in files and data fed via data streams.
3.2.3 Initializing Model Parameters
Before we can begin optimizing our modelʼs parameters by minibatch stochastic gradient descent,
we need to have some parameters in the first place. In the following code, we initialize weights
by sampling random numbers from a normal distribution with mean 0 and a standard deviation
of 0.01, and setting the bias to 0.
w = np.random.normal(0, 0.01, (2, 1))

b = np.zeros(1)
w.attach_grad()
b.attach_grad()

After initializing our parameters, our next task is to update them until they fit our data sufficiently
well. Each update requires taking the gradient of our loss function with respect to the parameters.
Given this gradient, we can update each parameter in the direction that may reduce the loss.
Since nobody wants to compute gradients explicitly (this is tedious and error prone), we use au-
tomatic differentiation, as introduced in Section 2.5, to compute the gradient.
3.2.4 Defining the Model
Next, we must define our model, relating its inputs and parameters to its outputs. Recall that to
calculate the output of the linear model, we simply take the matrix-vector dot product of the input
features X and the model weights w, and add the offset b to each example. Note that below Xw is
a vector and b is a scalar. Recall the broadcasting mechanism as described in Section 2.1.3. When
we add a vector and a scalar, the scalar is added to each component of the vector.
def linreg(X, w, b): #@save

"""The linear regression model."""
return np.dot(X, w) + b
3.2.5 Defining the Loss Function
Since updating our model requires taking the gradient of our loss function, we ought to define the
loss function first. Here we will use the squared loss function as described in Section 3.1. In the
implementation, we need to transform the true value y into the predicted valueʼs shape y_hat. The
result returned by the following function will also have the same shape as y_hat.
def squared_loss(y_hat, y): #@save

"""Squared loss."""
return (y_hat - y.reshape(y_hat.shape)) ** 2 / 2
3.2.6 Defining the Optimization Algorithm
As we discussed in Section 3.1, linear regression has a closed-form solution. However, this is not
a book about linear regression: it is a book about deep learning. Since none of the other models
that this book introduces can be solved analytically, we will take this opportunity to introduce your
first working example of minibatch stochastic gradient descent.
At each step, using one minibatch randomly drawn from our dataset, we will estimate the gradient
of the loss with respect to our parameters. Next, we will update our parameters in the direction
that may reduce the loss. The following code applies the minibatch stochastic gradient descent
update, given a set of parameters, a learning rate, and a batch size. The size of the update step is
determined by the learning rate lr. Because our loss is calculated as a sum over the minibatch of
examples, we normalize our step size by the batch size (batch_size), so that the magnitude of a
typical step size does not depend heavily on our choice of the batch size.
def sgd(params, lr, batch_size): #@save

"""Minibatch stochastic gradient descent."""
for param in params:
param[:] = param - lr * param.grad / batch_size

3.2.7 Training
Now that we have all of the parts in place, we are ready to implement the main training loop. It
is crucial that you understand this code because you will see nearly identical training loops over
and over again throughout your career in deep learning.
In each iteration, we will grab a minibatch of training examples, and pass them through our model
to obtain a set of predictions. After calculating the loss, we initiate the backwards pass through
the network, storing the gradients with respect to each parameter. Finally, we will call the opti-
mization algorithm sgd to update the model parameters.
In summary, we will execute the following loop:
• Initialize parameters (w, b)
• Repeat until done
∑
– Compute gradient g ← ∂(w,b) |B|
1
i∈B l(x(i) , y (i) , w, b)
– Update parameters (w, b) ← (w, b) − ηg

In each epoch, we will iterate through the entire dataset (using the data_iter function) once pass-
ing through every example in the training dataset (assuming that the number of examples is di-
visible by the batch size). The number of epochs num_epochs and the learning rate lr are both
hyperparameters, which we set here to 3 and 0.03, respectively. Unfortunately, setting hyperpa-
rameters is tricky and requires some adjustment by trial and error. We elide these details for now
but revise them later in Chapter 11.
lr = 0.03
num_epochs = 3
net = linreg
loss = squared_loss
for epoch in range(num_epochs):

for X, y in data_iter(batch_size, features, labels):
l = loss(net(X, w, b), y) # Minibatch loss in `X` and `y`
# Because `l` has a shape (`batch_size`, 1) and is not a scalar
# variable, the elements in `l` are added together to obtain a new
# variable, on which gradients with respect to [`w`, `b`] are computed
l.backward()
sgd([w, b], lr, batch_size) # Update parameters using their gradient
train_l = loss(net(features, w, b), labels)
print(f'epoch {epoch + 1}, loss {float(train_l.mean()):f}')
epoch 1, loss 0.025238

In this case, because we synthesized the dataset ourselves, we know precisely what the true pa-
rameters are. Thus, we can evaluate our success in training by comparing the true parameters
with those that we learned through our training loop. Indeed they turn out to be very close to
each other.

print(f'error in estimating w: {true_w - w.reshape(true_w.shape)}')
print(f'error in estimating b: {true_b - b}')
error in estimating w: [ 0.00068665 -0.00013518]

error in estimating b: [0.00068951]
Note that we should not take it for granted that we are able to recover the parameters perfectly.
However, in machine learning, we are typically less concerned with recovering true underlying
parameters, and more concerned with parameters that lead to highly accurate prediction. For-
tunately, even on difficult optimization problems, stochastic gradient descent can often find re-
markably good solutions, owing partly to the fact that, for deep networks, there exist many con-
figurations of the parameters that lead to highly accurate prediction.
Summary
• We saw how a deep network can be implemented and optimized from scratch, using just
tensors and auto differentiation, without any need for defining layers or fancy optimizers.
• This section only scratches the surface of what is possible. In the following sections, we will
describe additional models based on the concepts that we have just introduced and learn
how to implement them more concisely.
Exercises
1. What would happen if we were to initialize the weights to zero. Would the algorithm still
work?
2. Assume that you are Georg Simon Ohm52 trying to come up with a model between voltage
and current. Can you use auto differentiation to learn the parameters of your model?
3. Can you use Planckʼs Law53 to determine the temperature of an object using spectral energy
density?
4. What are the problems you might encounter if you wanted to compute the second deriva-
tives? How would you fix them?
5. Why is the reshape function needed in the squared_loss function?
6. Experiment using different learning rates to find out how fast the loss function value drops.
7. If the number of examples cannot be divided by the batch size, what happens to the
data_iter functionʼs behavior?
Discussions54
52
https://en.wikipedia.org/wiki/Georg_Ohm
53
https://en.wikipedia.org/wiki/Planck%27s_law
54

3.3 Concise Implementation of Linear Regression
Broad and intense interest in deep learning for the past several years has inspired companies,
academics, and hobbyists to develop a variety of mature open source frameworks for automating
the repetitive work of implementing gradient-based learning algorithms. In Section 3.2, we relied
only on (i) tensors for data storage and linear algebra; and (ii) auto differentiation for calculat-
ing gradients. In practice, because data iterators, loss functions, optimizers, and neural network
layers are so common, modern libraries implement these components for us as well.
In this section, we will show you how to implement the linear regression model from Section 3.2
concisely by using high-level APIs of deep learning frameworks.
3.3.1 Generating the Dataset
To start, we will generate the same dataset as in Section 3.2.

from mxnet import autograd, gluon, np, npx
npx.set_np()
true_w = np.array([2, -3.4])

true_b = 4.2
features, labels = d2l.synthetic_data(true_w, true_b, 1000)
Rather than rolling our own iterator, we can call upon the existing API in a framework to read
data. We pass in features and labels as arguments and specify batch_size when instantiating
a data iterator object. Besides, the boolean value is_train indicates whether or not we want the
data iterator object to shuffle the data on each epoch (pass through the dataset).
def load_array(data_arrays, batch_size, is_train=True): #@save

"""Construct a Gluon data iterator."""
dataset = gluon.data.ArrayDataset(*data_arrays)
return gluon.data.DataLoader(dataset, batch_size, shuffle=is_train)
batch_size = 10
data_iter = load_array((features, labels), batch_size)
Now we can use data_iter in much the same way as we called the data_iter function in Section
3.2. To verify that it is working, we can read and print the first minibatch of examples. Comparing
with Section 3.2, here we use iter to construct a Python iterator and use next to obtain the first
item from the iterator.
next(iter(data_iter))
3.3. Concise Implementation of Linear Regression 105

[array([[ 0.5443711 , 0.7681698 ],
[-0.47063652, -0.59975994],
[-0.47508195, -0.24207895],
[ 0.76801354, -0.69288605],
[-2.7625482 , 1.7799325 ],
[-0.55819434, 0.33346984],
[-0.4121327 , -1.5466391 ],
[ 0.02849233, -0.71190894],
[-0.9567186 , -1.6827176 ],
[-0.51572496, 0.07394829]]),
array([[ 2.6835713],
[ 5.289975 ],
[ 4.045882 ],
[ 8.095631 ],
[-7.4035244],
[ 1.9458812],
[ 8.63532 ],
[ 6.669124 ],
[ 8.01163 ],
[ 2.9113574]])]
When we implemented linear regression from scratch in Section 3.2, we defined our model pa-
rameters explicitly and coded up the calculations to produce output using basic linear algebra
operations. You should know how to do this. But once your models get more complex, and once
you have to do this nearly every day, you will be glad for the assistance. The situation is similar
to coding up your own blog from scratch. Doing it once or twice is rewarding and instructive, but
you would be a lousy web developer if every time you needed a blog you spent a month reinventing
the wheel.
For standard operations, we can use a frameworkʼs predefined layers, which allow us to focus espe-
cially on the layers used to construct the model rather than having to focus on the implementation.
We will first define a model variable net, which will refer to an instance of the Sequential class.
The Sequential class defines a container for several layers that will be chained together. Given
input data, a Sequential instance passes it through the first layer, in turn passing the output as
the second layerʼs input and so forth. In the following example, our model consists of only one
layer, so we do not really need Sequential. But since nearly all of our future models will involve
multiple layers, we will use it anyway just to familiarize you with the most standard workflow.
Recall the architecture of a single-layer network as shown in Fig. 3.1.2. The layer is said to be fully-
connected because each of its inputs is connected to each of its outputs by means of a matrix-vector
multiplication.
In Gluon, the fully-connected layer is defined in the Dense class. Since we only want to generate a
single scalar output, we set that number to 1.
It is worth noting that, for convenience, Gluon does not require us to specify the input shape for
each layer. So here, we do not need to tell Gluon how many inputs go into this linear layer. When
we first try to pass data through our model, e.g., when we execute net(X) later, Gluon will auto-
matically infer the number of inputs to each layer. We will describe how this works in more detail
later.

# `nn` is an abbreviation for neural networks
from mxnet.gluon import nn
net = nn.Sequential()
net.add(nn.Dense(1))
Before using net, we need to initialize the model parameters, such as the weights and bias in the
linear regression model. Deep learning frameworks often have a predefined way to initialize the
parameters. Here we specify that each weight parameter should be randomly sampled from a nor-
mal distribution with mean 0 and standard deviation 0.01. The bias parameter will be initialized
to zero.
We will import the initializer module from MXNet. This module provides various methods
for model parameter initialization. Gluon makes init available as a shortcut (abbreviation) to
access the initializer package. We only specify how to initialize the weight by calling init.
Normal(sigma=0.01). Bias parameters are initialized to zero by default.
from mxnet import init

net.initialize(init.Normal(sigma=0.01))
The code above may look straightforward but you should note that something strange is happening
here. We are initializing parameters for a network even though Gluon does not yet know how
many dimensions the input will have! It might be 2 as in our example or it might be 2000. Gluon
lets us get away with this because behind the scene, the initialization is actually deferred. The
real initialization will take place only when we for the first time attempt to pass data through the
network. Just be careful to remember that since the parameters have not been initialized yet, we
cannot access or manipulate them.
In Gluon, the loss module defines various loss functions. In this example, we will use the Gluon
implementation of squared loss (L2Loss).
loss = gluon.loss.L2Loss()
3.3.6 Defining the Optimization Algorithm
Minibatch stochastic gradient descent is a standard tool for optimizing neural networks and thus
Gluon supports it alongside a number of variations on this algorithm through its Trainer class.
When we instantiate Trainer, we will specify the parameters to optimize over (obtainable from
our model net via net.collect_params()), the optimization algorithm we wish to use (sgd), and
a dictionary of hyperparameters required by our optimization algorithm. Minibatch stochastic
gradient descent just requires that we set the value learning_rate, which is set to 0.03 here.
from mxnet import gluon

trainer = gluon.Trainer(net.collect_params(), 'sgd', {'learning_rate': 0.03})
3.3. Concise Implementation of Linear Regression 107

3.3.7 Training
You might have noticed that expressing our model through high-level APIs of a deep learning
framework requires comparatively few lines of code. We did not have to individually allocate
parameters, define our loss function, or implement minibatch stochastic gradient descent. Once
we start working with much more complex models, advantages of high-level APIs will grow con-
siderably. However, once we have all the basic pieces in place, the training loop itself is strikingly
similar to what we did when implementing everything from scratch.
To refresh your memory: for some number of epochs, we will make a complete pass over the
dataset (train_data), iteratively grabbing one minibatch of inputs and the corresponding ground-
truth labels. For each minibatch, we go through the following ritual:
• Generate predictions by calling net(X) and calculate the loss l (the forward propagation).
• Calculate gradients by running the backpropagation.
• Update the model parameters by invoking our optimizer.
For good measure, we compute the loss after each epoch and print it to monitor progress.
num_epochs = 3
for X, y in data_iter:
l = loss(net(X), y)
l.backward()
trainer.step(batch_size)
l = loss(net(features), labels)
print(f'epoch {epoch + 1}, loss {l.mean().asnumpy():f}')

Below, we compare the model parameters learned by training on finite data and the actual param-
eters that generated our dataset. To access parameters, we first access the layer that we need from
net and then access that layerʼs weights and bias. As in our from-scratch implementation, note
that our estimated parameters are close to their ground-truth counterparts.
w = net[0].weight.data()
print(f'error in estimating w: {true_w - w.reshape(true_w.shape)}')
b = net[0].bias.data()
print(f'error in estimating b: {true_b - b}')
error in estimating w: [4.3487549e-04 6.7949295e-05]

error in estimating b: [0.00079966]

Summary
• Using Gluon, we can implement models much more concisely.

• In Gluon, the data module provides tools for data processing, the nn module defines a large
number of neural network layers, and the loss module defines many common loss func-
tions.
• MXNetʼs module initializer provides various methods for model parameter initialization.
• Dimensionality and storage are automatically inferred, but be careful not to attempt to ac-
cess parameters before they have been initialized.
Exercises
1. If we replace l = loss(output, y) with l = loss(output, y).mean(), we need to change

trainer.step(batch_size) to trainer.step(1) for the code to behave identically. Why?
2. Review the MXNet documentation to see what loss functions and initialization methods are
provided in the modules gluon.loss and init. Replace the loss by Huberʼs loss.
3. How do you access the gradient of dense.weight?
Discussions55
3.4 Softmax Regression
In Section 3.1, we introduced linear regression, working through implementations from scratch
in Section 3.2 and again using high-level APIs of a deep learning framework in Section 3.3 to do
the heavy lifting.
Regression is the hammer we reach for when we want to answer how much? or how many? ques-
tions. If you want to predict the number of dollars (price) at which a house will be sold, or the
number of wins a baseball team might have, or the number of days that a patient will remain
hospitalized before being discharged, then you are probably looking for a regression model.
In practice, we are more often interested in classification: asking not “how much” but “which one”:
• Does this email belong in the spam folder or the inbox?
• Is this customer more likely to sign up or not to sign up for a subscription service?
• Does this image depict a donkey, a dog, a cat, or a rooster?
• Which movie is Aston most likely to watch next?
Colloquially, machine learning practitioners overload the word classification to describe two subtly
different problems: (i) those where we are interested only in hard assignments of examples to
categories (classes); and (ii) those where we wish to make soft assignments, i.e., to assess the
probability that each category applies. The distinction tends to get blurred, in part, because often,
even when we only care about hard assignments, we still use models that make soft assignments.
55
3.4. Softmax Regression 109

3.4.1 Classification Problem
To get our feet wet, let us start off with a simple image classification problem. Here, each input
consists of a 2 × 2 grayscale image. We can represent each pixel value with a single scalar, giving
us four features x1 , x2 , x3 , x4 . Further, let us assume that each image belongs to one among the
categories “cat”, “chicken”, and “dog”.
Next, we have to choose how to represent the labels. We have two obvious choices. Per-
haps the most natural impulse would be to choose y ∈ {1, 2, 3}, where the integers repre-
sent {dog, cat, chicken} respectively. This is a great way of storing such information on a com-
puter. If the categories had some natural ordering among them, say if we were trying to predict
{baby, toddler, adolescent, young adult, adult, geriatric}, then it might even make sense to cast
this problem as regression and keep the labels in this format.
But general classification problems do not come with natural orderings among the classes. For-
tunately, statisticians long ago invented a simple way to represent categorical data: the one-hot
encoding. A one-hot encoding is a vector with as many components as we have categories. The
component corresponding to particular instanceʼs category is set to 1 and all other components
are set to 0. In our case, a label y would be a three-dimensional vector, with (1, 0, 0) corresponding
to “cat”, (0, 1, 0) to “chicken”, and (0, 0, 1) to “dog”:
y ∈ {(1, 0, 0), (0, 1, 0), (0, 0, 1)}. (3.4.1)
3.4.2 Network Architecture
In order to estimate the conditional probabilities associated with all the possible classes, we need
a model with multiple outputs, one per class. To address classification with linear models, we will
need as many affine functions as we have outputs. Each output will correspond to its own affine
function. In our case, since we have 4 features and 3 possible output categories, we will need 12
scalars to represent the weights (w with subscripts), and 3 scalars to represent the biases (b with
subscripts). We compute these three logits, o1 , o2 , and o3 , for each input:
o1 = x1 w11 + x2 w12 + x3 w13 + x4 w14 + b1 ,
o2 = x1 w21 + x2 w22 + x3 w23 + x4 w24 + b2 , (3.4.2)
o3 = x1 w31 + x2 w32 + x3 w33 + x4 w34 + b3 .
We can depict this calculation with the neural network diagram shown in Fig. 3.4.1. Just as in lin-
ear regression, softmax regression is also a single-layer neural network. And since the calculation
of each output, o1 , o2 , and o3 , depends on all inputs, x1 , x2 , x3 , and x4 , the output layer of softmax
regression can also be described as fully-connected layer.
Fig. 3.4.1: Softmax regression is a single-layer neural network.
To express the model more compactly, we can use linear algebra notation. In vector form, we
arrive at o = Wx + b, a form better suited both for mathematics, and for writing code. Note that

we have gathered all of our weights into a 3×4 matrix and that for features of a given data example
x, our outputs are given by a matrix-vector product of our weights by our input features plus our
biases b.
3.4.3 Softmax Operation
The main approach that we are going to take here is to interpret the outputs of our model as proba-
bilities. We will optimize our parameters to produce probabilities that maximize the likelihood of
the observed data. Then, to generate predictions, we will set a threshold, for example, choosing
the label with the maximum predicted probabilities.
Put formally, we would like any output ŷj to be interpreted as the probability that a given item
belongs to class j. Then we can choose the class with the largest output value as our prediction
argmaxj yj . For example, if ŷ1 , ŷ2 , and ŷ3 are 0.1, 0.8, and 0.1, respectively, then we predict cate-
gory 2, which (in our example) represents “chicken”.
You might be tempted to suggest that we interpret the logits o directly as our outputs of interest.
However, there are some problems with directly interpreting the output of the linear layer as a
probability. On one hand, nothing constrains these numbers to sum to 1. On the other hand,
depending on the inputs, they can take negative values. These violate basic axioms of probability
presented in Section 2.6
To interpret our outputs as probabilities, we must guarantee that (even on new data), they will be
nonnegative and sum up to 1. Moreover, we need a training objective that encourages the model
to estimate faithfully probabilities. Of all instances when a classifier outputs 0.5, we hope that half
of those examples will actually belong to the predicted class. This is a property called calibration.
The softmax function, invented in 1959 by the social scientist R. Duncan Luce in the context of
choice models, does precisely this. To transform our logits such that they become nonnegative and
sum to 1, while requiring that the model remains differentiable, we first exponentiate each logit
(ensuring non-negativity) and then divide by their sum (ensuring that they sum to 1):
exp(oj )
ŷ = softmax(o) where ŷj = ∑ . (3.4.3)
k exp(ok )
It is easy to see ŷ1 + ŷ2 + ŷ3 = 1 with 0 ≤ ŷj ≤ 1 for all j. Thus, ŷ is a proper probability distribution
whose element values can be interpreted accordingly. Note that the softmax operation does not
change the ordering among the logits o, which are simply the pre-softmax values that determine
the probabilities assigned to each class. Therefore, during prediction we can still pick out the most
likely class by
argmax ŷj = argmax oj . (3.4.4)

j j
Although softmax is a nonlinear function, the outputs of softmax regression are still determined
by an affine transformation of input features; thus, softmax regression is a linear model.

3.4.4 Vectorization for Minibatches
To improve computational efficiency and take advantage of GPUs, we typically carry out vector
calculations for minibatches of data. Assume that we are given a minibatch X of examples with
feature dimensionality (number of inputs) d and batch size n. Moreover, assume that we have q
categories in the output. Then the minibatch features X are in Rn×d , weights W ∈ Rd×q , and the
bias satisfies b ∈ R1×q .
O = XW + b,
(3.4.5)
Ŷ = softmax(O).
This accelerates the dominant operation into a matrix-matrix product XW vs. the matrix-vector
products we would be executing if we processed one example at a time. Since each row in X rep-
resents a data example, the softmax operation itself can be computed rowwise: for each row of
O, exponentiate all entries and then normalize them by the sum. Triggering broadcasting during
the summation XW + b in (3.4.5), both the minibatch logits O and output probabilities Ŷ are n × q
matrices.
3.4.5 Loss Function
Next, we need a loss function to measure the quality of our predicted probabilities. We will rely
on maximum likelihood estimation, the very same concept that we encountered when providing a
probabilistic justification for the mean squared error objective in linear regression (Section 3.1.3).
Log-Likelihood
The softmax function gives us a vector ŷ, which we can interpret as estimated conditional prob-
abilities of each class given any input x, e.g., ŷ1 = P (y = cat | x). Suppose that the entire dataset
{X, Y} has n examples, where the example indexed by i consists of a feature vector x(i) and a one-
hot label vector y(i) . We can compare the estimates with reality by checking how probable the
actual classes are according to our model, given the features:
∏
n
P (Y | X) = P (y(i) | x(i) ). (3.4.6)
i=1
According to maximum likelihood estimation, we maximize P (Y | X), which is equivalent to min-

imizing the negative log-likelihood:
∑
n ∑
n
− log P (Y | X) = − log P (y (i)
|x )=
(i)
l(y(i) , ŷ(i) ), (3.4.7)
i=1 i=1
where for any pair of label y and model prediction ŷ over q classes, the loss function l is
∑
q
l(y, ŷ) = − yj log ŷj . (3.4.8)
j=1
For reasons explained later on, the loss function in (3.4.8) is commonly called the cross-entropy loss.
Since y is a one-hot vector of length q, the sum over all its coordinates j vanishes for all but one
term. Since all ŷj are predicted probabilities, their logarithm is never larger than 0. Consequently,
the loss function cannot be minimized any further if we correctly predict the actual label with

certainty, i.e., if the predicted probability P (y | x) = 1 for the actual label y. Note that this is
often impossible. For example, there might be label noise in the dataset (some examples may be
mislabeled). It may also not be possible when the input features are not sufficiently informative
to classify every example perfectly.
Softmax and Derivatives
Since the softmax and the corresponding loss are so common, it is worth understanding a bit
better how it is computed. Plugging (3.4.3) into the definition of the loss in (3.4.8) and using the
definition of the softmax we obtain:
∑
q
exp(oj )
l(y, ŷ) = − yj log ∑q
j=1 k=1 exp(ok )
∑
q ∑
q ∑
q
= yj log exp(ok ) − yj oj (3.4.9)
j=1 k=1 j=1
∑
q ∑
q
= log exp(ok ) − yj o j .
k=1 j=1
To understand a bit better what is going on, consider the derivative with respect to any logit oj .
We get
exp(oj )
∂oj l(y, ŷ) = ∑q − yj = softmax(o)j − yj . (3.4.10)
k=1 exp(ok )
In other words, the derivative is the difference between the probability assigned by our model,
as expressed by the softmax operation, and what actually happened, as expressed by elements in
the one-hot label vector. In this sense, it is very similar to what we saw in regression, where the
gradient was the difference between the observation y and estimate ŷ. This is not coincidence.
In any exponential family (see the online appendix on distributions56 ) model, the gradients of
the log-likelihood are given by precisely this term. This fact makes computing gradients easy in
practice.
Cross-Entropy Loss
Now consider the case where we observe not just a single outcome but an entire distribution over
outcomes. We can use the same representation as before for the label y. The only difference
is that rather than a vector containing only binary entries, say (0, 0, 1), we now have a generic
probability vector, say (0.1, 0.2, 0.7). The math that we used previously to define the loss l in
:eqref:eq_l_cross_entropy still works out fine, just that the interpretation is slightly more general.
It is the expected value of the loss for a distribution over labels. This loss is called the cross-entropy
loss and it is one of the most commonly used losses for classification problems. We can demystify
the name by introducing just the basics of information theory. If you wish to understand more
details of information theory, you may further refer to the online appendix on information the-
ory57 .
56
https://d2l.ai/chapter_appendix-mathematics-for-deep-learning/distributions.html
57
https://d2l.ai/chapter_appendix-mathematics-for-deep-learning/information-theory.html

3.4.6 Information Theory Basics
Information theory deals with the problem of encoding, decoding, transmitting, and manipulating
information (also known as data) in as concise form as possible.
Entropy
The central idea in information theory is to quantify the information content in data. This quantity
places a hard limit on our ability to compress the data. In information theory, this quantity is
called the entropy of a distribution p, and it is captured by the following equation:
∑
H[p] = −p(j) log p(j). (3.4.11)
j
One of the fundamental theorems of information theory states that in order to encode data drawn
randomly from the distribution p, we need at least H[p] “nats” to encode it. If you wonder what a
“nat” is, it is the equivalent of bit but when using a code with base e rather than one with base 2.
1
Thus, one nat is log(2) ≈ 1.44 bit.
Surprisal
You might be wondering what compression has to do with prediction. Imagine that we have a
stream of data that we want to compress. If it is always easy for us to predict the next token, then
this data is easy to compress! Take the extreme example where every token in the stream always
takes the same value. That is a very boring data stream! And not only it is boring, but it is also
easy to predict. Because they are always the same, we do not have to transmit any information to
communicate the contents of the stream. Easy to predict, easy to compress.
However if we cannot perfectly predict every event, then we might sometimes be surprised. Our
surprise is greater when we assigned an event lower probability. Claude Shannon settled on
log P 1(j) = − log P (j) to quantify oneʼs surprisal at observing an event j having assigned it a (sub-
jective) probability P (j). The entropy defined in (3.4.11) is then the expected surprisal when one
assigned the correct probabilities that truly match the data-generating process.
Cross-Entropy Revisited
So if entropy is level of surprise experienced by someone who knows the true probability, then
you might be wondering, what is cross-entropy? The cross-entropy from p to q, denoted H(p, q),
is the expected surprisal of an observer with subjective probabilities q upon seeing data that were
actually generated according to probabilities p. The lowest possible cross-entropy is achieved
when p = q. In this case, the cross-entropy from p to q is H(p, p) = H(p).
In short, we can think of the cross-entropy classification objective in two ways: (i) as maximizing
the likelihood of the observed data; and (ii) as minimizing our surprisal (and thus the number of
bits) required to communicate the labels.

3.4.7 Model Prediction and Evaluation
After training the softmax regression model, given any example features, we can predict the prob-
ability of each output class. Normally, we use the class with the highest predicted probability as
the output class. The prediction is correct if it is consistent with the actual class (label). In the next
part of the experiment, we will use accuracy to evaluate the modelʼs performance. This is equal to
the ratio between the number of correct predictions and the total number of predictions.
Summary
• The softmax operation takes a vector and maps it into probabilities.

• Softmax regression applies to classification problems. It uses the probability distribution of
the output class in the softmax operation.
• Cross-entropy is a good measure of the difference between two probability distributions. It
measures the number of bits needed to encode the data given our model.
Exercises
1. We can explore the connection between exponential families and the softmax in some more
depth.
1. Compute the second derivative of the cross-entropy loss l(y, ŷ) for the softmax.
2. Compute the variance of the distribution given by softmax(o) and show that it matches
the second derivative computed above.
2. Assume that we have three classes which occur with equal probability, i.e., the probability
vector is ( 13 , 31 , 13 ).
1. What is the problem if we try to design a binary code for it?
2. Can you design a better code? Hint: what happens if we try to encode two independent
observations? What if we encode n observations jointly?
3. Softmax is a misnomer for the mapping introduced above (but everyone in deep learning
uses it). The real softmax is defined as RealSoftMax(a, b) = log(exp(a) + exp(b)).
1. Prove that RealSoftMax(a, b) > max(a, b).
2. Prove that this holds for λ−1 RealSoftMax(λa, λb), provided that λ > 0.
3. Show that for λ → ∞ we have λ−1 RealSoftMax(λa, λb) → max(a, b).
4. What does the soft-min look like?
5. Extend this to more than two numbers.
Discussions58
58

3.5 The Image Classification Dataset
One of the widely used dataset for image classification is the MNIST dataset (LeCun et al., 1998).
While it had a good run as a benchmark dataset, even simple models by todayʼs standards achieve
classification accuracy over 95%, making it unsuitable for distinguishing between stronger models
and weaker ones. Today, MNIST serves as more of sanity checks than as a benchmark. To up the
ante just a bit, we will focus our discussion in the coming sections on the qualitatively similar, but
comparatively complex Fashion-MNIST dataset (Xiao et al., 2017), which was released in 2017.
%matplotlib inline
from mxnet import gluon
import sys
d2l.use_svg_display()
We can download and read the Fashion-MNIST dataset into memory via the build-in functions in
the framework.
mnist_train = gluon.data.vision.FashionMNIST(train=True)
mnist_test = gluon.data.vision.FashionMNIST(train=False)
Fashion-MNIST consists of images from 10 categories, each represented by 6000 images in the
training dataset and by 1000 in the test dataset. A test dataset (or test set) is used for evaluating
model performance and not for training. Consequently the training set and the test set contain
60000 and 10000 images, respectively.
len(mnist_train), len(mnist_test)
(60000, 10000)
The height and width of each input image are both 28 pixels. Note that the dataset consists of
grayscale images, whose number of channels is 1. For brevity, throughout this book we store the
shape of any image with height h width w pixels as h × w or (h, w).
mnist_train[0][0].shape
(28, 28, 1)
The images in Fashion-MNIST are associated with the following categories: t-shirt, trousers,
pullover, dress, coat, sandal, shirt, sneaker, bag, and ankle boot. The following function converts
between numeric label indices and their names in text.
def get_fashion_mnist_labels(labels): #@save

"""Return text labels for the Fashion-MNIST dataset."""
text_labels = ['t-shirt', 'trouser', 'pullover', 'dress', 'coat',

'sandal', 'shirt', 'sneaker', 'bag', 'ankle boot']
return [text_labels[int(i)] for i in labels]
We can now create a function to visualize these examples.
def show_images(imgs, num_rows, num_cols, titles=None, scale=1.5): #@save

"""Plot a list of images."""
figsize = (num_cols * scale, num_rows * scale)
_, axes = d2l.plt.subplots(num_rows, num_cols, figsize=figsize)
axes = axes.flatten()
for i, (ax, img) in enumerate(zip(axes, imgs)):
ax.imshow(d2l.numpy(img))
ax.axes.get_xaxis().set_visible(False)
ax.axes.get_yaxis().set_visible(False)
if titles:
ax.set_title(titles[i])
return axes
Here are the images and their corresponding labels (in text) for the first few examples in the train-
ing dataset.
X, y = mnist_train[:18]
show_images(X.squeeze(axis=-1), 2, 9, titles=get_fashion_mnist_labels(y));
3.5.2 Reading a Minibatch
To make our life easier when reading from the training and test sets, we use the built-in data
iterator rather than creating one from scratch. Recall that at each iteration, a data loader reads a
minibatch of data with size batch_size each time. We also randomly shuffle the examples for the
training data iterator.
batch_size = 256
def get_dataloader_workers(): #@save

"""Use 4 processes to read the data except for Windows."""
return 0 if sys.platform.startswith('win') else 4
# `ToTensor` converts the image data from uint8 to 32-bit floating point. It
# divides all numbers by 255 so that all pixel values are between 0 and 1
transformer = gluon.data.vision.transforms.ToTensor()
3.5. The Image Classification Dataset 117

train_iter = gluon.data.DataLoader(mnist_train.transform_first(transformer),
batch_size, shuffle=True,
num_workers=get_dataloader_workers())
Let us look at the time it takes to read the training data.
timer = d2l.Timer()
for X, y in train_iter:
continue
'1.53 sec'
3.5.3 Putting All Things Together
Now we define the load_data_fashion_mnist function that obtains and reads the Fashion-MNIST
dataset. It returns the data iterators for both the training set and validation set. In addition, it
accepts an optional argument to resize images to another shape.
def load_data_fashion_mnist(batch_size, resize=None): #@save

"""Download the Fashion-MNIST dataset and then load it into memory."""
dataset = gluon.data.vision
trans = [dataset.transforms.ToTensor()]
if resize:
trans.insert(0, dataset.transforms.Resize(resize))
trans = dataset.transforms.Compose(trans)
mnist_train = dataset.FashionMNIST(train=True).transform_first(trans)
mnist_test = dataset.FashionMNIST(train=False).transform_first(trans)
return (gluon.data.DataLoader(mnist_train, batch_size, shuffle=True,
num_workers=get_dataloader_workers()),
gluon.data.DataLoader(mnist_test, batch_size, shuffle=False,
num_workers=get_dataloader_workers()))
Below we test the image resizing feature of the load_data_fashion_mnist function by specifying
the resize argument.
train_iter, test_iter = load_data_fashion_mnist(32, resize=64)

print(X.shape, X.dtype, y.shape, y.dtype)
break
(32, 1, 64, 64) <class 'numpy.float32'> (32,) <class 'numpy.int32'>
We are now ready to work with the Fashion-MNIST dataset in the sections that follow.

Summary
• Fashion-MNIST is an apparel classification dataset consisting of images representing 10 cat-

egories. We will use this dataset in subsequent sections and chapters to evaluate various
classification algorithms.
• We store the shape of any image with height h width w pixels as h × w or (h, w).
• Data iterators are a key component for efficient performance. Rely on well-implemented
data iterators that exploit high-performance computing to avoid slowing down your training
loop.
Exercises
1. Does reducing the batch_size (for instance, to 1) affect the reading performance?
2. The data iterator performance is important. Do you think the current implementation is fast
enough? Explore various options to improve it.
3. Check out the frameworkʼs online API documentation. Which other datasets are available?
Discussions59
3.6 Implementation of Softmax Regression from Scratch
Just as we implemented linear regression from scratch, we believe that softmax regression is sim-
ilarly fundamental and you ought to know the gory details of how to implement it yourself. We
will work with the Fashion-MNIST dataset, just introduced in Section 3.5, setting up a data iterator
with batch size 256.

from mxnet import autograd, np, npx, gluon
npx.set_np()
batch_size = 256
train_iter, test_iter = d2l.load_data_fashion_mnist(batch_size)
As in our linear regression example, each example here will be represented by a fixed-length vec-
tor. Each example in the raw dataset is a 28 × 28 image. In this section, we will flatten each image,
treating them as vectors of length 784. In the future, we will talk about more sophisticated strate-
gies for exploiting the spatial structure in images, but for now we treat each pixel location as just
another feature.
Recall that in softmax regression, we have as many outputs as there are classes. Because our
dataset has 10 classes, our network will have an output dimension of 10. Consequently, our weights
will constitute a 784 × 10 matrix and the biases will constitute a 1 × 10 row vector. As with linear
59
3.6. Implementation of Softmax Regression from Scratch 119

regression, we will initialize our weights W with Gaussian noise and our biases to take the initial
value 0.
num_inputs = 784
num_outputs = 10
W = np.random.normal(0, 0.01, (num_inputs, num_outputs))

b = np.zeros(num_outputs)
W.attach_grad()
b.attach_grad()
3.6.2 Defining the Softmax Operation
Before implementing the softmax regression model, let us briefly review how the sum operator
work along specific dimensions in a tensor, as discussed in Section 2.3.6 and Section 2.3.6. Given
a matrix X we can sum over all elements (by default) or only over elements in the same axis, i.e.,
the same column (axis 0) or the same row (axis 1). Note that if X is an tensor with shape (2, 3)
and we sum over the columns, the result will be a vector with shape (3,). When invoking the sum
operator, we can specify to keep the number of axes in the original tensor, rather than collapsing
out the dimension that we summed over. This will result in a two-dimensional tensor with shape
(1, 3).
X = np.array([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])

X.sum(0, keepdims=True), X.sum(1, keepdims=True)
(array([[5., 7., 9.]]),

array([[ 6.],
[15.]]))
We are now ready to implement the softmax operation. Recall that softmax consists of three steps:
i) we exponentiate each term (using exp); ii) we sum over each row (we have one row per example
in the batch) to get the normalization constant for each example; iii) we divide each row by its
normalization constant, ensuring that the result sums to 1. Before looking at the code, let us
recall how this looks expressed as an equation:
exp(Xij )
softmax(X)ij = ∑ . (3.6.1)
k exp(Xik )
The denominator, or normalization constant, is also sometimes called the partition function (and
its logarithm is called the log-partition function). The origins of that name are in statistical
physics60 where a related equation models the distribution over an ensemble of particles.
def softmax(X):
X_exp = np.exp(X)
partition = X_exp.sum(1, keepdims=True)
return X_exp / partition # The broadcasting mechanism is applied here
As you can see, for any random input, we turn each element into a non-negative number. More-
over, each row sums up to 1, as is required for a probability.
60
https://en.wikipedia.org/wiki/Partition_function_(statistical_mechanics)

X = np.random.normal(0, 1, (2, 5))
X_prob = softmax(X)
X_prob, X_prob.sum(1)
(array([[0.22376052, 0.06659239, 0.06583703, 0.29964197, 0.3441681 ],

[0.63209665, 0.03179282, 0.194987 , 0.09209415, 0.04902935]]),
array([1. , 0.99999994]))
Note that while this looks correct mathematically, we were a bit sloppy in our implementation
because we failed to take precautions against numerical overflow or underflow due to large or
very small elements of the matrix.
Now that we have defined the softmax operation, we can implement the softmax regression model.
The below code defines how the input is mapped to the output through the network. Note that we
flatten each original image in the batch into a vector using the reshape function before passing
the data through our model.
def net(X):
return softmax(np.dot(X.reshape((-1, W.shape[0])), W) + b)
Next, we need to implement the cross-entropy loss function, as introduced in Section 3.4. This
may be the most common loss function in all of deep learning because, at the moment, classifi-
cation problems far outnumber regression problems.
Recall that cross-entropy takes the negative log-likelihood of the predicted probability assigned
to the true label. Rather than iterating over the predictions with a Python for-loop (which tends
to be inefficient), we can pick all elements by a single operator. Below, we create a toy data y_hat
with 2 examples of predicted probabilities over 3 classes. Then we pick the probability of the first
class in the first example and the probability of the third class in the second example.
y = np.array([0, 2])
y_hat = np.array([[0.1, 0.3, 0.6], [0.3, 0.2, 0.5]])
y_hat[[0, 1], y]
array([0.1, 0.5])
Now we can implement the cross-entropy loss function efficiently with just one line of code.
def cross_entropy(y_hat, y):

return - np.log(y_hat[range(len(y_hat)), y])
cross_entropy(y_hat, y)

array([2.3025851, 0.6931472])
3.6.5 Classification Accuracy
Given the predicted probability distribution y_hat, we typically choose the class with the highest
predicted probability whenever we must output a hard prediction. Indeed, many applications
require that we make a choice. Gmail must categorize an email into “Primary”, “Social”, “Updates”,
or “Forums”. It might estimate probabilities internally, but at the end of the day it has to choose
one among the classes.
When predictions are consistent with the label class y, they are correct. The classification ac-
curacy is the fraction of all predictions that are correct. Although it can be difficult to optimize
accuracy directly (it is not differentiable), it is often the performance measure that we care most
about, and we will nearly always report it when training classifiers.
To compute accuracy we do the following. First, if y_hat is a matrix, we assume that the second
dimension stores prediction scores for each class. We use argmax to obtain the predicted class by
the index for the largest entry in each row. Then we compare the predicted class with the ground-
truth y elementwise. Since the equality operator == is sensitive to data types, we convert y_hatʼs
data type to match that of y. The result is a tensor containing entries of 0 (false) and 1 (true).
Taking the sum yields the number of correct predictions.
def accuracy(y_hat, y): #@save

"""Compute the number of correct predictions."""
if len(y_hat.shape) > 1 and y_hat.shape[1] > 1:
y_hat = y_hat.argmax(axis=1)
cmp = y_hat.astype(y.dtype) == y
return float(d2l.reduce_sum(cmp.astype(y.dtype)))
We will continue to use the variables y_hat and y defined before as the predicted probability dis-
tributions and labels, respectively. We can see that the first exampleʼs prediction class is 2 (the
largest element of the row is 0.6 with the index 2), which is inconsistent with the actual label, 0.
The second exampleʼs prediction class is 2 (the largest element of the row is 0.5 with the index of
2), which is consistent with the actual label, 2. Therefore, the classification accuracy rate for these
two examples is 0.5.
accuracy(y_hat, y) / len(y)
0.5
Similarly, we can evaluate the accuracy for any model net on a dataset that is accessed via the data
iterator data_iter.
def evaluate_accuracy(net, data_iter): #@save

"""Compute the accuracy for a model on a dataset."""
metric = Accumulator(2) # No. of correct predictions, no. of predictions
for _, (X, y) in enumerate(data_iter):
metric.add(accuracy(net(X), y), y.size)
return metric[0] / metric[1]

Here Accumulator is a utility class to accumulate sums over multiple variables. In the above eval-
uate_accuracy function, we create 2 variables in the Accumulator instance for storing both the
number of correct predictions and the number of predictions, respectively. Both will be accumu-
lated over time as we iterate over the dataset.
class Accumulator: #@save

"""For accumulating sums over `n` variables."""
def __init__(self, n):
self.data = [0.0] * n
def add(self, *args):

self.data = [a + float(b) for a, b in zip(self.data, args)]
def reset(self):
self.data = [0.0] * len(self.data)
def __getitem__(self, idx):

return self.data[idx]
Because we initialized the net model with random weights, the accuracy of this model should be
close to random guessing, i.e., 0.1 for 10 classes.
evaluate_accuracy(net, test_iter)
0.0811
3.6.6 Training
The training loop for softmax regression should look strikingly familiar if you read through our
implementation of linear regression in Section 3.2. Here we refactor the implementation to make
it reusable. First, we define a function to train for one epoch. Note that updater is a general
function to update the model parameters, which accepts the batch size as an argument. It can be
either a wrapper of the d2l.sgd function or a frameworkʼs built-in optimization function.
def train_epoch_ch3(net, train_iter, loss, updater): #@save

"""Train a model within one epoch (defined in Chapter 3)."""
# Sum of training loss, sum of training accuracy, no. of examples
metric = Accumulator(3)
if isinstance(updater, gluon.Trainer):
updater = updater.step
# Compute gradients and update parameters
y_hat = net(X)
l = loss(y_hat, y)
l.backward()
updater(X.shape[0])
metric.add(float(l.sum()), accuracy(y_hat, y), y.size)
# Return training loss and training accuracy
return metric[0] / metric[2], metric[1] / metric[2]
Before showing the implementation of the training function, we define a utility class that plot data
in animation. Again, it aims to simplify code in the rest of the book.

class Animator: #@save
"""For plotting data in animation."""
def __init__(self, xlabel=None, ylabel=None, legend=None, xlim=None,
ylim=None, xscale='linear', yscale='linear',
fmts=('-', 'm--', 'g-.', 'r:'), nrows=1, ncols=1,
figsize=(3.5, 2.5)):
# Incrementally plot multiple lines
if legend is None:
legend = []
self.fig, self.axes = d2l.plt.subplots(nrows, ncols, figsize=figsize)
if nrows * ncols == 1:
self.axes = [self.axes, ]
# Use a lambda function to capture arguments
self.config_axes = lambda: d2l.set_axes(
self.axes[0], xlabel, ylabel, xlim, ylim, xscale, yscale, legend)
self.X, self.Y, self.fmts = None, None, fmts
def add(self, x, y):

# Add multiple data points into the figure
if not hasattr(y, "__len__"):
y = [y]
n = len(y)
if not hasattr(x, "__len__"):
x = [x] * n
if not self.X:
self.X = [[] for _ in range(n)]
if not self.Y:
self.Y = [[] for _ in range(n)]
for i, (a, b) in enumerate(zip(x, y)):
if a is not None and b is not None:
self.X[i].append(a)
self.Y[i].append(b)
self.axes[0].cla()
for x, y, fmt in zip(self.X, self.Y, self.fmts):
self.axes[0].plot(x, y, fmt)
self.config_axes()
display.display(self.fig)
display.clear_output(wait=True)
The following training function then trains a model net on a training dataset accessed via
train_iter for multiple epochs, which is specified by num_epochs. At the end of each epoch, the
model is evaluated on a testing dataset accessed via test_iter. We will leverage the Animator class
to visualize the training progress.
def train_ch3(net, train_iter, test_iter, loss, num_epochs, updater): #@save

"""Train a model (defined in Chapter 3)."""
animator = Animator(xlabel='epoch', xlim=[1, num_epochs], ylim=[0.3, 0.9],
legend=['train loss', 'train acc', 'test acc'])
train_metrics = train_epoch_ch3(net, train_iter, loss, updater)
test_acc = evaluate_accuracy(net, test_iter)
animator.add(epoch + 1, train_metrics + (test_acc,))
train_loss, train_acc = train_metrics
assert train_loss < 0.5, train_loss

assert train_acc <= 1 and train_acc > 0.7, train_acc
assert test_acc <= 1 and test_acc > 0.7, test_acc
As an implementation from scratch, we use the minibatch stochastic gradient descent defined in
Section 3.2 to optimize the loss function of the model with a learning rate 0.1.
lr = 0.1
def updater(batch_size):
return d2l.sgd([W, b], lr, batch_size)
Now we train the model with 10 epochs. Note that both the number of epochs (num_epochs), and
learning rate (lr) are adjustable hyperparameters. By changing their values, we may be able to
increase the classification accuracy of the model.
num_epochs = 10
train_ch3(net, train_iter, test_iter, cross_entropy, num_epochs, updater)
3.6.7 Prediction
Now that training is complete, our model is ready to classify some images. Given a series of im-
ages, we will compare their actual labels (first line of text output) and the predictions from the
model (second line of text output).
def predict_ch3(net, test_iter, n=6): #@save

"""Predict labels (defined in Chapter 3)."""
for X, y in test_iter:
break
trues = d2l.get_fashion_mnist_labels(y)
preds = d2l.get_fashion_mnist_labels(d2l.argmax(net(X), axis=1))
titles = [true +'\n' + pred for true, pred in zip(trues, preds)]
d2l.show_images(d2l.reshape(X[0:n], (n, 28, 28)), 1, n, titles=titles[0:n])
predict_ch3(net, test_iter)

Summary
• With softmax regression, we can train models for multiclass classification.

• The training loop of softmax regression is very similar to that in linear regression: retrieve
and read data, define models and loss functions, then train models using optimization algo-
rithms. As you will soon find out, most common deep learning models have similar training
procedures.
Exercises
1. In this section, we directly implemented the softmax function based on the mathematical
definition of the softmax operation. What problems might this cause? Hint: try to calculate
the size of exp(50).
2. The function cross_entropy in this section was implemented according to the definition of
the cross-entropy loss function. What could be the problem with this implementation? Hint:
consider the domain of the logarithm.
3. What solutions you can think of to fix the two problems above?
4. Is it always a good idea to return the most likely label? For example, would you do this for
medical diagnosis?
5. Assume that we want to use softmax regression to predict the next word based on some
features. What are some problems that might arise from a large vocabulary?
Discussions61
3.7 Concise Implementation of Softmax Regression
Just as high-level APIs of deep learning frameworks made it much easier to implement linear re-
gression in Section 3.3, we will find it similarly (or possibly more) convenient for implementing
classification models. Let us stick with the Fashion-MNIST dataset and keep the batch size at 256
as in Section 3.6.

from mxnet import gluon, init, npx
npx.set_np()
61

batch_size = 256
As mentioned in Section 3.4, the output layer of softmax regression is a fully-connected layer.
Therefore, to implement our model, we just need to add one fully-connected layer with 10 outputs
to our Sequential. Again, here, the Sequential is not really necessary, but we might as well form
the habit since it will be ubiquitous when implementing deep models. Again, we initialize the
weights at random with zero mean and standard deviation 0.01.
3.7.2 Softmax Implementation Revisited
In the previous example of Section 3.6, we calculated our modelʼs output and then ran this output
through the cross-entropy loss. Mathematically, that is a perfectly reasonable thing to do. How-
ever, from a computational perspective, exponentiation can be a source of numerical stability
issues.
Recall that the softmax function calculates ŷj = ∑exp(oj ) , where ŷj is the j th element of the
k exp(ok )
predicted probability distribution ŷ and oj is the j th element of the logits o. Ifsome of the ok
are very large (i.e., very positive), then exp(ok ) might be larger than the largest number we can
have for certain data types (i.e., overflow). This would make the denominator (and/or numerator)
inf (infinity) and we wind up encountering either 0, inf, or nan (not a number) for ŷj . In these
situations we do not get a well-defined return value for cross entropy.
One trick to get around this is to first subtract max(ok ) from all ok before proceeding with the
softmax calculation. You can verify that this shifting of each ok by constant factor does not change
the return value of softmax. After the subtraction and normalization step, it might be possible that
some oj have large negative values and thus that the corresponding exp(oj ) will take values close
to zero. These might be rounded to zero due to finite precision (i.e., underflow), making ŷj zero and
giving us -inf for log(ŷj ). A few steps down the road in backpropagation, we might find ourselves
faced with a screenful of the dreaded nan results.
Fortunately, we are saved by the fact that even though we are computing exponential functions,
we ultimately intend to take their log (when calculating the cross-entropy loss). By combining
these two operators softmax and cross entropy together, we can escape the numerical stability
issues that might otherwise plague us during backpropagation. As shown in the equation below,
we avoid calculating exp(oj ) and can use instead oj directly due to the canceling in log(exp(·)).
( )
exp(oj )
log (ŷj ) = log ∑
k exp(ok )
( )
∑
= log (exp(oj )) − log exp(ok ) (3.7.1)
k
( )
∑
= oj − log exp(ok ) .
k
3.7. Concise Implementation of Softmax Regression 127

We will want to keep the conventional softmax function handy in case we ever want to evaluate
the output probabilities by our model. But instead of passing softmax probabilities into our new
loss function, we will just pass the logits and compute the softmax and its log all at once inside the
cross entropy loss function, which does smart things like the “LogSumExp trick”62 .
loss = gluon.loss.SoftmaxCrossEntropyLoss()
3.7.3 Optimization Algorithm
Here, we use minibatch stochastic gradient descent with a learning rate of 0.1 as the optimiza-
tion algorithm. Note that this is the same as we applied in the linear regression example and it
illustrates the general applicability of the optimizers.
trainer = gluon.Trainer(net.collect_params(), 'sgd', {'learning_rate': 0.1})
3.7.4 Training
Next we call the training function defined in Section 3.6 to train the model.
num_epochs = 10
d2l.train_ch3(net, train_iter, test_iter, loss, num_epochs, trainer)
As before, this algorithm converges to a solution that achieves a decent accuracy, albeit this time
with fewer lines of code than before.
62
https://en.wikipedia.org/wiki/LogSumExp

Summary
• Using high-level APIs, we can implement softmax regression much more concisely.
• From a computational perspective, implementing softmax regression has intricacies. Note
that in many cases, a deep learning framework takes additional precautions beyond these
most well-known tricks to ensure numerical stability, saving us from even more pitfalls that
we would encounter if we tried to code all of our models from scratch in practice.
Exercises
1. Try adjusting the hyperparameters, such as the batch size, number of epochs, and learning
rate, to see what the results are.
2. Increase the numper of epochs for training. Why might the test accuracy decrease after a
while? How could we fix this?
Discussions63
63
3.7. Concise Implementation of Softmax Regression 129

4 | Multilayer Perceptrons
In this chapter, we will introduce your first truly deep network. The simplest deep networks are
called multilayer perceptrons, and they consist of multiple layers of neurons each fully connected
to those in the layer below (from which they receive input) and those above (which they, in turn,
influence). When we train high-capacity models we run the risk of overfitting. Thus, we will
need to provide your first rigorous introduction to the notions of overfitting, underfitting, and
model selection. To help you combat these problems, we will introduce regularization techniques
such as weight decay and dropout. We will also discuss issues relating to numerical stability and
parameter initialization that are key to successfully training deep networks. Throughout, we aim
to give you a firm grasp not just of the concepts but also of the practice of using deep networks.
At the end of this chapter, we apply what we have introduced so far to a real case: house price
prediction. We punt matters relating to the computational performance, scalability, and efficiency
of our models to subsequent chapters.
4.1 Multilayer Perceptrons
In Chapter 3, we introduced softmax regression (Section 3.4), implementing the algorithm from
scratch (Section 3.6) and using high-level APIs (Section 3.7), and training classifiers to recognize 10
categories of clothing from low-resolution images. Along the way, we learned how to wrangle data,
coerce our outputs into a valid probability distribution, apply an appropriate loss function, and
minimize it with respect to our modelʼs parameters. Now that we have mastered these mechanics
in the context of simple linear models, we can launch our exploration of deep neural networks,
the comparatively rich class of models with which this book is primarily concerned.
4.1.1 Hidden Layers
We have described the affine transformation in Section 3.1.1, which is a linear transformation
added by a bias. To begin, recall the model architecture corresponding to our softmax regression
example, illustrated in Fig. 3.4.1. This model mapped our inputs directly to our outputs via a
single affine transformation, followed by a softmax operation. If our labels truly were related to
our input data by an affine transformation, then this approach would be sufficient. But linearity
in affine transformations is a strong assumption.
131
Linear Models May Go Wrong
For example, linearity implies the weaker assumption of monotonicity: that any increase in our
feature must either always cause an increase in our modelʼs output (if the corresponding weight
is positive), or always cause a decrease in our modelʼs output (if the corresponding weight is neg-
ative). Sometimes that makes sense. For example, if we were trying to predict whether an indi-
vidual will repay a loan, we might reasonably imagine that holding all else equal, an applicant
with a higher income would always be more likely to repay than one with a lower income. While
monotonic, this relationship likely is not linearly associated with the probability of repayment.
An increase in income from 0 to 50 thousand likely corresponds to a bigger increase in likelihood
of repayment than an increase from 1 million to 1.05 million. One way to handle this might be
to preprocess our data such that linearity becomes more plausible, say, by using the logarithm of
income as our feature.
Note that we can easily come up with examples that violate monotonicity. Say for example that we
want to predict probability of death based on body temperature. For individuals with a body tem-
perature above 37°C (98.6°F), higher temperatures indicate greater risk. However, for individuals
with body temperatures below 37° C, higher temperatures indicate lower risk! In this case too, we
might resolve the problem with some clever preprocessing. Namely, we might use the distance
from 37°C as our feature.
But what about classifying images of cats and dogs? Should increasing the intensity of the pixel
at location (13, 17) always increase (or always decrease) the likelihood that the image depicts a
dog? Reliance on a linear model corresponds to the implicit assumption that the only requirement
for differentiating cats vs. dogs is to assess the brightness of individual pixels. This approach is
doomed to fail in a world where inverting an image preserves the category.
And yet despite the apparent absurdity of linearity here, as compared with our previous exam-
ples, it is less obvious that we could address the problem with a simple preprocessing fix. That
is because the significance of any pixel depends in complex ways on its context (the values of the
surrounding pixels). While there might exist a representation of our data that would take into
account the relevant interactions among our features, on top of which a linear model would be
suitable, we simply do not know how to calculate it by hand. With deep neural networks, we used
observational data to jointly learn both a representation via hidden layers and a linear predictor
that acts upon that representation.
Incorporating Hidden Layers
We can overcome these limitations of linear models and handle a more general class of functions
by incorporating one or more hidden layers. The easiest way to do this is to stack many fully-
connected layers on top of each other. Each layer feeds into the layer above it, until we generate
outputs. We can think of the first L−1 layers as our representation and the final layer as our linear
predictor. This architecture is commonly called a multilayer perceptron, often abbreviated as MLP.
Below, we depict an MLP diagrammatically (Fig. 4.1.1).
132 Chapter 4. Multilayer Perceptrons

Fig. 4.1.1: An MLP with a hidden layer of 5 hidden units.
This MLP has 4 inputs, 3 outputs, and its hidden layer contains 5 hidden units. Since the input layer
does not involve any calculations, producing outputs with this network requires implementing the
computations for both the hidden and output layers; thus, the number of layers in this MLP is 2.
Note that these layers are both fully connected. Every input influences every neuron in the hidden
layer, and each of these in turn influences every neuron in the output layer.
From Linear to Nonlinear
As before, by the matrix X ∈ Rn×d , we denote a minibatch of n examples where each example has
d inputs (features). For a one-hidden-layer MLP whose hidden layer has h hidden units, denote
by H ∈ Rn×h the outputs of the hidden layer, which are hidden representations. In mathematics or
code, H is also known as a hidden-layer variable or a hidden variable. Since the hidden and output
layers are both fully connected, we have hidden-layer weights W(1) ∈ Rd×h and biases b(1) ∈ R1×h
and output-layer weights W(2) ∈ Rh×q and biases b(2) ∈ R1×q . Formally, we calculate the outputs
O ∈ Rn×q of the one-hidden-layer MLP as follows:
H = XW(1) + b(1) ,
(4.1.1)
O = HW(2) + b(2) .
Note that after adding the hidden layer, our model now requires us to track and update additional
sets of parameters. So what have we gained in exchange? You might be surprised to find out that—
in the model defined above—we gain nothing for our troubles! The reason is plain. The hidden units
above are given by an affine function of the inputs, and the outputs (pre-softmax) are just an affine
function of the hidden units. An affine function of an affine function is itself an affine function.
Moreover, our linear model was already capable of representing any affine function.
We can view the equivalence formally by proving that for any values of the weights, we can just
collapse out the hidden layer, yielding an equivalent single-layer model with parameters W =
W(1) W(2) and b = b(1) W(2) + b(2) :
O = (XW(1) + b(1) )W(2) + b(2) = XW(1) W(2) + b(1) W(2) + b(2) = XW + b. (4.1.2)
In order to realize the potential of multilayer architectures, we need one more key ingredient:
a nonlinear activation function σ to be applied to each hidden unit following the affine transfor-
mation. The outputs of activation functions (e.g., σ(·)) are called activations. In general, with
activation functions in place, it is no longer possible to collapse our MLP into a linear model:
H = σ(XW(1) + b(1) ),
(4.1.3)
O = HW(2) + b(2) .
4.1. Multilayer Perceptrons 133

Since each row in X corresponds to an example in the minibatch, with some abuse of notation,
we define the nonlinearity σ to apply to its inputs in a rowwise fashion, i.e., one example at a
time. Note that we used the notation for softmax in the same way to denote a rowwise operation
in Section 3.4.4. Often, as in this section, the activation functions that we apply to hidden layers
are not merely rowwise, but elementwise. That means that after computing the linear portion of
the layer, we can calculate each activation without looking at the values taken by the other hidden
units. This is true for most activation functions.
To build more general MLPs, we can continue stacking such hidden layers, e.g., H(1) = σ1 (XW(1) +
b(1) ) and H(2) = σ2 (H(1) W(2) + b(2) ), one atop another, yielding ever more expressive models.
Universal Approximators
MLPs can capture complex interactions among our inputs via their hidden neurons, which depend
on the values of each of the inputs. We can easily design hidden nodes to perform arbitrary com-
putation, for instance, basic logic operations on a pair of inputs. Moreover, for certain choices
of the activation function, it is widely known that MLPs are universal approximators. Even with
a single-hidden-layer network, given enough nodes (possibly absurdly many), and the right set of
weights, we can model any function, though actually learning that function is the hard part. You
might think of your neural network as being a bit like the C programming language. The language,
like any other modern language, is capable of expressing any computable program. But actually
coming up with a program that meets your specifications is the hard part.
Moreover, just because a single-hidden-layer network can learn any function does not mean that
you should try to solve all of your problems with single-hidden-layer networks. In fact, we can
approximate many functions much more compactly by using deeper (vs. wider) networks. We
will touch upon more rigorous arguments in subsequent chapters.
%matplotlib inline
npx.set_np()
4.1.2 Activation Functions
Activation functions decide whether a neuron should be activated or not by calculating the
weighted sum and further adding bias with it. They are differentiable operators to transform
input signals to outputs, while most of them add non-linearity. Because activation functions are
fundamental to deep learning, let us briefly survey some common activation functions.
ReLU Function
The most popular choice, due to both simplicity of implementation and its good performance on a
variety of predictive tasks, is the rectified linear unit (ReLU). ReLU provides a very simple nonlinear
transformation. Given an element x, the function is defined as the maximum of that element and
0:
ReLU(x) = max(x, 0). (4.1.4)

Informally, the ReLU function retains only positive elements and discards all negative elements
by setting the corresponding activations to 0. To gain some intuition, we can plot the function. As
you can see, the activation function is piecewise linear.
x = np.arange(-8.0, 8.0, 0.1)

x.attach_grad()
y = npx.relu(x)
d2l.plot(x, y, 'x', 'relu(x)', figsize=(5, 2.5))
When the input is negative, the derivative of the ReLU function is 0, and when the input is positive,
the derivative of the ReLU function is 1. Note that the ReLU function is not differentiable when the
input takes value precisely equal to 0. In these cases, we default to the left-hand-side derivative
and say that the derivative is 0 when the input is 0. We can get away with this because the input
may never actually be zero. There is an old adage that if subtle boundary conditions matter, we
are probably doing (real) mathematics, not engineering. That conventional wisdom may apply
here. We plot the derivative of the ReLU function plotted below.
y.backward()
d2l.plot(x, x.grad, 'x', 'grad of relu', figsize=(5, 2.5))
The reason for using ReLU is that its derivatives are particularly well behaved: either they vanish

or they just let the argument through. This makes optimization better behaved and it mitigated
the well-documented problem of vanishing gradients that plagued previous versions of neural
networks (more on this later).
Note that there are many variants to the ReLU function, including the parameterized ReLU (pReLU)
function (He et al., 2015). This variation adds a linear term to ReLU, so some information still gets
through, even when the argument is negative:
pReLU(x) = max(0, x) + α min(0, x). (4.1.5)
Sigmoid Function
The sigmoid function transforms its inputs, for which values lie in the domain R, to outputs that lie
on the interval (0, 1). For that reason, the sigmoid is often called a squashing function: it squashes
any input in the range (-inf, inf) to some value in the range (0, 1):
1
sigmoid(x) = . (4.1.6)
1 + exp(−x)
In the earliest neural networks, scientists were interested in modeling biological neurons which
either fire or do not fire. Thus the pioneers of this field, going all the way back to McCulloch and
Pitts, the inventors of the artificial neuron, focused on thresholding units. A thresholding activa-
tion takes value 0 when its input is below some threshold and value 1 when the input exceeds the
threshold.
When attention shifted to gradient based learning, the sigmoid function was a natural choice be-
cause it is a smooth, differentiable approximation to a thresholding unit. Sigmoids are still widely
used as activation functions on the output units, when we want to interpret the outputs as prob-
abilities for binary classification problems (you can think of the sigmoid as a special case of the
softmax). However, the sigmoid has mostly been replaced by the simpler and more easily train-
able ReLU for most use in hidden layers. In later chapters on recurrent neural networks, we will
describe architectures that leverage sigmoid units to control the flow of information across time.
Below, we plot the sigmoid function. Note that when the input is close to 0, the sigmoid function
approaches a linear transformation.
y = npx.sigmoid(x)
d2l.plot(x, y, 'x', 'sigmoid(x)', figsize=(5, 2.5))

The derivative of the sigmoid function is given by the following equation:
d exp(−x)
sigmoid(x) = = sigmoid(x) (1 − sigmoid(x)) . (4.1.7)
dx (1 + exp(−x))2
The derivative of the sigmoid function is plotted below. Note that when the input is 0, the deriva-
tive of the sigmoid function reaches a maximum of 0.25. As the input diverges from 0 in either
direction, the derivative approaches 0.
y.backward()
d2l.plot(x, x.grad, 'x', 'grad of sigmoid', figsize=(5, 2.5))
Tanh Function
Like the sigmoid function, the tanh (hyperbolic tangent) function also squashes its inputs, trans-
forming them into elements on the interval between -1 and 1:
1 − exp(−2x)
tanh(x) = . (4.1.8)
1 + exp(−2x)
We plot the tanh function below. Note that as the input nears 0, the tanh function approaches a lin-
ear transformation. Although the shape of the function is similar to that of the sigmoid function,
the tanh function exhibits point symmetry about the origin of the coordinate system.
y = np.tanh(x)
d2l.plot(x, y, 'x', 'tanh(x)', figsize=(5, 2.5))

The derivative of the tanh function is:
d
tanh(x) = 1 − tanh2 (x). (4.1.9)
dx
The derivative of tanh function is plotted below. As the input nears 0, the derivative of the tanh
function approaches a maximum of 1. And as we saw with the sigmoid function, as the input
moves away from 0 in either direction, the derivative of the tanh function approaches 0.
y.backward()
d2l.plot(x, x.grad, 'x', 'grad of tanh', figsize=(5, 2.5))
In summary, we now know how to incorporate nonlinearities to build expressive multilayer neu-
ral network architectures. As a side note, your knowledge already puts you in command of a simi-
lar toolkit to a practitioner circa 1990. In some ways, you have an advantage over anyone working
in the 1990s, because you can leverage powerful open-source deep learning frameworks to build
models rapidly, using only a few lines of code. Previously, training these networks required re-
searchers to code up thousands of lines of C and Fortran.

Summary
• MLP adds one or multiple fully-connected hidden layers between the output and input layers
and transforms the output of the hidden layer via an activation function.
• Commonly-used activation functions include the ReLU function, the sigmoid function, and
the tanh function.
Exercises
1. Compute the derivative of the pReLU activation function.

2. Show that an MLP using only ReLU (or pReLU) constructs a continuous piecewise linear
function.
3. Show that tanh(x) + 1 = 2 sigmoid(2x).
4. Assume that we have a nonlinearity that applies to one minibatch at a time. What kinds of
problems do you expect this to cause?
Discussions64
4.2 Implementation of Multilayer Perceptrons from Scratch
Now that we have characterized multilayer perceptrons (MLPs) mathematically, let us try to im-
plement one ourselves. To compare against our previous results achieved with softmax regression
(Section 3.6), we will continue to work with the Fashion-MNIST image classification dataset (Sec-
tion 3.5).

from mxnet import gluon, np, npx
npx.set_np()
batch_size = 256
Recall that Fashion-MNIST contains 10 classes, and that each image consists of a 28×28 = 784 grid
of grayscale pixel values. Again, we will disregard the spatial structure among the pixels for now,
so we can think of this as simply a classification dataset with 784 input features and 10 classes. To
begin, we will implement an MLP with one hidden layer and 256 hidden units. Note that we can
regard both of these quantities as hyperparameters. Typically, we choose layer widths in powers of
2, which tend to be computationally efficient because of how memory is allocated and addressed
in hardware.
Again, we will represent our parameters with several tensors. Note that for every layer, we must
keep track of one weight matrix and one bias vector. As always, we allocate memory for the gra-
dients of the loss with respect to these parameters.
64
4.2. Implementation of Multilayer Perceptrons from Scratch 139

num_inputs, num_outputs, num_hiddens = 784, 10, 256
W1 = np.random.normal(scale=0.01, size=(num_inputs, num_hiddens))

b1 = np.zeros(num_hiddens)
W2 = np.random.normal(scale=0.01, size=(num_hiddens, num_outputs))
b2 = np.zeros(num_outputs)
params = [W1, b1, W2, b2]

param.attach_grad()
4.2.2 Activation Function
To make sure we know how everything works, we will implement the ReLU activation ourselves
using the maximum function rather than invoking the built-in relu function directly.
def relu(X):
return np.maximum(X, 0)
4.2.3 Model
Because we are disregarding spatial structure, we reshape each two-dimensional image into a flat
vector of length num_inputs. Finally, we implement our model with just a few lines of code.
def net(X):
X = X.reshape((-1, num_inputs))
H = relu(np.dot(X, W1) + b1)
return np.dot(H, W2) + b2
4.2.4 Loss Function
To ensure numerical stability, and because we already implemented the softmax function from
scratch (Section 3.6), we leverage the integrated function from high-level APIs for calculating the
softmax and cross-entropy loss. Recall our earlier discussion of these intricacies in Section 3.7.2.
We encourage the interested reader to examine the source code for the loss function to deepen
their knowledge of implementation details.

4.2.5 Training
Fortunately, the training loop for MLPs is exactly the same as for softmax regression. Leveraging
the d2l package again, we call the train_ch3 function (see Section 3.6), setting the number of
epochs to 10 and the learning rate to 0.5.
num_epochs, lr = 10, 0.1

d2l.train_ch3(net, train_iter, test_iter, loss, num_epochs,
lambda batch_size: d2l.sgd(params, lr, batch_size))
To evaluate the learned model, we apply it on some test data.
d2l.predict_ch3(net, test_iter)
Summary
• We saw that implementing a simple MLP is easy, even when done manually.
• However, with a large number of layers, implementing MLPs from scratch can still get messy
(e.g., naming and keeping track of our modelʼs parameters).
4.2. Implementation of Multilayer Perceptrons from Scratch 141

Exercises
1. Change the value of the hyperparameter num_hiddens and see how this hyperparameter in-
fluences your results. Determine the best value of this hyperparameter, keeping all others
constant.
2. Try adding an additional hidden layer to see how it affects the results.
3. How does changing the learning rate alter your results? Fixing the model architecture and
other hyperparameters (including number of epochs), what learning rate gives you the best
results?
4. What is the best result you can get by optimizing over all the hyperparameters (learning rate,
number of epochs, number of hidden layers, number of hidden units per layer) jointly?
5. Describe why it is much more challenging to deal with multiple hyperparameters.
6. What is the smartest strategy you can think of for structuring a search over multiple hyper-
parameters?
Discussions65
4.3 Concise Implementation of Multilayer Perceptrons
As you might expect, by relying on the high-level APIs, we can implement MLPs even more con-
cisely.

from mxnet import gluon, init, npx
npx.set_np()
4.3.1 Model
As compared with our concise implementation of softmax regression implementation (Section

3.7), the only difference is that we add two fully-connected layers (previously, we added one). The
first is our hidden layer, which contains 256 hidden units and applies the ReLU activation function.
The second is our output layer.
net.add(nn.Dense(256, activation='relu'),
nn.Dense(10))
The training loop is exactly the same as when we implemented softmax regression. This modu-
larity enables us to separate matters concerning the model architecture from orthogonal consid-
erations.
65

batch_size, lr, num_epochs = 256, 0.1, 10
trainer = gluon.Trainer(net.collect_params(), 'sgd', {'learning_rate': lr})

Summary
• Using high-level APIs, we can implement MLPs much more concisely.

• For the same classification problem, the implementation of an MLP is the same as that of
softmax regression except for additional hidden layers with activation functions.
Exercises
1. Try adding different numbers of hidden layers (you may also modify the learning rate). What
setting works best?
2. Try out different activation functions. Which one works best?
3. Try different schemes for initializing the weights. What method works best?
Discussions66
66
4.3. Concise Implementation of Multilayer Perceptrons 143

4.4 Model Selection, Underfitting, and Overfitting
As machine learning scientists, our goal is to discover patterns. But how can we be sure that we
have truly discovered a general pattern and not simply memorized our data? For example, imagine
that we wanted to hunt for patterns among genetic markers linking patients to their dementia
status, where the labels are drawn from the set {dementia, mild cognitive impairment, healthy}.
Because each personʼs genes identify them uniquely (ignoring identical siblings), it is possible to
memorize the entire dataset.
We do not want our model to say “That’s Bob! I remember him! He has dementia!” The reason why
is simple. When we deploy the model in the future, we will encounter patients that the model has
never seen before. Our predictions will only be useful if our model has truly discovered a general
pattern.
To recapitulate more formally, our goal is to discover patterns that capture regularities in the un-
derlying population from which our training set was drawn. If we are successful in this endeavor,
then we could successfully assess risk even for individuals that we have never encountered before.
This problem—how to discover patterns that generalize—is the fundamental problem of machine
learning.
The danger is that when we train models, we access just a small sample of data. The largest public
image datasets contain roughly one million images. More often, we must learn from only thou-
sands or tens of thousands of data examples. In a large hospital system, we might access hundreds
of thousands of medical records. When working with finite samples, we run the risk that we might
discover apparent associations that turn out not to hold up when we collect more data.
The phenomenon of fitting our training data more closely than we fit the underlying distribution
is called overfitting, and the techniques used to combat overfitting are called regularization. In
the previous sections, you might have observed this effect while experimenting with the Fashion-
MNIST dataset. If you altered the model structure or the hyperparameters during the experiment,
you might have noticed that with enough neurons, layers, and training epochs, the model can
eventually reach perfect accuracy on the training set, even as the accuracy on test data deterio-
rates.
4.4.1 Training Error and Generalization Error
In order to discuss this phenomenon more formally, we need to differentiate between training
error and generalization error. The training error is the error of our model as calculated on the
training dataset, while generalization error is the expectation of our modelʼs error were we to apply
it to an infinite stream of additional data examples drawn from the same underlying data distri-
bution as our original sample.
Problematically, we can never calculate the generalization error exactly. That is because the
stream of infinite data is an imaginary object. In practice, we must estimate the generalization
error by applying our model to an independent test set constituted of a random selection of data
examples that were withheld from our training set.
The following three thought experiments will help illustrate this situation better. Consider a col-
lege student trying to prepare for his final exam. A diligent student will strive to practice well and
test his abilities using exams from previous years. Nonetheless, doing well on past exams is no
guarantee that he will excel when it matters. For instance, the student might try to prepare by rote
learning the answers to the exam questions. This requires the student to memorize many things.
She might even remember the answers for past exams perfectly. Another student might prepare

by trying to understand the reasons for giving certain answers. In most cases, the latter student
will do much better.
Likewise, consider a model that simply uses a lookup table to answer questions. If the set of allow-
able inputs is discrete and reasonably small, then perhaps after viewing many training examples,
this approach would perform well. Still this model has no ability to do better than random guess-
ing when faced with examples that it has never seen before. In reality the input spaces are far too
large to memorize the answers corresponding to every conceivable input. For example, consider
the black and white 28 × 28 images. If each pixel can take one among 256 grayscale values, then
there are 256784 possible images. That means that there are far more low-resolution grayscale
thumbnail-sized images than there are atoms in the universe. Even if we could encounter such
data, we could never afford to store the lookup table.
Last, consider the problem of trying to classify the outcomes of coin tosses (class 0: heads, class
1: tails) based on some contextual features that might be available. Suppose that the coin is fair.
No matter what algorithm we come up with, the generalization error will always be 21 . However,
for most algorithms, we should expect our training error to be considerably lower, depending on
the luck of the draw, even if we did not have any features! Consider the dataset {0, 1, 1, 1, 0, 1}.
Our feature-less algorithm would have to fall back on always predicting the majority class, which
appears from our limited sample to be 1. In this case, the model that always predicts class 1 will
incur an error of 31 , considerably better than our generalization error. As we increase the amount
of data, the probability that the fraction of heads will deviate significantly from 12 diminishes, and
our training error would come to match the generalization error.
Statistical Learning Theory
Since generalization is the fundamental problem in machine learning, you might not be surprised
to learn that many mathematicians and theorists have dedicated their lives to developing formal
theories to describe this phenomenon. In their eponymous theorem67 , Glivenko and Cantelli de-
rived the rate at which the training error converges to the generalization error. In a series of semi-
nal papers, Vapnik and Chervonenkis68 extended this theory to more general classes of functions.
This work laid the foundations of statistical learning theory.
In the standard supervised learning setting, which we have addressed up until now and will stick
with throughout most of this book, we assume that both the training data and the test data are
drawn independently from identical distributions. This is commonly called the i.i.d. assumption,
which means that the process that samples our data has no memory. In other words, the second
example drawn and the third drawn are no more correlated than the second and the two-millionth
sample drawn.
Being a good machine learning scientist requires thinking critically, and already you should be
poking holes in this assumption, coming up with common cases where the assumption fails. What
if we train a mortality risk predictor on data collected from patients at UCSF Medical Center, and
apply it on patients at Massachusetts General Hospital? These distributions are simply not identi-
cal. Moreover, draws might be correlated in time. What if we are classifying the topics of Tweets?
The news cycle would create temporal dependencies in the topics being discussed, violating any
assumptions of independence.
Sometimes we can get away with minor violations of the i.i.d. assumption and our models will
continue to work remarkably well. After all, nearly every real-world application involves at least
67
https://en.wikipedia.org/wiki/Glivenko%E2%80%93Cantelli_theorem
68
https://en.wikipedia.org/wiki/Vapnik%E2%80%93Chervonenkis_theory
4.4. Model Selection, Underfitting, and Overfitting 145

some minor violation of the i.i.d. assumption, and yet we have many useful tools for various ap-
plications such as face recognition, speech recognition, and language translation.
Other violations are sure to cause trouble. Imagine, for example, if we try to train a face recog-
nition system by training it exclusively on university students and then want to deploy it as a tool
for monitoring geriatrics in a nursing home population. This is unlikely to work well since college
students tend to look considerably different from the elderly.
In subsequent chapters, we will discuss problems arising from violations of the i.i.d. assump-
tion. For now, even taking the i.i.d. assumption for granted, understanding generalization is a
formidable problem. Moreover, elucidating the precise theoretical foundations that might ex-
plain why deep neural networks generalize as well as they do continues to vex the greatest minds
in learning theory.
When we train our models, we attempt to search for a function that fits the training data as well as
possible. If the function is so flexible that it can catch on to spurious patterns just as easily as to
true associations, then it might perform too well without producing a model that generalizes well
to unseen data. This is precisely what we want to avoid or at least control. Many of the techniques
in deep learning are heuristics and tricks aimed at guarding against overfitting.
Model Complexity
When we have simple models and abundant data, we expect the generalization error to resemble
the training error. When we work with more complex models and fewer examples, we expect the
training error to go down but the generalization gap to grow. What precisely constitutes model
complexity is a complex matter. Many factors govern whether a model will generalize well. For
example a model with more parameters might be considered more complex. A model whose
parameters can take a wider range of values might be more complex. Often with neural networks,
we think of a model that takes more training iterations as more complex, and one subject to early
stopping (fewer training iterations) as less complex.
It can be difficult to compare the complexity among members of substantially different model
classes (say, decision trees vs. neural networks). For now, a simple rule of thumb is quite useful:
a model that can readily explain arbitrary facts is what statisticians view as complex, whereas one
that has only a limited expressive power but still manages to explain the data well is probably
closer to the truth. In philosophy, this is closely related to Popperʼs criterion of falsifiability of a
scientific theory: a theory is good if it fits data and if there are specific tests that can be used to
disprove it. This is important since all statistical estimation is post hoc, i.e., we estimate after we
observe the facts, hence vulnerable to the associated fallacy. For now, we will put the philosophy
aside and stick to more tangible issues.
In this section, to give you some intuition, we will focus on a few factors that tend to influence the
generalizability of a model class:
1. The number of tunable parameters. When the number of tunable parameters, sometimes
called the degrees of freedom, is large, models tend to be more susceptible to overfitting.
2. The values taken by the parameters. When weights can take a wider range of values, models
can be more susceptible to overfitting.
3. The number of training examples. It is trivially easy to overfit a dataset containing only
one or two examples even if your model is simple. But overfitting a dataset with millions of
examples requires an extremely flexible model.

4.4.2 Model Selection
In machine learning, we usually select our final model after evaluating several candidate models.
This process is called model selection. Sometimes the models subject to comparison are fundamen-
tally different in nature (say, decision trees vs. linear models). At other times, we are comparing
members of the same class of models that have been trained with different hyperparameter set-
tings.
With MLPs, for example, we may wish to compare models with different numbers of hidden layers,
different numbers of hidden units, and various choices of the activation functions applied to each
hidden layer. In order to determine the best among our candidate models, we will typically employ
a validation dataset.
Validation Dataset
In principle we should not touch our test set until after we have chosen all our hyperparameters.
Were we to use the test data in the model selection process, there is a risk that we might overfit
the test data. Then we would be in serious trouble. If we overfit our training data, there is always
the evaluation on test data to keep us honest. But if we overfit the test data, how would we ever
know?
Thus, we should never rely on the test data for model selection. And yet we cannot rely solely on
the training data for model selection either because we cannot estimate the generalization error
on the very data that we use to train the model.
In practical applications, the picture gets muddier. While ideally we would only touch the test
data once, to assess the very best model or to compare a small number of models to each other,
real-world test data is seldom discarded after just one use. We can seldom afford a new test set for
each round of experiments.
The common practice to address this problem is to split our data three ways, incorporating a vali-
dation dataset (or validation set) in addition to the training and test datasets. The result is a murky
practice where the boundaries between validation and test data are worryingly ambiguous. Un-
less explicitly stated otherwise, in the experiments in this book we are really working with what
should rightly be called training data and validation data, with no true test sets. Therefore, the
accuracy reported in each experiment of the book is really the validation accuracy and not a true
test set accuracy.
K-Fold Cross-Validation
When training data is scarce, we might not even be able to afford to hold out enough data to con-
stitute a proper validation set. One popular solution to this problem is to employ K-fold cross-
validation. Here, the original training data is split into K non-overlapping subsets. Then model
training and validation are executed K times, each time training on K − 1 subsets and validat-
ing on a different subset (the one not used for training in that round). Finally, the training and
validation errors are estimated by averaging over the results from the K experiments.

4.4.3 Underfitting or Overfitting?
When we compare the training and validation errors, we want to be mindful of two common situ-
ations. First, we want to watch out for cases when our training error and validation error are both
substantial but there is a little gap between them. If the model is unable to reduce the training
error, that could mean that our model is too simple (i.e., insufficiently expressive) to capture the
pattern that we are trying to model. Moreover, since the generalization gap between our train-
ing and validation errors is small, we have reason to believe that we could get away with a more
complex model. This phenomenon is known as underfitting.
On the other hand, as we discussed above, we want to watch out for the cases when our train-
ing error is significantly lower than our validation error, indicating severe overfitting. Note that
overfitting is not always a bad thing. With deep learning especially, it is well known that the best
predictive models often perform far better on training data than on holdout data. Ultimately, we
usually care more about the validation error than about the gap between the training and valida-
tion errors.
Whether we overfit or underfit can depend both on the complexity of our model and the size of
the available training datasets, two topics that we discuss below.
Model Complexity
To illustrate some classical intuition about overfitting and model complexity, we give an example
using polynomials. Given training data consisting of a single feature x and a corresponding real-
valued label y, we try to find the polynomial of degree d
∑
d
ŷ = xi wi (4.4.1)
i=0
to estimate the labels y. This is just a linear regression problem where our features are given by
the powers of x, the modelʼs weights are given by wi , and the bias is given by w0 since x0 = 1 for all
x. Since this is just a linear regression problem, we can use the squared error as our loss function.
A higher-order polynomial function is more complex than a lower-order polynomial function,
since the higher-order polynomial has more parameters and the model functionʼs selection range
is wider. Fixing the training dataset, higher-order polynomial functions should always achieve
lower (at worst, equal) training error relative to lower degree polynomials. In fact, whenever the
data examples each have a distinct value of x, a polynomial function with degree equal to the
number of data examples can fit the training set perfectly. We visualize the relationship between
polynomial degree and underfitting vs. overfitting in Fig. 4.4.1.

Fig. 4.4.1: Influence of model complexity on underfitting and overfitting
Dataset Size
The other big consideration to bear in mind is the dataset size. Fixing our model, the fewer sam-
ples we have in the training dataset, the more likely (and more severely) we are to encounter over-
fitting. As we increase the amount of training data, the generalization error typically decreases.
Moreover, in general, more data never hurt. For a fixed task and data distribution, there is typi-
cally a relationship between model complexity and dataset size. Given more data, we might prof-
itably attempt to fit a more complex model. Absent sufficient data, simpler models may be more
difficult to beat. For many tasks, deep learning only outperforms linear models when many thou-
sands of training examples are available. In part, the current success of deep learning owes to
the current abundance of massive datasets due to Internet companies, cheap storage, connected
devices, and the broad digitization of the economy.
4.4.4 Polynomial Regression
We can now explore these concepts interactively by fitting polynomials to data.

import math
npx.set_np()

Generating the Dataset
First we need data. Given x, we will use the following cubic polynomial to generate the labels on
training and test data:
x2 x3
y = 5 + 1.2x − 3.4 + 5.6 + ϵ where ϵ ∼ N (0, 0.12 ). (4.4.2)
2! 3!
The noise term ϵ obeys a normal distribution with a mean of 0 and a standard deviation of 0.1. For
optimization, we typically want to avoid very large values of gradients or losses. This is why the
i
features are rescaled from xi to xi! . It allows us to avoid very large values for large exponents i. We
will synthesize 100 samples each for the training set and test set.
max_degree = 20 # Maximum degree of the polynomial

n_train, n_test = 100, 100 # Training and test dataset sizes
true_w = np.zeros(max_degree) # Allocate lots of empty space
true_w[0:4] = np.array([5, 1.2, -3.4, 5.6])
features = np.random.normal(size=(n_train + n_test, 1))

np.random.shuffle(features)
poly_features = np.power(features, np.arange(max_degree).reshape(1, -1))
for i in range(max_degree):
poly_features[:, i] /= math.gamma(i + 1) # `gamma(n)` = (n-1)!
# Shape of `labels`: (`n_train` + `n_test`,)
labels = np.dot(poly_features, true_w)
labels += np.random.normal(scale=0.1, size=labels.shape)
Again, monomials stored in poly_features are rescaled by the gamma function, where Γ(n) =
(n − 1)!. Take a look at the first 2 samples from the generated dataset. The value 1 is technically a
feature, namely the constant feature corresponding to the bias.
features[:2], poly_features[:2, :], labels[:2]
(array([[2.2122064],
[1.1630787]]),
array([[1.00000000e+00, 2.21220636e+00, 2.44692850e+00, 1.80437028e+00,
9.97909844e-01, 4.41516489e-01, 1.62787601e-01, 5.14456779e-02,
1.42260585e-02, 3.49677517e-03, 7.73558859e-04, 1.55570160e-04,
2.86794402e-05, 4.88037267e-06, 7.71170789e-07, 1.13732597e-07,
1.57249964e-08, 2.04629069e-09, 2.51489857e-10, 2.92814419e-11],
[1.00000000e+00, 1.16307867e+00, 6.76375985e-01, 2.62226164e-01,
7.62474164e-02, 1.77363474e-02, 3.43812793e-03, 5.71259065e-04,
8.30524004e-05, 1.07329415e-05, 1.24832559e-06, 1.31990987e-07,
1.27929916e-08, 1.14455811e-09, 9.50865081e-11, 7.37287228e-12,
5.35951925e-13, 3.66678970e-14, 2.36931378e-15, 1.45036750e-16]]),
array([9.629796, 5.51997 ]))

Training and Testing the Model
Let us first implement a function to evaluate the loss on a given dataset.
def evaluate_loss(net, data_iter, loss): #@save

"""Evaluate the loss of a model on the given dataset."""
metric = d2l.Accumulator(2) # Sum of losses, no. of examples
l = loss(net(X), y)
metric.add(d2l.reduce_sum(l), l.size)
Now define the training function.
def train(train_features, test_features, train_labels, test_labels,

num_epochs=400):
# Switch off the bias since we already catered for it in the polynomial
# features
net.add(nn.Dense(1, use_bias=False))
net.initialize()
batch_size = min(10, train_labels.shape[0])
train_iter = d2l.load_array((train_features, train_labels), batch_size)
test_iter = d2l.load_array((test_features, test_labels), batch_size,
is_train=False)
trainer = gluon.Trainer(net.collect_params(), 'sgd',
{'learning_rate': 0.01})
animator = d2l.Animator(xlabel='epoch', ylabel='loss', yscale='log',
xlim=[1, num_epochs], ylim=[1e-3, 1e2],
legend=['train', 'test'])
d2l.train_epoch_ch3(net, train_iter, loss, trainer)
if epoch == 0 or (epoch + 1) % 20 == 0:
animator.add(epoch + 1, (evaluate_loss(net, train_iter, loss),
evaluate_loss(net, test_iter, loss)))
print('weight:', net[0].weight.data().asnumpy())
Third-Order Polynomial Function Fitting (Normal)
We will begin by first using a third-order polynomial function, which is the same order as that
of the data generation function. The results show that this modelʼs training and test losses can
be both effectively reduced. The learned model parameters are also close to the true values w =
[5, 1.2, −3.4, 5.6].
# Pick the first four dimensions, i.e., 1, x, x^2/2!, x^3/3! from the
# polynomial features
train(poly_features[:n_train, :4], poly_features[n_train:, :4],
labels[:n_train], labels[n_train:])
weight: [[ 4.998482 1.2171441 -3.3890193 5.60067 ]]

Linear Function Fitting (Underfitting)
Let us take another look at linear function fitting. After the decline in early epochs, it becomes
difficult to further decrease this modelʼs training loss. After the last epoch iteration has been
completed, the training loss is still high. When used to fit nonlinear patterns (like the third-order
polynomial function here) linear models are liable to underfit.
# Pick the first two dimensions, i.e., 1, x, from the polynomial features
train(poly_features[:n_train, :2], poly_features[n_train:, :2],
labels[:n_train], labels[n_train:])
weight: [[2.6284926 4.6981072]]

Higher-Order Polynomial Function Fitting (Overfitting)
Now let us try to train the model using a polynomial of too high degree. Here, there are insufficient
data to learn that the higher-degree coefficients should have values close to zero. As a result, our
overly-complex model is so susceptible that it is being influenced by noise in the training data.
Though the training loss can be effectively reduced, the test loss is still much higher. It shows that
the complex model overfits the data.
# Pick all the dimensions from the polynomial features

train(poly_features[:n_train, :], poly_features[n_train:, :],
labels[:n_train], labels[n_train:], num_epochs=1500)
weight: [[ 4.974571 1.3419539 -3.3172207 5.0192237 -0.15747286 1.4558134

0.17068975 0.21751356 0.06558682 -0.01592065 0.00982275 -0.05115314
-0.02413951 -0.01499531 -0.04940702 0.06389923 -0.04761839 -0.04380166
-0.05188227 0.05655775]]
In the subsequent sections, we will continue to discuss overfitting problems and methods for deal-
ing with them, such as weight decay and dropout.
Summary
• Since the generalization error cannot be estimated based on the training error, simply min-
imizing the training error will not necessarily mean a reduction in the generalization error.
Machine learning models need to be careful to safeguard against overfitting so as to mini-
mize the generalization error.
• A validation set can be used for model selection, provided that it is not used too liberally.
• Underfitting means that a model is not able to reduce the training error. When training error
is much lower than validation error, there is overfitting.
• We should choose an appropriately complex model and avoid using insufficient training
samples.

Exercises
1. Can you solve the polynomial regression problem exactly? Hint: use linear algebra.
2. Consider model selection for polynomials:
• Plot the training loss vs. model complexity (degree of the polynomial). What do you
observe? What degree of polynomial do you need to reduce the training loss to 0?
• Plot the test loss in this case.
• Generate the same plot as a function of the amount of data.
3. What happens if you drop the normalization (1/i!) of the polynomial features xi ? Can you
fix this in some other way?
4. Can you ever expect to see zero generalization error?
Discussions69
4.5 Weight Decay
Now that we have characterized the problem of overfitting, we can introduce some standard tech-
niques for regularizing models. Recall that we can always mitigate overfitting by going out and
collecting more training data. That can be costly, time consuming, or entirely out of our control,
making it impossible in the short run. For now, we can assume that we already have as much
high-quality data as our resources permit and focus on regularization techniques.
Recall that in our polynomial regression example (Section 4.4) we could limit our modelʼs capacity
simply by tweaking the degree of the fitted polynomial. Indeed, limiting the number of features
is a popular technique to mitigate overfitting. However, simply tossing aside features can be too
blunt an instrument for the job. Sticking with the polynomial regression example, consider what
might happen with high-dimensional inputs. The natural extensions of polynomials to multivari-
ate data are called monomials, which are simply products of powers of variables. The degree of a
monomial is the sum of the powers. For example, x21 x2 , and x3 x25 are both monomials of degree
3.
Note that the number of terms with degree d blows up rapidly as d(grows)larger. Given k variables,
the number of monomials of degree d (i.e., k multichoose d) is k−1+dk−1 . Even small changes in
degree, say from 2 to 3, dramatically increase the complexity of our model. Thus we often need a
more fine-grained tool for adjusting function complexity.
4.5.1 Norms and Weight Decay
We have described both the L2 norm and the L1 norm, which are special cases of the more general
Lp norm in Section 2.3.10. Weight decay (commonly called L2 regularization), might be the most
widely-used technique for regularizing parametric machine learning models. The technique is
motivated by the basic intuition that among all functions f , the function f = 0 (assigning the
value 0 to all inputs) is in some sense the simplest, and that we can measure the complexity of a
function by its distance from zero. But how precisely should we measure the distance between
a function and zero? There is no single right answer. In fact, entire branches of mathematics,
69

including parts of functional analysis and the theory of Banach spaces, are devoted to answering
this issue.
One simple interpretation might be to measure the complexity of a linear function f (x) = w⊤ x
by some norm of its weight vector, e.g., ∥w∥2 . The most common method for ensuring a small
weight vector is to add its norm as a penalty term to the problem of minimizing the loss. Thus we
replace our original objective, minimizing the prediction loss on the training labels, with new objec-
tive, minimizing the sum of the prediction loss and the penalty term. Now, if our weight vector grows
too large, our learning algorithm might focus on minimizing the weight norm ∥w∥2 vs. minimiz-
ing the training error. That is exactly what we want. To illustrate things in code, let us revive our
previous example from Section 3.1 for linear regression. There, our loss was given by
1 ∑ 1 ( ⊤ (i) )2
n
L(w, b) = w x + b − y (i) . (4.5.1)
n 2
i=1
Recall that x(i) are the features, y (i) are labels for all data examples i, and (w, b) are the weight
and bias parameters, respectively. To penalize the size of the weight vector, we must somehow
add ∥w∥2 to the loss function, but how should the model trade off the standard loss for this new
additive penalty? In practice, we characterize this tradeoff via the regularization constant λ, a non-
negative hyperparameter that we fit using validation data:
λ
L(w, b) + ∥w∥2 , (4.5.2)
2
For λ = 0, we recover our original loss function. For λ > 0, we restrict the size of ∥w∥. We divide
by 2 by convention: when we take the derivative of a quadratic function, the 2 and 1/2 cancel out,
ensuring that the expression for the update looks nice and simple. The astute reader might wonder
why we work with the squared norm and not the standard norm (i.e., the Euclidean distance). We
do this for computational convenience. By squaring the L2 norm, we remove the square root,
leaving the sum of squares of each component of the weight vector. This makes the derivative of
the penalty easy to compute: the sum of derivatives equals the derivative of the sum.
Moreover, you might ask why we work with the L2 norm in the first place and not, say, the L1
norm. In fact, other choices are valid and popular throughout statistics. While L2 -regularized
linear models constitute the classic ridge regression algorithm, L1 -regularized linear regression is
a similarly fundamental model in statistics, which is popularly known as lasso regression.
One reason to work with the L2 norm is that it places an outsize penalty on large components of the
weight vector. This biases our learning algorithm towards models that distribute weight evenly
across a larger number of features. In practice, this might make them more robust to measure-
ment error in a single variable. By contrast, L1 penalties lead to models that concentrate weights
on a small set of features by clearing the other weights to zero. This is called feature selection,
which may be desirable for other reasons.
Using the same notation in (3.1.10), the minibatch stochastic gradient descent updates for L2 -
regularized regression follow:
η ∑ (i) ( ⊤ (i) )
w ← (1 − ηλ) w − x w x + b − y (i) . (4.5.3)
|B|
i∈B
As before, we update w based on the amount by which our estimate differs from the observation.
However, we also shrink the size of w towards zero. That is why the method is sometimes called
“weight decay”: given the penalty term alone, our optimization algorithm decays the weight at each
step of training. In contrast to feature selection, weight decay offers us a continuous mechanism
4.5. Weight Decay 155

for adjusting the complexity of a function. Smaller values of λ correspond to less constrained w,
whereas larger values of λ constrain w more considerably.
Whether we include a corresponding bias penalty b2 can vary across implementations, and may
vary across layers of a neural network. Often, we do not regularize the bias term of a networkʼs
output layer.
4.5.2 High-Dimensional Linear Regression
We can illustrate the benefits of weight decay through a simple synthetic example.
%matplotlib inline
from mxnet import autograd, gluon, init, np, npx
npx.set_np()
First, we generate some data as before
∑
d
y = 0.05 + 0.01xi + ϵ where ϵ ∼ N (0, 0.012 ). (4.5.4)
i=1
We choose our label to be a linear function of our inputs, corrupted by Gaussian noise with zero
mean and standard deviation 0.01. To make the effects of overfitting pronounced, we can increase
the dimensionality of our problem to d = 200 and work with a small training set containing only
20 examples.
n_train, n_test, num_inputs, batch_size = 20, 100, 200, 5

true_w, true_b = np.ones((num_inputs, 1)) * 0.01, 0.05
train_data = d2l.synthetic_data(true_w, true_b, n_train)
train_iter = d2l.load_array(train_data, batch_size)
test_data = d2l.synthetic_data(true_w, true_b, n_test)
test_iter = d2l.load_array(test_data, batch_size, is_train=False)
4.5.3 Implementation from Scratch
In the following, we will implement weight decay from scratch, simply by adding the squared L2
penalty to the original target function.
Initializing Model Parameters
First, we will define a function to randomly initialize our model parameters.
def init_params():
w = np.random.normal(scale=1, size=(num_inputs, 1))
b = np.zeros(1)
w.attach_grad()
b.attach_grad()
return [w, b]

Defining L2 Norm Penalty
Perhaps the most convenient way to implement this penalty is to square all terms in place and
sum them up.
def l2_penalty(w):
return (w**2).sum() / 2
Defining the Training Loop
The following code fits a model on the training set and evaluates it on the test set. The linear
network and the squared loss have not changed since Chapter 3, so we will just import them via
d2l.linreg and d2l.squared_loss. The only change here is that our loss now includes the penalty
term.
def train(lambd):
w, b = init_params()
net, loss = lambda X: d2l.linreg(X, w, b), d2l.squared_loss
animator = d2l.Animator(xlabel='epochs', ylabel='loss', yscale='log',
xlim=[5, num_epochs], legend=['train', 'test'])
# The L2 norm penalty term has been added, and broadcasting
# makes `l2_penalty(w)` a vector whose length is `batch_size`
l = loss(net(X), y) + lambd * l2_penalty(w)
l.backward()
d2l.sgd([w, b], lr, batch_size)
if (epoch + 1) % 5 == 0:
animator.add(epoch + 1, (d2l.evaluate_loss(net, train_iter, loss),
d2l.evaluate_loss(net, test_iter, loss)))
print('L2 norm of w:', np.linalg.norm(w))
Training without Regularization
We now run this code with lambd = 0, disabling weight decay. Note that we overfit badly, decreas-
ing the training error but not the test error—a textook case of overfitting.
train(lambd=0)
L2 norm of w: 13.25939

Using Weight Decay
Below, we run with substantial weight decay. Note that the training error increases but the test
error decreases. This is precisely the effect we expect from regularization.
train(lambd=3)
L2 norm of w: 0.38248405
4.5.4 Concise Implementation
Because weight decay is ubiquitous in neural network optimization, the deep learning framework
makes it especially convenient, integrating weight decay into the optimization algorithm itself for
easy use in combination with any loss function. Moreover, this integration serves a computational
benefit, allowing implementation tricks to add weight decay to the algorithm, without any addi-
tional computational overhead. Since the weight decay portion of the update depends only on the
current value of each parameter, the optimizer must touch each parameter once anyway.
In the following code, we specify the weight decay hyperparameter directly through wd when in-
stantiating our Trainer. By default, Gluon decays both weights and biases simultaneously. Note

that the hyperparameter wd will be multiplied by wd_mult when updating model parameters. Thus,
if we set wd_mult to zero, the bias parameter b will not decay.
def train_concise(wd):
net.initialize(init.Normal(sigma=1))
{'learning_rate': lr, 'wd': wd})
# The bias parameter has not decayed. Bias names generally end with "bias"
net.collect_params('.*bias').setattr('wd_mult', 0)
animator = d2l.Animator(xlabel='epochs', ylabel='loss', yscale='log',
xlim=[5, num_epochs], legend=['train', 'test'])
l = loss(net(X), y)
l.backward()
if (epoch + 1) % 5 == 0:
animator.add(epoch + 1, (d2l.evaluate_loss(net, train_iter, loss),
d2l.evaluate_loss(net, test_iter, loss)))
print('L2 norm of w:', np.linalg.norm(net[0].weight.data()))
The plots look identical to those when we implemented weight decay from scratch. However, they
run appreciably faster and are easier to implement, a benefit that will become more pronounced
for larger problems.
train_concise(0)
L2 norm of w: 15.014069
train_concise(3)

L2 norm of w: 0.3399175
So far, we only touched upon one notion of what constitutes a simple linear function. Moreover,
what constitutes a simple nonlinear function can be an even more complex question. For in-
stance, reproducing kernel Hilbert space (RKHS)70 allows one to apply tools introduced for lin-
ear functions in a nonlinear context. Unfortunately, RKHS-based algorithms tend to scale poorly
to large, high-dimensional data. In this book we will default to the simple heuristic of applying
weight decay on all layers of a deep network.
Summary
• Regularization is a common method for dealing with overfitting. It adds a penalty term to
the loss function on the training set to reduce the complexity of the learned model.
• One particular choice for keeping the model simple is weight decay using an L2 penalty. This
leads to weight decay in the update steps of the learning algorithm.
• The weight decay functionality is provided in optimizers from deep learning frameworks.
• Different sets of parameters can have different update behaviors within the same training
loop.
Exercises
1. Experiment with the value of λ in the estimation problem in this section. Plot training and
test accuracy as a function of λ. What do you observe?
2. Use a validation set to find the optimal value of λ. Is it really the optimal value? Does this
matter?
∑
3. What would the update equations look like if instead of ∥w∥2 we used i |wi | as our penalty
of choice (L1 regularization)?
4. We know that ∥w∥2 = w⊤ w. Can you find a similar equation for matrices (see the Frobenius
norm in Section 2.3.10)?
70
https://en.wikipedia.org/wiki/Reproducing_kernel_Hilbert_space

5. Review the relationship between training error and generalization error. In addition to
weight decay, increased training, and the use of a model of suitable complexity, what other
ways can you think of to deal with overfitting?
6. In Bayesian statistics we use the product of prior and likelihood to arrive at a posterior via
P (w | x) ∝ P (x | w)P (w). How can you identify P (w) with regularization?
Discussions71
4.6 Dropout
In Section 4.5, we introduced the classical approach to regularizing statistical models by penal-
izing the L2 norm of the weights. In probabilistic terms, we could justify this technique by ar-
guing that we have assumed a prior belief that weights take values from a Gaussian distribution
with mean zero. More intuitively, we might argue that we encouraged the model to spread out its
weights among many features rather than depending too much on a small number of potentially
spurious associations.
4.6.1 Overfitting Revisited
Faced with more features than examples, linear models tend to overfit. But given more examples
than features, we can generally count on linear models not to overfit. Unfortunately, the reliability
with which linear models generalize comes at a cost. Naively applied, linear models do not take
into account interactions among features. For every feature, a linear model must assign either a
positive or a negative weight, ignoring context.
In traditional texts, this fundamental tension between generalizability and flexibility is described
as the bias-variance tradeoff. Linear models have high bias: they can only represent a small class
of functions. However, these models have low variance: they give similar results across different
random samples of the data.
Deep neural networks inhabit the opposite end of the bias-variance spectrum. Unlike linear mod-
els, neural networks are not confined to looking at each feature individually. They can learn in-
teractions among groups of features. For example, they might infer that “Nigeria” and “Western
Union” appearing together in an email indicates spam but that separately they do not.
Even when we have far more examples than features, deep neural networks are capable of over-
fitting. In 2017, a group of researchers demonstrated the extreme flexibility of neural networks by
training deep nets on randomly-labeled images. Despite the absence of any true pattern linking
the inputs to the outputs, they found that the neural network optimized by stochastic gradient de-
scent could label every image in the training set perfectly. Consider what this means. If the labels
are assigned uniformly at random and there are 10 classes, then no classifier can do better than
10% accuracy on holdout data. The generalization gap here is a whopping 90%. If our models are
so expressive that they can overfit this badly, then when should we expect them not to overfit?
The mathematical foundations for the puzzling generalization properties of deep networks re-
main open research questions, and we encourage the theoretically-oriented reader to dig deeper
into the topic. For now, we turn to the investigation of practical tools that tend to empirically
improve the generalization of deep nets.
71
4.6. Dropout 161

4.6.2 Robustness through Perturbations
Let us think briefly about what we expect from a good predictive model. We want it to peform well
on unseen data. Classical generalization theory suggests that to close the gap between train and
test performance, we should aim for a simple model. Simplicity can come in the form of a small
number of dimensions. We explored this when discussing the monomial basis functions of linear
models in Section 4.4. Additionally, as we saw when discussing weight decay (L2 regularization)
in Section 4.5, the (inverse) norm of the parameters also represents a useful measure of simplicity.
Another useful notion of simplicity is smoothness, i.e., that the function should not be sensitive to
small changes to its inputs. For instance, when we classify images, we would expect that adding
some random noise to the pixels should be mostly harmless.
In 1995, Christopher Bishop formalized this idea when he proved that training with input noise is
equivalent to Tikhonov regularization (Bishop, 1995). This work drew a clear mathematical con-
nection between the requirement that a function be smooth (and thus simple), and the require-
ment that it be resilient to perturbations in the input.
Then, in 2014, Srivastava et al. (Srivastava et al., 2014) developed a clever idea for how to apply
Bishopʼs idea to the internal layers of a network, too. Namely, they proposed to inject noise into
each layer of the network before calculating the subsequent layer during training. They realized
that when training a deep network with many layers, injecting noise enforces smoothness just on
the input-output mapping.
Their idea, called dropout, involves injecting noise while computing each internal layer during
forward propagation, and it has become a standard technique for training neural networks. The
method is called dropout because we literally drop out some neurons during training. Throughout
training, on each iteration, standard dropout consists of zeroing out some fraction of the nodes in
each layer before calculating the subsequent layer.
To be clear, we are imposing our own narrative with the link to Bishop. The original paper on
dropout offers intuition through a surprising analogy to sexual reproduction. The authors argue
that neural network overfitting is characterized by a state in which each layer relies on a specifc
pattern of activations in the previous layer, calling this condition co-adaptation. Dropout, they
claim, breaks up co-adaptation just as sexual reproduction is argued to break up co-adapted genes.
The key challenge then is how to inject this noise. One idea is to inject the noise in an unbiased
manner so that the expected value of each layer—while fixing the others—equals to the value it
would have taken absent noise.
In Bishopʼs work, he added Gaussian noise to the inputs to a linear model. At each training iter-
ation, he added noise sampled from a distribution with mean zero ϵ ∼ N (0, σ 2 ) to the input x,
yielding a perturbed point x′ = x + ϵ. In expectation, E[x′ ] = x.
In standard dropout regularization, one debiases each layer by normalizing by the fraction of
nodes that were retained (not dropped out). In other words, with dropout probability p, each inter-
mediate activation h is replaced by a random variable h′ as follows:
{
0 with probability p
h′ = h
(4.6.1)
1−p otherwise
By design, the expectation remains unchanged, i.e., E[h′ ] = h.

4.6.3 Dropout in Practice
Recall the MLP with a hidden layer and 5 hidden units in Fig. 4.1.1. When we apply dropout to
a hidden layer, zeroing out each hidden unit with probability p, the result can be viewed as a
network containing only a subset of the original neurons. In Fig. 4.6.1, h2 and h5 are removed.
Consequently, the calculation of the outputs no longer depends on h2 or h5 and their respective
gradient also vanishes when performing backpropagation. In this way, the calculation of the out-
put layer cannot be overly dependent on any one element of h1 , . . . , h5 .
Fig. 4.6.1: MLP before and after dropout.
Typically, we disable dropout at test time. Given a trained model and a new example, we do not
drop out any nodes and thus do not need to normalize. However, there are some exceptions: some
researchers use dropout at test time as a heuristic for estimating the uncertainty of neural network
predictions: if the predictions agree across many different dropout masks, then we might say that
the network is more confident.
To implement the dropout function for a single layer, we must draw as many samples from a
Bernoulli (binary) random variable as our layer has dimensions, where the random variable takes
value 1 (keep) with probability 1 − p and 0 (drop) with probability p. One easy way to implement
this is to first draw samples from the uniform distribution U [0, 1]. Then we can keep those nodes
for which the corresponding sample is greater than p, dropping the rest.
In the following code, we implement a dropout_layer function that drops out the elements in the
tensor input X with probability dropout, rescaling the remainder as described above: dividing the
survivors by 1.0-dropout.

npx.set_np()
def dropout_layer(X, dropout):

assert 0 <= dropout <= 1
# In this case, all elements are dropped out
4.6. Dropout 163

if dropout == 1:
return np.zeros_like(X)
# In this case, all elements are kept
if dropout == 0:
return X
mask = np.random.uniform(0, 1, X.shape) > dropout
return mask.astype(np.float32) * X / (1.0 - dropout)
We can test out the dropout_layer function on a few examples. In the following lines of code, we
pass our input X through the dropout operation, with probabilities 0, 0.5, and 1, respectively.
print(dropout_layer(X, 0))
print(dropout_layer(X, 0.5))
print(dropout_layer(X, 1))
[[ 0. 1. 2. 3. 4. 5. 6. 7.]
[ 8. 9. 10. 11. 12. 13. 14. 15.]]
[[ 0. 2. 4. 6. 8. 10. 12. 14.]
[ 0. 18. 20. 0. 0. 0. 28. 0.]]
[[0. 0. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0.]]
Defining Model Parameters
Again, we work with the Fashion-MNIST dataset introduced in Section 3.5. We define an MLP with
two hidden layers containing 256 units each.
num_inputs, num_outputs, num_hiddens1, num_hiddens2 = 784, 10, 256, 256
W1 = np.random.normal(scale=0.01, size=(num_inputs, num_hiddens1))

b1 = np.zeros(num_hiddens1)
W2 = np.random.normal(scale=0.01, size=(num_hiddens1, num_hiddens2))
b2 = np.zeros(num_hiddens2)
W3 = np.random.normal(scale=0.01, size=(num_hiddens2, num_outputs))
b3 = np.zeros(num_outputs)
params = [W1, b1, W2, b2, W3, b3]

param.attach_grad()

Defining the Model
The model below applies dropout to the output of each hidden layer (following the activation func-
tion). We can set dropout probabilities for each layer separately. A common trend is to set a lower
dropout probability closer to the input layer. Below we set it to 0.2 and 0.5 for the first and second
hidden layers, respectively. We ensure that dropout is only active during training.
dropout1, dropout2 = 0.2, 0.5
def net(X):
X = X.reshape(-1, num_inputs)
H1 = npx.relu(np.dot(X, W1) + b1)
# Use dropout only when training the model
if autograd.is_training():
# Add a dropout layer after the first fully connected layer
H1 = dropout_layer(H1, dropout1)
H2 = npx.relu(np.dot(H1, W2) + b2)
# Add a dropout layer after the second fully connected layer
H2 = dropout_layer(H2, dropout2)
return np.dot(H2, W3) + b3
Training and Testing
This is similar to the training and testing of MLPs described previously.
num_epochs, lr, batch_size = 10, 0.5, 256

d2l.train_ch3(net, train_iter, test_iter, loss, num_epochs,
lambda batch_size: d2l.sgd(params, lr, batch_size))
4.6. Dropout 165

With high-level APIs, all we need to do is add a Dropout layer after each fully-connected layer,
passing in the dropout probability as the only argument to its constructor. During training, the
Dropout layer will randomly drop out outputs of the previous layer (or equivalently, the inputs to
the subsequent layer) according to the specified dropout probability. When not in training mode,
the Dropout layer simply passes the data through during testing.
net.add(nn.Dense(256, activation="relu"),
# Add a dropout layer after the first fully connected layer
nn.Dropout(dropout1),
nn.Dense(256, activation="relu"),
# Add a dropout layer after the second fully connected layer
nn.Dropout(dropout2),
nn.Dense(10))
Next, we train and test the model.

Summary
• Beyond controlling the number of dimensions and the size of the weight vector, dropout is
yet another tool to avoid overfitting. Often they are used jointly.
• Dropout replaces an activation h with a random variable with expected value h.
• Dropout is only used during training.

Exercises
1. What happens if you change the dropout probabilities for the first and second layers? In
particular, what happens if you switch the ones for both layers? Design an experiment to
answer these questions, describe your results quantitatively, and summarize the qualitative
takeaways.
2. Increase the number of epochs and compare the results obtained when using dropout with
those when not using it.
3. What is the variance of the activations in each hidden layer when dropout is and is not ap-
plied? Draw a plot to show how this quantity evolves over time for both models.
4. Why is dropout not typically used at test time?
5. Using the model in this section as an example, compare the effects of using dropout and
weight decay. What happens when dropout and weight decay are used at the same time?
Are the results additive? Are there diminished returns (or worse)? Do they cancel each other
out?
6. What happens if we apply dropout to the individual weights of the weight matrix rather than
the activations?
7. Invent another technique for injecting random noise at each layer that is different from the
standard dropout technique. Can you develop a method that outperforms dropout on the
Fashion-MNIST dataset (for a fixed architecture)?
Discussions72
4.7 Forward Propagation, Backward Propagation, and Computational

Graphs
So far, we have trained our models with minibatch stochastic gradient descent. However, when
we implemented the algorithm, we only worried about the calculations involved in forward prop-
agation through the model. When it came time to calculate the gradients, we just invoked the
backpropagation function provided by the deep learning framework.
The automatic calculation of gradients (automatic differentiation) profoundly simplifies the im-
plementation of deep learning algorithms. Before automatic differentiation, even small changes
to complicated models required recalculating complicated derivatives by hand. Surprisingly of-
ten, academic papers had to allocate numerous pages to deriving update rules. While we must
continue to rely on automatic differentiation so we can focus on the interesting parts, you ought
to know how these gradients are calculated under the hood if you want to go beyond a shallow
understanding of deep learning.
In this section, we take a deep dive into the details of backward propagation (more commonly called
backpropagation). To convey some insight for both the techniques and their implementations, we
rely on some basic mathematics and computational graphs. To start, we focus our exposition on
a one-hidden-layer MLP with weight decay (L2 regularization).
72
4.7. Forward Propagation, Backward Propagation, and Computational Graphs 167

4.7.1 Forward Propagation
Forward propagation (or forward pass) refers to the calculation and storage of intermediate variables
(including outputs) for a neural network in order from the input layer to the output layer. We now
work step-by-step through the mechanics of a neural network with one hidden layer. This may
seem tedious but in the eternal words of funk virtuoso James Brown, you must “pay the cost to be
the boss”.
For the sake of simplicity, let us assume that the input example is x ∈ Rd and that our hidden layer
does not include a bias term. Here the intermediate variable is:
z = W(1) x, (4.7.1)
where W(1) ∈ Rh×d is the weight parameter of the hidden layer. After running the intermediate
variable z ∈ Rh through the activation function ϕ we obtain our hidden activation vector of length
h,
h = ϕ(z). (4.7.2)
The hidden variable h is also an intermediate variable. Assuming that the parameters of the output
layer only possess a weight of W(2) ∈ Rq×h , we can obtain an output layer variable with a vector
of length q:
o = W(2) h. (4.7.3)
Assuming that the loss function is l and the example label is y, we can then calculate the loss term
for a single data example,
L = l(o, y). (4.7.4)
According to the definition of L2 regularization, given the hyperparameter λ, the regularization

term is
λ ( (1) 2 )
s= ∥W ∥F + ∥W(2) ∥2F , (4.7.5)
2
where the Frobenius norm of the matrix is simply the L2 norm applied after flattening the matrix
into a vector. Finally, the modelʼs regularized loss on a given data example is:
J = L + s. (4.7.6)
We refer to J as the objective function in the following discussion.
4.7.2 Computational Graph of Forward Propagation
Plotting computational graphs helps us visualize the dependencies of operators and variables
within the calculation. Fig. 4.7.1 contains the graph associated with the simple network described
above, where squares denote variables and circles denote operators. The lower-left corner signi-
fies the input and the upper-right corner is the output. Notice that the directions of the arrows
(which illustrate data flow) are primarily rightward and upward.

Fig. 4.7.1: Computational graph of forward propagation.
4.7.3 Backpropagation
Backpropagation refers to the method of calculating the gradient of neural network parameters.
In short, the method traverses the network in reverse order, from the output to the input layer,
according to the chain rule from calculus. The algorithm stores any intermediate variables (partial
derivatives) required while calculating the gradient with respect to some parameters. Assume that
we have functions Y = f (X) and Z = g(Y), in which the input and the output X, Y, Z are tensors of
arbitrary shapes. By using the chain rule, we can compute the derivative of Z with respect to X via
( )
∂Z ∂Z ∂Y
= prod , . (4.7.7)
∂X ∂Y ∂X
Here we use the prod operator to multiply its arguments after the necessary operations, such as
transposition and swapping input positions, have been carried out. For vectors, this is straight-
forward: it is simply matrix-matrix multiplication. For higher dimensional tensors, we use the
appropriate counterpart. The operator prod hides all the notation overhead.
Recall that the parameters of the simple network with one hidden layer, whose computational
graph is in Fig. 4.7.1, are W(1) and W(2) . The objective of backpropagation is to calculate the gra-
dients ∂J/∂W(1) and ∂J/∂W(2) . To accomplish this, we apply the chain rule and calculate, in turn,
the gradient of each intermediate variable and parameter. The order of calculations are reversed
relative to those performed in forward propagation, since we need to start with the outcome of the
computational graph and work our way towards the parameters. The first step is to calculate the
gradients of the objective function J = L + s with respect to the loss term L and the regularization
term s.
∂J ∂J
= 1 and = 1. (4.7.8)
∂L ∂s
Next, we compute the gradient of the objective function with respect to variable of the output layer
o according to the chain rule:
( )
∂J ∂J ∂L ∂L
= prod , = ∈ Rq . (4.7.9)
∂o ∂L ∂o ∂o
Next, we calculate the gradients of the regularization term with respect to both parameters:
∂s ∂s
(1)
= λW(1) and = λW(2) . (4.7.10)
∂W ∂W(2)
Now we are able to calculate the gradient ∂J/∂W(2) ∈ Rq×h of the model parameters closest to
the output layer. Using the chain rule yields:
( ) ( )
∂J ∂J ∂o ∂J ∂s ∂J ⊤
(2)
= prod , (2)
+ prod , (2)
= h + λW(2) . (4.7.11)
∂W ∂o ∂W ∂s ∂W ∂o
4.7. Forward Propagation, Backward Propagation, and Computational Graphs 169

To obtain the gradient with respect to W(1) we need to continue backpropagation along the output
layer to the hidden layer. The gradient with respect to the hidden layerʼs outputs ∂J/∂h ∈ Rh is
given by
( )
∂J ∂J ∂o ⊤ ∂J
= prod , = W(2) . (4.7.12)
∂h ∂o ∂h ∂o
Since the activation function ϕ applies elementwise, calculating the gradient ∂J/∂z ∈ Rh of the
intermediate variable z requires that we use the elementwise multiplication operator, which we
denote by ⊙:
( )
∂J ∂J ∂h ∂J
= prod , = ⊙ ϕ′ (z) . (4.7.13)
∂z ∂h ∂z ∂h
Finally, we can obtain the gradient ∂J/∂W(1) ∈ Rh×d of the model parameters closest to the input
layer. According to the chain rule, we get
( ) ( )
∂J ∂J ∂z ∂J ∂s ∂J ⊤
(1)
= prod , (1)
+ prod , (1)
= x + λW(1) . (4.7.14)
∂W ∂z ∂W ∂s ∂W ∂z
4.7.4 Training Neural Networks
When training neural networks, forward and backward propagation depend on each other. In
particular, for forward propagation, we traverse the computational graph in the direction of de-
pendencies and compute all the variables on its path. These are then used for backpropagation
where the compute order on the graph is reversed.
Take the aforementioned simple network as an example to illustrate. On one hand, comput-
ing the regularization term (4.7.5) during forward propagation depends on the current values of
model parameters W(1) and W(2) . They are given by the optimization algorithm according to back-
propagation in the latest iteration. On the other hand, the gradient calculation for the parameter
eq_backprop-J-h during backpropagation depends on the current value of the hidden variable h,
which is given by forward propagation.
Therefore when training neural networks, after model parameters are initialized, we alternate
forward propagation with backpropagation, updating model parameters using gradients given by
backpropagation. Note that backpropagation reuses the stored intermediate values from forward
propagation to avoid duplicate calculations. One of the consequences is that we need to retain
the intermediate values until backpropagation is complete. This is also one of the reasons why
training requires significantly more memory than plain prediction. Besides, the size of such in-
termediate values is roughly proportional to the number of network layers and the batch size.
Thus, training deeper networks using larger batch sizes more easily leads to out of memory errors.
Summary
• Forward propagation sequentially calculates and stores intermediate variables within the
computational graph defined by the neural network. It proceeds from the input to the output
layer.
• Backpropagation sequentially calculates and stores the gradients of intermediate variables
and parameters within the neural network in the reversed order.

• When training deep learning models, forward propagation and back propagation are inter-
dependent.
• Training requires significantly more memory than prediction.
Exercises
1. Assume that the inputs X to some scalar function f are n × m matrices. What is the dimen-
sionality of the gradient of f with respect to X?
2. Add a bias to the hidden layer of the model described in this section (you do not need to
include bias in the regularization term).
1. Draw the corresponding computational graph.
2. Derive the forward and backward propagation equations.
3. Compute the memory footprint for training and prediction in the model described in this
section.
4. Assume that you want to compute second derivatives. What happens to the computational
graph? How long do you expect the calculation to take?
5. Assume that the computational graph is too large for your GPU.
1. Can you partition it over more than one GPU?
2. What are the advantages and disadvantages over training on a smaller minibatch?
Discussions73
4.8 Numerical Stability and Initialization
Thus far, every model that we have implemented required that we initialize its parameters accord-
ing to some pre-specified distribution. Until now, we took the initialization scheme for granted,
glossing over the details of how these choices are made. You might have even gotten the impres-
sion that these choices are not especially important. To the contrary, the choice of initialization
scheme plays a significant role in neural network learning, and it can be crucial for maintaining
numerical stability. Moreover, these choices can be tied up in interesting ways with the choice of
the nonlinear activation function. Which function we choose and how we initialize parameters
can determine how quickly our optimization algorithm converges. Poor choices here can cause us
to encounter exploding or vanishing gradients while training. In this section, we delve into these
topics with greater detail and discuss some useful heuristics that you will find useful throughout
your career in deep learning.
73
4.8. Numerical Stability and Initialization 171

4.8.1 Vanishing and Exploding Gradients
Consider a deep network with L layers, input x and output o. With each layer l defined by a trans-
formation fl parameterized by weights W(l) , whose hidden variable is h(l) (let h(0) = x), our net-
work can be expressed as:
h(l) = fl (h(l−1) ) and thus o = fL ◦ . . . ◦ f1 (x). (4.8.1)
If all the hidden variables and the input are vectors, we can write the gradient of o with respect to
any set of parameters W(l) as follows:
∂W(l) o = ∂h(L−1) h(L) · . . . · ∂h(l) h(l+1) ∂W(l) h(l) .

| {z } | {z } | {z } (4.8.2)
def def def
M(L) = M(l+1) = v(l) =
In other words, this gradient is the product of L − l matrices M(L) · . . . · M(l+1) and the gradient vec-
tor v(l) . Thus we are susceptible to the same problems of numerical underflow that often crop up
when multiplying together too many probabilities. When dealing with probabilities, a common
trick is to switch into log-space, i.e., shifting pressure from the mantissa to the exponent of the
numerical representation. Unfortunately, our problem above is more serious: initially the matri-
ces M(l) may have a wide variety of eigenvalues. They might be small or large, and their product
might be very large or very small.
The risks posed by unstable gradients go beyond numerical representation. Gradients of unpre-
dictable magnitude also threaten the stability of our optimization algorithms. We may be fac-
ing parameter updates that are either (i) excessively large, destroying our model (the exploding
gradient problem); or (ii) excessively small (the vanishing gradient problem), rendering learning
impossible as parameters hardly move on each update.
Vanishing Gradients
One frequent culprit causing the vanishing gradient problem is the choice of the activation func-
tion σ that is appended following each layerʼs linear operations. Historically, the sigmoid function
1/(1+exp(−x)) (introduced in :numref:sec_mlp) was popular because it resembles a thresholding
function. Since early artificial neural networks were inspired by biological neural networks, the
idea of neurons that fire either fully or not at all (like biological neurons) seemed appealing. Let
us take a closer look at the sigmoid to see why it can cause vanishing gradients.
%matplotlib inline
npx.set_np()
x = np.arange(-8.0, 8.0, 0.1)

x.attach_grad()
y = npx.sigmoid(x)
y.backward()
d2l.plot(x, [y, x.grad], legend=['sigmoid', 'gradient'], figsize=(4.5, 2.5))

As you can see, the sigmoidʼs gradient vanishes both when its inputs are large and when they are
small. Moreover, when backpropagating through many layers, unless we are in the Goldilocks
zone, where the inputs to many of the sigmoids are close to zero, the gradients of the overall
product may vanish. When our network boasts many layers, unless we are careful, the gradient
will likely be cut off at some layer. Indeed, this problem used to plague deep network training.
Consequently, ReLUs, which are more stable (but less neurally plausible), have emerged as the
default choice for practitioners.
Exploding Gradients
The opposite problem, when gradients explode, can be similarly vexing. To illustrate this a bit
better, we draw 100 Gaussian random matrices and multiply them with some initial matrix. For
the scale that we picked (the choice of the variance σ 2 = 1), the matrix product explodes. When
this happens due to the initialization of a deep network, we have no chance of getting a gradient
descent optimizer to converge.
M = np.random.normal(size=(4, 4))
print('a single matrix', M)
for i in range(100):
M = np.dot(M, np.random.normal(size=(4, 4)))
print('after multiplying 100 matrices', M)
a single matrix [[ 2.2122064 1.1630787 0.7740038 0.4838046 ]

[ 1.0434405 0.29956347 1.1839255 0.15302546]
[ 1.8917114 -1.1688148 -1.2347414 1.5580711 ]
[-1.771029 -0.5459446 -0.45138445 -2.3556297 ]]
after multiplying 100 matrices [[ 3.4459714e+23 -7.8040680e+23 5.9973287e+23 4.5229990e+23]
[ 2.5275089e+23 -5.7240326e+23 4.3988473e+23 3.3174740e+23]
[ 1.3731286e+24 -3.1097155e+24 2.3897773e+24 1.8022959e+24]
[-4.4951040e+23 1.0180033e+24 -7.8232281e+23 -5.9000354e+23]]

Breaking the Symmetry
Another problem in neural network design is the symmetry inherent in their parametrization.
Assume that we have a simple MLP with one hidden layer and two units. In this case, we could
permute the weights W(1) of the first layer and likewise permute the weights of the output layer
to obtain the same function. There is nothing special differentiating the first hidden unit vs. the
second hidden unit. In other words, we have permutation symmetry among the hidden units of
each layer.
This is more than just a theoretical nuisance. Consider the aforementioned one-hidden-layer MLP
with two hidden units. For illustration, suppose that the output layer transforms the two hidden
units into only one output unit. Imagine what would happen if we initialized all of the parameters
of the hidden layer as W(1) = c for some constant c. In this case, during forward propagation
either hidden unit takes the same inputs and parameters, producing the same activation, which
is fed to the output unit. During backpropagation, differentiating the output unit with respect to
parameters W(1) gives a gradient whose elements all take the same value. Thus, after gradient-
based iteration (e.g., minibatch stochastic gradient descent), all the elements of W(1) still take the
same value. Such iterations would never break the symmetry on its own and we might never be
able to realize the networkʼs expressive power. The hidden layer would behave as if it had only a
single unit. Note that while minibatch stochastic gradient descent would not break this symmetry,
dropout regularization would!
4.8.2 Parameter Initialization
One way of addressing—or at least mitigating—the issues raised above is through careful initial-
ization. Additional care during optimization and suitable regularization can further enhance sta-
bility.
Default Initialization
In the previous sections, e.g., in Section 3.3, we used a normal distribution to initialize the values
of our weights. If we do not specify the initialization method, the framework will use a default
random initialization method, which often works well in practice for moderate problem sizes.
Xavier Initialization
Let us look at the scale distribution of an output (e.g., a hidden variable) oi for some fully-
connected layer without nonlinearities. With nin inputs xj and their associated weights wij for this
layer, an output is given by
∑
nin
oi = wij xj . (4.8.3)
j=1
The weights wij are all drawn independently from the same distribution. Furthermore, let us
assume that this distribution has zero mean and variance σ 2 . Note that this does not mean that
the distribution has to be Gaussian, just that the mean and variance need to exist. For now, let
us assume that the inputs to the layer xj also have zero mean and variance γ 2 and that they are

independent of wij and independent of each other. In this case, we can compute the mean and
variance of oi as follows:
∑
nin
E[oi ] = E[wij xj ]
j=1
∑
nin
= E[wij ]E[xj ]
j=1
= 0,
Var[oi ] = E[o2i ] − (E[oi ])2 (4.8.4)
∑
nin
= 2 2
E[wij xj ] − 0
j=1
∑
nin
2
= E[wij ]E[x2j ]
j=1
= nin σ 2 γ 2 .
One way to keep the variance fixed is to set nin σ 2 = 1. Now consider backpropagation. There
we face a similar problem, albeit with gradients being propagated from the layers closer to the
output. Using the same reasoning as for forward propagation, we see that the gradientsʼ variance
can blow up unless nout σ 2 = 1, where nout is the number of outputs of this layer. This leaves us in
a dilemma: we cannot possibly satisfy both conditions simultaneously. Instead, we simply try to
satisfy:
√
1 2 2
(nin + nout )σ = 1 or equivalently σ = . (4.8.5)
2 nin + nout
This is the reasoning underlying the now-standard and practically beneficial Xavier initialization,
named after the first author of its creators (Glorot & Bengio, 2010). Typically, the Xavier initializa-
2
tion samples weights from a Gaussian distribution with zero mean and variance σ 2 = nin +n out
. We
can also adapt Xavierʼs intuition to choose the variance when sampling weights from a uniform
2 2
distribution. Note that the uniform distribution U (−a, a) has variance a3 . Plugging a3 into our
condition on σ 2 yields the suggestion to initialize according to
( √ √ )
6 6
U − , . (4.8.6)
nin + nout nin + nout
Though the assumption for nonexistence of nonlinearities in the above mathematical reasoning
can be easily violated in neural networks, the Xavier initialization method turns out to work well
in practice.
Beyond
The reasoning above barely scratches the surface of modern approaches to parameter initializa-
tion. A deep learning framework often implements over a dozen different heuristics. Moreover,
parameter initialization continues to be a hot area of fundamental research in deep learning.
Among these are heuristics specialized for tied (shared) parameters, super-resolution, sequence
models, and other situations. For instance, Xiao et al. demonstrated the possibility of training
10000-layer neural networks without architectural tricks by using a carefully-designed initializa-
tion method (Xiao et al., 2018).

If the topic interests you we suggest a deep dive into this moduleʼs offerings, reading the papers
that proposed and analyzed each heuristic, and then exploring the latest publications on the topic.
Perhaps you will stumble across or even invent a clever idea and contribute an implementation to
deep learning frameworks.
Summary
• Vanishing and exploding gradients are common issues in deep networks. Great care in pa-
rameter initialization is required to ensure that gradients and parameters remain well con-
trolled.
• Initialization heuristics are needed to ensure that the initial gradients are neither too large
nor too small.
• ReLU activation functions mitigate the vanishing gradient problem. This can accelerate con-
vergence.
• Random initialization is key to ensure that symmetry is broken before optimization.
• Xavier initialization suggests that, for each layer, variance of any output is not affected by
the number of inputs, and variance of any gradient is not affected by the number of outputs.
Exercises
1. Can you design other cases where a neural network might exhibit symmetry requiring break-
ing besides the permutation symmetry in an MLPʼs layers?
2. Can we initialize all weight parameters in linear regression or in softmax regression to the
same value?
3. Look up analytic bounds on the eigenvalues of the product of two matrices. What does this
tell you about ensuring that gradients are well conditioned?
4. If we know that some terms diverge, can we fix this after the fact? Look at the paper on
layer-wise adaptive rate scaling for inspiration (You et al., 2017).
Discussions74
4.9 Environment and Distribution Shift
In the previous sections, we worked through a number of hands-on applications of machine learn-
ing, fitting models to a variety of datasets. And yet, we never stopped to contemplate either where
data come from in the first place or what we plan to ultimately do with the outputs from our mod-
els. Too often, machine learning developers in possession of data rush to develop models without
pausing to consider these fundamental issues.
Many failed machine learning deployments can be traced back to this pattern. Sometimes mod-
els appear to perform marvelously as measured by test set accuracy but fail catastrophically in
deployment when the distribution of data suddenly shifts. More insidiously, sometimes the very
deployment of a model can be the catalyst that perturbs the data distribution. Say, for example,
that we trained a model to predict who will repay vs. default on a loan, finding that an applicantʼs
74

choice of footwear was associated with the risk of default (Oxfords indicate repayment, sneakers
indicate default). We might be inclined to thereafter grant loans to all applicants wearing Oxfords
and to deny all applicants wearing sneakers.
In this case, our ill-considered leap from pattern recognition to decision-making and our failure
to critically consider the environment might have disastrous consequences. For starters, as soon
as we began making decisions based on footwear, customers would catch on and change their
behavior. Before long, all applicants would be wearing Oxfords, without any coinciding improve-
ment in credit-worthiness. Take a minute to digest this because similar issues abound in many
applications of machine learning: by introducing our model-based decisions to the environment,
we might break the model.
While we cannot possibly give these topics a complete treatment in one section, we aim here to
expose some common concerns, and to stimulate the critical thinking required to detect these
situations early, mitigate damage, and use machine learning responsibly. Some of the solutions
are simple (ask for the “right” data), some are technically difficult (implement a reinforcement
learning system), and others require that we step outside the realm of statistical prediction al-
together and grapple with difficult philosophical questions concerning the ethical application of
algorithms.
4.9.1 Types of Distribution Shift
To begin, we stick with the passive prediction setting considering the various ways that data distri-
butions might shift and what might be done to salvage model performance. In one classic setup,
we assume that our training data were sampled from some distribution pS (x, y) but that our test
data will consist of unlabeled examples drawn from some different distribution pT (x, y). Already,
we must confront a sobering reality. Absent any assumptions on how pS and pT relate to each
other, learning a robust classifier is impossible.
Consider a binary classification problem, where we wish to distinguish between dogs and cats.
If the distribution can shift in arbitrary ways, then our setup permits the pathological case in
which the distribution over inputs remains constant: pS (x) = pT (x), but the labels are all flipped:
pS (y|x) = 1 − pT (y|x). In other words, if God can suddenly decide that in the future all “cats”
are now dogs and what we previously called “dogs” are now cats—without any change in the dis-
tribution of inputs p(x), then we cannot possibly distinguish this setting from one in which the
distribution did not change at all.
Fortunately, under some restricted assumptions on the ways our data might change in the future,
principled algorithms can detect shift and sometimes even adapt on the fly, improving on the
accuracy of the original classifier.
Covariate Shift
Among categories of distribution shift, covariate shift may be the most widely studied. Here, we
assume that while the distribution of inputs may change over time, the labeling function, i.e., the
conditional distribution P (y | x) does not change. Statisticians call this covariate shift because
the problem arises due to a shift in the distribution of the covariates (features). While we can
sometimes reason about distribution shift without invoking causality, we note that covariate shift
is the natural assumption to invoke in settings where we believe that x causes y.
Consider the challenge of distinguishing cats and dogs. Our training data might consist of images
of the kind in Fig. 4.9.1.
4.9. Environment and Distribution Shift 177

Fig. 4.9.1: Training data for distinguishing cats and dogs.
At test time we are asked to classify the images in Fig. 4.9.2.
Fig. 4.9.2: Test data for distinguishing cats and dogs.
The training set consists of photos, while the test set contains only cartoons. Training on a dataset
with substantially different characteristics from the test set can spell trouble absent a coherent
plan for how to adapt to the new domain.
Label Shift
Label shift describes the converse problem. Here, we assume that the label marginal P (y) can
change but the class-conditional distribution P (x | y) remains fixed across domains. Label shift is
a reasonable assumption to make when we believe that y causes x. For example, we may want to
predict diagnoses given their symptoms (or other manifestations), even as the relative prevalence
of diagnoses are changing over time. Label shift is the appropriate assumption here because dis-
eases cause symptoms. In some degenerate cases the label shift and covariate shift assumptions
can hold simultaneously. For example, when the label is deterministic, the covariate shift assump-
tion will be satisfied, even when y causes x. Interestingly, in these cases, it is often advantageous
to work with methods that flow from the label shift assumption. That is because these methods
tend to involve manipulating objects that look like labels (often low-dimensional), as opposed to
objects that look like inputs, which tend to be high-dimensional in deep learning.

Concept Shift
We may also encounter the related problem of concept shift, which arises when the very definitions
of labels can change. This sounds weird—a cat is a cat, no? However, other categories are subject to
changes in usage over time. Diagnostic criteria for mental illness, what passes for fashionable, and
job titles, are all subject to considerable amounts of concept shift. It turns out that if we navigate
around the United States, shifting the source of our data by geography, we will find considerable
concept shift regarding the distribution of names for soft drinks as shown in Fig. 4.9.3.
Fig. 4.9.3: Concept shift on soft drink names in the United States.
If we were to build a machine translation system, the distribution P (y | x) might be different de-
pending on our location. This problem can be tricky to spot. We might hope to exploit knowledge
that shift only takes place gradually either in a temporal or geographic sense.
4.9.2 Examples of Distribution Shift
Before delving into formalism and algorithms, we can discuss some concrete situations where
covariate or concept shift might not be obvious.
Medical Diagnostics
Imagine that you want to design an algorithm to detect cancer. You collect data from healthy and
sick people and you train your algorithm. It works fine, giving you high accuracy and you conclude
that you are ready for a successful career in medical diagnostics. Not so fast.
The distributions that gave rise to the training data and those you will encounter in the wild might
differ considerably. This happened to an unfortunate startup that some of us (authors) worked
with years ago. They were developing a blood test for a disease that predominantly affects older
men and hoped to study it using blood samples that they had collected from patients. However, it
is considerably more difficult to obtain blood samples from healthy men than sick patients already
in the system. To compensate, the startup solicited blood donations from students on a university
campus to serve as healthy controls in developing their test. Then they asked whether we could
help them to build a classifier for detecting the disease.

As we explained to them, it would indeed be easy to distinguish between the healthy and sick
cohorts with near-perfect accuracy. However, that is because the test subjects differed in age,
hormone levels, physical activity, diet, alcohol consumption, and many more factors unrelated
to the disease. This was unlikely to be the case with real patients. Due to their sampling proce-
dure, we could expect to encounter extreme covariate shift. Moreover, this case was unlikely to
be correctable via conventional methods. In short, they wasted a significant sum of money.
Self-Driving Cars
Say a company wanted to leverage machine learning for developing self-driving cars. One key
component here is a roadside detector. Since real annotated data are expensive to get, they had
the (smart and questionable) idea to use synthetic data from a game rendering engine as additional
training data. This worked really well on “test data” drawn from the rendering engine. Alas, inside
a real car it was a disaster. As it turned out, the roadside had been rendered with a very simplis-
tic texture. More importantly, all the roadside had been rendered with the same texture and the
roadside detector learned about this “feature” very quickly.
A similar thing happened to the US Army when they first tried to detect tanks in the forest. They
took aerial photographs of the forest without tanks, then drove the tanks into the forest and took
another set of pictures. The classifier appeared to work perfectly. Unfortunately, it had merely
learned how to distinguish trees with shadows from trees without shadows—the first set of pictures
was taken in the early morning, the second set at noon.
Nonstationary Distributions
A much more subtle situation arises when the distribution changes slowly (also known as nonsta-
tionary distribution) and the model is not updated adequately. Below are some typical cases.
• We train a computational advertising model and then fail to update it frequently (e.g., we
forget to incorporate that an obscure new device called an iPad was just launched).
• We build a spam filter. It works well at detecting all spam that we have seen so far. But then
the spammers wisen up and craft new messages that look unlike anything we have seen
before.
• We build a product recommendation system. It works throughout the winter but then con-
tinues to recommend Santa hats long after Christmas.
More Anecdotes
• We build a face detector. It works well on all benchmarks. Unfortunately it fails on test
data—the offending examples are close-ups where the face fills the entire image (no such
data were in the training set).
• We build a Web search engine for the US market and want to deploy it in the UK.
• We train an image classifier by compiling a large dataset where each among a large set of
classes is equally represented in the dataset, say 1000 categories, represented by 1000 images
each. Then we deploy the system in the real world, where the actual label distribution of
photographs is decidedly non-uniform.

4.9.3 Correction of Distribution Shift
As we have discussed, there are many cases where training and test distributions P (x, y) are dif-
ferent. In some cases, we get lucky and the models work despite covariate, label, or concept shift.
In other cases, we can do better by employing principled strategies to cope with the shift. The re-
mainder of this section grows considerably more technical. The impatient reader could continue
on to the next section as this material is not prerequisite to subsequent concepts.
Empirical Risk and True Risk
Let us first reflect about what exactly is happening during model training: we iterate over features
and associated labels of training data {(x1 , y1 ), . . . , (xn , yn )} and update the parameters of a model
f after every minibatch. For simplicity we do not consider regularization, so we largely minimize
the loss on the training:
1∑
n
minimize l(f (xi ), yi ), (4.9.1)
f n
i=1
where l is the loss function measuring “how bad” the prediction f (xi ) is given the associated la-
bel yi . Statisticians call the term in (4.9.1) empirical risk. Empirical risk is an average loss over
the training data to approximate the true risk, which is the expectation of the loss over the entire
population of data drawn from their true distribution p(x, y):
∫ ∫
Ep(x,y) [l(f (x), y)] = l(f (x), y)p(x, y) dxdy. (4.9.2)
However, in practice we typically cannot obtain the entire population of data. Thus, empirical risk
minimization, which is minimizing empirical risk in (4.9.1), is a practical strategy for machine
learning, with the hope to approximate minimizing true risk.
Covariate Shift Correction
Assume that we want to estimate some dependency P (y | x) for which we have labeled data (xi , yi ).
Unfortunately, the observations xi are drawn from some source distribution q(x) rather than the
target distribution p(x). Fortunately, the dependency assumption means that the conditional dis-
tribution does not change: p(y | x) = q(y | x). If the source distribution q(x) is “wrong”, we can
correct for that by using the following simple identity in true risk:
∫ ∫ ∫ ∫
p(x)
l(f (x), y)p(y | x)p(x) dxdy = l(f (x), y)q(y | x)q(x) dxdy. (4.9.3)
q(x)
In other words, we need to reweigh each data example by the ratio of the probability that it would
have been drawn from the correct distribution to that from the wrong one:
def p(xi )
βi = . (4.9.4)
q(xi )
Plugging in the weight βi for each data example (xi , yi ) we can train our model using weighted
empirical risk minimization:
1∑
n
minimize βi l(f (xi ), yi ). (4.9.5)
f n
i=1

Alas, we do not know that ratio, so before we can do anything useful we need to estimate it. Many
methods are available, including some fancy operator-theoretic approaches that attempt to recali-
brate the expectation operator directly using a minimum-norm or a maximum entropy principle.
Note that for any such approach, we need samples drawn from both distributions—the “true” p,
e.g., by access to test data, and the one used for generating the training set q (the latter is trivially
available). Note however, that we only need features x ∼ p(x); we do not need to access labels
y ∼ p(y).
In this case, there exists a very effective approach that will give almost as good results as the orig-
inal: logistic regression, which is a special case of softmax regression for binary classification.
This is all that is needed to compute estimated probability ratios. We learn a classifier to distin-
guish between data drawn from p(x) and data drawn from q(x). If it is impossible to distinguish
between the two distributions then it means that the associated instances are equally likely to
come from either one of the two distributions. On the other hand, any instances that can be well
discriminated should be significantly overweighted or underweighted accordingly. For simplic-
ityʼs sake assume that we have an equal number of instances from both distributions p(x) and q(x),
respectively. Now denote by z labels that are 1 for data drawn from p and −1 for data drawn from
q. Then the probability in a mixed dataset is given by
p(x) P (z = 1 | x) p(x)
P (z = 1 | x) = and hence = . (4.9.6)
p(x) + q(x) P (z = −1 | x) q(x)
Thus, if we use a logistic regression approach, where P (z = 1 | x) = 1

1+exp(−h(x)) (h is a parame-
terized function), it follows that
1/(1 + exp(−h(xi )))

βi = = exp(h(xi )). (4.9.7)
exp(−h(xi ))/(1 + exp(−h(xi )))
As a result, we need to solve two problems: first one to distinguish between data drawn from both
distributions, and then a weighted empirical risk minimization problem in (4.9.5) where we weigh
terms by βi .
Now we are ready to describe a correction algorithm. Suppose that we have a training set
{(x1 , y1 ), . . . , (xn , yn )} and an unlabeled test set {u1 , . . . , um }. For covariate shift, we assume that
xi for all 1 ≤ i ≤ n are drawn from some source distribution and ui for all 1 ≤ i ≤ m are drawn
from the target distribution. Here is a prototypical algorithm for correcting covariate shift:
1. Generate a binary-classification training set: {(x1 , −1), . . . , (xn , −1), (u1 , 1), . . . , (um , 1)}.
2. Train a binary classifier using logistic regression to get function h.
3. Weigh training data using βi = exp(h(xi )) or better βi = min(exp(h(xi )), c) for some con-
stant c.
4. Use weights βi for training on {(x1 , y1 ), . . . , (xn , yn )} in (4.9.5).
Note that the above algorithm relies on a crucial assumption. For this scheme to work, we need
that each data example in the target (e.g., test time) distribution had nonzero probability of oc-
curring at training time. If we find a point where p(x) > 0 but q(x) = 0, then the corresponding
importance weight should be infinity.

Label Shift Correction
Assume that we are dealing with a classification task with k categories. Using the same notation in
Section 4.9.3, q and p are the source distribution (e.g., training time) and target distribution (e.g.,
test time), respectively. Assume that the distribution of labels shifts over time: q(y) ̸= p(y), but
the class-conditional distribution stays the same: q(x | y) = p(x | y). If the source distribution
q(y) is “wrong”, we can correct for that according to the following identity in true risk as defined
in (4.9.2):
∫ ∫ ∫ ∫
p(y)
l(f (x), y)p(x | y)p(y) dxdy = l(f (x), y)q(x | y)q(y) dxdy. (4.9.8)
q(y)
Here, our importance weights will correspond to the label likelihood ratios
def p(yi )
βi = . (4.9.9)
q(yi )
One nice thing about label shift is that if we have a reasonably good model on the source distribu-
tion, then we can get consistent estimates of these weights without ever having to deal with the
ambient dimension. In deep learning, the inputs tend to be high-dimensional objects like images,
while the labels are often simpler objects like categories.
To estimate the target label distribution, we first take our reasonably good off-the-shelf classifier
(typically trained on the training data) and compute its confusion matrix using the validation set
(also from the training distribution). The confusion matrix, C, is simply a k × k matrix, where
each column corresponds to the label category (ground truth) and each row corresponds to our
modelʼs predicted category. Each cellʼs value cij is the fraction of total predictions on the validation
set where the true label was j and our model predicted i.
Now, we cannot calculate the confusion matrix on the target data directly, because we do not get
to see the labels for the examples that we see in the wild, unless we invest in a complex real-time
annotation pipeline. What we can do, however, is average all of our models predictions at test
time together, yielding the mean model outputs µ(ŷ) ∈ Rk , whose ith element µ(ŷi ) is the fraction
of total predictions on the test set where our model predicted i.
It turns out that under some mild conditions—if our classifier was reasonably accurate in the first
place, and if the target data contain only categories that we have seen before, and if the label shift
assumption holds in the first place (the strongest assumption here), then we can estimate the test
set label distribution by solving a simple linear system
Cp(y) = µ(ŷ), (4.9.10)

∑
because as an estimate kj=1 cij p(yj ) = µ(ŷi ) holds for all 1 ≤ i ≤ k, where p(yj ) is the j th element
of the k-dimensional label distribution vector p(y). If our classifier is sufficiently accurate to begin
with, then the confusion matrix C will be invertible, and we get a solution p(y) = C−1 µ(ŷ).
Because we observe the labels on the source data, it is easy to estimate the distribution q(y). Then
for any training example i with label yi , we can take the ratio of our estimated p(yi )/q(yi ) to cal-
culate the weight βi , and plug this into weighted empirical risk minimization in (4.9.5).

Concept Shift Correction
Concept shift is much harder to fix in a principled manner. For instance, in a situation where
suddenly the problem changes from distinguishing cats from dogs to one of distinguishing white
from black animals, it will be unreasonable to assume that we can do much better than just col-
lecting new labels and training from scratch. Fortunately, in practice, such extreme shifts are
rare. Instead, what usually happens is that the task keeps on changing slowly. To make things
more concrete, here are some examples:
• In computational advertising, new products are launched, old products become less popular.
This means that the distribution over ads and their popularity changes gradually and any
click-through rate predictor needs to change gradually with it.
• Traffic camera lenses degrade gradually due to environmental wear, affecting image quality
progressively.
• News content changes gradually (i.e., most of the news remains unchanged but new stories
appear).
In such cases, we can use the same approach that we used for training networks to make them
adapt to the change in the data. In other words, we use the existing network weights and simply
perform a few update steps with the new data rather than training from scratch.
4.9.4 A Taxonomy of Learning Problems
Armed with knowledge about how to deal with changes in distributions, we can now consider
some other aspects of machine learning problem formulation.
Batch Learning
In batch learning, we have access to training features and labels {(x1 , y1 ), . . . , (xn , yn )}, which we
use to train a model f (x). Later on, we deploy this model to score new data (x, y) drawn from the
same distribution. This is the default assumption for any of the problems that we discuss here. For
instance, we might train a cat detector based on lots of pictures of cats and dogs. Once we trained
it, we ship it as part of a smart catdoor computer vision system that lets only cats in. This is then
installed in a customerʼs home and is never updated again (barring extreme circumstances).
Online Learning
Now imagine that the data (xi , yi ) arrives one sample at a time. More specifically, assume that
we first observe xi , then we need to come up with an estimate f (xi ) and only once we have done
this, we observe yi and with it, we receive a reward or incur a loss, given our decision. Many
real problems fall into this category. For example, we need to predict tomorrowʼs stock price, this
allows us to trade based on that estimate and at the end of the day we find out whether our estimate
allowed us to make a profit. In other words, in online learning, we have the following cycle where
we are continuously improving our model given new observations.
model ft −→ data xt −→ estimate ft (xt ) −→ observation yt −→ loss l(yt , ft (xt )) −→ model ft+1
(4.9.11)

Bandits
Bandits are a special case of the problem above. While in most learning problems we have a con-
tinuously parametrized function f where we want to learn its parameters (e.g., a deep network),
in a bandit problem we only have a finite number of arms that we can pull, i.e., a finite number
of actions that we can take. It is not very surprising that for this simpler problem stronger theo-
retical guarantees in terms of optimality can be obtained. We list it mainly since this problem is
often (confusingly) treated as if it were a distinct learning setting.
Control
In many cases the environment remembers what we did. Not necessarily in an adversarial manner
but it will just remember and the response will depend on what happened before. For instance, a
coffee boiler controller will observe different temperatures depending on whether it was heating
the boiler previously. PID (proportional-integral-derivative) controller algorithms are a popular
choice there. Likewise, a userʼs behavior on a news site will depend on what we showed him
previously (e.g., he will read most news only once). Many such algorithms form a model of the
environment in which they act such as to make their decisions appear less random. Recently,
control theory (e.g., PID variants) has also been used to automatically tune hyperparameters to
achive better disentangling and reconstruction quality, and improve the diversity of generated
text and the reconstruction quality of generated images (Shao et al., 2020).
Reinforcement Learning
In the more general case of an environment with memory, we may encounter situations where
the environment is trying to cooperate with us (cooperative games, in particular for non-zero-sum
games), or others where the environment will try to win. Chess, Go, Backgammon, or StarCraft
are some of the cases in reinforcement learning. Likewise, we might want to build a good controller
for autonomous cars. The other cars are likely to respond to the autonomous carʼs driving style in
nontrivial ways, e.g., trying to avoid it, trying to cause an accident, and trying to cooperate with
it.
Considering the Environment
One key distinction between the different situations above is that the same strategy that might have
worked throughout in the case of a stationary environment, might not work throughout when the
environment can adapt. For instance, an arbitrage opportunity discovered by a trader is likely to
disappear once he starts exploiting it. The speed and manner at which the environment changes
determines to a large extent the type of algorithms that we can bring to bear. For instance, if we
know that things may only change slowly, we can force any estimate to change only slowly, too. If
we know that the environment might change instantaneously, but only very infrequently, we can
make allowances for that. These types of knowledge are crucial for the aspiring data scientist to
deal with concept shift, i.e., when the problem that he is trying to solve changes over time.

4.9.5 Fairness, Accountability, and Transparency in Machine Learning
Finally, it is important to remember that when you deploy machine learning systems you are not
merely optimizing a predictive model—you are typically providing a tool that will be used to (par-
tially or fully) automate decisions. These technical systems can impact the lives of individuals
subject to the resulting decisions. The leap from considering predictions to decisions raises not
only new technical questions, but also a slew of ethical questions that must be carefully consid-
ered. If we are deploying a medical diagnostic system, we need to know for which populations
it may work and which it may not. Overlooking foreseeable risks to the welfare of a subpopula-
tion could cause us to administer inferior care. Moreover, once we contemplate decision-making
systems, we must step back and reconsider how we evaluate our technology. Among other con-
sequences of this change of scope, we will find that accuracy is seldom the right measure. For
instance, when translating predictions into actions, we will often want to take into account the
potential cost sensitivity of erring in various ways. If one way of misclassifying an image could
be perceived as a racial sleight of hand, while misclassification to a different category would be
harmless, then we might want to adjust our thresholds accordingly, accounting for societal val-
ues in designing the decision-making protocol. We also want to be careful about how prediction
systems can lead to feedback loops. For example, consider predictive policing systems, which al-
locate patrol officers to areas with high forecasted crime. It is easy to see how a worrying pattern
can emerge:
1. Neighborhoods with more crime get more patrols.
2. Consequently, more crimes are discovered in these neighborhoods, entering the training
data available for future iterations.
3. Exposed to more positives, the model predicts yet more crime in these neighborhoods.
4. In the next iteration, the updated model targets the same neighborhood even more heavily
leading to yet more crimes discovered, etc.
Often, the various mechanisms by which a modelʼs predictions become coupled to its training data
are unaccounted for in the modeling process. This can lead to what researchers call runaway feed-
back loops. Additionally, we want to be careful about whether we are addressing the right problem
in the first place. Predictive algorithms now play an outsize role in mediating the dissemination of
information. Should the news that an individual encounters be determined by the set of Facebook
pages they have Liked? These are just a few among the many pressing ethical dilemmas that you
might encounter in a career in machine learning.
Summary
• In many cases training and test sets do not come from the same distribution. This is called
distribution shift.
• True risk is the expectation of the loss over the entire population of data drawn from their
true distribution. However, this entire population is usually unavailable. Empirical risk is
an average loss over the training data to approximate the true risk. In practice, we perform
empirical risk minimization.
• Under the corresponding assumptions, covariate and label shift can be detected and cor-
rected for at test time. Failure to account for this bias can become problematic at test time.
• In some cases, the environment may remember automated actions and respond in surpris-
ing ways. We must account for this possibility when building models and continue to mon-

itor live systems, open to the possibility that our models and the environment will become
entangled in unanticipated ways.
Exercises
1. What could happen when we change the behavior of a search engine? What might the users
do? What about the advertisers?
2. Implement a covariate shift detector. Hint: build a classifier.
3. Implement a covariate shift corrector.
4. Besides distribution shift, what else could affect how empirical risk approximates true risk?
Discussions75
4.10 Predicting House Prices on Kaggle
Now that we have introduced some basic tools for building and training deep networks and reg-
ularizing them with techniques including weight decay and dropout, we are ready to put all this
knowledge into practice by participating in a Kaggle competition. The house price prediction
competition is a great place to start. The data are fairly generic and do not exhibit exotic structure
that might require specialized models (as audio or video might). This dataset, collected by Bart de
Cock in 2011 (DeCock, 2011), covers house prices in Ames, IA from the period of 2006–2010. It is
considerably larger than the famous Boston housing dataset76 of Harrison and Rubinfeld (1978),
boasting both more examples and more features.
In this section, we will walk you through details of data preprocessing, model design, and hyper-
parameter selection. We hope that through a hands-on approach, you will gain some intuitions
that will guide you in your career as a data scientist.
4.10.1 Downloading and Caching Datasets
Throughout the book, we will train and test models on various downloaded datasets. Here, we
implement several utility functions to facilitate data downloading. First, we maintain a dictionary
DATA_HUB that maps a string (the name of the dataset) to a tuple containing both the URL to locate
the dataset and the SHA-1 key that verifies the integrity of the file. All such datasets are hosted at
the site whose address is DATA_URL.
import os
import requests
import zipfile
import tarfile
import hashlib
DATA_HUB = dict() #@save

DATA_URL = 'http://d2l-data.s3-accelerate.amazonaws.com/' #@save
75
76
https://archive.ics.uci.edu/ml/machine-learning-databases/housing/housing.names
4.10. Predicting House Prices on Kaggle 187

The following download function downloads a dataset, caching it in a local directory (../data by
default) and returns the name of the downloaded file. If a file corresponding to this dataset already
exists in the cache directory and its SHA-1 matches the one stored in DATA_HUB, our code will use
the cached file to avoid clogging up your internet with redundant downloads.
def download(name, cache_dir=os.path.join('..', 'data')): #@save

"""Download a file inserted into DATA_HUB, return the local filename."""
assert name in DATA_HUB, f"{name} does not exist in {DATA_HUB}."
url, sha1_hash = DATA_HUB[name]
d2l.mkdir_if_not_exist(cache_dir)
fname = os.path.join(cache_dir, url.split('/')[-1])
if os.path.exists(fname):
sha1 = hashlib.sha1()
with open(fname, 'rb') as f:
while True:
data = f.read(1048576)
if not data:
break
sha1.update(data)
if sha1.hexdigest() == sha1_hash:
return fname # Hit cache
print(f'Downloading {fname} from {url}...')
r = requests.get(url, stream=True, verify=True)
with open(fname, 'wb') as f:
f.write(r.content)
return fname
We also implement two additional utility functions: one is to download and extract a zip or tar
file and the other to download all the datasets used in this book from DATA_HUB into the cache
directory.
def download_extract(name, folder=None): #@save

"""Download and extract a zip/tar file."""
fname = download(name)
base_dir = os.path.dirname(fname)
data_dir, ext = os.path.splitext(fname)
if ext == '.zip':
fp = zipfile.ZipFile(fname, 'r')
elif ext in ('.tar', '.gz'):
fp = tarfile.open(fname, 'r')
else:
assert False, 'Only zip/tar files can be extracted.'
fp.extractall(base_dir)
return os.path.join(base_dir, folder) if folder else data_dir
def download_all(): #@save

"""Download all files in the DATA_HUB."""
for name in DATA_HUB:
download(name)

4.10.2 Kaggle
Kaggle77 is a popular platform that hosts machine learning competitions. Each competition cen-
ters on a dataset and many are sponsored by stakeholders who offer prizes to the winning solu-
tions. The platform helps users to interact via forums and shared code, fostering both collabo-
ration and competition. While leaderboard chasing often spirals out of control, with researchers
focusing myopically on preprocessing steps rather than asking fundamental questions, there is
also tremendous value in the objectivity of a platform that facilitates direct quantitative compar-
isons among competing approaches as well as code sharing so that everyone can learn what did
and did not work. If you want to participate in a Kaggle competition, you will first need to register
for an account (see Fig. 4.10.1).
Fig. 4.10.1: The Kaggle website.
On the house price prediction competition page, as illustrated in Fig. 4.10.2, you can find the
dataset (under the “Data” tab), submit predictions, and see your ranking, The URL is right here:
https://www.kaggle.com/c/house-prices-advanced-regression-techniques
Fig. 4.10.2: The house price prediction competition page.

77
https://www.kaggle.com

4.10.3 Accessing and Reading the Dataset
Note that the competition data is separated into training and test sets. Each record includes the
property value of the house and attributes such as street type, year of construction, roof type,
basement condition, etc. The features consist of various data types. For example, the year of
construction is represented by an integer, the roof type by discrete categorical assignments, and
other features by floating point numbers. And here is where reality complicates things: for some
examples, some data are altogether missing with the missing value marked simply as “na”. The
price of each house is included for the training set only (it is a competition after all). We will want
to partition the training set to create a validation set, but we only get to evaluate our models on
the official test set after uploading predictions to Kaggle. The “Data” tab on the competition tab in
Fig. 4.10.2 has links to download the data.
To get started, we will read in and process the data using pandas, which we have introduced in
Section 2.2. So, you will want to make sure that you have pandas installed before proceeding fur-
ther. Fortunately, if you are reading in Jupyter, we can install pandas without even leaving the
notebook.
# If pandas is not installed, please uncomment the following line:

# !pip install pandas
%matplotlib inline
from mxnet import gluon, autograd, init, np, npx
import pandas as pd
npx.set_np()
For convenience, we can download and cache the Kaggle housing dataset using the script we de-
fined above.
DATA_HUB['kaggle_house_train'] = ( #@save
DATA_URL + 'kaggle_house_pred_train.csv',
'585e9cc93e70b39160e7921475f9bcd7d31219ce')
DATA_HUB['kaggle_house_test'] = ( #@save
DATA_URL + 'kaggle_house_pred_test.csv',
'fa19780a7b011d9b009e8bff8e99922a8ee2eb90')
We use pandas to load the two csv files containing training and test data respectively.
train_data = pd.read_csv(download('kaggle_house_train'))
test_data = pd.read_csv(download('kaggle_house_test'))
Downloading ../data/kaggle_house_pred_train.csv from http://d2l-data.s3-accelerate.amazonaws.

,→com/kaggle_house_pred_train.csv...
Downloading ../data/kaggle_house_pred_test.csv from http://d2l-data.s3-accelerate.amazonaws.
,→com/kaggle_house_pred_test.csv...
The training dataset includes 1460 examples, 80 features, and 1 label, while the test data contains
1459 examples and 80 features.

print(train_data.shape)
print(test_data.shape)
(1460, 81)
(1459, 80)
Let us take a look at the first four and last two features as well as the label (SalePrice) from the first
four examples.
print(train_data.iloc[0:4, [0, 1, 2, 3, -3, -2, -1]])
Id MSSubClass MSZoning LotFrontage SaleType SaleCondition SalePrice

0 1 60 RL 65.0 WD Normal 208500
1 2 20 RL 80.0 WD Normal 181500
2 3 60 RL 68.0 WD Normal 223500
3 4 70 RL 60.0 WD Abnorml 140000
We can see that in each example, the first feature is the ID. This helps the model identify each
training example. While this is convenient, it does not carry any information for prediction pur-
poses. Hence, we remove it from the dataset before feeding the data into the model.
all_features = pd.concat((train_data.iloc[:, 1:-1], test_data.iloc[:, 1:]))
4.10.4 Data Preprocessing
As stated above, we have a wide variety of data types. We will need to preprocess the data before we
can start modeling. Let us start with the numerical features. First, we apply a heuristic, replacing
all missing values by the corresponding featureʼs mean. Then, to put all features on a common
scale, we standardize the data by rescaling features to zero mean and unit variance:
x−µ
x← . (4.10.1)
σ
To verify that this indeed transforms our feature (variable) such that it has zero mean and unit
variance, note that E[ x−µ µ−µ
σ ] = σ = 0 and that E[(x − µ) ] = (σ + µ ) − 2µ + µ = σ . Intuitively,
2 2 2 2 2 2
we standardize the data for two reasons. First, it proves convenient for optimization. Second,
because we do not know a priori which features will be relevant, we do not want to penalize coef-
ficients assigned to one feature more than on any other.
numeric_features = all_features.dtypes[all_features.dtypes != 'object'].index

all_features[numeric_features] = all_features[numeric_features].apply(
lambda x: (x - x.mean()) / (x.std()))
# After standardizing the data all means vanish, hence we can set missing
# values to 0
all_features[numeric_features] = all_features[numeric_features].fillna(0)
Next we deal with discrete values. This includes features such as “MSZoning”. We replace them by
a one-hot encoding in the same way that we previously transformed multiclass labels into vectors
(see Section 3.4.1). For instance, “MSZoning” assumes the values “RL” and “RM”. Dropping the
“MSZoning” feature, two new indicator features “MSZoning_RL” and “MSZoning_RM” are created

with values being either 0 or 1. According to one-hot encoding, if the original value of “MSZon-
ing” is “RL”, then “MSZoning_RL” is 1 and “MSZoning_RM” is 0. The pandas package does this
automatically for us.
# `Dummy_na=True` considers "na" (missing value) as a valid feature value, and

# creates an indicator feature for it
all_features = pd.get_dummies(all_features, dummy_na=True)
all_features.shape
(2919, 331)
You can see that this conversion increases the number of features from 79 to 331. Finally, via the
values attribute, we can extract the NumPy format from the pandas format and convert it into the
tensor representation for training.
n_train = train_data.shape[0]
train_features = np.array(all_features[:n_train].values, dtype=np.float32)
test_features = np.array(all_features[n_train:].values, dtype=np.float32)
train_labels = np.array(
train_data.SalePrice.values.reshape(-1, 1), dtype=np.float32)
4.10.5 Training
To get started we train a linear model with squared loss. Not surprisingly, our linear model will
not lead to a competition-winning submission but it provides a sanity check to see whether there
is meaningful information in the data. If we cannot do better than random guessing here, then
there might be a good chance that we have a data processing bug. And if things work, the linear
model will serve as a baseline giving us some intuition about how close the simple model gets
to the best reported models, giving us a sense of how much gain we should expect from fancier
models.
def get_net():
net.initialize()
return net
With house prices, as with stock prices, we care about relative quantities more than absolute quan-
tities. Thus we tend to care more about the relative error y−ŷ
y than about the absolute error y − ŷ.
For instance, if our prediction is off by USD 100,000 when estimating the price of a house in Rural
Ohio, where the value of a typical house is 125,000 USD, then we are probably doing a horrible job.
On the other hand, if we err by this amount in Los Altos Hills, California, this might represent a
stunningly accurate prediction (there, the median house price exceeds 4 million USD).
One way to address this problem is to measure the discrepancy in the logarithm of the price esti-
mates. In fact, this is also the official error measure used by the competition to evaluate the quality
of submissions. After all, a small value δ for | log y − log ŷ| ≤ δ translates into e−δ ≤ yŷ ≤ eδ . This
leads to the following root-mean-squared-error between the logarithm of the predicted price and

the logarithm of the label price:
v
u n
u1 ∑
t (log yi − log ŷi )2 . (4.10.2)
n
i=1
def log_rmse(net, features, labels):

# To further stabilize the value when the logarithm is taken, set the
# value less than 1 as 1
clipped_preds = np.clip(net(features), 1, float('inf'))
return np.sqrt(2 * loss(np.log(clipped_preds), np.log(labels)).mean())
Unlike in previous sections, our training functions will rely on the Adam optimizer (we will de-
scribe it in greater detail later). The main appeal of this optimizer is that, despite doing no better
(and sometimes worse) given unlimited resources for hyperparameter optimization, people tend
to find that it is significantly less sensitive to the initial learning rate.
def train(net, train_features, train_labels, test_features, test_labels,

num_epochs, learning_rate, weight_decay, batch_size):
train_ls, test_ls = [], []
train_iter = d2l.load_array((train_features, train_labels), batch_size)
# The Adam optimization algorithm is used here
trainer = gluon.Trainer(net.collect_params(), 'adam', {
'learning_rate': learning_rate, 'wd': weight_decay})
l = loss(net(X), y)
l.backward()
train_ls.append(log_rmse(net, train_features, train_labels))
if test_labels is not None:
test_ls.append(log_rmse(net, test_features, test_labels))
return train_ls, test_ls
4.10.6 K-Fold Cross-Validation
You might recall that we introduced K-fold cross-validation in the section where we discussed how
to deal with model selection (Section 4.4). We will put this to good use to select the model design
and to adjust the hyperparameters. We first need a function that returns the ith fold of the data in
a K-fold cross-validation procedure. It proceeds by slicing out the ith segment as validation data
and returning the rest as training data. Note that this is not the most efficient way of handling data
and we would definitely do something much smarter if our dataset was considerably larger. But
this added complexity might obfuscate our code unnecessarily so we can safely omit it here owing
to the simplicity of our problem.
def get_k_fold_data(k, i, X, y):

assert k > 1
fold_size = X.shape[0] // k
X_train, y_train = None, None
for j in range(k):
idx = slice(j * fold_size, (j + 1) * fold_size)

X_part, y_part = X[idx, :], y[idx]
if j == i:
X_valid, y_valid = X_part, y_part
elif X_train is None:
X_train, y_train = X_part, y_part
else:
X_train = np.concatenate([X_train, X_part], 0)
y_train = np.concatenate([y_train, y_part], 0)
return X_train, y_train, X_valid, y_valid
The training and verification error averages are returned when we train K times in the K-fold
cross-validation.
def k_fold(k, X_train, y_train, num_epochs,

learning_rate, weight_decay, batch_size):
train_l_sum, valid_l_sum = 0, 0
for i in range(k):
data = get_k_fold_data(k, i, X_train, y_train)
net = get_net()
train_ls, valid_ls = train(net, *data, num_epochs, learning_rate,
weight_decay, batch_size)
train_l_sum += train_ls[-1]
valid_l_sum += valid_ls[-1]
if i == 0:
d2l.plot(list(range(1, num_epochs+1)), [train_ls, valid_ls],
xlabel='epoch', ylabel='rmse',
legend=['train', 'valid'], yscale='log')
print(f'fold {i + 1}, train log rmse {float(train_ls[-1]):f}, '
f'valid log rmse {float(valid_ls[-1]):f}')
return train_l_sum / k, valid_l_sum / k
4.10.7 Model Selection
In this example, we pick an untuned set of hyperparameters and leave it up to the reader to im-
prove the model. Finding a good choice can take time, depending on how many variables one
optimizes over. With a large enough dataset, and the normal sorts of hyperparameters, K-fold
cross-validation tends to be reasonably resilient against multiple testing. However, if we try an
unreasonably large number of options we might just get lucky and find that our validation perfor-
mance is no longer representative of the true error.
k, num_epochs, lr, weight_decay, batch_size = 5, 100, 5, 0, 64

train_l, valid_l = k_fold(k, train_features, train_labels, num_epochs, lr,
weight_decay, batch_size)
print(f'{k}-fold validation: avg train log rmse: {float(train_l):f}, '
f'avg valid log rmse: {float(valid_l):f}')
fold 1, train log rmse 0.169762, valid log rmse 0.156920


5-fold validation: avg train log rmse: 0.165232, avg valid log rmse: 0.170282
Notice that sometimes the number of training errors for a set of hyperparameters can be very
low, even as the number of errors on K-fold cross-validation is considerably higher. This indi-
cates that we are overfitting. Throughout training you will want to monitor both numbers. Less
overfitting might indicate that our data can support a more powerful model. Massive overfitting
might suggest that we can gain by incorporating regularization techniques.
4.10.8 Submitting Predictions on Kaggle
Now that we know what a good choice of hyperparameters should be, we might as well use all the
data to train on it (rather than just 1 − 1/K of the data that are used in the cross-validation slices).
The model that we obtain in this way can then be applied to the test set. Saving the predictions in
a csv file will simplify uploading the results to Kaggle.
def train_and_pred(train_features, test_feature, train_labels, test_data,

num_epochs, lr, weight_decay, batch_size):
net = get_net()
train_ls, _ = train(net, train_features, train_labels, None, None,
num_epochs, lr, weight_decay, batch_size)
d2l.plot(np.arange(1, num_epochs + 1), [train_ls], xlabel='epoch',
ylabel='log rmse', yscale='log')
print(f'train log rmse {float(train_ls[-1]):f}')
# Apply the network to the test set
preds = d2l.numpy(net(test_features))
# Reformat it to export to Kaggle
test_data['SalePrice'] = pd.Series(preds.reshape(1, -1)[0])
submission = pd.concat([test_data['Id'], test_data['SalePrice']], axis=1)
submission.to_csv('submission.csv', index=False)
One nice sanity check is to see whether the predictions on the test set resemble those of the K-fold
cross-validation process. If they do, it is time to upload them to Kaggle. The following code will
generate a file called submission.csv.

train_and_pred(train_features, test_features, train_labels, test_data,
num_epochs, lr, weight_decay, batch_size)
train log rmse 0.162304
Next, as demonstrated in Fig. 4.10.3, we can submit our predictions on Kaggle and see how they
compare with the actual house prices (labels) on the test set. The steps are quite simple:
• Log in to the Kaggle website and visit the house price prediction competition page.
• Click the “Submit Predictions” or “Late Submission” button (as of this writing, the button is
located on the right).
• Click the “Upload Submission File” button in the dashed box at the bottom of the page and
select the prediction file you wish to upload.
• Click the “Make Submission” button at the bottom of the page to view your results.
Fig. 4.10.3: Submitting data to Kaggle

Summary
• Real data often contain a mix of different data types and need to be preprocessed.
• Rescaling real-valued data to zero mean and unit variance is a good default. So is replacing
missing values with their mean.
• Transforming categorical features into indicator features allows us to treat them like one-hot
vectors.
• We can use K-fold cross-validation to select the model and adjust the hyperparameters.
• Logarithms are useful for relative errors.
Exercises
1. Submit your predictions for this section to Kaggle. How good are your predictions?
2. Can you improve your model by minimizing the logarithm of prices directly? What happens
if you try to predict the logarithm of the price rather than the price?
3. Is it always a good idea to replace missing values by their mean? Hint: can you construct a
situation where the values are not missing at random?
4. Improve the score on Kaggle by tuning the hyperparameters through K-fold cross-
validation.
5. Improve the score by improving the model (e.g., layers, weight decay, and dropout).
6. What happens if we do not standardize the continuous numerical features like what we have
done in this section?
Discussions78
78

5 | Deep Learning Computation
Alongside giant datasets and powerful hardware, great software tools have played an indispens-
able role in the rapid progress of deep learning. Starting with the pathbreaking Theano library
released in 2007, flexible open-source tools have enabled researchers to rapidly prototype models,
avoiding repetitive work when recycling standard components while still maintaining the ability
to make low-level modifications. Over time, deep learningʼs libraries have evolved to offer in-
creasingly coarse abstractions. Just as semiconductor designers went from specifying transistors
to logical circuits to writing code, neural networks researchers have moved from thinking about
the behavior of individual artificial neurons to conceiving of networks in terms of whole layers,
and now often design architectures with far coarser blocks in mind.
So far, we have introduced some basic machine learning concepts, ramping up to fully-functional
deep learning models. In the last chapter, we implemented each component of an MLP from
scratch and even showed how to leverage high-level APIs to roll out the same models effortlessly.
To get you that far that fast, we called upon the libraries, but skipped over more advanced details
about how they work. In this chapter, we will peel back the curtain, digging deeper into the key
components of deep learning computation, namely model construction, parameter access and
initialization, designing custom layers and blocks, reading and writing models to disk, and lever-
aging GPUs to achieve dramatic speedups. These insights will move you from end user to power
user, giving you the tools needed to reap the benefits of a mature deep learning library while re-
taining the flexibility to implement more complex models, including those you invent yourself!
While this chapter does not introduce any new models or datasets, the advanced modeling chap-
ters that follow rely heavily on these techniques.
5.1 Layers and Blocks
When we first introduced neural networks, we focused on linear models with a single output.
Here, the entire model consists of just a single neuron. Note that a single neuron (i) takes some
set of inputs; (ii) generates a corresponding scalar output; and (iii) has a set of associated param-
eters that can be updated to optimize some objective function of interest. Then, once we started
thinking about networks with multiple outputs, we leveraged vectorized arithmetic to characterize
an entire layer of neurons. Just like individual neurons, layers (i) take a set of inputs, (ii) generate
corresponding outputs, and (iii) are described by a set of tunable parameters. When we worked
through softmax regression, a single layer was itself the model. However, even when we subse-
quently introduced MLPs, we could still think of the model as retaining this same basic structure.
Interestingly, for MLPs, both the entire model and its constituent layers share this structure. The
entire model takes in raw inputs (the features), generates outputs (the predictions), and possesses
parameters (the combined parameters from all constituent layers). Likewise, each individual layer
ingests inputs (supplied by the previous layer) generates outputs (the inputs to the subsequent
199
layer), and possesses a set of tunable parameters that are updated according to the signal that
flows backwards from the subsequent layer.
While you might think that neurons, layers, and models give us enough abstractions to go about
our business, it turns out that we often find it convenient to speak about components that are
larger than an individual layer but smaller than the entire model. For example, the ResNet-152
architecture, which is wildly popular in computer vision, possesses hundreds of layers. These
layers consist of repeating patterns of groups of layers. Implementing such a network one layer at
a time can grow tedious. This concern is not just hypothetical—such design patterns are common
in practice. The ResNet architecture mentioned above won the 2015 ImageNet and COCO com-
puter vision competitions for both recognition and detection (He et al., 2016a) and remains a go-to
architecture for many vision tasks. Similar architectures in which layers are arranged in various
repeating patterns are now ubiquitous in other domains, including natural language processing
and speech.
To implement these complex networks, we introduce the concept of a neural network block. A
block could describe a single layer, a component consisting of multiple layers, or the entire model
itself! One benefit of working with the block abstraction is that they can be combined into larger
artifacts, often recursively. This is illustrated in Fig. 5.1.1. By defining code to generate blocks
of arbitrary complexity on demand, we can write surprisingly compact code and still implement
complex neural networks.
Fig. 5.1.1: Multiple layers are combined into blocks, forming repeating patterns of larger models.
From a programing standpoint, a block is represented by a class. Any subclass of it must define a
forward propagation function that transforms its input into output and must store any necessary
parameters. Note that some blocks do not require any parameters at all. Finally a block must pos-
sess a backpropagation function, for purposes of calculating gradients. Fortunately, due to some
behind-the-scenes magic supplied by the auto differentiation (introduced in Section 2.5) when
defining our own block, we only need to worry about parameters and the forward propagation
function.
To begin, we revisit the code that we used to implement MLPs (Section 4.3). The following code
generates a network with one fully-connected hidden layer with 256 units and ReLU activation,
followed by a fully-connected output layer with 10 units (no activation function).
200 Chapter 5. Deep Learning Computation

npx.set_np()
net.add(nn.Dense(256, activation='relu'))
net.initialize()
X = np.random.uniform(size=(2, 20))
net(X)
array([[ 0.06240272, -0.03268593, 0.02582653, 0.02254182, -0.03728798,

-0.04253786, 0.00540613, -0.01364186, -0.09915452, -0.02272738],
[ 0.02816677, -0.03341204, 0.03565666, 0.02506382, -0.04136416,
-0.04941845, 0.01738528, 0.01081961, -0.09932579, -0.01176298]])
In this example, we constructed our model by instantiating an nn.Sequential, assigning the re-
turned object to the net variable. Next, we repeatedly call its add function, appending layers in
the order that they should be executed. In short, nn.Sequential defines a special kind of Block,
the class that presents a block in Gluon. It maintains an ordered list of constituent Blocks. The
add function simply facilitates the addition of each successive Block to the list. Note that each
layer is an instance of the Dense class which is itself a subclass of Block. The forward propagation
(forward) function is also remarkably simple: it chains each Block in the list together, passing the
output of each as the input to the next. Note that until now, we have been invoking our models via
the construction net(X) to obtain their outputs. This is actually just shorthand for net.forward(X),
a slick Python trick achieved via the Block classʼs __call__ function.
5.1.1 A Custom Block
Perhaps the easiest way to develop intuition about how a block works is to implement one our-
selves. Before we implement our own custom block, we briefly summarize the basic functionality
that each block must provide:
1. Ingest input data as arguments to its forward propagation function.
2. Generate an output by having the forward propagation function return a value. Note that
the output may have a different shape from the input. For example, the first fully-connected
layer in our model above ingests an input of arbitrary dimension but returns an output of
dimension 256.
3. Calculate the gradient of its output with respect to its input, which can be accessed via its
backpropagation function. Typically this happens automatically.
4. Store and provide access to those parameters necessary to execute the forward propagation
computation.
5. Initialize model parameters as needed.
In the following snippet, we code up a block from scratch corresponding to an MLP with one hid-
den layer with 256 hidden units, and a 10-dimensional output layer. Note that the MLP class below
inherits the class that represents a block. We will heavily rely on the parent classʼs functions, sup-
plying only our own constructor (the __init__ function in Python) and the forward propagation
function.
5.1. Layers and Blocks 201

class MLP(nn.Block):
# Declare a layer with model parameters. Here, we declare two
# fully-connected layers
def __init__(self, **kwargs):
# Call the constructor of the `MLP` parent class `Block` to perform
# the necessary initialization. In this way, other function arguments
# can also be specified during class instantiation, such as the model
# parameters, `params` (to be described later)
super().__init__(**kwargs)
self.hidden = nn.Dense(256, activation='relu') # Hidden layer
self.out = nn.Dense(10) # Output layer
# Define the forward propagation of the model, that is, how to return the
# required model output based on the input `X`
def forward(self, X):
return self.out(self.hidden(X))
Let us first focus on the forward propagation function. Note that it takes X as the input, calculates
the hidden representation with the activation function applied, and outputs its logits. In this MLP
implementation, both layers are instance variables. To see why this is reasonable, imagine instan-
tiating two MLPs, net1 and net2, and training them on different data. Naturally, we would expect
them to represent two different learned models.
We instantiate the MLPʼs layers in the constructor and subsequently invoke these layers on each
call to the forward propagation function. Note a few key details. First, our customized __init__
function invokes the parent classʼs __init__ function via super().__init__() sparing us the
pain of restating boilerplate code applicable to most blocks. We then instantiate our two fully-
connected layers, assigning them to self.hidden and self.out. Note that unless we implement a
new operator, we need not worry about the backpropagation function or parameter initialization.
The system will generate these functions automatically. Let us try this out.
net = MLP()
net.initialize()
net(X)
array([[-0.03989594, -0.1041471 , 0.06799038, 0.05245074, 0.02526059,

-0.00640342, 0.04182098, -0.01665319, -0.02067346, -0.07863817],
[-0.03612847, -0.07210436, 0.09159479, 0.07890771, 0.02494172,
-0.01028665, 0.01732428, -0.02843242, 0.03772651, -0.06671704]])
A key virtue of the block abstraction is its versatility. We can subclass a block to create layers (such
as the fully-connected layer class), entire models (such as the MLP class above), or various compo-
nents of intermediate complexity. We exploit this versatility throughout the following chapters,
such as when addressing convolutional neural networks.

5.1.2 The Sequential Block
We can now take a closer look at how the Sequential class works. Recall that Sequential was
designed to daisy-chain other blocks together. To build our own simplified MySequential, we just
need to define two key function: 1. A function to append blocks one by one to a list. 2. A forward
propagation function to pass an input through the chain of blocks, in the same order as they were
appended.
The following MySequential class delivers the same functionality of the default Sequential class.
class MySequential(nn.Block):
def add(self, block):
# Here, `block` is an instance of a `Block` subclass, and we assume
# that it has a unique name. We save it in the member variable
# `_children` of the `Block` class, and its type is OrderedDict. When
# the `MySequential` instance calls the ìnitialize` function, the
# system automatically initializes all members of `_children`
self._children[block.name] = block

# OrderedDict guarantees that members will be traversed in the order
# they were added
for block in self._children.values():
X = block(X)
return X
The add function adds a single block to the ordered dictionary _children. You might wonder why
every Gluon Block possesses a _children attribute and why we used it rather than just define a
Python list ourselves. In short the chief advantage of _children is that during our blockʼs param-
eter initialization, Gluon knows to look inside the _children dictionary to find sub-blocks whose
parameters also need to be initialized.
When our MySequentialʼs forward propagation function is invoked, each added block is executed
in the order in which they were added. We can now reimplement an MLP using our MySequential
class.
net = MySequential()
net.initialize()
net(X)
array([[-0.07645682, -0.01130233, 0.04952145, -0.04651389, -0.04131573,

-0.05884133, -0.0621381 , 0.01311472, -0.01379425, -0.02514282],
[-0.05124625, 0.00711231, -0.00155935, -0.07555379, -0.06675334,
-0.01762914, 0.00589084, 0.01447191, -0.04330775, 0.03317726]])
Note that this use of MySequential is identical to the code we previously wrote for the Sequential
class (as described in Section 4.3).

5.1.3 Executing Code in the Forward Propagation Function
The Sequential class makes model construction easy, allowing us to assemble new architectures
without having to define our own class. However, not all architectures are simple daisy chains.
When greater flexibility is required, we will want to define our own blocks. For example, we
might want to execute Pythonʼs control flow within the forward propagation function. Moreover,
we might want to perform arbitrary mathematical operations, not simply relying on predefined
neural network layers.
You might have noticed that until now, all of the operations in our networks have acted upon our
networkʼs activations and its parameters. Sometimes, however, we might want to incorporate
terms that are neither the result of previous layers nor updatable parameters. We call these con-
stant parameters. Say for example that we want a layer that calculates the function f (x, w) = c·w⊤ x,
where x is the input, w is our parameter, and c is some specified constant that is not updated dur-
ing optimization. So we implement a FixedHiddenMLP class as follows.
class FixedHiddenMLP(nn.Block):
# Random weight parameters created with the `get_constant` function
# are not updated during training (i.e., constant parameters)
self.rand_weight = self.params.get_constant(
'rand_weight', np.random.uniform(size=(20, 20)))
self.dense = nn.Dense(20, activation='relu')

X = self.dense(X)
# Use the created constant parameters, as well as the `relu` and `dot`
# functions
X = npx.relu(np.dot(X, self.rand_weight.data()) + 1)
# Reuse the fully-connected layer. This is equivalent to sharing
# parameters with two fully-connected layers
X = self.dense(X)
# Control flow
while np.abs(X).sum() > 1:
X /= 2
return X.sum()
In this FixedHiddenMLP model, we implement a hidden layer whose weights (self.rand_weight)

are initialized randomly at instantiation and are thereafter constant. This weight is not a model
parameter and thus it is never updated by backpropagation. The network then passes the output
of this “fixed” layer through a fully-connected layer.
Note that before returning the output, our model did something unusual. We ran a while-loop,
testing on the condition its L1 norm is larger than 1, and dividing our output vector by 2 until it
satisfied the condition. Finally, we returned the sum of the entries in X. To our knowledge, no
standard neural network performs this operation. Note that this particular operation may not be
useful in any real-world task. Our point is only to show you how to integrate arbitrary code into
the flow of your neural network computations.
net = FixedHiddenMLP()
net.initialize()
net(X)

array(0.52637565)
We can mix and match various ways of assembling blocks together. In the following example, we
nest blocks in some creative ways.
class NestMLP(nn.Block):
self.net = nn.Sequential()
self.net.add(nn.Dense(64, activation='relu'),
nn.Dense(32, activation='relu'))
self.dense = nn.Dense(16, activation='relu')

return self.dense(self.net(X))
chimera = nn.Sequential()
chimera.add(NestMLP(), nn.Dense(20), FixedHiddenMLP())
chimera.initialize()
chimera(X)
array(0.97720534)
5.1.4 Compilation
The avid reader might start to worry about the efficiency of some of these operations. After all, we
have lots of dictionary lookups, code execution, and lots of other Pythonic things taking place in
what is supposed to be a high-performance deep learning library. The problems of Pythonʼs global
interpreter lock79 are well known. In the context of deep learning, we worry that our extremely
fast GPU(s) might have to wait until a puny CPU runs Python code before it gets another job to run.
The best way to speed up Python is by avoiding it altogether.
One way that Gluon does this is by allowing for hybridization, which will be described later. Here,
the Python interpreter executes a block the first time it is invoked. The Gluon runtime records
what is happening and the next time around it short-circuits calls to Python. This can accelerate
things considerably in some cases but care needs to be taken when control flow (as above) leads
down different branches on different passes through the net. We recommend that the interested
reader checks out the hybridization section (Section 12.1) to learn about compilation after finish-
ing the current chapter.
79
https://wiki.python.org/moin/GlobalInterpreterLock

Summary
• Layers are blocks.

• Many layers can comprise a block.
• Many blocks can comprise a block.
• A block can contain code.
• Blocks take care of lots of housekeeping, including parameter initialization and backpropa-
gation.
• Sequential concatenations of layers and blocks are handled by the Sequential block.
Exercises
1. What kinds of problems will occur if you change MySequential to store blocks in a Python
list?
2. Implement a block that takes two blocks as an argument, say net1 and net2 and returns
the concatenated output of both networks in the forward propagation. This is also called a
parallel block.
3. Assume that you want to concatenate multiple instances of the same network. Implement
a factory function that generates multiple instances of the same block and build a larger
network from it.
Discussions80
5.2 Parameter Management
Once we have chosen an architecture and set our hyperparameters, we proceed to the training
loop, where our goal is to find parameter values that minimize our loss function. After training, we
will need these parameters in order to make future predictions. Additionally, we will sometimes
wish to extract the parameters either to reuse them in some other context, to save our model
to disk so that it may be executed in other software, or for examination in the hope of gaining
scientific understanding.
Most of the time, we will be able to ignore the nitty-gritty details of how parameters are declared
and manipulated, relying on deep learning frameworks to do the heavy lifting. However, when we
move away from stacked architectures with standard layers, we will sometimes need to get into
the weeds of declaring and manipulating parameters. In this section, we cover the following:
• Accessing parameters for debugging, diagnostics, and visualizations.
• Parameter initialization.
• Sharing parameters across different model components.
We start by focusing on an MLP with one hidden layer.
80

from mxnet import init, np, npx
npx.set_np()
net.initialize() # Use the default initialization method
net(X) # Forward computation
array([[0.0054572 ],
[0.00488594]])
5.2.1 Parameter Access
Let us start with how to access parameters from the models that you already know. When a model
is defined via the Sequential class, we can first access any layer by indexing into the model as
though it were a list. Each layerʼs parameters are conveniently located in its attribute. We can
inspect the parameters of the second fully-connected layer as follows.
print(net[1].params)
dense1_ (
Parameter dense1_weight (shape=(1, 8), dtype=float32)
Parameter dense1_bias (shape=(1,), dtype=float32)
)
The output tells us a few important things. First, this fully-connected layer contains two parame-
ters, corresponding to that layerʼs weights and biases, respectively. Both are stored as single pre-
cision floats (float32). Note that the names of the parameters allow us to uniquely identify each
layerʼs parameters, even in a network containing hundreds of layers.
Targeted Parameters
Note that each parameter is represented as an instance of the parameter class. To do anything
useful with the parameters, we first need to access the underlying numerical values. There are
several ways to do this. Some are simpler while others are more general. The following code
extracts the bias from the second neural network layer, which returns a parameter class instance,
and further accesses that parameterʼs value.
print(type(net[1].bias))
print(net[1].bias)
print(net[1].bias.data())
<class 'mxnet.gluon.parameter.Parameter'>
[0.]
5.2. Parameter Management 207

Parameters are complex objects, containing values, gradients, and additional information. Thatʼs
why we need to request the value explicitly.
In addition to the value, each parameter also allows us to access the gradient. Because we have
not invoked backpropagation for this network yet, it is in its initial state.
net[1].weight.grad()
array([[0., 0., 0., 0., 0., 0., 0., 0.]])
All Parameters at Once
When we need to perform operations on all parameters, accessing them one-by-one can grow
tedious. The situation can grow especially unwieldy when we work with more complex blocks
(e.g., nested blocks), since we would need to recurse through the entire tree to extract each sub-
blockʼs parameters. Below we demonstrate accessing the parameters of the first fully-connected
layer vs. accessing all layers.
print(net[0].collect_params())
print(net.collect_params())
dense0_ (
)
sequential0_ (
)
This provides us with another way of accessing the parameters of the network as follows.
net.collect_params()['dense1_bias'].data()
array([0.])
Collecting Parameters from Nested Blocks
Let us see how the parameter naming conventions work if we nest multiple blocks inside each
other. For that we first define a function that produces blocks (a block factory, so to speak) and
then combine these inside yet larger blocks.
def block1():

return net
def block2():
for _ in range(4):
# Nested here
net.add(block1())
return net
rgnet = nn.Sequential()
rgnet.add(block2())
rgnet.add(nn.Dense(10))
rgnet.initialize()
rgnet(X)
array([[-6.3465841e-09, -1.1096747e-09, 6.4161774e-09, 6.6354131e-09,

-1.1265502e-09, 1.3285141e-10, 9.3619361e-09, 3.2229091e-09,
5.9429857e-09, 8.8181418e-09],
[-8.6219423e-09, -7.5150813e-10, 8.3133243e-09, 8.9321119e-09,
-1.6739999e-09, 3.2406069e-10, 1.2115976e-08, 4.4926454e-09,
8.0741742e-09, 1.2075874e-08]])
Now that we have designed the network, let us see how it is organized.
print(rgnet.collect_params)
print(rgnet.collect_params())
<bound method Block.collect_params of Sequential(

(0): Sequential(
(0): Sequential(
(0): Dense(4 -> 32, Activation(relu))
)
(1): Sequential(
)
(2): Sequential(
)
(3): Sequential(
)
)
(1): Dense(16 -> 10, linear)
)>
sequential1_ (

)
Since the layers are hierarchically nested, we can also access them as though indexing through
nested lists. For instance, we can access the first major block, within it the second sub-block, and
within that the bias of the first layer, with as follows.
rgnet[0][1][0].bias.data()
array([0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.])
5.2.2 Parameter Initialization
Now that we know how to access the parameters, let us look at how to initialize them properly.
We discussed the need for proper initialization in Section 4.8. The deep learning framework pro-
vides default random initializations to its layers. However, we often want to initialize our weights
according to various other protocols. The framework provides most commonly used protocols,
and also allows to create a custom initializer.
By default, MXNet initializes weight parameters by randomly drawing from a uniform distribution
U (−0.07, 0.07), clearing bias parameters to zero. MXNetʼs init module provides a variety of preset
initialization methods.
Built-in Initialization
Let us begin by calling on built-in initializers. The code below initializes all weight parameters as
Gaussian random variables with standard deviation 0.01, while bias parameters cleared to zero.
# Here `force_reinit` ensures that parameters are freshly initialized even if

# they were already initialized previously
net.initialize(init=init.Normal(sigma=0.01), force_reinit=True)
net[0].weight.data()[0]
array([-0.00324057, -0.00895028, -0.00698632, 0.01030831])

We can also initialize all the parameters to a given constant value (say, 1).
net.initialize(init=init.Constant(1), force_reinit=True)
array([1., 1., 1., 1.])
We can also apply different initializers for certain blocks. For example, below we initialize the
first layer with the Xavier initializer and initialize the second layer to a constant value of 42.
net[0].weight.initialize(init=init.Xavier(), force_reinit=True)
net[1].initialize(init=init.Constant(42), force_reinit=True)
print(net[0].weight.data()[0])
print(net[1].weight.data())
[-0.17594433 0.02314097 -0.1992535 0.09509248]

[[42. 42. 42. 42. 42. 42. 42. 42.]]
Custom Initialization
Sometimes, the initialization methods we need are not provided by the deep learning framework.
In the example below, we define an initializer for any weight parameter w using the following
strange distribution:


U (5, 10) with probability 14
w∼ 0 with probability 12 (5.2.1)


U (−10, −5) with probability 4 1
Here we define a subclass of the Initializer class. Usually, we only need to implement the
_init_weight function which takes a tensor argument (data) and assigns to it the desired initial-
ized values.
class MyInit(init.Initializer):
def _init_weight(self, name, data):
print('Init', name, data.shape)
data[:] = np.random.uniform(-10, 10, data.shape)
data *= np.abs(data) >= 5
net.initialize(MyInit(), force_reinit=True)
net[0].weight.data()[:2]
Init dense0_weight (8, 4)

Init dense1_weight (1, 8)
array([[ 0. , -0. , -0. , 8.522827 ],

[ 0. , -8.828651 , -0. , -5.6012006]])
Note that we always have the option of setting parameters directly.

net[0].weight.data()[:] += 1
net[0].weight.data()[0, 0] = 42
array([42. , 1. , 1. , 9.522827])
A note for advanced users: if you want to adjust parameters within an autograd scope, you need
to use set_data to avoid confusing the automatic differentiation mechanics.
5.2.3 Tied Parameters
Often, we want to share parameters across multiple layers. Let us see how to do this elegantly.
In the following we allocate a dense layer and then use its parameters specifically to set those of
another layer.
# We need to give the shared layer a name so that we can refer to its
# parameters
shared = nn.Dense(8, activation='relu')
shared,
nn.Dense(8, activation='relu', params=shared.params),
nn.Dense(10))
net.initialize()
net(X)
# Check whether the parameters are the same

print(net[1].weight.data()[0] == net[2].weight.data()[0])
net[1].weight.data()[0, 0] = 100
# Make sure that they are actually the same object rather than just having the
# same value
print(net[1].weight.data()[0] == net[2].weight.data()[0])
[ True True True True True True True True]

[ True True True True True True True True]
This example shows that the parameters of the second and third layer are tied. They are not just
equal, they are represented by the same exact tensor. Thus, if we change one of the parameters,
the other one changes, too. You might wonder, when parameters are tied what happens to the
gradients? Since the model parameters contain gradients, the gradients of the second hidden
layer and the third hidden layer are added together during backpropagation.

Summary
• We have several ways to access, initialize, and tie model parameters.

• We can use custom initialization.
Exercises
1. Use the FancyMLP model defined in Section 5.1 and access the parameters of the various lay-
ers.
2. Look at the initialization module document to explore different initializers.
3. Construct an MLP containing a shared parameter layer and train it. During the training
process, observe the model parameters and gradients of each layer.
4. Why is sharing parameters a good idea?
Discussions81
5.3 Deferred Initialization
So far, it might seem that we got away with being sloppy in setting up our networks. Specifically,
we did the following unintuitive things, which might not seem like they should work:
• We defined the network architectures without specifying the input dimensionality.
• We added layers without specifying the output dimension of the previous layer.
• We even “initialized” these parameters before providing enough information to determine
how many parameters our models should contain.
You might be surprised that our code runs at all. After all, there is no way the deep learning
framework could tell what the input dimensionality of a network would be. The trick here is that
the framework defers initialization, waiting until the first time we pass data through the model, to
infer the sizes of each layer on the fly.
Later on, when working with convolutional neural networks, this technique will become even
more convenient since the input dimensionality (i.e., the resolution of an image) will affect the
dimensionality of each subsequent layer. Hence, the ability to set parameters without the need
to know, at the time of writing the code, what the dimensionality is can greatly simplify the task
of specifying and subsequently modifying our models. Next, we go deeper into the mechanics of
initialization.
81
5.3. Deferred Initialization 213

5.3.1 Instantiating a Network
To begin, let us instantiate an MLP.
from mxnet import init, np, npx

npx.set_np()
def get_net():
return net
net = get_net()
At this point, the network cannot possibly know the dimensions of the input layerʼs weights be-
cause the input dimension remains unknown. Consequently the framework has not yet initialized
any parameters. We confirm by attempting to access the parameters below.
print(net.collect_params)
print(net.collect_params())
<bound method Block.collect_params of Sequential(

(0): Dense(-1 -> 256, Activation(relu))
(1): Dense(-1 -> 10, linear)
)>
sequential0_ (
Parameter dense0_weight (shape=(256, -1), dtype=float32)
)
Note that while the parameter objects exist, the input dimension to each layer is listed as -1. MXNet
uses the special value -1 to indicate that the parameter dimension remains unknown. At this point,
attempts to access net[0].weight.data() would trigger a runtime error stating that the network
must be initialized before the parameters can be accessed. Now let us see what happens when we
attempt to initialize parameters via the initialize function.
net.initialize()
net.collect_params()
sequential0_ (
)
As we can see, nothing has changed. When input dimensions are unknown, calls to initialize do
not truly initialize the parameters. Instead, this call registers to MXNet that we wish (and option-
ally, according to which distribution) to initialize the parameters.

Next let us pass data through the network to make the framework finally initialize parameters.
net(X)
net.collect_params()
sequential0_ (
)
As soon as we know the input dimensionality, 20, the framework can identify the shape of the
first layerʼs weight matrix by plugging in the value of 20. Having recognized the first layerʼs shape,
the framework proceeds to the second layer, and so on through the computational graph until all
shapes are known. Note that in this case, only the first layer requires deferred initialization, but
the framework initializes sequentially. Once all parameter shapes are known, the framework can
finally initialize the parameters.
Summary
• Deferred initialization can be convenient, allowing the framework to infer parameter shapes
automatically, making it easy to modify architectures and eliminating one common source
of errors.
• We can pass data through the model to make the framework finally initialize parameters.
Exercises
1. What happens if you specify the input dimensions to the first layer but not to subsequent
layers? Do you get immediate initialization?
2. What happens if you specify mismatching dimensions?
3. What would you need to do if you have input of varying dimensionality? Hint: look at the
parameter tying.
Discussions82
82
5.3. Deferred Initialization 215

5.4 Custom Layers
One factor behind deep learningʼs success is the availability of a wide range of layers that can be
composed in creative ways to design architectures suitable for a wide variety of tasks. For instance,
researchers have invented layers specifically for handling images, text, looping over sequential
data, and performing dynamic programming. Sooner or later, you will encounter or invent a layer
that does not exist yet in the deep learning framework. In these cases, you must build a custom
layer. In this section, we show you how.
5.4.1 Layers without Parameters
To start, we construct a custom layer that does not have any parameters of its own. This should
look familiar if you recall our introduction to block in Section 5.1. The following CenteredLayer
class simply subtracts the mean from its input. To build it, we simply need to inherit from the
base layer class and implement the forward propagation function.

npx.set_np()
class CenteredLayer(nn.Block):

return X - X.mean()
Let us verify that our layer works as intended by feeding some data through it.
layer = CenteredLayer()
layer(np.array([1, 2, 3, 4, 5]))
array([-2., -1., 0., 1., 2.])
We can now incorporate our layer as a component in constructing more complex models.
net.add(nn.Dense(128), CenteredLayer())
net.initialize()
As an extra sanity check, we can send random data through the network and check that the mean
is in fact 0. Because we are dealing with floating point numbers, we may still see a very small
nonzero number due to quantization.
Y = net(np.random.uniform(size=(4, 8)))
Y.mean()
array(3.783498e-10)

5.4.2 Layers with Parameters
Now that we know how to define simple layers, let us move on to defining layers with parameters
that can be adjusted through training. We can use built-in functions to create parameters, which
provide some basic housekeeping functionality. In particular, they govern access, initialization,
sharing, saving, and loading model parameters. This way, among other benefits, we will not need
to write custom serialization routines for every custom layer.
Now let us implement our own version of the fully-connected layer. Recall that this layer requires
two parameters, one to represent the weight and the other for the bias. In this implementation,
we bake in the ReLU activation as a default. This layer requires to input arguments: in_units and
units, which denote the number of inputs and outputs, respectively.
class MyDense(nn.Block):
def __init__(self, units, in_units, **kwargs):
self.weight = self.params.get('weight', shape=(in_units, units))
self.bias = self.params.get('bias', shape=(units,))
def forward(self, x):

linear = np.dot(x, self.weight.data(ctx=x.ctx)) + self.bias.data(
ctx=x.ctx)
return npx.relu(linear)
Next, we instantiate the MyDense class and access its model parameters.
dense = MyDense(units=3, in_units=5)

dense.params
mydense0_ (
Parameter mydense0_weight (shape=(5, 3), dtype=<class 'numpy.float32'>)
Parameter mydense0_bias (shape=(3,), dtype=<class 'numpy.float32'>)
)
We can directly carry out forward propagation calculations using custom layers.
dense.initialize()
dense(np.random.uniform(size=(2, 5)))
array([[0. , 0.01633355, 0. ],
[0. , 0.01581812, 0. ]])
We can also construct models using custom layers. Once we have that we can use it just like the
built-in fully-connected layer.
net.add(MyDense(8, in_units=64),
MyDense(1, in_units=8))
net.initialize()
net(np.random.uniform(size=(2, 64)))
5.4. Custom Layers 217

array([[0.06508517],
[0.0615553 ]])
Summary
• We can design custom layers via the basic layer class. This allows us to define flexible new
layers that behave differently from any existing layers in the library.
• Once defined, custom layers can be invoked in arbitrary contexts and architectures.
• Layers can have local parameters, which can be created through built-in functions.
Exercises
1. Design
∑ a layer that takes an input and computes a tensor reduction, i.e., it returns yk =
i,j W ijk xi xj .
2. Design a layer that returns the leading half of the Fourier coefficients of the data.
Discussions83
5.5 File I/O
So far we discussed how to process data and how to build, train, and test deep learning models.
However, at some point, we will hopefully be happy enough with the learned models that we will
want to save the results for later use in various contexts (perhaps even to make predictions in de-
ployment). Additionally, when running a long training process, the best practice is to periodically
save intermediate results (checkpointing) to ensure that we do not lose several days worth of com-
putation if we trip over the power cord of our server. Thus it is time to learn how to load and store
both individual weight vectors and entire models. This section addresses both issues.
5.5.1 Loading and Saving Tensors
For individual tensors, we can directly invoke the load and save functions to read and write them
respectively. Both functions require that we supply a name, and save requires as input the variable
to be saved.

npx.set_np()
x = np.arange(4)
npx.save('x-file', x)
We can now read the data from the stored file back into memory.
83

x2 = npx.load('x-file')
x2
[array([0., 1., 2., 3.])]
We can store a list of tensors and read them back into memory.
y = np.zeros(4)
npx.save('x-files', [x, y])
x2, y2 = npx.load('x-files')
(x2, y2)
(array([0., 1., 2., 3.]), array([0., 0., 0., 0.]))
We can even write and read a dictionary that maps from strings to tensors. This is convenient
when we want to read or write all the weights in a model.
mydict = {'x': x, 'y': y}

npx.save('mydict', mydict)
mydict2 = npx.load('mydict')
mydict2
{'x': array([0., 1., 2., 3.]), 'y': array([0., 0., 0., 0.])}
5.5.2 Loading and Saving Model Parameters
Saving individual weight vectors (or other tensors) is useful, but it gets very tedious if we want
to save (and later load) an entire model. After all, we might have hundreds of parameter groups
sprinkled throughout. For this reason the deep learning framework provides built-in function-
alities to load and save entire networks. An important detail to note is that this saves model pa-
rameters and not the entire model. For example, if we have a 3-layer MLP, we need to specify the
architecture separately. The reason for this is that the models themselves can contain arbitrary
code, hence they cannot be serialized as naturally. Thus, in order to reinstate a model, we need
to generate the architecture in code and then load the parameters from disk. Let us start with our
familiar MLP.
class MLP(nn.Block):
super(MLP, self).__init__(**kwargs)
self.hidden = nn.Dense(256, activation='relu')
self.output = nn.Dense(10)

return self.output(self.hidden(x))
net = MLP()
net.initialize()
Y = net(X)
5.5. File I/O 219

Next, we store the parameters of the model as a file with the name “mlp.params”.
net.save_parameters('mlp.params')
To recover the model, we instantiate a clone of the original MLP model. Instead of randomly
initializing the model parameters, we read the parameters stored in the file directly.
clone = MLP()
clone.load_parameters('mlp.params')
Since both instances have the same model parameters, the computational result of the same input
X should be the same. Let us verify this.
Y_clone = clone(X)
Y_clone == Y
array([[ True, True, True, True, True, True, True, True, True,
True],
[ True, True, True, True, True, True, True, True, True,
True]])
Summary
• The save and load functions can be used to perform file I/O for tensor objects.
• We can save and load the entire sets of parameters for a network via a parameter dictionary.
• Saving the architecture has to be done in code rather than in parameters.
Exercises
1. Even if there is no need to deploy trained models to a different device, what are the practical
benefits of storing model parameters?
2. Assume that we want to reuse only parts of a network to be incorporated into a network
of a different architecture. How would you go about using, say the first two layers from a
previous network in a new network?
3. How would you go about saving the network architecture and parameters? What restrictions
would you impose on the architecture?
Discussions84
84

5.6 GPUs
In Table 1.5.1, we discussed the rapid growth of computation over the past two decades. In a
nutshell, GPU performance has increased by a factor of 1000 every decade since 2000. This offers
great opportunities but it also suggests a significant need to provide such performance.
In this section, we begin to discuss how to harness this computational performance for your re-
search. First by using single GPUs and at a later point, how to use multiple GPUs and multiple
servers (with multiple GPUs).
Specifically, we will discuss how to use a single NVIDIA GPU for calculations. First, make sure
you have at least one NVIDIA GPU installed. Then, download the NVIDIA driver and CUDA85 and
follow the prompts to set the appropriate path. Once these preparations are complete, the nvidia-
smi command can be used to view the graphics card information.
!nvidia-smi
Sun Aug 9 22:24:36 2020

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 418.67 Driver Version: 418.67 CUDA Version: 10.1 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
|===============================+======================+======================|
| 0 Tesla V100-SXM2... Off | 00000000:00:1B.0 Off | 0 |
| N/A 45C P0 51W / 300W | 0MiB / 16130MiB | 0% Default |
+-------------------------------+----------------------+----------------------+
| 1 Tesla V100-SXM2... Off | 00000000:00:1C.0 Off | 0 |
+-------------------------------+----------------------+----------------------+
| 2 Tesla V100-SXM2... Off | 00000000:00:1D.0 Off | 0 |
+-------------------------------+----------------------+----------------------+
| 3 Tesla V100-SXM2... Off | 00000000:00:1E.0 Off | 0 |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: GPU Memory |
| GPU PID Type Process name Usage |
|=============================================================================|
| 2 47745 C ...iconda3/envs/d2l-en-master-1/bin/python 2425MiB |
+-----------------------------------------------------------------------------+
You might have noticed that a MXNet tensor looks almost identical to a NumPy ndarray. But there
are a few crucial differences. One of the key features that distinguishes MXNet from NumPy is its
support for diverse hardware devices.
In MXNet, every array has a context. So far, by default, all variables and associated computation
have been assigned to the CPU. Typically, other contexts might be various GPUs. Things can get
even hairier when we deploy jobs across multiple servers. By assigning arrays to contexts intel-
ligently, we can minimize the time spent transferring data between devices. For example, when
85
5.6. GPUs 221

training neural networks on a server with a GPU, we typically prefer for the modelʼs parameters
to live on the GPU.
Next, we need to confirm that the GPU version of MXNet is installed. If a CPU version of MXNet
is already installed, we need to uninstall it first. For example, use the pip uninstall mxnet com-
mand, then install the corresponding MXNet version according to your CUDA version. Assuming
you have CUDA 10.0 installed, you can install the MXNet version that supports CUDA 10.0 via pip
install mxnet-cu100.
To run the programs in this section, you need at least two GPUs. Note that this might be extravagant
for most desktop computers but it is easily available in the cloud, e.g., by using the AWS EC2 multi-
GPU instances. Almost all other sections do not require multiple GPUs. Instead, this is simply to
illustrate how data flow between different devices.
5.6.1 Computing Devices
We can specify devices, such as CPUs and GPUs, for storage and calculation. By default, tensors
are created in the main memory and then use the CPU to calculate it.
In MXNet, the CPU and GPU can be indicated by cpu() and gpu(). It should be noted that cpu() (or
any integer in the parentheses) means all physical CPUs and memory. This means that MXNetʼs
calculations will try to use all CPU cores. However, gpu() only represents one card and the cor-
responding memory. If there are multiple GPUs, we use gpu(i) to represent the ith GPU (i starts
from 0). Also, gpu(0) and gpu() are equivalent.

npx.set_np()
npx.cpu(), npx.gpu(), npx.gpu(1)
(cpu(0), gpu(0), gpu(1))
We can query the number of available GPUs.
npx.num_gpus()
Now we define two convenient functions that allow us to run code even if the requested GPUs do
not exist.
def try_gpu(i=0): #@save

"""Return gpu(i) if exists, otherwise return cpu()."""
return npx.gpu(i) if npx.num_gpus() >= i + 1 else npx.cpu()
def try_all_gpus(): #@save

"""Return all available GPUs, or [cpu()] if no GPU exists."""
devices = [npx.gpu(i) for i in range(npx.num_gpus())]
return devices if devices else [npx.cpu()]
try_gpu(), try_gpu(10), try_all_gpus()

(gpu(0), cpu(0), [gpu(0), gpu(1)])
5.6.2 Tensors and GPUs
By default, tensors are created on the CPU. We can query the device where the tensor is located.
x = np.array([1, 2, 3])
x.ctx
cpu(0)
It is important to note that whenever we want to operate on multiple terms, they need to be on the
same device. For instance, if we sum two tensors, we need to make sure that both arguments live
on the same device—otherwise the framework would not know where to store the result or even
how to decide where to perform the computation.
Storage on the GPU
There are several ways to store a tensor on the GPU. For example, we can specify a storage device
when creating a tensor. Next, we create the tensor variable X on the first gpu. The tensor created
on a GPU only consumes the memory of this GPU. We can use the nvidia-smi command to view
GPU memory usage. In general, we need to make sure that we do not create data that exceed the
GPU memory limit.
X = np.ones((2, 3), ctx=try_gpu())

X
array([[1., 1., 1.],

[1., 1., 1.]], ctx=gpu(0))
Assuming that you have at least two GPUs, the following code will create a random tensor on the
second GPU.
Y = np.random.uniform(size=(2, 3), ctx=try_gpu(1))

Y
array([[0.67478997, 0.07540122, 0.9956977 ],

[0.09488854, 0.415456 , 0.11231736]], ctx=gpu(1))
5.6. GPUs 223

Copying
If we want to compute X + Y, we need to decide where to perform this operation. For instance,
as shown in Fig. 5.6.1, we can transfer X to the second GPU and perform the operation there. Do
not simply add X and Y, since this will result in an exception. The runtime engine would not know
what to do: it cannot find data on the same device and it fails. Since Y lives on the second GPU, we
need to move X there before we can add the two.
Fig. 5.6.1: Copy data to perform an operation on the same device.
Z = X.copyto(try_gpu(1))
print(X)
print(Z)
[[1. 1. 1.]
[1. 1. 1.]] @gpu(0)
[[1. 1. 1.]
[1. 1. 1.]] @gpu(1)
Now that the data are on the same GPU (both Z and Y are), we can add them up.
Y + Z
array([[1.6747899, 1.0754012, 1.9956977],

[1.0948886, 1.415456 , 1.1123173]], ctx=gpu(1))
Imagine that your variable Z already lives on your second GPU. What happens if we still call Z.
copyto(gpu(1))? It will make a copy and allocate new memory, even though that variable already
lives on the desired device. There are times where, depending on the environment our code is
running in, two variables may already live on the same device. So we want to make a copy only
if the variables currently live in different devices. In these cases, we can call as_in_ctx. If the
variable already live in the specified device then this is a no-op. Unless you specifically want to
make a copy, as_in_ctx is the method of choice.
Z.as_in_ctx(try_gpu(1)) is Z
True

Side Notes
People use GPUs to do machine learning because they expect them to be fast. But transferring
variables between devices is slow. So we want you to be 100% certain that you want to do some-
thing slow before we let you do it. If the deep learning framework just did the copy automatically
without crashing then you might not realize that you had written some slow code.
Also, transferring data between devices (CPU, GPUs, and other machines) is something that is
much slower than computation. It also makes parallelization a lot more difficult, since we have to
wait for data to be sent (or rather to be received) before we can proceed with more operations. This
is why copy operations should be taken with great care. As a rule of thumb, many small operations
are much worse than one big operation. Moreover, several operations at a time are much better
than many single operations interspersed in the code unless you know what you are doing. This
is the case since such operations can block if one device has to wait for the other before it can do
something else. It is a bit like ordering your coffee in a queue rather than pre-ordering it by phone
and finding out that it is ready when you are.
Last, when we print tensors or convert tensors to the NumPy format, if the data is not in the main
memory, the framework will copy it to the main memory first, resulting in additional transmis-
sion overhead. Even worse, it is now subject to the dreaded global interpreter lock that makes
everything wait for Python to complete.
5.6.3 Neural Networks and GPUs
Similarly, a neural network model can specify devices. The following code puts the model param-
eters on the GPU.
net.initialize(ctx=try_gpu())
We will see many more examples of how to run models on GPUs in the following chapters, simply
since they will become somewhat more computationally intensive.
When the input is a tensor on the GPU, the model will calculate the result on the same GPU.
net(X)
array([[0.04995865],
[0.04995865]], ctx=gpu(0))
Let us confirm that the model parameters are stored on the same GPU.
net[0].weight.data().ctx
gpu(0)
In short, as long as all data and parameters are on the same device, we can learn models efficiently.
In the following chapters we will see several such examples.
5.6. GPUs 225

Summary
• We can specify devices for storage and calculation, such as the CPU or GPU. By default, data
are created in the main memory and then use the CPU for calculations.
• The deep learning framework requires all input data for calculation to be on the same device,
be it CPU or the same GPU.
• You can lose significant performance by moving data without care. A typical mistake is as
follows: computing the loss for every minibatch on the GPU and reporting it back to the user
on the command line (or logging it in a NumPy ndarray) will trigger a global interpreter lock
which stalls all GPUs. It is much better to allocate memory for logging inside the GPU and
only move larger logs.
Exercises
1. Try a larger computation task, such as the multiplication of large matrices, and see the dif-
ference in speed between the CPU and GPU. What about a task with a small amount of cal-
culations?
2. How should we read and write model parameters on the GPU?
3. Measure the time it takes to compute 1000 matrix-matrix multiplications of 100 × 100 matri-
ces and log the Frobenius norm of the output matrix one result at a time vs. keeping a log on
the GPU and transferring only the final result.
4. Measure how much time it takes to perform two matrix-matrix multiplications on two GPUs
at the same time vs. in sequence on one GPU. Hint: you should see almost linear scaling.
Discussions86
86

6 | Convolutional Neural Networks
In earlier chapters, we came up against image data, for which each example consists of a two-
dimensional grid of pixels. Depending on whether we are handling black-and-white or color im-
ages, each pixel location might be associated with either one or multiple numerical values, respec-
tively. Until now, our way of dealing with this rich structure was deeply unsatisfying. We simply
discarded each imageʼs spatial structure by flattening them into one-dimensional vectors, feeding
them through a fully-connected MLP. Because these networks are invariant to the order of the fea-
tures, we could get similar results regardless of whether we preserve an order corresponding to
the spatial structure of the pixels or if we permute the columns of our design matrix before fitting
the MLPʼs parameters. Preferably, we would leverage our prior knowledge that nearby pixels are
typically related to each other, to build efficient models for learning from image data.
This chapter introduces convolutional neural networks (CNNs), a powerful family of neural networks
that are designed for precisely this purpose. CNN-based architectures are now ubiquitous in the
field of computer vision, and have become so dominant that hardly anyone today would develop a
commercial application or enter a competition related to image recognition, object detection, or
semantic segmentation, without building off of this approach.
Modern CNNs, as they are called colloquially owe their design to inspirations from biology, group
theory, and a healthy dose of experimental tinkering. In addition to their sample efficiency in
achieving accurate models, CNNs tend to be computationally efficient, both because they require
fewer parameters than fully-connected architectures and because convolutions are easy to par-
allelize across GPU cores. Consequently, practitioners often apply CNNs whenever possible, and
increasingly they have emerged as credible competitors even on tasks with a one-dimensional se-
quence structure, such as audio, text, and time series analysis, where recurrent neural networks
are conventionally used. Some clever adaptations of CNNs have also brought them to bear on
graph-structured data and in recommender systems.
First, we will walk through the basic operations that comprise the backbone of all convolu-
tional networks. These include the convolutional layers themselves, nitty-gritty details includ-
ing padding and stride, the pooling layers used to aggregate information across adjacent spatial
regions, the use of multiple channels at each layer, and a careful discussion of the structure of
modern architectures. We will conclude the chapter with a full working example of LeNet, the
first convolutional network successfully deployed, long before the rise of modern deep learning.
In the next chapter, we will dive into full implementations of some popular and comparatively
recent CNN architectures whose designs represent most of the techniques commonly used by
modern practitioners.
227
6.1 From Fully-Connected Layers to Convolutions
To this day, the models that we have discussed so far remain appropriate options when we are
dealing with tabular data. By tabular, we mean that the data consist of rows corresponding to
examples and columns corresponding to features. With tabular data, we might anticipate that
the patterns we seek could involve interactions among the features, but we do not assume any
structure a priori concerning how the features interact.
Sometimes, we truly lack knowledge to guide the construction of craftier architectures. In these
cases, an MLP may be the best that we can do. However, for high-dimensional perceptual data,
such structure-less networks can grow unwieldy.
For instance, let us return to our running example of distinguishing cats from dogs. Say that we do
a thorough job in data collection, collecting an annotated dataset of one-megapixel photographs.
This means that each input to the network has one million dimensions. Even an aggressive reduc-
tion to one thousand hidden dimensions would require a fully-connected layer characterized by
106 × 103 = 109 parameters. Unless we have lots of GPUs, a talent for distributed optimization,
and an extraordinary amount of patience, learning the parameters of this network may turn out
to be infeasible.
A careful reader might object to this argument on the basis that one megapixel resolution may not
be necessary. However, while we might be able to get away with one hundred thousand pixels,
our hidden layer of size 1000 grossly underestimates the number of hidden units that it takes to
learn good representations of images, so a practical system will still require billions of parame-
ters. Moreover, learning a classifier by fitting so many parameters might require collecting an
enormous dataset. And yet today both humans and computers are able to distinguish cats from
dogs quite well, seemingly contradicting these intuitions. That is because images exhibit rich
structure that can be exploited by humans and machine learning models alike. Convolutional
neural networks (CNNs) are one creative way that machine learning has embraced for exploiting
some of the known structure in natural images.
6.1.1 Invariance
Imagine that you want to detect an object in an image. It seems reasonable that whatever method
we use to recognize objects should not be overly concerned with the precise location of the ob-
ject in the image. Ideally, our system should exploit this knowledge. Pigs usually do not fly and
planes usually do not swim. Nonetheless, we should still recognize a pig were one to appear at the
top of the image. We can draw some inspiration here from the childrenʼs game “Whereʼs Waldo”
(depicted in Fig. 6.1.1). The game consists of a number of chaotic scenes bursting with activities.
Waldo shows up somewhere in each, typically lurking in some unlikely location. The readerʼs goal
is to locate him. Despite his characteristic outfit, this can be surprisingly difficult, due to the large
number of distractions. However, what Waldo looks like does not depend upon where Waldo is lo-
cated. We could sweep the image with a Waldo detector that could assign a score to each patch,
indicating the likelihood that the patch contains Waldo. CNNs systematize this idea of spatial in-
variance, exploiting it to learn useful representations with fewer parameters.
228 Chapter 6. Convolutional Neural Networks

Fig. 6.1.1: An image of the “Whereʼs Waldo” game.
We can now make these intuitions more concrete by enumerating a few desiderata to guide our
design of a neural network architecture suitable for computer vision:
1. In the earliest layers, our network should respond similarly to the same patch, regardless of
where it appears in the image. This principle is called translation invariance.
2. The earliest layers of the network should focus on local regions, without regard for the con-
tents of the image in distant regions. This is the locality principle. Eventually, these local
representations can be aggregated to make predictions at the whole image level.
Let us see how this translates into mathematics.
6.1.2 Constraining the MLP
To start off, we can consider an MLP with two-dimensional images X as inputs and their imme-
diate hidden representations H similarly represented as matrices in mathematics and as two-
dimensional tensors in code, where both X and H have the same shape. Let that sink in. We now
conceive of not only the inputs but also the hidden representations as possessing spatial structure.
Let [X]i,j and [H]i,j denote the pixel at location (i, j) in the input image and hidden representation,
respectively. Consequently, to have each of the hidden units receive input from each of the input
pixels, we would switch from using weight matrices (as we did previously in MLPs) to represent-
ing our parameters as fourth-order weight tensors W. Suppose that U contains biases, we could
formally express the fully-connected layer as
∑∑
[H]i,j = [U]i,j + [W]i,j,k,l [X]k,l
k l
∑∑ , (6.1.1)
= [U]i,j + [V]i,j,a,b [X]i+a,j+b .
a b
where the switch from W to V is entirely cosmetic for now since there is a one-to-one correspon-
dence between coefficients in both fourth-order tensors. We simply re-index the subscripts (k, l)
such that k = i + a and l = j + b. In other words, we set [V]i,j,a,b = [W]i,j,i+a,j+b . The indices a and
6.1. From Fully-Connected Layers to Convolutions 229

b run over both positive and negative offsets, covering the entire image. For any given location (i,
j) in the hidden representation [H]i,j , we compute its value by summing over pixels in x, centered
around (i, j) and weighted by [V]i,j,a,b .
Translation Invariance
Now let us invoke the first principle established above: translation invariance. This implies that
a shift in the input X should simply lead to a shift in the hidden representation H. This is only
possible if V and U do not actually depend on (i, j), i.e., we have [V]i,j,a,b = [V]a,b and U is a constant,
say u. As a result, we can simplify the definition for H:
∑∑
[H]i,j = u + [V]a,b [X]i+a,j+b . (6.1.2)
a b
This is a convolution! We are effectively weighting pixels at (i + a, j + b) in the vicinity of location

(i, j) with coefficients [V]a,b to obtain the value [H]i,j . Note that [V]a,b needs many fewer coefficients
than [V]i,j,a,b since it no longer depends on the location within the image. We have made significant
progress!
Locality
Now let us invoke the second principle: locality. As motivated above, we believe that we should
not have to look very far away from location (i, j) in order to glean relevant information to assess
what is going on at [H]i,j . This means that outside some range |a| > ∆ or |b| > ∆, we should set
[V]a,b = 0. Equivalently, we can rewrite [H]i,j as
∑
∆ ∑
∆
[H]i,j = u + [V]a,b [X]i+a,j+b . (6.1.3)
a=−∆ b=−∆
Note that (6.1.3), in a nutshell, is a convolutional layer. Convolutional neural networks (CNNs) are
a special family of neural networks that contain convolutional layers. In the deep learning re-
search community, V is referred to as a convolution kernel, a filter, or simply the layerʼs weights that
are often learnable parameters. When the local region is small, the difference as compared with
a fully-connected network can be dramatic. While previously, we might have required billions
of parameters to represent just a single layer in an image-processing network, we now typically
need just a few hundred, without altering the dimensionality of either the inputs or the hidden
representations. The price paid for this drastic reduction in parameters is that our features are
now translation invariant and that our layer can only incorporate local information, when de-
termining the value of each hidden activation. All learning depends on imposing inductive bias.
When that bias agrees with reality, we get sample-efficient models that generalize well to unseen
data. But of course, if those biases do not agree with reality, e.g., if images turned out not to be
translation invariant, our models might struggle even to fit our training data.

6.1.3 Convolutions
Before going further, we should briefly review why the above operation is called a convolution. In
mathematics, the convolution between two functions, say f, g : Rd → R is defined as
∫
(f ∗ g)(x) = f (z)g(x − z)dz. (6.1.4)
That is, we measure the overlap between f and g when one function is “flipped” and shifted by x.
Whenever we have discrete objects, the integral turns into a sum. For instance, for vectors from
the set of square summable infinite dimensional vectors with index running over Z we obtain the
following definition:
∑
(f ∗ g)(i) = f (a)g(i − a). (6.1.5)
a
For two-dimensional tensors, we have a corresponding sum with indices (a, b) for f and (i−a, j−b)
for g, respectively:
∑∑
(f ∗ g)(i, j) = f (a, b)g(i − a, j − b). (6.1.6)
a b
This looks similar to (6.1.3), with one major difference. Rather than using (i+a, j +b), we are using
the difference instead. Note, though, that this distinction is mostly cosmetic since we can always
match the notation between (6.1.3) and (6.1.6). Our original definition in (6.1.3) more properly
describes a cross-correlation. We will come back to this in the following section.
6.1.4 “Where’s Waldo” Revisited
Returning to our Waldo detector, let us see what this looks like. The convolutional layer picks
windows of a given size and weighs intensities according to the filter V, as demonstrated in Fig.
6.1.2. We might aim to learn a model so that wherever the “waldoness” is highest, we should find
a peak in the hidden layer representations.
Fig. 6.1.2: Detect Waldo.
6.1. From Fully-Connected Layers to Convolutions 231

Channels
There is just one problem with this approach. So far, we blissfully ignored that images consist
of 3 channels: red, green, and blue. In reality, images are not two-dimensional objects but rather
third-order tensors, characterized by a height, width, and channel, e.g., with shape 1024×1024×3
pixels. While the first two of these axes concern spatial relationships, the third can be regarded
as assigning a multidimensional representation to each pixel location.
We thus index X as [X]i,j,k . The convolutional filter has to adapt accordingly. Instead of [V]a,b , we
now have [V]a,b,c .
Moreover, just as our input consists of a third-order tensor, it turns out to be a good idea to similarly
formulate our hidden representations as third-order tensors H. In other words, rather than just
having a single hidden representation corresponding to each spatial location, we want an entire
vector of hidden representations corresponding to each spatial location. We could think of the
hidden representations as comprising a number of two-dimensional grids stacked on top of each
other. As in the inputs, these are sometimes called channels. They are also sometimes called feature
maps, as each provides a spatialized set of learned features to the subsequent layer. Intuitively,
you might imagine that at lower layers that are closer to inputs, some channels could become
specialized to recognize edges while others could recognize textures.
To support multiple channels in both inputs (X) and hidden representations (H), we can add a
fourth coordinate to V: [V]a,b,c,d . Putting everything together we have:
∑
∆ ∑
∆ ∑
[H]i,j,d = [V]a,b,c,d [X]i+a,j+b,c , (6.1.7)
a=−∆ b=−∆ c
where d indexes the output channels in the hidden representations H. The subsequent convolu-
tional layer will go on to take a third-order tensor, H, as the input. Being more general, (6.1.7) is
the definition of a convolutional layer for multiple channels, where V is a kernel or filter of the
layer.
There are still many operations that we need to address. For instance, we need to figure out how to
combine all the hidden representations to a single output, e.g., whether there is a Waldo anywhere
in the image. We also need to decide how to compute things efficiently, how to combine multi-
ple layers, appropriate activation functions, and how to make reasonable design choices to yield
networks that are effective in practice. We turn to these issues in the remainder of the chapter.
Summary
• Translation invariance in images implies that all patches of an image will be treated in the
same manner.
• Locality means that only a small neighborhood of pixels will be used to compute the corre-
sponding hidden representations.
• In image processing, convolutional layers typically require many fewer parameters than
fully-connected layers.
• CNNS are a special family of neural networks that contain convolutional layers.
• Channels on input and output allow our model to capture multiple aspects of an image at
each spatial location.

Exercises
1. Assume that the size of the convolution kernel is ∆ = 0. Show that in this case the convolu-
tion kernel implements an MLP independently for each set of channels.
2. Why might translation invariance not be a good idea after all?
3. What problems must we deal with when deciding how to treat hidden representations cor-
responding to pixel locations at the boundary of an image?
4. Describe an analogous convolutional layer for audio.
5. Do you think that convolutional layers might also be applicable for text data? Why or why
not?
6. Prove that f ∗ g = g ∗ f .
Discussions87
6.2 Convolutions for Images
Now that we understand how convolutional layers work in theory, we are ready to see how they
work in practice. Building on our motivation of convolutional neural networks as efficient archi-
tectures for exploring structure in image data, we stick with images as our running example.
6.2.1 The Cross-Correlation Operation
Recall that strictly speaking, convolutional layers are a misnomer, since the operations they ex-
press are more accurately described as cross-correlations. Based on our descriptions of convolu-
tional layers in Section 6.1, in such a layer, an input tensor and a kernel tensor are combined to
produce an output tensor through a cross-correlation operation.
Let us ignore channels for now and see how this works with two-dimensional data and hidden
representations. In Fig. 6.2.1, the input is a two-dimensional tensor with a height of 3 and width
of 3. We mark the shape of the tensor as 3 × 3 or (3, 3). The height and width of the kernel are
both 2. The shape of the kernel window (or convolution window) is given by the height and width of
the kernel (here it is 2 × 2).
Fig. 6.2.1: Two-dimensional cross-correlation operation. The shaded portions are the first output
element as well as the input and kernel tensor elements used for the output computation: 0 × 0 +
1 × 1 + 3 × 2 + 4 × 3 = 19.
In the two-dimensional cross-correlation operation, we begin with the convolution window posi-
tioned at the top-left corner of the input tensor and slide it across the input tensor, both from left
87
6.2. Convolutions for Images 233

to right and top to bottom. When the convolution window slides to a certain position, the input
subtensor contained in that window and the kernel tensor are multiplied elementwise and the re-
sulting tensor is summed up yielding a single scalar value. This result gives the value of the output
tensor at the corresponding location. Here, the output tensor has a height of 2 and width of 2 and
the four elements are derived from the two-dimensional cross-correlation operation:
0 × 0 + 1 × 1 + 3 × 2 + 4 × 3 = 19,
1 × 0 + 2 × 1 + 4 × 2 + 5 × 3 = 25,
(6.2.1)
3 × 0 + 4 × 1 + 6 × 2 + 7 × 3 = 37,
4 × 0 + 5 × 1 + 7 × 2 + 8 × 3 = 43.
Note that along each axis, the output size is slightly smaller than the input size. Because the kernel
has width and height greater than one, we can only properly compute the cross-correlation for
locations where the kernel fits wholly within the image, the output size is given by the input size
nh × nw minus the size of the convolution kernel kh × kw via
(nh − kh + 1) × (nw − kw + 1). (6.2.2)
This is the case since we need enough space to “shift” the convolution kernel across the image.
Later we will see how to keep the size unchanged by padding the image with zeros around its
boundary so that there is enough space to shift the kernel. Next, we implement this process in
the corr2d function, which accepts an input tensor X and a kernel tensor K and returns an output
tensor Y.

npx.set_np()
def corr2d(X, K): #@save

"""Compute 2D cross-correlation."""
h, w = K.shape
Y = np.zeros((X.shape[0] - h + 1, X.shape[1] - w + 1))
for i in range(Y.shape[0]):
for j in range(Y.shape[1]):
Y[i, j] = d2l.reduce_sum((X[i: i + h, j: j + w] * K))
return Y
We can construct the input tensor X and the kernel tensor K from Fig. 6.2.1 to validate the output
of the above implementation of the two-dimensional cross-correlation operation.
X = np.array([[0.0, 1.0, 2.0], [3.0, 4.0, 5.0], [6.0, 7.0, 8.0]])

K = np.array([[0.0, 1.0], [2.0, 3.0]])
corr2d(X, K)
array([[19., 25.],
[37., 43.]])

6.2.2 Convolutional Layers
A convolutional layer cross-correlates the input and kernel and adds a scalar bias to produce an
output. The two parameters of a convolutional layer are the kernel and the scalar bias. When
training models based on convolutional layers, we typically initialize the kernels randomly, just
as we would with a fully-connected layer.
We are now ready to implement a two-dimensional convolutional layer based on the corr2d func-
tion defined above. In the __init__ constructor function, we declare weight and bias as the two
model parameters. The forward propagation function calls the corr2d function and adds the bias.
class Conv2D(nn.Block):
def __init__(self, kernel_size, **kwargs):
self.weight = self.params.get('weight', shape=kernel_size)
self.bias = self.params.get('bias', shape=(1,))

return corr2d(x, self.weight.data()) + self.bias.data()
In h × w convolution or a h × w convolution kernel, the height and width of the convolution kernel
are h and w, respectively. We also refer to a convolutional layer with a h × w convolution kernel
simply as a h × w convolutional layer.
6.2.3 Object Edge Detection in Images
Let us take a moment to parse a simple application of a convolutional layer: detecting the edge of
an object in an image by finding the location of the pixel change. First, we construct an “image”
of 6 × 8 pixels. The middle four columns are black (0) and the rest are white (1).
X = np.ones((6, 8))
X[:, 2:6] = 0
X
array([[1., 1., 0., 0., 0., 0., 1., 1.],

[1., 1., 0., 0., 0., 0., 1., 1.],
[1., 1., 0., 0., 0., 0., 1., 1.],
[1., 1., 0., 0., 0., 0., 1., 1.],
[1., 1., 0., 0., 0., 0., 1., 1.],
[1., 1., 0., 0., 0., 0., 1., 1.]])
Next, we construct a kernel K with a height of 1 and a width of 2. When we perform the cross-
correlation operation with the input, if the horizontally adjacent elements are the same, the output
is 0. Otherwise, the output is non-zero.
K = np.array([[1.0, -1.0]])
We are ready to perform the cross-correlation operation with arguments X (our input) and K (our
kernel). As you can see, we detect 1 for the edge from white to black and -1 for the edge from black
to white. All other outputs take value 0.

Y = corr2d(X, K)
Y
array([[ 0., 1., 0., 0., 0., -1., 0.],

[ 0., 1., 0., 0., 0., -1., 0.],
[ 0., 1., 0., 0., 0., -1., 0.],
[ 0., 1., 0., 0., 0., -1., 0.],
[ 0., 1., 0., 0., 0., -1., 0.],
[ 0., 1., 0., 0., 0., -1., 0.]])
We can now apply the kernel to the transposed image. As expected, it vanishes. The kernel K only
detects vertical edges.
corr2d(X.T, K)
array([[0., 0., 0., 0., 0.],

[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.]])
6.2.4 Learning a Kernel
Designing an edge detector by finite differences [1, -1] is neat if we know this is precisely what
we are looking for. However, as we look at larger kernels, and consider successive layers of con-
volutions, it might be impossible to specify precisely what each filter should be doing manually.
Now let us see whether we can learn the kernel that generated Y from X by looking at the input–
output pairs only. We first construct a convolutional layer and initialize its kernel as a random
tensor. Next, in each iteration, we will use the squared error to compare Y with the output of
the convolutional layer. We can then calculate the gradient to update the kernel. For the sake of
simplicity, in the following we use the built-in class for two-dimensional convolutional layers and
ignore the bias.
# Construct a two-dimensional convolutional layer with 1 output channel and a

# kernel of shape (1, 2). For the sake of simplicity, we ignore the bias here
conv2d = nn.Conv2D(1, kernel_size=(1, 2), use_bias=False)
conv2d.initialize()
# The two-dimensional convolutional layer uses four-dimensional input and

# output in the format of (example, channel, height, width), where the batch
# size (number of examples in the batch) and the number of channels are both 1
X = X.reshape(1, 1, 6, 8)
Y = Y.reshape(1, 1, 6, 7)
for i in range(10):
Y_hat = conv2d(X)

l = (Y_hat - Y) ** 2
l.backward()
# Update the kernel
conv2d.weight.data()[:] -= 3e-2 * conv2d.weight.grad()
if (i + 1) % 2 == 0:
print(f'batch {i+1}, loss {float(l.sum()):.3f}')
batch 2, loss 4.949

batch 4, loss 0.831
batch 6, loss 0.140
batch 8, loss 0.024
batch 10, loss 0.004
Note that the error has dropped to a small value after 10 iterations. Now we will take a look at the
kernel tensor we learned.
d2l.reshape(conv2d.weight.data(), (1, 2))
array([[ 0.9895 , -0.9873705]])
Indeed, the learned kernel tensor is remarkably close to the kernel tensor K we defined earlier.
6.2.5 Cross-Correlation and Convolution
Recall our observation from Section 6.1 of the correspondence between the cross-correlation and
convolution operations. Here let us continue to consider two-dimensional convolutional layers.
What if such layers perform strict convolution operations as defined in (6.1.6) instead of cross-
correlations? In order to obtain the output of the strict convolution operation, we only need to flip
the two-dimensional kernel tensor both horizontally and vertically, and then perform the cross-
correlation operation with the input tensor.
It is noteworthy that since kernels are learned from data in deep learning, the outputs of con-
volutional layers remain unaffected no matter such layers perform either the strict convolution
operations or the cross-correlation operations.
To illustrate this, suppose that a convolutional layer performs cross-correlation and learns the ker-
nel in Fig. 6.2.1, which is denoted as the matrix K here. Assuming that other conditions remain
unchanged, when this layer performs strict convolution instead, the learned kernel K′ will be the
same as K after K′ is flipped both horizontally and vertically. That is to say, when the convolu-
tional layer performs strict convolution for the input in Fig. 6.2.1 and K′ , the same output in Fig.
6.2.1 (cross-correlation of the input and K) will be obtained.
In keeping with standard terminology with deep learning literature, we will continue to refer to the
cross-correlation operation as a convolution even though, strictly-speaking, it is slightly different.
Besides, we use the term element to refer to an entry (or component) of any tensor representing a
layer representation or a convolution kernel.

6.2.6 Feature Map and Receptive Field
As described in Section 6.1.4, the convolutional layer output in Fig. 6.2.1 is sometimes called a fea-
ture map, as it can be regarded as the learned representations (features) in the spatial dimensions
(e.g., width and height) to the subsequent layer. In CNNs, for any element x of some layer, its re-
ceptive field refers to all the elements (from all the previous layers) that may affect the calculation
of x during the forward propagation. Note that the receptive field may be larger than the actual
size of the input.
Let us continue to use Fig. 6.2.1 to explain the receptive field. Given the 2 × 2 convolution kernel,
the receptive field of the shaded output element (of value 19) is the four elements in the shaded
portion of the input. Now let us denote the 2 × 2 output as Y and consider a deeper CNN with an
additional 2 × 2 convolutional layer that takes Y as its input, outputting a single element z. In this
case, the receptive field of z on Y includes all the four elements of Y, while the receptive field on
the input includes all the nine input elements. Thus, when any element in a feature map needs a
larger receptive field to detect input features over a broader area, we can build a deeper network.
Summary
• The core computation of a two-dimensional convolutional layer is a two-dimensional cross-

correlation operation. In its simplest form, this performs a cross-correlation operation on
the two-dimensional input data and the kernel, and then adds a bias.
• We can design a kernel to detect edges in images.
• We can learn the kernelʼs parameters from data.
• With kernels learned from data, the outputs of convolutional layers remain unaffected
regardless of such layersʼ performed operations (either strict convolution or cross-
correlation).
• When any element in a feature map needs a larger receptive field to detect broader features
on the input, a deeper network can be considered.
Exercises
1. Construct an image X with diagonal edges.

1. What happens if you apply the kernel K in this section to it?
2. What happens if you transpose X?
3. What happens if you transpose K?
2. When you try to automatically find the gradient for the Conv2D class we created, what kind
of error message do you see?
3. How do you represent a cross-correlation operation as a matrix multiplication by changing
the input and kernel tensors?
4. Design some kernels manually.
1. What is the form of a kernel for the second derivative?
2. What is the kernel for an integral?

3. What is the minimum size of a kernel to obtain a derivative of degree d?
Discussions88
6.3 Padding and Stride
In the previous example of Fig. 6.2.1, our input had both a height and width of 3 and our convo-
lution kernel had both a height and width of 2, yielding an output representation with dimension
2 × 2. As we generalized in :numref:sec_conv_layer, assuming that the input shape is nh × nw and
the convolution kernel shape is kh × kw , then the output shape will be (nh − kh + 1) × (nw − kw + 1).
Therefore, the output shape of the convolutional layer is determined by the shape of the input and
the shape of the convolution kernel.
In several cases, we incorporate techniques, including padding and strided convolutions, that af-
fect the size of the output. As motivation, note that since kernels generally have width and height
greater than 1, after applying many successive convolutions, we tend to wind up with outputs that
are considerably smaller than our input. If we start with a 240 × 240 pixel image, 10 layers of 5 × 5
convolutions reduce the image to 200 × 200 pixels, slicing off 30% of the image and with it oblit-
erating any interesting information on the boundaries of the original image. Padding is the most
popular tool for handling this issue.
In other cases, we may want to reduce the dimensionality drastically, e.g., if we find the original
input resolution to be unwieldy. Strided convolutions are a popular technique that can help in these
instances.
6.3.1 Padding
As described above, one tricky issue when applying convolutional layers is that we tend to lose
pixels on the perimeter of our image. Since we typically use small kernels, for any given convo-
lution, we might only lose a few pixels, but this can add up as we apply many successive convolu-
tional layers. One straightforward solution to this problem is to add extra pixels of filler around
the boundary of our input image, thus increasing the effective size of the image. Typically, we
set the values of the extra pixels to zero. In Fig. 6.3.1, we pad a 3 × 3 input, increasing its size to
5 × 5. The corresponding output then increases to a 4 × 4 matrix. The shaded portions are the first
output element as well as the input and kernel tensor elements used for the output computation:
0 × 0 + 0 × 1 + 0 × 2 + 0 × 3 = 0.
Fig. 6.3.1: Two-dimensional cross-correlation with padding.

88
6.3. Padding and Stride 239

In general, if we add a total of ph rows of padding (roughly half on top and half on bottom) and
a total of pw columns of padding (roughly half on the left and half on the right), the output shape
will be
(nh − kh + ph + 1) × (nw − kw + pw + 1). (6.3.1)
This means that the height and width of the output will increase by ph and pw , respectively.
In many cases, we will want to set ph = kh − 1 and pw = kw − 1 to give the input and output the
same height and width. This will make it easier to predict the output shape of each layer when
constructing the network. Assuming that kh is odd here, we will pad ph /2 rows on both sides of
the height. If kh is even, one possibility is to pad ⌈ph /2⌉ rows on the top of the input and ⌊ph /2⌋
rows on the bottom. We will pad both sides of the width in the same way.
CNNs commonly use convolution kernels with odd height and width values, such as 1, 3, 5, or 7.
Choosing odd kernel sizes has the benefit that we can preserve the spatial dimensionality while
padding with the same number of rows on top and bottom, and the same number of columns on
left and right.
Moreover, this practice of using odd kernels and padding to precisely preserve dimensionality
offers a clerical benefit. For any two-dimensional tensor X, when the kernelʼs size is odd and the
number of padding rows and columns on all sides are the same, producing an output with the same
height and width as the input, we know that the output Y[i, j] is calculated by cross-correlation
of the input and convolution kernel with the window centered on X[i, j].
In the following example, we create a two-dimensional convolutional layer with a height and width
of 3 and apply 1 pixel of padding on all sides. Given an input with a height and width of 8, we find
that the height and width of the output is also 8.

npx.set_np()
# For convenience, we define a function to calculate the convolutional layer.

# This function initializes the convolutional layer weights and performs
# corresponding dimensionality elevations and reductions on the input and
# output
def comp_conv2d(conv2d, X):
conv2d.initialize()
# Here (1, 1) indicates that the batch size and the number of channels
# are both 1
X = X.reshape((1, 1) + X.shape)
Y = conv2d(X)
# Exclude the first two dimensions that do not interest us: examples and
# channels
return Y.reshape(Y.shape[2:])
# Note that here 1 row or column is padded on either side, so a total of 2

# rows or columns are added
conv2d = nn.Conv2D(1, kernel_size=3, padding=1)
comp_conv2d(conv2d, X).shape
(8, 8)

When the height and width of the convolution kernel are different, we can make the output and
input have the same height and width by setting different padding numbers for height and width.
# Here, we use a convolution kernel with a height of 5 and a width of 3. The

# padding numbers on either side of the height and width are 2 and 1,
# respectively
conv2d = nn.Conv2D(1, kernel_size=(5, 3), padding=(2, 1))
(8, 8)
6.3.2 Stride
When computing the cross-correlation, we start with the convolution window at the top-left cor-
ner of the input tensor, and then slide it over all locations both down and to the right. In previous
examples, we default to sliding one element at a time. However, sometimes, either for computa-
tional efficiency or because we wish to downsample, we move our window more than one element
at a time, skipping the intermediate locations.
We refer to the number of rows and columns traversed per slide as the stride. So far, we have used
strides of 1, both for height and width. Sometimes, we may want to use a larger stride. Fig. 6.3.2
shows a two-dimensional cross-correlation operation with a stride of 3 vertically and 2 horizon-
tally. The shaded portions are the output elements as well as the input and kernel tensor elements
used for the output computation: 0 × 0 + 0 × 1 + 1 × 2 + 2 × 3 = 8, 0 × 0 + 6 × 1 + 0 × 2 + 0 × 3 = 6.
We can see that when the second element of the first column is outputted, the convolution win-
dow slides down three rows. The convolution window slides two columns to the right when the
second element of the first row is outputted. When the convolution window continues to slide
two columns to the right on the input, there is no output because the input element cannot fill the
window (unless we add another column of padding).
Fig. 6.3.2: Cross-correlation with strides of 3 and 2 for height and width, respectively.
In general, when the stride for the height is sh and the stride for the width is sw , the output shape
is
⌊(nh − kh + ph + sh )/sh ⌋ × ⌊(nw − kw + pw + sw )/sw ⌋. (6.3.2)
If we set ph = kh − 1 and pw = kw − 1, then the output shape will be simplified to ⌊(nh + sh −
1)/sh ⌋ × ⌊(nw + sw − 1)/sw ⌋. Going a step further, if the input height and width are divisible by
the strides on the height and width, then the output shape will be (nh /sh ) × (nw /sw ).
Below, we set the strides on both the height and width to 2, thus halving the input height and width.
6.3. Padding and Stride 241

conv2d = nn.Conv2D(1, kernel_size=3, padding=1, strides=2)
(4, 4)
Next, we will look at a slightly more complicated example.
conv2d = nn.Conv2D(1, kernel_size=(3, 5), padding=(0, 1), strides=(3, 4))

(2, 2)
For the sake of brevity, when the padding number on both sides of the input height and width are
ph and pw respectively, we call the padding (ph , pw ). Specifically, when ph = pw = p, the padding is
p. When the strides on the height and width are sh and sw , respectively, we call the stride (sh , sw ).
Specifically, when sh = sw = s, the stride is s. By default, the padding is 0 and the stride is 1.
In practice, we rarely use inhomogeneous strides or padding, i.e., we usually have ph = pw and
sh = sw .
Summary
• Padding can increase the height and width of the output. This is often used to give the output
the same height and width as the input.
• The stride can reduce the resolution of the output, for example reducing the height and width
of the output to only 1/n of the height and width of the input (n is an integer greater than 1).
• Padding and stride can be used to adjust the dimensionality of the data effectively.
Exercises
1. For the last example in this section, use mathematics to calculate the output shape to see if
it is consistent with the experimental result.
2. Try other padding and stride combinations on the experiments in this section.
3. For audio signals, what does a stride of 2 correspond to?
4. What are the computational benefits of a stride larger than 1?
Discussions89
89

6.4 Multiple Input and Multiple Output Channels
While we have described the multiple channels that comprise each image (e.g., color images have
the standard RGB channels to indicate the amount of red, green and blue) and convolutional layers
for multiple channels in Section 6.1.4, until now, we simplified all of our numerical examples by
working with just a single input and a single output channel. This has allowed us to think of our
inputs, convolution kernels, and outputs each as two-dimensional tensors.
When we add channels into the mix, our inputs and hidden representations both become three-
dimensional tensors. For example, each RGB input image has shape 3×h×w. We refer to this axis,
with a size of 3, as the channel dimension. In this section, we will take a deeper look at convolution
kernels with multiple input and multiple output channels.
6.4.1 Multiple Input Channels
When the input data contain multiple channels, we need to construct a convolution kernel with
the same number of input channels as the input data, so that it can perform cross-correlation with
the input data. Assuming that the number of channels for the input data is ci , the number of input
channels of the convolution kernel also needs to be ci . If our convolution kernelʼs window shape
is kh × kw , then when ci = 1, we can think of our convolution kernel as just a two-dimensional
tensor of shape kh × kw .
However, when ci > 1, we need a kernel that contains a tensor of shape kh × kw for every input
channel. Concatenating these ci tensors together yields a convolution kernel of shape ci × kh × kw .
Since the input and convolution kernel each have ci channels, we can perform a cross-correlation
operation on the two-dimensional tensor of the input and the two-dimensional tensor of the con-
volution kernel for each channel, adding the ci results together (summing over the channels) to
yield a two-dimensional tensor. This is the result of a two-dimensional cross-correlation between
a multi-channel input and a multi-input-channel convolution kernel.
In Fig. 6.4.1, we demonstrate an example of a two-dimensional cross-correlation with two input
channels. The shaded portions are the first output element as well as the input and kernel tensor
elements used for the output computation: (1×1+2×2+4×3+5×4)+(0×0+1×1+3×2+4×3) = 56.
Fig. 6.4.1: Cross-correlation computation with 2 input channels.
To make sure we really understand what is going on here, we can implement cross-correlation
operations with multiple input channels ourselves. Notice that all we are doing is performing one
cross-correlation operation per channel and then adding up the results.
6.4. Multiple Input and Multiple Output Channels 243

npx.set_np()
def corr2d_multi_in(X, K):

# First, iterate through the 0th dimension (channel dimension) of `X` and
# `K`. Then, add them together
return sum(d2l.corr2d(x, k) for x, k in zip(X, K))
We can construct the input tensor X and the kernel tensor K corresponding to the values in Fig.
6.4.1 to validate the output of the cross-correlation operation.
X = np.array([[[0.0, 1.0, 2.0], [3.0, 4.0, 5.0], [6.0, 7.0, 8.0]],

[[1.0, 2.0, 3.0], [4.0, 5.0, 6.0], [7.0, 8.0, 9.0]]])
K = np.array([[[0.0, 1.0], [2.0, 3.0]], [[1.0, 2.0], [3.0, 4.0]]])
corr2d_multi_in(X, K)
array([[ 56., 72.],

[104., 120.]])
6.4.2 Multiple Output Channels
Regardless of the number of input channels, so far we always ended up with one output channel.
However, as we discussed in Section 6.1.4, it turns out to be essential to have multiple channels
at each layer. In the most popular neural network architectures, we actually increase the channel
dimension as we go higher up in the neural network, typically downsampling to trade off spatial
resolution for greater channel depth. Intuitively, you could think of each channel as responding to
some different set of features. Reality is a bit more complicated than the most naive interpreta-
tions of this intuition since representations are not learned independent but are rather optimized
to be jointly useful. So it may not be that a single channel learns an edge detector but rather that
some direction in channel space corresponds to detecting edges.
Denote by ci and co the number of input and output channels, respectively, and let kh and kw
be the height and width of the kernel. To get an output with multiple channels, we can create a
kernel tensor of shape ci × kh × kw for every output channel. We concatenate them on the output
channel dimension, so that the shape of the convolution kernel is co × ci × kh × kw . In cross-
correlation operations, the result on each output channel is calculated from the convolution kernel
corresponding to that output channel and takes input from all channels in the input tensor.
We implement a cross-correlation function to calculate the output of multiple channels as shown
below.
def corr2d_multi_in_out(X, K):

# Iterate through the 0th dimension of `K`, and each time, perform
# cross-correlation operations with input `X`. All of the results are
# stacked together
return np.stack([corr2d_multi_in(X, k) for k in K], 0)
We construct a convolution kernel with 3 output channels by concatenating the kernel tensor K
with K+1 (plus one for each element in K) and K+2.

K = np.stack((K, K + 1, K + 2), 0)
K.shape
(3, 2, 2, 2)
Below, we perform cross-correlation operations on the input tensor X with the kernel tensor K.
Now the output contains 3 channels. The result of the first channel is consistent with the result of
the previous input tensor X and the multi-input channel, single-output channel kernel.
corr2d_multi_in_out(X, K)
array([[[ 56., 72.],

[104., 120.]],
[[ 76., 100.],
[148., 172.]],
[[ 96., 128.],
[192., 224.]]])
6.4.3 1 × 1 Convolutional Layer
At first, a 1 × 1 convolution, i.e., kh = kw = 1, does not seem to make much sense. After all, a
convolution correlates adjacent pixels. A 1 × 1 convolution obviously does not. Nonetheless, they
are popular operations that are sometimes included in the designs of complex deep networks. Let
us see in some detail what it actually does.
Because the minimum window is used, the 1 × 1 convolution loses the ability of larger convo-
lutional layers to recognize patterns consisting of interactions among adjacent elements in the
height and width dimensions. The only computation of the 1 × 1 convolution occurs on the chan-
nel dimension.
Fig. 6.4.2 shows the cross-correlation computation using the 1 × 1 convolution kernel with 3 input
channels and 2 output channels. Note that the inputs and outputs have the same height and width.
Each element in the output is derived from a linear combination of elements at the same position in
the input image. You could think of the 1 × 1 convolutional layer as constituting a fully-connected
layer applied at every single pixel location to transform the ci corresponding input values into co
output values. Because this is still a convolutional layer, the weights are tied across pixel location.
Thus the 1 × 1 convolutional layer requires co × ci weights (plus the bias).
Fig. 6.4.2: The cross-correlation computation uses the 1 × 1 convolution kernel with 3 input chan-
nels and 2 output channels. The input and output have the same height and width.
6.4. Multiple Input and Multiple Output Channels 245

Let us check whether this works in practice: we implement a 1 × 1 convolution using a fully-
connected layer. The only thing is that we need to make some adjustments to the data shape
before and after the matrix multiplication.
def corr2d_multi_in_out_1x1(X, K):

c_i, h, w = X.shape
c_o = K.shape[0]
X = X.reshape((c_i, h * w))
K = K.reshape((c_o, c_i))
Y = np.dot(K, X) # Matrix multiplication in the fully-connected layer
return Y.reshape((c_o, h, w))
When performing 1 × 1 convolution, the above function is equivalent to the previously imple-
mented cross-correlation function corr2d_multi_in_out. Let us check this with some sample
data.
X = np.random.normal(0, 1, (3, 3, 3))

K = np.random.normal(0, 1, (2, 3, 1, 1))
Y1 = corr2d_multi_in_out_1x1(X, K)
Y2 = corr2d_multi_in_out(X, K)
assert float(d2l.reduce_sum(np.abs(Y1 - Y2))) < 1e-6
Summary
• Multiple channels can be used to extend the model parameters of the convolutional layer.
• The 1 × 1 convolutional layer is equivalent to the fully-connected layer, when applied on a
per pixel basis.
• The 1 × 1 convolutional layer is typically used to adjust the number of channels between
network layers and to control model complexity.
Exercises
1. Assume that we have two convolution kernels of size k1 and k2 , respectively (with no non-
linearity in between).
1. Prove that the result of the operation can be expressed by a single convolution.
2. What is the dimensionality of the equivalent single convolution?
3. Is the converse true?
2. Assume an input of shape ci × h × w and a convolution kernel of shape co × ci × kh × kw ,
padding of (ph , pw ), and stride of (sh , sw ).
1. What is the computational cost (multiplications and additions) for the forward propa-
gation?
2. What is the memory footprint?
3. What is the memory footprint for the backward computation?
4. What is the computational cost for the backpropagation?

3. By what factor does the number of calculations increase if we double the number of input
channels ci and the number of output channels co ? What happens if we double the padding?
4. If the height and width of a convolution kernel is kh = kw = 1, what is the computational
complexity of the forward propagation?
5. Are the variables Y1 and Y2 in the last example of this section exactly the same? Why?
6. How would you implement convolutions using matrix multiplication when the convolution
window is not 1 × 1?
Discussions90
6.5 Pooling
Often, as we process images, we want to gradually reduce the spatial resolution of our hidden
representations, aggregating information so that the higher up we go in the network, the larger
the receptive field (in the input) to which each hidden node is sensitive.
Often our ultimate task asks some global question about the image, e.g., does it contain a cat? So
typically the units of our final layer should be sensitive to the entire input. By gradually aggregat-
ing information, yielding coarser and coarser maps, we accomplish this goal of ultimately learning
a global representation, while keeping all of the advantages of convolutional layers at the inter-
mediate layers of processing.
Moreover, when detecting lower-level features, such as edges (as discussed in Section 6.2), we
often want our representations to be somewhat invariant to translation. For instance, if we take
the image X with a sharp delineation between black and white and shift the whole image by one
pixel to the right, i.e., Z[i, j] = X[i, j + 1], then the output for the new image Z might be vastly
different. The edge will have shifted by one pixel. In reality, objects hardly ever occur exactly at
the same place. In fact, even with a tripod and a stationary object, vibration of the camera due to
the movement of the shutter might shift everything by a pixel or so (high-end cameras are loaded
with special features to address this problem).
This section introduces pooling layers, which serve the dual purposes of mitigating the sensitivity
of convolutional layers to location and of spatially downsampling representations.
6.5.1 Maximum Pooling and Average Pooling
Like convolutional layers, pooling operators consist of a fixed-shape window that is slid over all
regions in the input according to its stride, computing a single output for each location traversed
by the fixed-shape window (sometimes known as the pooling window). However, unlike the cross-
correlation computation of the inputs and kernels in the convolutional layer, the pooling layer
contains no parameters (there is no kernel). Instead, pooling operators are deterministic, typically
calculating either the maximum or the average value of the elements in the pooling window. These
operations are called maximum pooling (max pooling for short) and average pooling, respectively.
In both cases, as with the cross-correlation operator, we can think of the pooling window as start-
ing from the top left of the input tensor and sliding across the input tensor from left to right and
90
6.5. Pooling 247

top to bottom. At each location that the pooling window hits, it computes the maximum or aver-
age value of the input subtensor in the window, depending on whether max or average pooling is
employed.
Fig. 6.5.1: Maximum pooling with a pooling window shape of 2 × 2. The shaded portions are
the first output element as well as the input tensor elements used for the output computation:
max(0, 1, 3, 4) = 4.
The output tensor in Fig. 6.5.1 has a height of 2 and a width of 2. The four elements are derived
from the maximum value in each pooling window:
max(0, 1, 3, 4) = 4,
max(1, 2, 4, 5) = 5,
(6.5.1)
max(3, 4, 6, 7) = 7,
max(4, 5, 7, 8) = 8.
A pooling layer with a pooling window shape of p × q is called a p × q pooling layer. The pooling
operation is called p × q pooling.
Let us return to the object edge detection example mentioned at the beginning of this section.
Now we will use the output of the convolutional layer as the input for 2 × 2 maximum pooling. Set
the convolutional layer input as X and the pooling layer output as Y. Whether or not the values of
X[i, j] and X[i, j + 1] are different, or X[i, j + 1] and X[i, j + 2] are different, the pooling
layer always outputs Y[i, j] = 1. That is to say, using the 2 × 2 maximum pooling layer, we can
still detect if the pattern recognized by the convolutional layer moves no more than one element
in height or width.
In the code below, we implement the forward propagation of the pooling layer in the pool2d func-
tion. This function is similar to the corr2d function in Section 6.2. However, here we have no
kernel, computing the output as either the maximum or the average of each region in the input.

npx.set_np()
def pool2d(X, pool_size, mode='max'):

p_h, p_w = pool_size
Y = np.zeros((X.shape[0] - p_h + 1, X.shape[1] - p_w + 1))
for j in range(Y.shape[1]):
if mode == 'max':
Y[i, j] = X[i: i + p_h, j: j + p_w].max()
elif mode == 'avg':
Y[i, j] = X[i: i + p_h, j: j + p_w].mean()
return Y

We can construct the input tensor X in Fig. 6.5.1 to validate the output of the two-dimensional
maximum pooling layer.
X = np.array([[0.0, 1.0, 2.0], [3.0, 4.0, 5.0], [6.0, 7.0, 8.0]])

pool2d(X, (2, 2))
array([[4., 5.],
[7., 8.]])
Also, we experiment with the average pooling layer.
pool2d(X, (2, 2), 'avg')
array([[2., 3.],
[5., 6.]])
6.5.2 Padding and Stride
As with convolutional layers, pooling layers can also change the output shape. And as before, we
can alter the operation to achieve a desired output shape by padding the input and adjusting the
stride. We can demonstrate the use of padding and strides in pooling layers via the built-in two-
dimensional maximum pooling layer from the deep learning framework. We first construct an
input tensor X whose shape has four dimensions, where the number of examples and number of
channels are both 1.
X = d2l.reshape(np.arange(16, dtype=np.float32), (1, 1, 4, 4))

X
array([[[[ 0., 1., 2., 3.],

[ 4., 5., 6., 7.],
[ 8., 9., 10., 11.],
[12., 13., 14., 15.]]]])
By default, the stride and the pooling window in the instance from the frameworkʼs built-in class
have the same shape. Below, we use a pooling window of shape (3, 3), so we get a stride shape
of (3, 3) by default.
pool2d = nn.MaxPool2D(3)
# Because there are no model parameters in the pooling layer, we do not need
# to call the parameter initialization function
pool2d(X)
array([[[[10.]]]])
The stride and padding can be manually specified.
pool2d = nn.MaxPool2D(3, padding=1, strides=2)

pool2d(X)
6.5. Pooling 249

array([[[[ 5., 7.],
[13., 15.]]]])
Of course, we can specify an arbitrary rectangular pooling window and specify the padding and
stride for height and width, respectively.
pool2d = nn.MaxPool2D((2, 3), padding=(1, 2), strides=(2, 3))

pool2d(X)
array([[[[ 0., 3.],

[ 8., 11.],
[12., 15.]]]])
6.5.3 Multiple Channels
When processing multi-channel input data, the pooling layer pools each input channel separately,
rather than summing the inputs up over channels as in a convolutional layer. This means that the
number of output channels for the pooling layer is the same as the number of input channels.
Below, we will concatenate tensors X and X + 1 on the channel dimension to construct an input
with 2 channels.
X = np.concatenate((X, X + 1), 1)
X
array([[[[ 0., 1., 2., 3.],

[ 4., 5., 6., 7.],
[ 8., 9., 10., 11.],
[12., 13., 14., 15.]],
[[ 1., 2., 3., 4.],

[ 5., 6., 7., 8.],
[ 9., 10., 11., 12.],
[13., 14., 15., 16.]]]])
As we can see, the number of output channels is still 2 after pooling.
pool2d = nn.MaxPool2D(3, padding=1, strides=2)

pool2d(X)
array([[[[ 5., 7.],

[13., 15.]],
[[ 6., 8.],
[14., 16.]]]])

Summary
• Taking the input elements in the pooling window, the maximum pooling operation assigns
the maximum value as the output and the average pooling operation assigns the average
value as the output.
• One of the major benefits of a pooling layer is to alleviate the excessive sensitivity of the
convolutional layer to location.
• We can specify the padding and stride for the pooling layer.
• Maximum pooling, combined with a stride larger than 1 can be used to reduce the spatial
dimensions (e.g., width and height).
• The pooling layerʼs number of output channels is the same as the number of input channels.
Exercises
1. Can you implement average pooling as a special case of a convolution layer? If so, do it.
2. Can you implement max pooling as a special case of a convolution layer? If so, do it.
3. What is the computational cost of the pooling layer? Assume that the input to the pooling
layer is of size c × h × w, the pooling window has a shape of ph × pw with a padding of (ph , pw )
and a stride of (sh , sw ).
4. Why do you expect maximum pooling and average pooling to work differently?
5. Do we need a separate minimum pooling layer? Can you replace it with another operation?
6. Is there another operation between average and maximum pooling that you could consider
(hint: recall the softmax)? Why might it not be so popular?
Discussions91
6.6 Convolutional Neural Networks (LeNet)
We now have all the ingredients required to assemble a fully-functional CNN. In our earlier en-
counter with image data, we applied a softmax regression model (Section 3.6) and an MLP model
(Section 4.2) to pictures of clothing in the Fashion-MNIST dataset. To make such data amenable
to softmax regression and MLPs, we first flattened each image from a 28 × 28 matrix into a fixed-
length 784-dimensional vector, and thereafter processed them with fully-connected layers. Now
that we have a handle on convolutional layers, we can retain the spatial structure in our images. As
an additional benefit of replacing fully-connected layers with convolutional layers, we will enjoy
more parsimonious models that require far fewer parameters.
In this section, we will introduce LeNet, among the first published CNNs to capture wide attention
for its performance on computer vision tasks. The model was introduced by (and named for) Yann
LeCun, then a researcher at AT&T Bell Labs, for the purpose of recognizing handwritten digits
in images (LeCun et al., 1998). This work represented the culmination of a decade of research
developing the technology. In 1989, LeCun published the first study to successfully train CNNs via
backpropagation.
91
6.6. Convolutional Neural Networks (LeNet) 251

At the time LeNet achieved outstanding results matching the performance of support vector ma-
chines, then a dominant approach in supervised learning. LeNet was eventually adapted to rec-
ognize digits for processing deposits in ATM machines. To this day, some ATMs still run the code
that Yann and his colleague Leon Bottou wrote in the 1990s!
6.6.1 LeNet
At a high level, LeNet (LeNet-5) consists of two parts: (i) a convolutional encoder consisting of
two convolutional layers; and (ii) a dense block consisting of three fully-connected layers; The
architecture is summarized in Fig. 6.6.1.
Fig. 6.6.1: Data flow in LeNet. The input is a handwritten digit, the output a probability over 10
possible outcomes.
The basic units in each convolutional block are a convolutional layer, a sigmoid activation func-
tion, and a subsequent average pooling operation. Note that while ReLUs and max-pooling work
better, these discoveries had not yet been made in the 1990s. Each convolutional layer uses a 5 × 5
kernel and a sigmoid activation function. These layers map spatially arranged inputs to a number
of two-dimensional feature maps, typically increasing the number of channels. The first convolu-
tional layer has 6 output channels, while the second has 16. Each 2 × 2 pooling operation (stride 2)
reduces dimensionality by a factor of 4 via spatial downsampling. The convolutional block emits
an output with shape given by (batch size, number of channel, height, width).
In order to pass output from the convolutional block to the dense block, we must flatten each
example in the minibatch. In other words, we take this four-dimensional input and transform
it into the two-dimensional input expected by fully-connected layers: as a reminder, the two-
dimensional representation that we desire has uses the first dimension to index examples in the
minibatch and the second to give the flat vector representation of each example. LeNetʼs dense
block has three fully-connected layers, with 120, 84, and 10 outputs, respectively. Because we
are still performing classification, the 10-dimensional output layer corresponds to the number of
possible output classes.
While getting to the point where you truly understand what is going on inside LeNet may have
taken a bit of work, hopefully the following code snippet will convince you that implementing such
models with modern deep learning frameworks is remarkably simple. We need only to instantiate
a Sequential block and chain together the appropriate layers.

npx.set_np()
net.add(nn.Conv2D(channels=6, kernel_size=5, padding=2, activation='sigmoid'),
nn.AvgPool2D(pool_size=2, strides=2),
nn.Conv2D(channels=16, kernel_size=5, activation='sigmoid'),
nn.AvgPool2D(pool_size=2, strides=2),
# `Dense` will transform an input of the shape (batch size, number of
# channels, height, width) into an input of the shape (batch size,
# number of channels * height * width) automatically by default
nn.Dense(120, activation='sigmoid'),
nn.Dense(84, activation='sigmoid'),
nn.Dense(10))
We took a small liberty with the original model, removing the Gaussian activation in the final
layer. Other than that, this network matches the original LeNet-5 architecture.
By passing a single-channel (black and white) 28 × 28 image through the network and printing the
output shape at each layer, we can inspect the model to make sure that its operations line up with
what we expect from Fig. 6.6.2.
Fig. 6.6.2: Compressed notation for LeNet-5.
X = np.random.uniform(size=(1, 1, 28, 28))

net.initialize()
for layer in net:
X = layer(X)
print(layer.name, 'output shape:\t', X.shape)

conv0 output shape: (1, 6, 28, 28)
pool0 output shape: (1, 6, 14, 14)
dense0 output shape: (1, 120)
Note that the height and width of the representation at each layer throughout the convolutional
block is reduced (compared with the previous layer). The first convolutional layer uses 2 pixels of
padding to compensate for the reduction in height and width that would otherwise result from us-
ing a 5×5 kernel. In contrast, the second convolutional layer forgoes padding, and thus the height
and width are both reduced by 4 pixels. As we go up the stack of layers, the number of channels
increases layer-over-layer from 1 in the input to 6 after the first convolutional layer and 16 after
the second convolutional layer. However, each pooling layer halves the height and width. Finally,
each fully-connected layer reduces dimensionality, finally emitting an output whose dimension
matches the number of classes.
6.6.2 Training
Now that we have implemented the model, let us run an experiment to see how LeNet fares on
Fashion-MNIST.
batch_size = 256
train_iter, test_iter = d2l.load_data_fashion_mnist(batch_size=batch_size)
While CNNs have fewer parameters, they can still be more expensive to compute than similarly
deep MLPs because each parameter participates in many more multiplications. If you have access
to a GPU, this might be a good time to put it into action to speed up training.
For evaluation, we need to make a slight modification to the evaluate_accuracy function that we
described in Section 3.6. Since the full dataset is in the main memory, we need to copy it to the
GPU memory before the model uses GPU to compute with the dataset.
def evaluate_accuracy_gpu(net, data_iter, device=None): #@save

"""Compute the accuracy for a model on a dataset using a GPU."""
if not device: # Query the first device where the first parameter is on
device = list(net.collect_params().values())[0].list_ctx()[0]
# No. of correct predictions, no. of predictions
metric = d2l.Accumulator(2)
X, y = X.as_in_ctx(device), y.as_in_ctx(device)
metric.add(d2l.accuracy(net(X), y), y.size)
return metric[0]/metric[1]
We also need to update our training function to deal with GPUs. Unlike the train_epoch_ch3 de-
fined in Section 3.6, we now need to move each minibatch of data to our designated device (hope-
fully, the GPU) prior to making the forward and backward propagations.
The training function train_ch6 is also similar to train_ch3 defined in Section 3.6. Since we will
be implementing networks with many layers going forward, we will rely primarily on high-level
APIs. The following training function assumes a model created from high-level APIs as input and

is optimized accordingly. We initialize the model parameters on the device indicated by the device
argument, using Xavier initialization as introduced in Section 4.8.2. Just as with MLPs, our loss
function is cross-entropy, and we minimize it via minibatch stochastic gradient descent. Since
each epoch takes tens of seconds to run, we visualize the training loss more frequently.
#@save
def train_ch6(net, train_iter, test_iter, num_epochs, lr,
device=d2l.try_gpu()):
"""Train a model with a GPU (defined in Chapter 6)."""
net.initialize(force_reinit=True, ctx=device, init=init.Xavier())
trainer = gluon.Trainer(net.collect_params(),
'sgd', {'learning_rate': lr})
animator = d2l.Animator(xlabel='epoch', xlim=[0, num_epochs],
timer = d2l.Timer()
# Sum of training loss, sum of training accuracy, no. of examples
for i, (X, y) in enumerate(train_iter):
timer.start()
# Here is the major difference compared with `d2l.train_epoch_ch3`
y_hat = net(X)
l = loss(y_hat, y)
l.backward()
trainer.step(X.shape[0])
metric.add(l.sum(), d2l.accuracy(y_hat, y), X.shape[0])
timer.stop()
train_loss = metric[0] / metric[2]
train_acc = metric[1] / metric[2]
if (i + 1) % 50 == 0:
animator.add(epoch + i / len(train_iter),
(train_loss, train_acc, None))
test_acc = evaluate_accuracy_gpu(net, test_iter)
animator.add(epoch + 1, (None, None, test_acc))
print(f'loss {train_loss:.3f}, train acc {train_acc:.3f}, '
f'test acc {test_acc:.3f}')
print(f'{metric[2] * num_epochs / timer.sum():.1f} examples/sec '
f'on {str(device)}')
Now let us train and evaluate the LeNet-5 model.
lr, num_epochs = 0.9, 10

train_ch6(net, train_iter, test_iter, num_epochs, lr)
loss 0.471, train acc 0.822, test acc 0.824

49576.0 examples/sec on gpu(0)

Summary
• A CNN is a network that employs convolutional layers.

• In a CNN, we interleave convolutions, nonlinearities, and (often) pooling operations.
• In a CNN, convolutional layers are typically arranged so that they gradually decrease the
spatial resolution of the representations, while increasing the number of channels.
• In traditional CNNs, the representations encoded by the convolutional blocks are processed
by one or more fully-connected layers prior to emitting output.
• LeNet was arguably the first successful deployment of such a network.
Exercises
1. Replace the average pooling with max pooling. What happens?

2. Try to construct a more complex network based on LeNet to improve its accuracy.
1. Adjust the convolution window size.
2. Adjust the number of output channels.
3. Adjust the activation function (e.g., ReLU).
4. Adjust the number of convolution layers.
5. Adjust the number of fully connected layers.
6. Adjust the learning rates and other training details (e.g., initialization and number of
epochs.)
3. Try out the improved network on the original MNIST dataset.
4. Display the activations of the first and second layer of LeNet for different inputs (e.g.,
sweaters and coats).
Discussions92
92

7 | Modern Convolutional Neural Net-
works
Now that we understand the basics of wiring together CNNs, we will take you through a tour of
modern CNN architectures. In this chapter, each section corresponds to a significant CNN archi-
tecture that was at some point (or currently) the base model upon which many research projects
and deployed systems were built. Each of these networks was briefly a dominant architecture and
many were winners or runners-up in the ImageNet competition, which has served as a barometer
of progress on supervised learning in computer vision since 2010.
These models include AlexNet, the first large-scale network deployed to beat conventional com-
puter vision methods on a large-scale vision challenge; the VGG network, which makes use of a
number of repeating blocks of elements; the network in network (NiN) which convolves whole
neural networks patch-wise over inputs; GoogLeNet, which uses networks with parallel concate-
nations; residual networks (ResNet), which remain the most popular off-the-shelf architecture in
computer vision; and densely connected networks (DenseNet), which are expensive to compute
but have set some recent benchmarks.
While the idea of deep neural networks is quite simple (stack together a bunch of layers), perfor-
mance can vary wildly across architectures and hyperparameter choices. The neural networks
described in this chapter are the product of intuition, a few mathematical insights, and a whole
lot of trial and error. We present these models in chronological order, partly to convey a sense
of the history so that you can form your own intuitions about where the field is heading and per-
haps develop your own architectures. For instance, batch normalization and residual connections
described in this chapter have offered two popular ideas for training and designing deep models.
7.1 Deep Convolutional Neural Networks (AlexNet)
Although CNNs were well known in the computer vision and machine learning communities fol-
lowing the introduction of LeNet, they did not immediately dominate the field. Although LeNet
achieved good results on early small datasets, the performance and feasibility of training CNNs on
larger, more realistic datasets had yet to be established. In fact, for much of the intervening time
between the early 1990s and the watershed results of 2012, neural networks were often surpassed
by other machine learning methods, such as support vector machines.
For computer vision, this comparison is perhaps not fair. That is although the inputs to convolu-
tional networks consist of raw or lightly-processed (e.g., by centering) pixel values, practitioners
would never feed raw pixels into traditional models. Instead, typical computer vision pipelines
consisted of manually engineering feature extraction pipelines. Rather than learn the features, the
features were crafted. Most of the progress came from having more clever ideas for features, and
the learning algorithm was often relegated to an afterthought.
257
Although some neural network accelerators were available in the 1990s, they were not yet suffi-
ciently powerful to make deep multichannel, multilayer CNNs with a large number of parameters.
Moreover, datasets were still relatively small. Added to these obstacles, key tricks for training
neural networks including parameter initialization heuristics, clever variants of stochastic gra-
dient descent, non-squashing activation functions, and effective regularization techniques were
still missing.
Thus, rather than training end-to-end (pixel to classification) systems, classical pipelines looked
more like this:
1. Obtain an interesting dataset. In early days, these datasets required expensive sensors (at
the time, 1 megapixel images were state-of-the-art).
2. Preprocess the dataset with hand-crafted features based on some knowledge of optics, geom-
etry, other analytic tools, and occasionally on the serendipitous discoveries of lucky gradu-
ate students.
3. Feed the data through a standard set of feature extractors such as the SIFT (scale-invariant
feature transform) (Lowe, 2004), the SURF (speeded up robust features) (Bay et al., 2006), or
any number of other hand-tuned pipelines.
4. Dump the resulting representations into your favorite classifier, likely a linear model or ker-
nel method, to train a classifier.
If you spoke to machine learning researchers, they believed that machine learning was both im-
portant and beautiful. Elegant theories proved the properties of various classifiers. The field of
machine learning was thriving, rigorous, and eminently useful. However, if you spoke to a com-
puter vision researcher, you would hear a very different story. The dirty truth of image recogni-
tion, they would tell you, is that features, not learning algorithms, drove progress. Computer vi-
sion researchers justifiably believed that a slightly bigger or cleaner dataset or a slightly improved
feature-extraction pipeline mattered far more to the final accuracy than any learning algorithm.
7.1.1 Learning Representations
Another way to cast the state of affairs is that the most important part of the pipeline was the rep-
resentation. And up until 2012 the representation was calculated mechanically. In fact, engineer-
ing a new set of feature functions, improving results, and writing up the method was a prominent
genre of paper. SIFT (Lowe, 2004), SURF (Bay et al., 2006), HOG (histograms of oriented gradient)
(Dalal & Triggs, 2005), bags of visual words93 and similar feature extractors ruled the roost.
Another group of researchers, including Yann LeCun, Geoff Hinton, Yoshua Bengio, Andrew Ng,
Shun-ichi Amari, and Juergen Schmidhuber, had different plans. They believed that features
themselves ought to be learned. Moreover, they believed that to be reasonably complex, the fea-
tures ought to be hierarchically composed with multiple jointly learned layers, each with learn-
able parameters. In the case of an image, the lowest layers might come to detect edges, colors, and
textures. Indeed, Alex Krizhevsky, Ilya Sutskever, and Geoff Hinton proposed a new variant of a
CNN, AlexNet, that achieved excellent performance in the 2012 ImageNet challenge. AlexNet was
named after Alex Krizhevsky, the first author of the breakthrough ImageNet classification paper
(Krizhevsky et al., 2012).
Interestingly in the lowest layers of the network, the model learned feature extractors that resem-
bled some traditional filters. Fig. 7.1.1 is reproduced from the AlexNet paper (Krizhevsky et al.,
2012) and describes lower-level image descriptors.
93
https://en.wikipedia.org/wiki/Bag-of-words_model_in_computer_vision
258 Chapter 7. Modern Convolutional Neural Networks

Fig. 7.1.1: Image filters learned by the first layer of AlexNet.
Higher layers in the network might build upon these representations to represent larger struc-
tures, like eyes, noses, blades of grass, and so on. Even higher layers might represent whole ob-
jects like people, airplanes, dogs, or frisbees. Ultimately, the final hidden state learns a compact
representation of the image that summarizes its contents such that data belonging to different
categories be separated easily.
While the ultimate breakthrough for many-layered CNNs came in 2012, a core group of researchers
had dedicated themselves to this idea, attempting to learn hierarchical representations of visual
data for many years. The ultimate breakthrough in 2012 can be attributed to two key factors.
Missing Ingredient: Data
Deep models with many layers require large amounts of data in order to enter the regime where
they significantly outperform traditional methods based on convex optimizations (e.g., linear and
kernel methods). However, given the limited storage capacity of computers, the relative expense
of sensors, and the comparatively tighter research budgets in the 1990s, most research relied on
tiny datasets. Numerous papers addressed the UCI collection of datasets, many of which contained
only hundreds or (a few) thousands of images captured in unnatural settings with low resolution.
In 2009, the ImageNet dataset was released, challenging researchers to learn models from 1 mil-
lion examples, 1000 each from 1000 distinct categories of objects. The researchers, led by Fei-Fei
Li, who introduced this dataset leveraged Google Image Search to prefilter large candidate sets for
each category and employed the Amazon Mechanical Turk crowdsourcing pipeline to confirm for
each image whether it belonged to the associated category. This scale was unprecedented. The
associated competition, dubbed the ImageNet Challenge pushed computer vision and machine
learning research forward, challenging researchers to identify which models performed best at a
greater scale than academics had previously considered.
7.1. Deep Convolutional Neural Networks (AlexNet) 259

Missing Ingredient: Hardware
Deep learning models are voracious consumers of compute cycles. Training can take hundreds
of epochs, and each iteration requires passing data through many layers of computationally-
expensive linear algebra operations. This is one of the main reasons why in the 1990s and early
2000s, simple algorithms based on the more-efficiently optimized convex objectives were pre-
ferred.
Graphical processing units (GPUs) proved to be a game changer in making deep learning feasible.
These chips had long been developed for accelerating graphics processing to benefit computer
games. In particular, they were optimized for high throughput 4×4 matrix-vector products, which
are needed for many computer graphics tasks. Fortunately, this math is strikingly similar to that
required to calculate convolutional layers. Around that time, NVIDIA and ATI had begun optimiz-
ing GPUs for general computing operations, going as far as to market them as general-purpose GPUs
(GPGPU).
To provide some intuition, consider the cores of a modern microprocessor (CPU). Each of the
cores is fairly powerful running at a high clock frequency and sporting large caches (up to several
megabytes of L3). Each core is well-suited to executing a wide range of instructions, with branch
predictors, a deep pipeline, and other bells and whistles that enable it to run a large variety of
programs. This apparent strength, however, is also its Achilles heel: general-purpose cores are
very expensive to build. They require lots of chip area, a sophisticated support structure (memory
interfaces, caching logic between cores, high-speed interconnects, and so on), and they are com-
paratively bad at any single task. Modern laptops have up to 4 cores, and even high-end servers
rarely exceed 64 cores, simply because it is not cost effective.
By comparison, GPUs consist of 100 ∼ 1000 small processing elements (the details differ some-
what between NVIDIA, ATI, ARM and other chip vendors), often grouped into larger groups
(NVIDIA calls them warps). While each core is relatively weak, sometimes even running at sub-
1GHz clock frequency, it is the total number of such cores that makes GPUs orders of magnitude
faster than CPUs. For instance, NVIDIAʼs recent Volta generation offers up to 120 TFlops per chip
for specialized instructions (and up to 24 TFlops for more general-purpose ones), while floating
point performance of CPUs has not exceeded 1 TFlop to date. The reason for why this is possible is
actually quite simple: first, power consumption tends to grow quadratically with clock frequency.
Hence, for the power budget of a CPU core that runs 4 times faster (a typical number), you can use
16 GPU cores at 1/4 the speed, which yields 16 × 1/4 = 4 times the performance. Furthermore,
GPU cores are much simpler (in fact, for a long time they were not even able to execute general-
purpose code), which makes them more energy efficient. Last, many operations in deep learning
require high memory bandwidth. Again, GPUs shine here with buses that are at least 10 times as
wide as many CPUs.
Back to 2012. A major breakthrough came when Alex Krizhevsky and Ilya Sutskever implemented
a deep CNN that could run on GPU hardware. They realized that the computational bottlenecks
in CNNs, convolutions and matrix multiplications, are all operations that could be parallelized in
hardware. Using two NVIDIA GTX 580s with 3GB of memory, they implemented fast convolutions.
The code cuda-convnet94 was good enough that for several years it was the industry standard and
powered the first couple years of the deep learning boom.
94
https://code.google.com/archive/p/cuda-convnet/

7.1.2 AlexNet
AlexNet, which employed an 8-layer CNN, won the ImageNet Large Scale Visual Recognition Chal-
lenge 2012 by a phenomenally large margin. This network showed, for the first time, that the
features obtained by learning can transcend manually-designed features, breaking the previous
paradigm in computer vision.
The architectures of AlexNet and LeNet are very similar, as Fig. 7.1.2 illustrates. Note that we
provide a slightly streamlined version of AlexNet removing some of the design quirks that were
needed in 2012 to make the model fit on two small GPUs.
Fig. 7.1.2: From LeNet (left) to AlexNet (right).
The design philosophies of AlexNet and LeNet are very similar, but there are also significant dif-
ferences. First, AlexNet is much deeper than the comparatively small LeNet5. AlexNet consists of
eight layers: five convolutional layers, two fully-connected hidden layers, and one fully-connected
output layer. Second, AlexNet used the ReLU instead of the sigmoid as its activation function. Let
us delve into the details below.

Architecture
In AlexNetʼs first layer, the convolution window shape is 11 × 11. Since most images in ImageNet
are more than ten times higher and wider than the MNIST images, objects in ImageNet data tend
to occupy more pixels. Consequently, a larger convolution window is needed to capture the object.
The convolution window shape in the second layer is reduced to 5×5, followed by 3×3. In addition,
after the first, second, and fifth convolutional layers, the network adds maximum pooling layers
with a window shape of 3 × 3 and a stride of 2. Moreover, AlexNet has ten times more convolution
channels than LeNet.
After the last convolutional layer there are two fully-connected layers with 4096 outputs. These
two huge fully-connected layers produce model parameters of nearly 1 GB. Due to the limited
memory in early GPUs, the original AlexNet used a dual data stream design, so that each of their
two GPUs could be responsible for storing and computing only its half of the model. Fortunately,
GPU memory is comparatively abundant now, so we rarely need to break up models across GPUs
these days (our version of the AlexNet model deviates from the original paper in this aspect).
Activation Functions
Besides, AlexNet changed the sigmoid activation function to a simpler ReLU activation function.
On one hand, the computation of the ReLU activation function is simpler. For example, it does
not have the exponentiation operation found in the sigmoid activation function. On the other
hand, the ReLU activation function makes model training easier when using different parameter
initialization methods. This is because, when the output of the sigmoid activation function is very
close to 0 or 1, the gradient of these regions is almost 0, so that backpropagation cannot continue
to update some of the model parameters. In contrast, the gradient of the ReLU activation function
in the positive interval is always 1. Therefore, if the model parameters are not properly initialized,
the sigmoid function may obtain a gradient of almost 0 in the positive interval, so that the model
cannot be effectively trained.
Capacity Control and Preprocessing
AlexNet controls the model complexity of the fully-connected layer by dropout (Section 4.6), while
LeNet only uses weight decay. To augment the data even further, the training loop of AlexNet
added a great deal of image augmentation, such as flipping, clipping, and color changes. This
makes the model more robust and the larger sample size effectively reduces overfitting. We will
discuss data augmentation in greater detail in Section 13.1.

npx.set_np()
# Here, we use a larger 11 x 11 window to capture objects. At the same time,
# we use a stride of 4 to greatly reduce the height and width of the output.
# Here, the number of output channels is much larger than that in LeNet
net.add(nn.Conv2D(96, kernel_size=11, strides=4, activation='relu'),
nn.MaxPool2D(pool_size=3, strides=2),
# Make the convolution window smaller, set padding to 2 for consistent

# height and width across the input and output, and increase the
# number of output channels
nn.Conv2D(256, kernel_size=5, padding=2, activation='relu'),
# Use three successive convolutional layers and a smaller convolution
# window. Except for the final convolutional layer, the number of
# output channels is further increased. Pooling layers are not used to
# reduce the height and width of input after the first two
# convolutional layers
# Here, the number of outputs of the fully-connected layer is several
# times larger than that in LeNet. Use the dropout layer to mitigate
# overfitting
nn.Dense(4096, activation='relu'), nn.Dropout(0.5),
# Output layer. Since we are using Fashion-MNIST, the number of
# classes is 10, instead of 1000 as in the paper
nn.Dense(10))
We construct a single-channel data example with both height and width of 224 to observe the
output shape of each layer. It matches the AlexNet architecture in Fig. 7.1.2.

net.initialize()
for layer in net:
X = layer(X)

dropout0 output shape: (1, 4096)

Although AlexNet is trained on ImageNet in the paper, we use Fashion-MNIST here since training
an ImageNet model to convergence could take hours or days even on a modern GPU. One of the
problems with applying AlexNet directly on Fashion-MNIST is that its images have lower resolu-
tion (28 × 28 pixels) than ImageNet images. To make things work, we upsample them to 224 × 224
(generally not a smart practice, but we do it here to be faithful to the AlexNet architecture). We
perform this resizing with the resize argument in the d2l.load_data_fashion_mnist function.
batch_size = 128
train_iter, test_iter = d2l.load_data_fashion_mnist(batch_size, resize=224)
7.1.4 Training
Now, we can start training AlexNet. Compared with LeNet in Section 6.6, the main change here is
the use of a smaller learning rate and much slower training due to the deeper and wider network,
the higher image resolution, and the more costly convolutions.

d2l.train_ch6(net, train_iter, test_iter, num_epochs, lr)

Summary
• AlexNet has a similar structure to that of LeNet, but uses more convolutional layers and a
larger parameter space to fit the large-scale ImageNet dataset.
• Today AlexNet has been surpassed by much more effective architectures but it is a key step
from shallow to deep networks that are used nowadays.
• Although it seems that there are only a few more lines in AlexNetʼs implementation than in
LeNet, it took the academic community many years to embrace this conceptual change and

take advantage of its excellent experimental results. This was also due to the lack of efficient
computational tools.
• Dropout, ReLU, and preprocessing were the other key steps in achieving excellent perfor-
mance in computer vision tasks.
Exercises
1. Try increasing the number of epochs. Compared with LeNet, how are the results different?
Why?
2. AlexNet may be too complex for the Fashion-MNIST dataset.
1. Try simplifying the model to make the training faster, while ensuring that the accuracy
does not drop significantly.
2. Design a better model that works directly on 28 × 28 images.
3. Modify the batch size, and observe the changes in accuracy and GPU memory.
4. Analyze computational performance of AlexNet.
1. What is the dominant part for the memory footprint of AlexNet?
2. What is the dominant part for computation in AlexNet?
3. How about memory bandwidth when computing the results?
5. Apply dropout and ReLU to LeNet-5. Does it improve? How about preprocessing?
Discussions95
7.2 Networks Using Blocks (VGG)
While AlexNet offered empirical evidence that deep CNNs can achieve good results, it did not
provide a general template to guide subsequent researchers in designing new networks. In the
following sections, we will introduce several heuristic concepts commonly used to design deep
networks.
Progress in this field mirrors that in chip design where engineers went from placing transistors to
logical elements to logic blocks. Similarly, the design of neural network architectures had grown
progressively more abstract, with researchers moving from thinking in terms of individual neu-
rons to whole layers, and now to blocks, repeating patterns of layers.
The idea of using blocks first emerged from the Visual Geometry Group96 (VGG) at Oxford Univer-
sity, in their eponymously-named VGG network. It is easy to implement these repeated structures
in code with any modern deep learning framework by using loops and subroutines.
95
96
http://www.robots.ox.ac.uk/~vgg/
7.2. Networks Using Blocks (VGG) 265

7.2.1 VGG Blocks
The basic building block of classic CNNs is a sequence of the following: (i) a convolutional layer
with padding to maintain the resolution, (ii) a nonlinearity such as a ReLU, (iii) a pooling layer such
as a max pooling layer. One VGG block consists of a sequence of convolutional layers, followed by
a max pooling layer for spatial downsampling. In the original VGG paper (Simonyan & Zisserman,
2014), the authors employed convolutions with 3×3 kernels with padding of 1 (keeping height and
width) and 2 × 2 max pooling with stride of 2 (halving the resolution after each block). In the code
below, we define a function called vgg_block to implement one VGG block. The function takes
two arguments corresponding to the number of convolutional layers num_convs and the number
of output channels num_channels.

npx.set_np()
def vgg_block(num_convs, num_channels):

blk = nn.Sequential()
for _ in range(num_convs):
blk.add(nn.Conv2D(num_channels, kernel_size=3,
padding=1, activation='relu'))
blk.add(nn.MaxPool2D(pool_size=2, strides=2))
return blk
7.2.2 VGG Network
Like AlexNet and LeNet, the VGG Network can be partitioned into two parts: the first consisting
mostly of convolutional and pooling layers and the second consisting of fully-connected layers.
This is depicted in Fig. 7.2.1.

Fig. 7.2.1: From AlexNet to VGG that is designed from building blocks.
The convolutional part of the network connects several VGG blocks from Fig. 7.2.1 (also defined
in the vgg_block function) in succession. The following variable conv_arch consists of a list of
tuples (one per block), where each contains two values: the number of convolutional layers and
the number of output channels, which are precisely the arguments required to call the vgg_block
function. The fully-connected part of the VGG network is identical to that covered in AlexNet.
The original VGG network had 5 convolutional blocks, among which the first two have one convo-
lutional layer each and the latter three contain two convolutional layers each. The first block has
64 output channels and each subsequent block doubles the number of output channels, until that
number reaches 512. Since this network uses 8 convolutional layers and 3 fully-connected layers,
it is often called VGG-11.
conv_arch = ((1, 64), (1, 128), (2, 256), (2, 512), (2, 512))
The following code implements VGG-11. This is a simple matter of executing a for-loop over
conv_arch.
def vgg(conv_arch):
# The convolutional part
for (num_convs, num_channels) in conv_arch:
net.add(vgg_block(num_convs, num_channels))
# The fully-connected part
net.add(nn.Dense(4096, activation='relu'), nn.Dropout(0.5),
nn.Dense(10))
return net
net = vgg(conv_arch)

Next, we will construct a single-channel data example with a height and width of 224 to observe
the output shape of each layer.
net.initialize()
for blk in net:
X = blk(X)
print(blk.name, 'output shape:\t', X.shape)
sequential1 output shape: (1, 64, 112, 112)

As you can see, we halve height and width at each block, finally reaching a height and width of 7
before flattening the representations for processing by the fully-connected part of the network.
7.2.3 Training
Since VGG-11 is more computationally-heavy than AlexNet we construct a network with a smaller
number of channels. This is more than sufficient for training on Fashion-MNIST.
ratio = 4
small_conv_arch = [(pair[0], pair[1] // ratio) for pair in conv_arch]
net = vgg(small_conv_arch)
Apart from using a slightly larger learning rate, the model training process is similar to that of
AlexNet in Section 7.1.
lr, num_epochs, batch_size = 0.05, 10, 128



Summary
• VGG-11 constructs a network using reusable convolutional blocks. Different VGG models can
be defined by the differences in the number of convolutional layers and output channels in
each block.
• The use of blocks leads to very compact representations of the network definition. It allows
for efficient design of complex networks.
• In their VGG paper, Simonyan and Ziserman experimented with various architectures. In
particular, they found that several layers of deep and narrow convolutions (i.e., 3 × 3) were
more effective than fewer layers of wider convolutions.
Exercises
1. When printing out the dimensions of the layers we only saw 8 results rather than 11. Where
did the remaining 3 layer information go?
2. Compared with AlexNet, VGG is much slower in terms of computation, and it also needs
more GPU memory. Analyze the reasons for this.
3. Try changing the height and width of the images in Fashion-MNIST from 224 to 96. What
influence does this have on the experiments?
4. Refer to Table 1 in the VGG paper (Simonyan & Zisserman, 2014) to construct other common
models, such as VGG-16 or VGG-19.
Discussions97
97

7.3 Network in Network (NiN)
LeNet, AlexNet, and VGG all share a common design pattern: extract features exploiting spatial
structure via a sequence of convolution and pooling layers and then post-process the representa-
tions via fully-connected layers. The improvements upon LeNet by AlexNet and VGG mainly lie in
how these later networks widen and deepen these two modules. Alternatively, one could imagine
using fully-connected layers earlier in the process. However, a careless use of dense layers might
give up the spatial structure of the representation entirely, network in network (NiN) blocks offer
an alternative. They were proposed based on a very simple insight: to use an MLP on the channels
for each pixel separately (Lin et al., 2013).
7.3.1 NiN Blocks
Recall that the inputs and outputs of convolutional layers consist of four-dimensional tensors with
axes corresponding to the example, channel, height, and width. Also recall that the inputs and
outputs of fully-connected layers are typically two-dimensional tensors corresponding to the ex-
ample and feature. The idea behind NiN is to apply a fully-connected layer at each pixel location
(for each height and width). If we tie the weights across each spatial location, we could think of
this as a 1 × 1 convolutional layer (as described in Section 6.4) or as a fully-connected layer acting
independently on each pixel location. Another way to view this is to think of each element in the
spatial dimension (height and width) as equivalent to an example and a channel as equivalent to
a feature.
Fig. 7.3.1 illustrates the main structural differences between VGG and NiN, and their blocks. The
NiN block consists of one convolutional layer followed by two 1 × 1 convolutional layers that act
as per-pixel fully-connected layers with ReLU activations. The convolution window shape of the
first layer is typically set by the user. The subsequent window shapes are fixed to 1 × 1.

Fig. 7.3.1: Comparing architectures of VGG and NiN, and their blocks.

npx.set_np()
def nin_block(num_channels, kernel_size, strides, padding):

blk.add(nn.Conv2D(num_channels, kernel_size, strides, padding,
activation='relu'),
nn.Conv2D(num_channels, kernel_size=1, activation='relu'),
nn.Conv2D(num_channels, kernel_size=1, activation='relu'))
return blk
7.3. Network in Network (NiN) 271

7.3.2 NiN Model
The original NiN network was proposed shortly after AlexNet and clearly draws some inspiration.
NiN uses convolutional layers with window shapes of 11×11, 5×5, and 3×3, and the corresponding
numbers of output channels are the same as in AlexNet. Each NiN block is followed by a maximum
pooling layer with a stride of 2 and a window shape of 3 × 3.
One significant difference between NiN and AlexNet is that NiN avoids fully-connected layers al-
together. Instead, NiN uses an NiN block with a number of output channels equal to the number
of label classes, followed by a global average pooling layer, yielding a vector of logits. One ad-
vantage of NiNʼs design is that it significantly reduces the number of required model parameters.
However, in practice, this design sometimes requires increased model training time.
net.add(nin_block(96, kernel_size=11, strides=4, padding=0),
nin_block(256, kernel_size=5, strides=1, padding=2),
nn.Dropout(0.5),
# There are 10 label classes
# The global average pooling layer automatically sets the window shape
# to the height and width of the input
nn.GlobalAvgPool2D(),
# Transform the four-dimensional output into two-dimensional output
# with a shape of (batch size, 10)
nn.Flatten())
We create a data example to see the output shape of each block.

net.initialize()
for layer in net:
X = layer(X)

dropout0 output shape: (1, 384, 5, 5)
flatten0 output shape: (1, 10)

7.3.3 Training
As before we use Fashion-MNIST to train the model. NiNʼs training is similar to that for AlexNet
and VGG.


Summary
• NiN uses blocks consisting of a convolutional layer and multiple 1 × 1 convolutional layers.
This can be used within the convolutional stack to allow for more per-pixel nonlinearity.
• NiN removes the fully-connected layers and replaces them with global average pooling (i.e.,
summing over all locations) after reducing the number of channels to the desired number
of outputs (e.g., 10 for Fashion-MNIST).
• Removing the fully-connected layers reduces overfitting. NiN has dramatically fewer pa-
rameters.
• The NiN design influenced many subsequent CNN designs.
Exercises
1. Tune the hyperparameters to improve the classification accuracy.

2. Why are there two 1 × 1 convolutional layers in the NiN block? Remove one of them, and
then observe and analyze the experimental phenomena.
3. Calculate the resource usage for NiN.
1. What is the number of parameters?
2. What is the amount of computation?
7.3. Network in Network (NiN) 273

3. What is the amount of memory needed during training?
4. What is the amount of memory needed during prediction?
4. What are possible problems with reducing the 384 × 5 × 5 representation to a 10 × 5 × 5
representation in one step?
Discussions98
7.4 Networks with Parallel Concatenations (GoogLeNet)
In 2014, GoogLeNet won the ImageNet Challenge, proposing a structure that combined the
strengths of NiN and paradigms of repeated blocks (Szegedy et al., 2015). One focus of the paper
was to address the question of which sized convolution kernels are best. After all, previous popu-
lar networks employed choices as small as 1 × 1 and as large as 11 × 11. One insight in this paper
was that sometimes it can be advantageous to employ a combination of variously-sized kernels. In
this section, we will introduce GoogLeNet, presenting a slightly simplified version of the original
model: we omit a few ad-hoc features that were added to stabilize training but are unnecessary
now with better training algorithms available.
7.4.1 Inception Blocks
The basic convolutional block in GoogLeNet is called an Inception block, likely named due to a quote
from the movie Inception (“We Need To Go Deeper”), which launched a viral meme.
Fig. 7.4.1: Structure of the Inception block.
As depicted in Fig. 7.4.1, the inception block consists of four parallel paths. The first three paths
use convolutional layers with window sizes of 1 × 1, 3 × 3, and 5 × 5 to extract information from
different spatial sizes. The middle two paths perform a 1 × 1 convolution on the input to reduce
the number of channels, reducing the modelʼs complexity. The fourth path uses a 3 × 3 maximum
pooling layer, followed by a 1 × 1 convolutional layer to change the number of channels. The
four paths all use appropriate padding to give the input and output the same height and width.
Finally, the outputs along each path are concatenated along the channel dimension and comprise
the blockʼs output. The commonly-tuned hyperparameters of the Inception block are the number
of output channels per layer.
98

npx.set_np()
class Inception(nn.Block):
# `c1`--`c4` are the number of output channels for each path
def __init__(self, c1, c2, c3, c4, **kwargs):
super(Inception, self).__init__(**kwargs)
# Path 1 is a single 1 x 1 convolutional layer
self.p1_1 = nn.Conv2D(c1, kernel_size=1, activation='relu')
# Path 2 is a 1 x 1 convolutional layer followed by a 3 x 3
# convolutional layer
self.p2_1 = nn.Conv2D(c2[0], kernel_size=1, activation='relu')
self.p2_2 = nn.Conv2D(c2[1], kernel_size=3, padding=1,
activation='relu')
# Path 3 is a 1 x 1 convolutional layer followed by a 5 x 5
self.p3_1 = nn.Conv2D(c3[0], kernel_size=1, activation='relu')
self.p3_2 = nn.Conv2D(c3[1], kernel_size=5, padding=2,
activation='relu')
# Path 4 is a 3 x 3 maximum pooling layer followed by a 1 x 1
self.p4_1 = nn.MaxPool2D(pool_size=3, strides=1, padding=1)
self.p4_2 = nn.Conv2D(c4, kernel_size=1, activation='relu')

p1 = self.p1_1(x)
p2 = self.p2_2(self.p2_1(x))
# Concatenate the outputs on the channel dimension
return np.concatenate((p1, p2, p3, p4), axis=1)
To gain some intuition for why this network works so well, consider the combination of the filters.
They explore the image in varying ranges. This means that details at different extents can be
recognized efficiently by different filters. At the same time, we can allocate different amounts of
parameters for different ranges (e.g., more for short range but not ignore the long range entirely).
7.4.2 GoogLeNet Model
As shown in Fig. 7.4.2, GoogLeNet uses a stack of a total of 9 inception blocks and global aver-
age pooling to generate its estimates. Maximum pooling between inception blocks reduces the
dimensionality. The first module is similar to AlexNet and LeNet. The stack of blocks is inherited
from VGG and the global average pooling avoids a stack of fully-connected layers at the end.
7.4. Networks with Parallel Concatenations (GoogLeNet) 275

Fig. 7.4.2: The GoogLeNet architecture.
We can now implement GoogLeNet piece by piece. The first module uses a 64-channel 7 × 7 con-
volutional layer.
b1 = nn.Sequential()
b1.add(nn.Conv2D(64, kernel_size=7, strides=2, padding=3, activation='relu'),
nn.MaxPool2D(pool_size=3, strides=2, padding=1))
The second module uses two convolutional layers: first, a 64-channel 1 × 1 convolutional layer,
then a 3 × 3 convolutional layer that triples the number of channels. This corresponds to the
second path in the Inception block.
b2.add(nn.Conv2D(64, kernel_size=1, activation='relu'),
The third module connects two complete Inception blocks in series. The number of output chan-
nels of the first Inception block is 64 + 128 + 32 + 32 = 256, and the number-of-output-channel
ratio among the four paths is 64 : 128 : 32 : 32 = 2 : 4 : 1 : 1. The second and third paths first
reduce the number of input channels to 96/192 = 1/2 and 16/192 = 1/12, respectively, and then
connect the second convolutional layer. The number of output channels of the second Inception
block is increased to 128 + 192 + 96 + 64 = 480, and the number-of-output-channel ratio among
the four paths is 128 : 192 : 96 : 64 = 4 : 6 : 3 : 2. The second and third paths first reduce the
number of input channels to 128/256 = 1/2 and 32/256 = 1/8, respectively.

b3.add(Inception(64, (96, 128), (16, 32), 32),
Inception(128, (128, 192), (32, 96), 64),
The fourth module is more complicated. It connects five Inception blocks in series, and they have
192+208+48+64 = 512, 160+224+64+64 = 512, 128+256+64+64 = 512, 112+288+64+64 = 528,
and 256+320+128+128 = 832 output channels, respectively. The number of channels assigned to
these paths is similar to that in the third module: the second path with the 3×3 convolutional layer
outputs the largest number of channels, followed by the first path with only the 1×1 convolutional
layer, the third path with the 5×5 convolutional layer, and the fourth path with the 3×3 maximum
pooling layer. The second and third paths will first reduce the number of channels according to
the ratio. These ratios are slightly different in different Inception blocks.
b4.add(Inception(192, (96, 208), (16, 48), 64),
Inception(160, (112, 224), (24, 64), 64),
Inception(128, (128, 256), (24, 64), 64),
Inception(112, (144, 288), (32, 64), 64),
Inception(256, (160, 320), (32, 128), 128),
The fifth module has two Inception blocks with 256 + 320 + 128 + 128 = 832 and 384 + 384 + 128 +
128 = 1024 output channels. The number of channels assigned to each path is the same as that in
the third and fourth modules, but differs in specific values. It should be noted that the fifth block is
followed by the output layer. This block uses the global average pooling layer to change the height
and width of each channel to 1, just as in NiN. Finally, we turn the output into a two-dimensional
array followed by a fully-connected layer whose number of outputs is the number of label classes.
b5.add(Inception(256, (160, 320), (32, 128), 128),
Inception(384, (192, 384), (48, 128), 128),
nn.GlobalAvgPool2D())
net.add(b1, b2, b3, b4, b5, nn.Dense(10))
The GoogLeNet model is computationally complex, so it is not as easy to modify the number of
channels as in VGG. To have a reasonable training time on Fashion-MNIST, we reduce the input
height and width from 224 to 96. This simplifies the computation. The changes in the shape of the
output between the various modules are demonstrated below.

net.initialize()
for layer in net:
X = layer(X)

7.4. Networks with Parallel Concatenations (GoogLeNet) 277

7.4.3 Training
As before, we train our model using the Fashion-MNIST dataset. We transform it to 96 × 96 pixel
resolution before invoking the training procedure.


Summary
• The Inception block is equivalent to a subnetwork with four paths. It extracts information
in parallel through convolutional layers of different window shapes and maximum pooling
layers. 1 × 1 convolutions reduce channel dimensionality on a per-pixel level. Maximum
pooling reduces the resolution.
• GoogLeNet connects multiple well-designed Inception blocks with other layers in series.
The ratio of the number of channels assigned in the Inception block is obtained through a
large number of experiments on the ImageNet dataset.
• GoogLeNet, as well as its succeeding versions, was one of the most efficient models on Im-
ageNet, providing similar test accuracy with lower computational complexity.

Exercises
1. There are several iterations of GoogLeNet. Try to implement and run them. Some of them
include the following:
• Add a batch normalization layer (Ioffe & Szegedy, 2015), as described later in Section
7.5.
• Make adjustments to the Inception block (Szegedy et al., 2016).
• Use label smoothing for model regularization (Szegedy et al., 2016).
• Include it in the residual connection (Szegedy et al., 2017), as described later in Section
7.6.
2. What is the minimum image size for GoogLeNet to work?
3. Compare the model parameter sizes of AlexNet, VGG, and NiN with GoogLeNet. How do the
latter two network architectures significantly reduce the model parameter size?
4. Why do we need a long range convolution initially?
Discussions99
7.5 Batch Normalization
Training deep neural networks is difficult. And getting them to converge in a reasonable amount
of time can be tricky. In this section, we describe batch normalization, a popular and effective
technique that consistently accelerates the convergence of deep networks (Ioffe & Szegedy, 2015).
Together with residual blocks—covered later in Section 7.6—batch normalization has made it pos-
sible for practitioners to routinely train networks with over 100 layers.
7.5.1 Training Deep Networks
To motivate batch normalization, let us review a few practical challenges that arise when training
machine learning models and neural networks in particular.
First, choices regarding data preprocessing often make an enormous difference in the final re-
sults. Recall our application of MLPs to predicting house prices (Section 4.10). Our first step when
working with real data was to standardize our input features to each have a mean of zero and vari-
ance of one. Intuitively, this standardization plays nicely with our optimizers because it puts the
parameters a priori at a similar scale.
Second, for a typical MLP or CNN, as we train, the variables (e.g., affine transformation outputs in
MLP) in intermediate layers may take values with widely varying magnitudes: both along the lay-
ers from the input to the output, across units in the same layer, and over time due to our updates to
the model parameters. The inventors of batch normalization postulated informally that this drift
in the distribution of such variables could hamper the convergence of the network. Intuitively,
we might conjecture that if one layer has variable values that are 100 times that of another layer,
this might necessitate compensatory adjustments in the learning rates.
Third, deeper networks are complex and easily capable of overfitting. This means that regular-
ization becomes more critical.
99
7.5. Batch Normalization 279

Batch normalization is applied to individual layers (optionally, to all of them) and works as follows:
In each training iteration, we first normalize the inputs (of batch normalization) by subtracting
their mean and dividing by their standard deviation, where both are estimated based on the statis-
tics of the current minibatch. Next, we apply a scale coefficient and a scale offset. It is precisely
due to this normalization based on batch statistics that batch normalization derives its name.
Note that if we tried to apply batch normalization with minibatches of size 1, we would not be able
to learn anything. That is because after subtracting the means, each hidden unit would take value
0! As you might guess, since we are devoting a whole section to batch normalization, with large
enough minibatches, the approach proves effective and stable. One takeaway here is that when
applying batch normalization, the choice of batch size may be even more significant than without
batch normalization.
Formally, denoting by x ∈ B an input to batch normalization (BN) that is from a minibatch B,
batch normalization transforms x according to the following expression:
x − µ̂B
BN(x) = γ ⊙ + β. (7.5.1)
σ̂ B
In (7.5.1), µ̂B is the sample mean and σ̂ B is the sample standard deviation of the minibatch B.
After applying standardization, the resulting minibatch has zero mean and unit variance. Because
the choice of unit variance (vs. some other magic number) is an arbitrary choice, we commonly
include element-wise scale parameter γ and shift parameter β that have the same shape as x. Note
that γ and β are parameters that need to be learned jointly with the other model parameters.
Consequently, the variable magnitudes for intermediate layers cannot diverge during training be-
cause batch normalization actively centers and rescales them back to a given mean and size (via
µ̂B and σ̂ B ). One piece of practitionerʼs intuition or wisdom is that batch normalization seems to
allow for more aggressive learning rates.
Formally, we calculate µ̂B and σ̂ B in (7.5.1) as follows:
1 ∑
µ̂B = x,
|B|
x∈B
(7.5.2)
1 ∑
2
σ̂ B = (x − µ̂B )2 + ϵ.
|B|
x∈B
Note that we add a small constant ϵ > 0 to the variance estimate to ensure that we never attempt
division by zero, even in cases where the empirical variance estimate might vanish. The estimates
µ̂B and σ̂ B counteract the scaling issue by using noisy estimates of mean and variance. You might
think that this noisiness should be a problem. As it turns out, this is actually beneficial.
This turns out to be a recurring theme in deep learning. For reasons that are not yet well-
characterized theoretically, various sources of noise in optimization often lead to faster training
and less overfitting: this variation appears to act as a form of regularization. In some preliminary
research, (Teye et al., 2018) and (Luo et al., 2018) relate the properties of batch normalization to
Bayesian priors and penalties respectively. In particular, this sheds some light on the puzzle of
why batch normalization works best for moderate minibatches sizes in the 50 ∼ 100 range.
Fixing a trained model, you might think that we would prefer using the entire dataset to estimate
the mean and variance. Once training is complete, why would we want the same image to be clas-
sified differently, depending on the batch in which it happens to reside? During training, such
exact calculation is infeasible because the intermediate variables for all data examples change ev-
ery time we update our model. However, once the model is trained, we can calculate the means
and variances of each layerʼs variables based on the entire dataset. Indeed this is standard practice

for models employing batch normalization and thus batch normalization layers function differ-
ently in training mode (normalizing by minibatch statistics) and in prediction mode (normalizing by
dataset statistics).
We are now ready to take a look at how batch normalization works in practice.
7.5.2 Batch Normalization Layers
Batch normalization implementations for fully-connected layers and convolutional layers are
slightly different. We discuss both cases below. Recall that one key differences between batch
normalization and other layers is that because batch normalization operates on a full minibatch
at a time, we cannot just ignore the batch dimension as we did before when introducing other
layers.
Fully-Connected Layers
When applying batch normalization to fully-connected layers, the original paper inserts batch
normalization after the affine transformation and before the nonlinear activation function (later
applications may insert batch normalization right after activation functions) (Ioffe & Szegedy,
2015). Denoting the input to the fully-connected layer by x, the affine transformation by Wx + b
(with the weight parameter W and the bias parameter b), and the activation function by ϕ, we
can express the computation of a batch-normalization-enabled, fully-connected layer output h as
follows:
h = ϕ(BN(Wx + b)). (7.5.3)
Recall that mean and variance are computed on the same minibatch on which the transformation
is applied.
Convolutional Layers
Similarly, with convolutional layers, we can apply batch normalization after the convolution and
before the nonlinear activation function. When the convolution has multiple output channels, we
need to carry out batch normalization for each of the outputs of these channels, and each channel
has its own scale and shift parameters, both of which are scalars. Assume that our minibatches
contain m examples and that for each channel, the output of the convolution has height p and
width q. For convolutional layers, we carry out each batch normalization over the m ·p ·q elements
per output channel simultaneously. Thus, we collect the values over all spatial locations when
computing the mean and variance and consequently apply the same mean and variance within a
given channel to normalize the value at each spatial location.

Batch Normalization During Prediction
As we mentioned earlier, batch normalization typically behaves differently in training mode and
prediction mode. First, the noise in the sample mean and the sample variance arising from es-
timating each on minibatches are no longer desirable once we have trained the model. Second,
we might not have the luxury of computing per-batch normalization statistics. For example, we
might need to apply our model to make one prediction at a time.
Typically, after training, we use the entire dataset to compute stable estimates of the variable
statistics and then fix them at prediction time. Consequently, batch normalization behaves dif-
ferently during training and at test time. Recall that dropout also exhibits this characteristic.
Below, we implement a batch normalization layer with tensors from scratch.

from mxnet import autograd, np, npx, init
npx.set_np()
def batch_norm(X, gamma, beta, moving_mean, moving_var, eps, momentum):

# Use àutograd` to determine whether the current mode is training mode or
# prediction mode
if not autograd.is_training():
# If it is prediction mode, directly use the mean and variance
# obtained by moving average
X_hat = (X - moving_mean) / np.sqrt(moving_var + eps)
else:
assert len(X.shape) in (2, 4)
if len(X.shape) == 2:
# When using a fully-connected layer, calculate the mean and
# variance on the feature dimension
mean = X.mean(axis=0)
var = ((X - mean) ** 2).mean(axis=0)
else:
# When using a two-dimensional convolutional layer, calculate the
# mean and variance on the channel dimension (axis=1). Here we
# need to maintain the shape of `X`, so that the broadcasting
# operation can be carried out later
mean = X.mean(axis=(0, 2, 3), keepdims=True)
var = ((X - mean) ** 2).mean(axis=(0, 2, 3), keepdims=True)
# In training mode, the current mean and variance are used for the
# standardization
X_hat = (X - mean) / np.sqrt(var + eps)
# Update the mean and variance using moving average
moving_mean = momentum * moving_mean + (1.0 - momentum) * mean
moving_var = momentum * moving_var + (1.0 - momentum) * var
Y = gamma * X_hat + beta # Scale and shift
return Y, moving_mean, moving_var
We can now create a proper BatchNorm layer. Our layer will maintain proper parameters for scale
gamma and shift beta, both of which will be updated in the course of training. Additionally, our
layer will maintain moving averages of the means and variances for subsequent use during model
prediction.

Putting aside the algorithmic details, note the design pattern underlying our implementation of
the layer. Typically, we define the mathematics in a separate function, say batch_norm. We then
integrate this functionality into a custom layer, whose code mostly addresses bookkeeping mat-
ters, such as moving data to the right device context, allocating and initializing any required vari-
ables, keeping track of moving averages (here for mean and variance), and so on. This pattern
enables a clean separation of mathematics from boilerplate code. Also note that for the sake of
convenience we did not worry about automatically inferring the input shape here, thus we need
to specify the number of features throughout. Do not worry, the high-level batch normalization
APIs in the deep learning framework will care of this for us and we will demonstrate that later.
class BatchNorm(nn.Block):
# `num_features`: the number of outputs for a fully-connected layer
# or the number of output channels for a convolutional layer. `num_dims`:
# 2 for a fully-connected layer and 4 for a convolutional layer
def __init__(self, num_features, num_dims, **kwargs):
if num_dims == 2:
shape = (1, num_features)
else:
shape = (1, num_features, 1, 1)
# The scale parameter and the shift parameter (model parameters) are
# initialized to 1 and 0, respectively
self.gamma = self.params.get('gamma', shape=shape, init=init.One())
self.beta = self.params.get('beta', shape=shape, init=init.Zero())
# The variables that are not model parameters are initialized to 0
self.moving_mean = np.zeros(shape)
self.moving_var = np.zeros(shape)

# If `X` is not on the main memory, copy `moving_mean` and
# `moving_var` to the device where `X` is located
if self.moving_mean.ctx != X.ctx:
self.moving_mean = self.moving_mean.copyto(X.ctx)
self.moving_var = self.moving_var.copyto(X.ctx)
# Save the updated `moving_mean` and `moving_var`
Y, self.moving_mean, self.moving_var = batch_norm(
X, self.gamma.data(), self.beta.data(), self.moving_mean,
self.moving_var, eps=1e-12, momentum=0.9)
return Y
7.5.4 Applying Batch Normalization in LeNet
To see how to apply BatchNorm in context, below we apply it to a traditional LeNet model (Section
6.6). Recall that batch normalization is applied after the convolutional layers or fully-connected
layers but before the corresponding activation functions.
net.add(nn.Conv2D(6, kernel_size=5),
BatchNorm(6, num_dims=4),
nn.Activation('sigmoid'),
nn.Conv2D(16, kernel_size=5),

nn.Dense(120),
nn.Dense(84),
nn.Dense(10))
As before, we will train our network on the Fashion-MNIST dataset. This code is virtually identical
to that when we first trained LeNet (Section 6.6). The main difference is the considerably larger
learning rate.


Let us have a look at the scale parameter gamma and the shift parameter beta learned from the first
batch normalization layer.
net[1].gamma.data().reshape(-1,), net[1].beta.data().reshape(-1,)
(array([1.956041 , 0.8407571, 3.1203268, 1.492053 , 1.2769177, 2.3011174], ctx=gpu(0)),

array([ 1.1319363 , 0.36316666, -3.2172477 , -0.7223122 , -0.57278776,
-0.21445163], ctx=gpu(0)))

Compared with the BatchNorm class, which we just defined ourselves, we can use the BatchNorm
class defined in high-level APIs from the deep learning framework directly. The code looks virtu-
ally identical to the application our implementation above.
net.add(nn.Conv2D(6, kernel_size=5),
nn.BatchNorm(),
nn.Conv2D(16, kernel_size=5),
nn.BatchNorm(),
nn.Dense(120),
nn.BatchNorm(),
nn.Dense(84),
nn.BatchNorm(),
nn.Dense(10))
Below, we use the same hyperparameters to train our model. Note that as usual, the high-level API
variant runs much faster because its code has been compiled to C++ or CUDA while our custom
implementation must be interpreted by Python.


7.5.6 Controversy
Intuitively, batch normalization is thought to make the optimization landscape smoother. How-
ever, we must be careful to distinguish between speculative intuitions and true explanations for
the phenomena that we observe when training deep models. Recall that we do not even know why
simpler deep neural networks (MLPs and conventional CNNs) generalize well in the first place.
Even with dropout and weight decay, they remain so flexible that their ability to generalize to
unseen data cannot be explained via conventional learning-theoretic generalization guarantees.
In the original paper proposing batch normalization, the authors, in addition to introducing a
powerful and useful tool, offered an explanation for why it works: by reducing internal covari-
ate shift. Presumably by internal covariate shift the authors meant something like the intuition
expressed above—the notion that the distribution of variable values changes over the course of
training. However, there were two problems with this explanation: i) This drift is very different
from covariate shift, rendering the name a misnomer. ii) The explanation offers an under-specified
intuition but leaves the question of why precisely this technique works an open question wanting for
a rigorous explanation. Throughout this book, we aim to convey the intuitions that practitioners
use to guide their development of deep neural networks. However, we believe that it is important
to separate these guiding intuitions from established scientific fact. Eventually, when you master
this material and start writing your own research papers you will want to be clear to delineate
between technical claims and hunches.
Following the success of batch normalization, its explanation in terms of internal covariate shift
has repeatedly surfaced in debates in the technical literature and broader discourse about how to
present machine learning research. In a memorable speech given while accepting a Test of Time
Award at the 2017 NeurIPS conference, Ali Rahimi used internal covariate shift as a focal point in
an argument likening the modern practice of deep learning to alchemy. Subsequently, the ex-
ample was revisited in detail in a position paper outlining troubling trends in machine learning
(Lipton & Steinhardt, 2018). Other authors have proposed alternative explanations for the success
of batch normalization, some claiming that batch normalizationʼs success comes despite exhibit-
ing behavior that is in some ways opposite to those claimed in the original paper (Santurkar et al.,
2018).
We note that the internal covariate shift is no more worthy of criticism than any of thousands of
similarly vague claims made every year in the technical machine learning literature. Likely, its
resonance as a focal point of these debates owes to its broad recognizability to the target audience.
Batch normalization has proven an indispensable method, applied in nearly all deployed image
classifiers, earning the paper that introduced the technique tens of thousands of citations.
Summary
• During model training, batch normalization continuously adjusts the intermediate output
of the neural network by utilizing the mean and standard deviation of the minibatch, so that
the values of the intermediate output in each layer throughout the neural network are more
stable.
• The batch normalization methods for fully-connected layers and convolutional layers are
slightly different.
• Like a dropout layer, batch normalization layers have different computation results in train-
ing mode and prediction mode.

• Batch normalization has many beneficial side effects, primarily that of regularization. On
the other hand, the original motivation of reducing internal covariate shift seems not to be
a valid explanation.
Exercises
1. Can we remove the bias parameter from the fully-connected layer or the convolutional layer
before the batch normalization? Why?
2. Compare the learning rates for LeNet with and without batch normalization.
1. Plot the increase in training and test accuracy.
2. How large can you make the learning rate?
3. Do we need batch normalization in every layer? Experiment with it?
4. Can you replace dropout by batch normalization? How does the behavior change?
5. Fix the parameters beta and gamma, and observe and analyze the results.
6. Review the online documentation for BatchNorm from the high-level APIs to see the other
applications for batch normalization.
7. Research ideas: think of other normalization transforms that you can apply? Can you apply
the probability integral transform? How about a full rank covariance estimate?
Discussions100
7.6 Residual Networks (ResNet)
As we design increasingly deeper networks it becomes imperative to understand how adding lay-
ers can increase the complexity and expressiveness of the network. Even more important is the
ability to design networks where adding layers makes networks strictly more expressive rather
than just different. To make some progress we need a bit of mathematics.
7.6.1 Function Classes
Consider F, the class of functions that a specific network architecture (together with learning
rates and other hyperparameter settings) can reach. That is, for all f ∈ F there exists some set of
parameters (e.g., weights and biases) that can be obtained through training on a suitable dataset.
Let us assume that f ∗ is the “truth” function that we really would like to find. If it is in F, we are in
good shape but typically we will not be quite so lucky. Instead, we will try to find some fF∗ which
is our best bet within F. For instance, given a dataset with features X and labels y, we might try
finding it by solving the following optimization problem:
fF∗ := argmin L(X, y, f ) subject to f ∈ F . (7.6.1)

f
It is only reasonable to assume that if we design a different and more powerful architecture F ′ we
should arrive at a better outcome. In other words, we would expect that fF∗ ′ is “better” than fF∗ .
However, if F ̸⊆ F ′ there is no guarantee that this should even happen. In fact, fF∗ ′ might well
100
7.6. Residual Networks (ResNet) 287

be worse. As illustrated by Fig. 7.6.1, for non-nested function classes, a larger function class does
not always move closer to the “truth” function f ∗ . For instance, on the left of Fig. 7.6.1, though
F3 is closer to f ∗ than F1 , F6 moves away and there is no guarantee that further increasing the
complexity can reduce the distance from f ∗ . With nested function classes where F1 ⊆ . . . ⊆ F6
on the right of Fig. 7.6.1, we can avoid the aforementioned issue from the non-nested function
classes.
Fig. 7.6.1: For non-nested function classes, a larger (indicated by area) function class does not
guarantee to get closer to the “truth” function (f ∗ ). This does not happen in nested function
classes.
Thus, only if larger function classes contain the smaller ones are we guaranteed that increasing
them strictly increases the expressive power of the network. For deep neural networks, if we can
train the newly-added layer into an identity function f (x) = x, the new model will be as effective
as the original model. As the new model may get a better solution to fit the training dataset, the
added layer might make it easier to reduce training errors.
This is the question that He et al. considered when working on very deep computer vision models
(He et al., 2016a). At the heart of their proposed residual network (ResNet) is the idea that every
additional layer should more easily contain the identity function as one of its elements. These
considerations are rather profound but they led to a surprisingly simple solution, a residual block.
With it, ResNet won the ImageNet Large Scale Visual Recognition Challenge in 2015. The design
had a profound influence on how to build deep neural networks.
7.6.2 Residual Blocks
Let us focus on a local part of a neural network, as depicted in Fig. 7.6.2. Denote the input by x.
We assume that the desired underlying mapping we want to obtain by learning is f (x), to be used
as the input to the activation function on the top. On the left of Fig. 7.6.2, the portion within the
dotted-line box must directly learn the mapping f (x). On the right, the portion within the dotted-
line box needs to learn the residual mapping f (x) − x, which is how the residual block derives its
name. If the identity mapping f (x) = x is the desired underlying mapping, the residual mapping
is easier to learn: we only need to push the weights and biases of the upper weight layer (e.g.,
fully-connected layer and convolutional layer) within the dotted-line box to zero. The right figure
in Fig. 7.6.2 illustrates the residual block of ResNet, where the solid line carrying the layer input x to

the addition operator is called a residual connection (or shortcut connection). With residual blocks,
inputs can forward propagate faster through the residual connections across layers.
Fig. 7.6.2: A regular block (left) and a residual block (right).
ResNet follows VGGʼs full 3 × 3 convolutional layer design. The residual block has two 3 × 3 con-
volutional layers with the same number of output channels. Each convolutional layer is followed
by a batch normalization layer and a ReLU activation function. Then, we skip these two convolu-
tion operations and add the input directly before the final ReLU activation function. This kind of
design requires that the output of the two convolutional layers has to be of the same shape as the
input, so that they can be added together. If we want to change the number of channels, we need
to introduce an additional 1 × 1 convolutional layer to transform the input into the desired shape
for the addition operation. Let us have a look at the code below.

npx.set_np()
class Residual(nn.Block): #@save

def __init__(self, num_channels, use_1x1conv=False, strides=1, **kwargs):
self.conv1 = nn.Conv2D(num_channels, kernel_size=3, padding=1,
strides=strides)
self.conv2 = nn.Conv2D(num_channels, kernel_size=3, padding=1)
if use_1x1conv:
self.conv3 = nn.Conv2D(num_channels, kernel_size=1,
strides=strides)
else:
self.conv3 = None
self.bn1 = nn.BatchNorm()

Y = npx.relu(self.bn1(self.conv1(X)))

Y = self.bn2(self.conv2(Y))
if self.conv3:
X = self.conv3(X)
return npx.relu(Y + X)
This code generates two types of networks: one where we add the input to the output before ap-
plying the ReLU nonlinearity whenever use_1x1conv=False, and one where we adjust channels
and resolution by means of a 1 × 1 convolution before adding. Fig. 7.6.3 illustrates this:
Fig. 7.6.3: ResNet block with and without 1 × 1 convolution.
Now let us look at a situation where the input and output are of the same shape.
blk = Residual(3)
blk.initialize()
blk(X).shape
(4, 3, 6, 6)
We also have the option to halve the output height and width while increasing the number of
output channels.
blk = Residual(6, use_1x1conv=True, strides=2)

blk.initialize()
blk(X).shape
(4, 6, 3, 3)

7.6.3 ResNet Model
The first two layers of ResNet are the same as those of the GoogLeNet we described before: the
7×7 convolutional layer with 64 output channels and a stride of 2 is followed by the 3×3 maximum
pooling layer with a stride of 2. The difference is the batch normalization layer added after each
convolutional layer in ResNet.
net.add(nn.Conv2D(64, kernel_size=7, strides=2, padding=3),
nn.BatchNorm(), nn.Activation('relu'),
GoogLeNet uses four modules made up of Inception blocks. However, ResNet uses four modules
made up of residual blocks, each of which uses several residual blocks with the same number
of output channels. The number of channels in the first module is the same as the number of
input channels. Since a maximum pooling layer with a stride of 2 has already been used, it is not
necessary to reduce the height and width. In the first residual block for each of the subsequent
modules, the number of channels is doubled compared with that of the previous module, and the
height and width are halved.
Now, we implement this module. Note that special processing has been performed on the first
module.
def resnet_block(num_channels, num_residuals, first_block=False):

for i in range(num_residuals):
if i == 0 and not first_block:
blk.add(Residual(num_channels, use_1x1conv=True, strides=2))
else:
blk.add(Residual(num_channels))
return blk
Then, we add all the modules to ResNet. Here, two residual blocks are used for each module.
net.add(resnet_block(64, 2, first_block=True),
resnet_block(128, 2),
resnet_block(512, 2))
Finally, just like GoogLeNet, we add a global average pooling layer, followed by the fully-connected
layer output.
net.add(nn.GlobalAvgPool2D(), nn.Dense(10))
There are 4 convolutional layers in each module (excluding the 1×1 convolutional layer). Together
with the first 7 × 7 convolutional layer and the final fully-connected layer, there are 18 layers in
total. Therefore, this model is commonly known as ResNet-18. By configuring different numbers
of channels and residual blocks in the module, we can create different ResNet models, such as
the deeper 152-layer ResNet-152. Although the main architecture of ResNet is similar to that of
GoogLeNet, ResNetʼs structure is simpler and easier to modify. All these factors have resulted in
the rapid and widespread use of ResNet. Fig. 7.6.4 depicts the full ResNet-18.

Fig. 7.6.4: The ResNet-18 architecture.
Before training ResNet, let us observe how the input shape changes across different modules in
ResNet. As in all the previous architectures, the resolution decreases while the number of chan-
nels increases up until the point where a global average pooling layer aggregates all features.

net.initialize()
for layer in net:
X = layer(X)

batchnorm4 output shape: (1, 64, 112, 112)
relu0 output shape: (1, 64, 112, 112)
7.6.4 Training
We train ResNet on the Fashion-MNIST dataset, just like before.


Summary
• Nested function classes are desirable. Learning an additional layer in deep neural networks
as an identity function (though this is an extreme case) should be made easy.
• The residual mapping can learn the identity function more easily, such as pushing parame-
ters in the weight layer to zero.
• We can train an effective deep neural network by having residual blocks. Inputs can forward
propagate faster through the residual connections across layers.
• ResNet had a major influence on the design of subsequent deep neural networks, both for
convolutional and sequential nature.

Exercises
1. What are the major differences between the Inception block in Fig. 7.4.1 and the residual
block? After removing some paths in the Inception block, how are they related to each other?
2. Refer to Table 1 in the ResNet paper (He et al., 2016a) to implement different variants.
3. For deeper networks, ResNet introduces a “bottleneck” architecture to reduce model com-
plexity. Try to implement it.
4. In subsequent versions of ResNet, the authors changed the “convolution, batch normal-
ization, and activation” structure to the “batch normalization, activation, and convolution”
structure. Make this improvement yourself. See Figure 1 in (He et al., 2016b) for details.
5. Why cannot we just increase the complexity of functions without bound, even if the function
classes are nested?
Discussions101
7.7 Densely Connected Networks (DenseNet)
ResNet significantly changed the view of how to parametrize the functions in deep networks.
DenseNet (dense convolutional network) is to some extent the logical extension of this (Huang
et al., 2017). To understand how to arrive at it, let us take a small detour to mathematics.
7.7.1 From ResNet to DenseNet
Recall the Taylor expansion for functions. For the point x = 0 it can be written as
f ′′ (0) 2 f ′′′ (0) 3

f (x) = f (0) + f ′ (0)x + x + x + .... (7.7.1)
2! 3!
The key point is that it decomposes a function into increasingly higher order terms. In a similar
vein, ResNet decomposes functions into
f (x) = x + g(x). (7.7.2)
That is, ResNet decomposes f into a simple linear term and a more complex nonlinear one. What
if we want to capture (not necessarily add) information beyond two terms? One solution was
DenseNet (Huang et al., 2017).
Fig. 7.7.1: The main difference between ResNet (left) and DenseNet (right) in cross-layer connec-
tions: use of addition and use of concatenation.
101

As shown in Fig. 7.7.1, the key difference between ResNet and DenseNet is that in the latter case
outputs are concatenated (denoted by [, ]) rather than added. As a result, we perform a mapping
from x to its values after applying an increasingly complex sequence of functions:
x → [x, f1 (x), f2 ([x, f1 (x)]), f3 ([x, f1 (x), f2 ([x, f1 (x)])]), . . .] . (7.7.3)
In the end, all these functions are combined in MLP to reduce the number of features again. In
terms of implementation this is quite simple: rather than adding terms, we concatenate them.
The name DenseNet arises from the fact that the dependency graph between variables becomes
quite dense. The last layer of such a chain is densely connected to all previous layers. The dense
connections are shown in Fig. 7.7.2.
Fig. 7.7.2: Dense connections in DenseNet.
The main components that compose a DenseNet are dense blocks and transition layers. The for-
mer define how the inputs and outputs are concatenated, while the latter control the number of
channels so that it is not too large.
7.7.2 Dense Blocks
DenseNet uses the modified “batch normalization, activation, and convolution” structure of
ResNet (see the exercise in Section 7.6). First, we implement this convolution block structure.

npx.set_np()
def conv_block(num_channels):
blk.add(nn.BatchNorm(),
nn.Activation('relu'),
nn.Conv2D(num_channels, kernel_size=3, padding=1))
return blk
A dense block consists of multiple convolution blocks, each using the same number of output chan-
nels. In the forward propagation, however, we concatenate the input and output of each convo-
lution block on the channel dimension.
class DenseBlock(nn.Block):
def __init__(self, num_convs, num_channels, **kwargs):
self.net = nn.Sequential()
7.7. Densely Connected Networks (DenseNet) 295

for _ in range(num_convs):
self.net.add(conv_block(num_channels))

for blk in self.net:
Y = blk(X)
# Concatenate the input and output of each block on the channel
# dimension
X = np.concatenate((X, Y), axis=1)
return X
In the following example, we define a DenseBlock instance with 2 convolution blocks of 10 output
channels. When using an input with 3 channels, we will get an output with 3+2×10 = 23 channels.
The number of convolution block channels controls the growth in the number of output channels
relative to the number of input channels. This is also referred to as the growth rate.
blk = DenseBlock(2, 10)

blk.initialize()
Y = blk(X)
Y.shape
(4, 23, 8, 8)
7.7.3 Transition Layers
Since each dense block will increase the number of channels, adding too many of them will lead
to an excessively complex model. A transition layer is used to control the complexity of the model.
It reduces the number of channels by using the 1 × 1 convolutional layer and halves the height
and width of the average pooling layer with a stride of 2, further reducing the complexity of the
model.
def transition_block(num_channels):
blk.add(nn.BatchNorm(), nn.Activation('relu'),
nn.Conv2D(num_channels, kernel_size=1),
nn.AvgPool2D(pool_size=2, strides=2))
return blk
Apply a transition layer with 10 channels to the output of the dense block in the previous example.
This reduces the number of output channels to 10, and halves the height and width.
blk = transition_block(10)
blk.initialize()
blk(Y).shape
(4, 10, 4, 4)

7.7.4 DenseNet Model
Next, we will construct a DenseNet model. DenseNet first uses the same single convolutional layer
and maximum pooling layer as in ResNet.
nn.BatchNorm(), nn.Activation('relu'),
Then, similar to the four modules made up of residual blocks that ResNet uses, DenseNet uses
four dense blocks. Similar to ResNet, we can set the number of convolutional layers used in each
dense block. Here, we set it to 4, consistent with the ResNet-18 model in Section 7.6. Furthermore,
we set the number of channels (i.e., growth rate) for the convolutional layers in the dense block
to 32, so 128 channels will be added to each dense block.
In ResNet, the height and width are reduced between each module by a residual block with a
stride of 2. Here, we use the transition layer to halve the height and width and halve the number
of channels.
# `num_channels`: the current number of channels

num_channels, growth_rate = 64, 32
num_convs_in_dense_blocks = [4, 4, 4, 4]
for i, num_convs in enumerate(num_convs_in_dense_blocks):

net.add(DenseBlock(num_convs, growth_rate))
# This is the number of output channels in the previous dense block
num_channels += num_convs * growth_rate
# A transition layer that halves the number of channels is added between
# the dense blocks
if i != len(num_convs_in_dense_blocks) - 1:
num_channels //= 2
net.add(transition_block(num_channels))
Similar to ResNet, a global pooling layer and a fully-connected layer are connected at the end to
produce the output.
net.add(nn.BatchNorm(),
nn.Activation('relu'),
nn.GlobalAvgPool2D(),
nn.Dense(10))
7.7.5 Training
Since we are using a deeper network here, in this section, we will reduce the input height and
width from 224 to 96 to simplify the computation.

7.7. Densely Connected Networks (DenseNet) 297

Summary
• In terms of cross-layer connections, unlike ResNet, where inputs and outputs are added to-
gether, DenseNet concatenates inputs and outputs on the channel dimension.
• The main components that compose DenseNet are dense blocks and transition layers.
• We need to keep the dimensionality under control when composing the network by adding
transition layers that shrink the number of channels again.
Exercises
1. Why do we use average pooling rather than maximum pooling in the transition layer?
2. One of the advantages mentioned in the DenseNet paper is that its model parameters are
smaller than those of ResNet. Why is this the case?
3. One problem for which DenseNet has been criticized is its high memory consumption.
1. Is this really the case? Try to change the input shape to 224 × 224 to see the actual GPU
memory consumption.
2. Can you think of an alternative means of reducing the memory consumption? How
would you need to change the framework?
4. Implement the various DenseNet versions presented in Table 1 of the DenseNet paper
(Huang et al., 2017).
5. Design an MLP-based model by applying the DenseNet idea. Apply it to the housing price
prediction task in Section 4.10.
Discussions102
102

8 | Recurrent Neural Networks
So far we encountered two types of data: generic vectors and images. For the latter we designed
specialized layers to take advantage of the regularity properties in them. In other words, if we were
to permute the pixels in an image, it would be much more difficult to reason about its content of
something that would look much like the background of a test pattern in the times of analog TV.
Most importantly, so far we tacitly assumed that our data is generated i.i.d., i.e., independently
and identically distributed, all drawn from some distribution. Unfortunately, this is not true for
most data. For instance, the words in this paragraph are written in sequence, and it would be
quite difficult to decipher its meaning if they were permuted randomly. Likewise, image frames
in a video, the audio signal in a conversation, or the browsing behavior on a website, all follow
sequential order. It is thus only reasonable to assume that specialized models for such data will
do better at describing it and at solving estimation problems.
Another issue arises from the fact that we might not only receive a sequence as an input but rather
might be expected to continue the sequence. For instance, the task could be to continue the series
2, 4, 6, 8, 10, … This is quite common in time series analysis, to predict the stock market, the fever
curve of a patient or the acceleration needed for a race car. Again we want to have models that
can handle such data.
In short, while convolutional neural networks can efficiently process spatial information, recur-
rent neural networks are designed to better handle sequential information. These networks in-
troduce state variables to store past information, together with the current inputs, to determine
the current outputs.
Many of the examples for using recurrent networks are based on text data. Hence, we will em-
phasize language models in this chapter. After a more formal review of sequence data we discuss
basic concepts of a language model and use this discussion as the inspiration for the design of re-
current neural networks. Next, we describe the gradient calculation method in recurrent neural
networks to explore problems that may be encountered in recurrent neural network training.
8.1 Sequence Models
Imagine that you are watching movies on Netflix. As a good Netflix user, you decide to rate each
of the movies religiously. After all, a good movie is a good movie, and you want to watch more of
them, right? As it turns out, things are not quite so simple. Peopleʼs opinions on movies can change
quite significantly over time. In fact, psychologists even have names for some of the effects:
• There is anchoring103 , based on someone elseʼs opinion. For instance after the Oscar awards,
ratings for the corresponding movie go up, even though it is still the same movie. This effect
103
https://en.wikipedia.org/wiki/Anchoring
299
persists for a few months until the award is forgotten. (Wu et al., 2017) showed that the effect
lifts rating by over half a point.
• There is the Hedonic adaptation104 , where humans quickly adapt to accept an improved (or
a bad) situation as the new normal. For instance, after watching many good movies, the
expectations that the next movie is equally good or better are high, hence even an average
movie might be considered a bad movie after many great ones.
• There is seasonality. Very few viewers like to watch a Santa Claus movie in August.
• In some cases movies become unpopular due to the misbehaviors of directors or actors in
the production.
• Some movies become cult movies, because they were almost comically bad. Plan 9 from
Outer Space and Troll 2 achieved a high degree of notoriety for this reason.
In short, ratings are anything but stationary. Using temporal dynamics helped (Koren, 2009) to
recommend movies more accurately. But it is not just about movies.
• Many users have highly particular behavior when it comes to the time when they open apps.
For instance, social media apps are much more popular after school with students. Stock
market trading apps are more commonly used when the markets are open.
• It is much harder to predict tomorrowʼs stock prices than to fill in the blanks for a stock price
we missed yesterday, even though both are just a matter of estimating one number. After
all, foresight is so much harder than hindsight. In statistics, the former (predicting beyond
the known observations) is called extrapolation whereas the latter (estimating between the
existing observations) is called interpolation.
• Music, speech, text, movies, steps, etc. are all sequential in nature. If we were to permute
them they would make little sense. The headline dog bites man is much less surprising than
man bites dog, even though the words are identical.
• Earthquakes are strongly correlated, i.e., after a massive earthquake there are very likely
several smaller aftershocks, much more so than without the strong quake. In fact, earth-
quakes are spatiotemporally correlated, i.e., the aftershocks typically occur within a short
time span and in close proximity.
• Humans interact with each other in a sequential nature, as can be seen in Twitter fights,
dance patterns and debates.
8.1.1 Statistical Tools
In short, we need statistical tools and new deep neural networks architectures to deal with se-
quence data. To keep things simple, we use the stock price illustrated in Fig. 8.1.1 as an example.
104
https://en.wikipedia.org/wiki/Hedonic_treadmill
300 Chapter 8. Recurrent Neural Networks

Fig. 8.1.1: FTSE 100 index over 30 years
Let us denote the prices by xt ≥ 0, i.e., at time t ∈ N we observe price xt . For a trader to do well
in the stock market on day t he should want to predict xt via
xt ∼ p(xt | xt−1 , . . . , x1 ). (8.1.1)
Autoregressive Models
In order to achieve this, our trader could use a regressor such as the one we trained in Section
3.3. There is just a major problem: the number of inputs, xt−1 , . . . , x1 varies, depending on t.
That is, the number increases with the amount of data that we encounter, and we will need an
approximation to make this computationally tractable. Much of what follows in this chapter will
revolve around how to estimate p(xt | xt−1 , . . . , x1 ) efficiently. In a nutshell it boils down to two
strategies:
1. Assume that the potentially rather long sequence xt−1 , . . . , x1 is not really necessary. In this
case we might content ourselves with some timespan τ and only use xt−1 , . . . , xt−τ observa-
tions. The immediate benefit is that now the number of arguments is always the same, at
least for t > τ . This allows us to train a deep network as indicated above. Such models will
be called autoregressive models, as they quite literally perform regression on themselves.
2. Another strategy, shown in Fig. 8.1.2, is to try and keep some summary ht of the past obser-
vations, at the same time update ht in addition to the prediction x̂t . This leads to models that
estimate xt with x̂t = p(xt | xt−1 , ht ) and moreover updates of the form ht = g(ht−1 , xt−1 ).
Since ht is never observed, these models are also called latent autoregressive models. LSTMs
and GRUs are examples of this.
Fig. 8.1.2: A latent autoregressive model.
8.1. Sequence Models 301

Both cases raise the obvious question of how to generate training data. One typically uses histor-
ical observations to predict the next observation given the ones up to right now. Obviously we do
not expect time to stand still. However, a common assumption is that while the specific values of
xt might change, at least the dynamics of the time series itself will not. This is reasonable, since
novel dynamics are just that, novel and thus not predictable using data that we have so far. Statis-
ticians call dynamics that do not change stationary. Regardless of what we do, we will thus get an
estimate of the entire time series via
∏
T
p(x1 , . . . , xT ) = p(xt | xt−1 , . . . , x1 ). (8.1.2)
t=1
Note that the above considerations still hold if we deal with discrete objects, such as words, rather
than numbers. The only difference is that in such a situation we need to use a classifier rather
than a regressor to estimate p(xt | xt−1 , . . . , x1 ).
Markov Model
Recall the approximation that in an autoregressive model we use only (xt−1 , . . . , xt−τ ) instead of
(xt−1 , . . . , x1 ) to estimate xt . Whenever this approximation is accurate we say that the sequence
satisfies a Markov condition. In particular, if τ = 1, we have a first order Markov model and p(x) is
given by
∏
T
p(x1 , . . . , xT ) = p(xt | xt−1 ). (8.1.3)
t=1
Such models are particularly nice whenever xt assumes only a discrete value, since in this case
dynamic programming can be used to compute values along the chain exactly. For instance, we
can compute p(xt+1 | xt−1 ) efficiently using the fact that we only need to take into account a very
short history of past observations:
∑
p(xt+1 | xt−1 ) = p(xt+1 | xt )p(xt | xt−1 ). (8.1.4)
xt
Going into details of dynamic programming is beyond the scope of this section, but we will intro-
duce it in Section 9.4. Control and reinforcement learning algorithms use such tools extensively.
Causality
In principle, there is nothing wrong with unfolding p(x1 , . . . , xT ) in reverse order. After all, by
conditioning we can always write it via
∏
1
p(x1 , . . . , xT ) = p(xt | xt+1 , . . . , xT ). (8.1.5)
t=T
In fact, if we have a Markov model, we can obtain a reverse conditional probability distribution,
too. In many cases, however, there exists a natural direction for the data, namely going forward
in time. It is clear that future events cannot influence the past. Hence, if we change xt , we may be
able to influence what happens for xt+1 going forward but not the converse. That is, if we change
xt , the distribution over past events will not change. Consequently, it ought to be easier to explain
p(xt+1 | xt ) rather than p(xt | xt+1 ). For instance, (Hoyer et al., 2009) show that in some cases we
can find xt+1 = f (xt ) + ϵ for some additive noise, whereas the converse is not true. This is great
news, since it is typically the forward direction that we are interested in estimating. For more on
this topic see e.g., the book by (Peters et al., 2017a). We are barely scratching the surface of it.

8.1.2 A Toy Example
After so much theory, let us try this out in practice. Let us begin by generating some data. To keep
things simple we generate our time series by using a sine function with some additive noise.
%matplotlib inline
from mxnet import autograd, np, npx, gluon, init
npx.set_np()
T = 1000 # Generate a total of 1000 points

time = np.arange(0, T, dtype=np.float32)
x = np.sin(0.01 * time) + np.random.normal(0, 0.2, (T,))
d2l.plot(time, [x])
Next we need to turn this time series into features and labels that the network can train on. Based
on the embedding dimension τ we map the data into pairs yt = xt and zt = (xt−1 , . . . , xt−τ ). The
astute reader might have noticed that this gives us τ fewer data examples, since we do not have
sufficient history for the first τ of them. A simple fix, in particular if the time series is long is to
discard those few terms. Alternatively we could pad the time series with zeros. The code below
is essentially identical to the training code in previous sections. We kept the architecture fairly
simple.
tau = 4
features = np.zeros((T-tau, tau))
for i in range(tau):
features[:, i] = x[i: T-tau+i]
labels = d2l.reshape(x[tau:], (-1, 1))
batch_size, n_train = 16, 600

train_iter = d2l.load_array((features[:n_train], labels[:n_train]),
batch_size, is_train=True)
A few layers of a fully connected network, ReLU activation and L2 loss. Since much of the modeling
is identical to the previous sections when we built regression estimators, we will not delve into
much detail.

# Vanilla MLP architecture
def get_net():
nn.Dense(1))
net.initialize(init.Xavier())
return net
# Least mean squares loss

Now we are ready to train.
def train_net(net, train_iter, loss, epochs, lr):

trainer = gluon.Trainer(net.collect_params(), 'adam',
{'learning_rate': lr})
for epoch in range(1, epochs + 1):
l = loss(net(X), y)
l.backward()
print(f'epoch {epoch}, '
f'loss: {d2l.evaluate_loss(net, train_iter, loss):f}')
net = get_net()
train_net(net, train_iter, loss, 10, 0.01)
epoch 1, loss: 0.039526

epoch 2, loss: 0.032300
epoch 3, loss: 0.028369
epoch 4, loss: 0.027100
epoch 5, loss: 0.026571
epoch 6, loss: 0.026366
epoch 7, loss: 0.025704
epoch 8, loss: 0.025523
epoch 9, loss: 0.026450
epoch 10, loss: 0.026086
8.1.3 Predictions
Since training loss is small, we would expect our model to work well. Let us see what this means
in practice. The first thing to check is how well the model is able to predict what happens in the
next timestep.
estimates = net(features)
d2l.plot([time, time[tau:]], [d2l.numpy(x), d2l.numpy(estimates)],
legend=['data', 'estimate'])

This looks nice, just as we expected it. Even beyond 600 observations the estimates still look rather
trustworthy. There is just one little problem to this: if we observe data only until timestep 600, we
cannot hope to receive the ground truth for all future predictions. Instead, we need to work our
way forward one step at a time:
x601 = f (x600 , . . . , x597 ),

x602 = f (x601 , . . . , x598 ), (8.1.6)
x603 = f (x602 , . . . , x599 ).
In other words, we will have to use our own predictions to make future predictions. Let us see
how well this goes.
predictions = np.zeros(T)
predictions[:n_train] = x[:n_train]
for i in range(n_train, T):
predictions[i] = net(
predictions[(i-tau):i].reshape(1, -1)).reshape(1)
d2l.plot([time, time[tau:], time[n_train:]],
[d2l.numpy(x), d2l.numpy(estimates),
d2l.numpy(predictions[n_train:])],
legend=['data', 'estimate', 'multistep'], figsize=(4.5, 2.5))
As the above example shows, this is a spectacular failure. The estimates decay to a constant pretty
quickly after a few prediction steps. Why did the algorithm work so poorly? This is ultimately due

to the fact that the errors build up. Let us say that after step 1 we have some error ϵ1 = ϵ̄. Now
the input for step 2 is perturbed by ϵ1 , hence we suffer some error in the order of ϵ2 = ϵ̄ + Lϵ1 ,
and so on. The error can diverge rather rapidly from the true observations. This is a common
phenomenon. For instance, weather forecasts for the next 24 hours tend to be pretty accurate but
beyond that the accuracy declines rapidly. We will discuss methods for improving this throughout
this chapter and beyond.
Let us verify this observation by computing the k-step predictions on the entire sequence.
k = 33 # Look up to k - tau steps ahead
features = np.zeros((k, T-k))

for i in range(tau): # Copy the first tau features from x
features[i] = x[i:T-k+i]
for i in range(tau, k): # Predict the (i-tau)-th step

features[i] = net(features[(i-tau):i].T).T
steps = (4, 8, 16, 32)

d2l.plot([time[i:T-k+i] for i in steps],
[d2l.numpy(features[i]) for i in steps],
legend=[f'step {i}' for i in steps], figsize=(4.5, 2.5))
This clearly illustrates how the quality of the estimates changes as we try to predict further into the
future. While the 8-step predictions are still pretty good, anything beyond that is pretty useless.
Summary
• Sequence models require specialized statistical tools for estimation. Two popular choices
are autoregressive models and latent-variable autoregressive models.
• As we predict further in time, the errors accumulate and the quality of the estimates de-
grades, often dramatically.
• There is quite a difference in difficulty between interpolation and extrapolation. Conse-
quently, if you have a time series, always respect the temporal order of the data when train-
ing, i.e., never train on future data.
• For causal models (e.g., time going forward), estimating the forward direction is typically a
lot easier than the reverse direction.

Exercises
1. Improve the above model.

• Incorporate more than the past 4 observations? How many do you really need?
• How many would you need if there was no noise? Hint: you can write sin and cos as a
differential equation.
• Can you incorporate older features while keeping the total number of features con-
stant? Does this improve accuracy? Why?
• Change the neural network architecture and see what happens.
2. An investor wants to find a good security to buy. He looks at past returns to decide which
one is likely to do well. What could possibly go wrong with this strategy?
3. Does causality also apply to text? To which extent?
4. Give an example for when a latent autoregressive model might be needed to capture the
dynamic of the data.
Discussions105
8.2 Text Preprocessing
Text is an important example of sequence data. An article can be simply viewed as a sequence of
words, or a sequence of characters. Given text data is a major data format besides images we are
using in this book, this section will dedicate to explain the common preprocessing steps for text
data. Such preprocessing often consists of four steps:
1. Load text as strings into memory.
2. Split strings into tokens, where a token could be a word or a character.
3. Build a vocabulary for these tokens to map them into numerical indices.
4. Map all the tokens in data into indices for ease of feeding into models.
import collections
import re
To get started we load text from H. G. Wellsʼ Time Machine106 . This is a fairly small corpus of just
over 30, 000 words, but for the purpose of what we want to illustrate this is just fine. More realistic
document collections contain many billions of words. The following function reads the dataset
into a list of sentences, each sentence is a string. Here we ignore punctuation and capitalization.
105
106
http://www.gutenberg.org/ebooks/35
8.2. Text Preprocessing 307

#@save
d2l.DATA_HUB['time_machine'] = (d2l.DATA_URL + 'timemachine.txt',
'090b5e7e70c295757f55df93cb0a180b9691891a')
def read_time_machine(): #@save

"""Load the time machine book into a list of sentences."""
with open(d2l.download('time_machine'), 'r') as f:
lines = f.readlines()
return [re.sub('[Â-Za-z]+', ' ', line.strip().lower())
for line in lines]
lines = read_time_machine()
f'# sentences {len(lines)}'
'# sentences 3221'
8.2.2 Tokenization
For each sentence, we split it into a list of tokens. A token is a data example the model will train
and predict. The following function supports splitting a sentence into words or characters, and
returns a list of split strings.
def tokenize(lines, token='word'): #@save

"""Split sentences into word or char tokens."""
if token == 'word':
return [line.split(' ') for line in lines]
elif token == 'char':
return [list(line) for line in lines]
else:
print('ERROR: unknown token type '+token)
tokens = tokenize(lines)
tokens[0:2]
[['the', 'time', 'machine', 'by', 'h', 'g', 'wells', ''], ['']]
8.2.3 Vocabulary
The string type of the token is inconvenient to be used by models, which take numerical inputs.
Now let us build a dictionary, often called vocabulary as well, to map string tokens into numerical
indices starting from 0. To do so, we first count the unique tokens in all documents, called cor-
pus, and then assign a numerical index to each unique token according to its frequency. Rarely
appeared tokens are often removed to reduce the complexity. A token does not exist in corpus
or has been removed is mapped into a special unknown (“<unk>”) token. We optionally add a list
of reserved tokens, such as “<pad>” a token for padding, “<bos>” to present the beginning for a
sentence, and “<eos>” for the ending of a sentence.
class Vocab: #@save

def __init__(self, tokens, min_freq=0, reserved_tokens=None):

if reserved_tokens is None:
reserved_tokens = []
# Sort according to frequencies
counter = count_corpus(tokens)
self.token_freqs = sorted(counter.items(), key=lambda x: x[0])
self.token_freqs.sort(key=lambda x: x[1], reverse=True)
self.unk, uniq_tokens = 0, ['<unk>'] + reserved_tokens
uniq_tokens += [token for token, freq in self.token_freqs
if freq >= min_freq and token not in uniq_tokens]
self.idx_to_token, self.token_to_idx = [], dict()
for token in uniq_tokens:
self.idx_to_token.append(token)
self.token_to_idx[token] = len(self.idx_to_token) - 1
def __len__(self):
return len(self.idx_to_token)
def __getitem__(self, tokens):

if not isinstance(tokens, (list, tuple)):
return self.token_to_idx.get(tokens, self.unk)
return [self.__getitem__(token) for token in tokens]
def to_tokens(self, indices):

if not isinstance(indices, (list, tuple)):
return self.idx_to_token[indices]
return [self.idx_to_token[index] for index in indices]
def count_corpus(sentences): #@save

# Flatten a list of token lists into a list of tokens
tokens = [tk for line in sentences for tk in line]
return collections.Counter(tokens)
We construct a vocabulary with the time machine dataset as the corpus, and then print the map
between a few tokens and their indices.
vocab = Vocab(tokens)
print(list(vocab.token_to_idx.items())[0:10])
[('<unk>', 0), ('the', 1), ('', 2), ('i', 3), ('and', 4), ('of', 5), ('a', 6), ('to', 7), (
,→'was', 8), ('in', 9)]
After that, we can convert each sentence into a list of numerical indices. To illustrate in detail, we
print two sentences with their corresponding indices.
for i in range(8, 10):

print('words:', tokens[i])
print('indices:', vocab[tokens[i]])
words: ['the', 'time', 'traveller', 'for', 'so', 'it', 'will', 'be', 'convenient', 'to',
,→'speak', 'of', 'him', '']
indices: [1, 20, 72, 17, 38, 12, 120, 43, 706, 7, 660, 5, 112, 2]
words: ['was', 'expounding', 'a', 'recondite', 'matter', 'to', 'us', 'his', 'grey', 'eyes',
,→'shone', 'and']
8.2. Text Preprocessing 309

indices: [8, 1654, 6, 3864, 634, 7, 131, 26, 344, 127, 484, 4]
Using the above functions, we package everything into the load_corpus_time_machine function,
which returns corpus, a list of token indices, and vocab, the vocabulary of the time machine cor-
pus. The modification we did here is that corpus is a single list, not a list of token lists, since we
do not keep the sequence information in the following models. Besides, we use character tokens
to simplify the training in later sections.
def load_corpus_time_machine(max_tokens=-1): #@save

lines = read_time_machine()
tokens = tokenize(lines, 'char')
vocab = Vocab(tokens)
corpus = [vocab[tk] for line in tokens for tk in line]
if max_tokens > 0:
corpus = corpus[:max_tokens]
return corpus, vocab
corpus, vocab = load_corpus_time_machine()

len(corpus), len(vocab)
(171489, 28)
Summary
• We preprocessed the documents by tokenizing them into words or characters and then map-
ping into indices.
Exercises
1. Tokenization is a key preprocessing step. It varies for different languages. Try to find another
3 commonly used methods to tokenize sentences.
Discussions107
8.3 Language Models and the Dataset
In Section 8.2, we see how to map text data into tokens, and these tokens can be viewed as a time
series of discrete observations. Assuming the tokens in a text of length T are in turn x1 , x2 , . . . , xT ,
then, in the discrete time series, xt (1 ≤ t ≤ T ) can be considered as the output or label of timestep
t. Given such a sequence, the goal of a language model is to estimate the probability
p(x1 , x2 , . . . , xT ). (8.3.1)
107

Language models are incredibly useful. For instance, an ideal language model would be able
to generate natural text just on its own, simply by drawing one word at a time wt ∼ p(wt |
wt−1 , . . . , w1 ). Quite unlike the monkey using a typewriter, all text emerging from such a model
would pass as natural language, e.g., English text. Furthermore, it would be sufficient for gener-
ating a meaningful dialog, simply by conditioning the text on previous dialog fragments. Clearly
we are still very far from designing such a system, since it would need to understand the text rather
than just generate grammatically sensible content.
Nonetheless language models are of great service even in their limited form. For instance, the
phrases “to recognize speech” and “to wreck a nice beach” sound very similar. This can cause am-
biguity in speech recognition, ambiguity that is easily resolved through a language model which
rejects the second translation as outlandish. Likewise, in a document summarization algorithm
it is worth while knowing that “dog bites man” is much more frequent than “man bites dog”, or
that “I want to eat grandma” is a rather disturbing statement, whereas “I want to eat, grandma” is
much more benign.
8.3.1 Estimating a Language Model
The obvious question is how we should model a document, or even a sequence of words. We can
take recourse to the analysis we applied to sequence models in the previous section. Let us start
by applying basic probability rules:
∏
T
p(w1 , w2 , . . . , wT ) = p(w1 ) p(wt | w1 , . . . , wt−1 ). (8.3.2)
t=2
For example, the probability of a text sequence containing four tokens consisting of words and
punctuation would be given as:
p(Statistics, is, fun, .) = p(Statistics)p(is | Statistics)p(fun | Statistics, is)p(. | Statistics, is, fun).
(8.3.3)
In order to compute the language model, we need to calculate the probability of words and the
conditional probability of a word given the previous few words, i.e., language model parameters.
Here, we assume that the training dataset is a large text corpus, such as all Wikipedia entries,
Project Gutenberg108 , or all text posted online on the web. The probability of words can be calcu-
lated from the relative word frequency of a given word in the training dataset.
For example, p(Statistics) can be calculated as the probability of any sentence starting with the
word “statistics”. A slightly less accurate approach would be to count all occurrences of the word
“statistics” and divide it by the total number of words in the corpus. This works fairly well, partic-
ularly for frequent words. Moving on, we could attempt to estimate
n(Statistics is)
p̂(is | Statistics) = . (8.3.4)
n(Statistics)
Here n(w) and n(w, w′ ) are the number of occurrences of singletons and pairs of words respec-
tively. Unfortunately, estimating the probability of a word pair is somewhat more difficult, since
the occurrences of “Statistics is” are a lot less frequent. In particular, for some unusual word
combinations it may be tricky to find enough occurrences to get accurate estimates. Things take
a turn for the worse for 3-word combinations and beyond. There will be many plausible 3-word
combinations that we likely will not see in our dataset. Unless we provide some solution to give
108
https://en.wikipedia.org/wiki/Project_Gutenberg
8.3. Language Models and the Dataset 311

such word combinations nonzero weight, we will not be able to use these as a language model. If
the dataset is small or if the words are very rare, we might not find even a single one of them.
A common strategy is to perform some form of Laplace smoothing. We already encountered this
in our discussion of naive Bayes in Section 18.9 where the solution was to add a small constant to
all counts. This helps with singletons, e.g., via
n(w) + ϵ1 /m
p̂(w) = ,
n + ϵ1
n(w, w′ ) + ϵ2 p̂(w′ )
p̂(w′ | w) = , (8.3.5)
n(w) + ϵ2
n(w, w′ , w′′ ) + ϵ3 p̂(w′ , w′′ )
p̂(w′′ | w′ , w) = .
n(w, w′ ) + ϵ3
Here the coefficients ϵi > 0 determine how much we use the estimate for a shorter sequence as
a fill-in for longer ones. Moreover, m is the total number of words we encounter. The above is
a rather primitive variant of what Kneser-Ney smoothing and Bayesian nonparametrics can ac-
complish. See e.g., (Wood et al., 2011) for more detail of how to accomplish this. Unfortunately,
models like this get unwieldy rather quickly for the following reasons. First, we need to store all
counts. Second, this entirely ignores the meaning of the words. For instance, “cat” and “feline”
should occur in related contexts. It is quite difficult to adjust such models to additional contexts,
whereas, deep learning based language models are well suited to take this into account. Last, long
word sequences are almost certain to be novel, hence a model that simply counts the frequency
of previously seen word sequences is bound to perform poorly there.
8.3.2 Markov Models and n-grams
Before we discuss solutions involving deep learning, we need some more terminology and con-
cepts. Recall our discussion of Markov Models in the previous section. Let us apply this to lan-
guage modeling. A distribution over sequences satisfies the Markov property of first order if
p(wt+1 | wt , . . . , w1 ) = p(wt+1 | wt ). Higher orders correspond to longer dependencies. This
leads to a number of approximations that we could apply to model a sequence:
p(w1 , w2 , w3 , w4 ) = p(w1 )p(w2 )p(w3 )p(w4 ),

p(w1 , w2 , w3 , w4 ) = p(w1 )p(w2 | w1 )p(w3 | w2 )p(w4 | w3 ), (8.3.6)
p(w1 , w2 , w3 , w4 ) = p(w1 )p(w2 | w1 )p(w3 | w1 , w2 )p(w4 | w2 , w3 ).
The probability formulae that involve one, two, and three variables are typically referred to as
unigram, bigram, and trigram models respectively. In the following, we will learn how to design
better models.
8.3.3 Natural Language Statistics
Let us see how this works on real data. We construct a vocabulary based on the time machine data
similar to Section 8.2 and print the top 10 most frequent words.

import random
npx.set_np()

tokens = d2l.tokenize(d2l.read_time_machine())
vocab = d2l.Vocab(tokens)
vocab.token_freqs[:10]
[('the', 2261),
('', 1282),
('i', 1267),
('and', 1245),
('of', 1155),
('a', 816),
('to', 695),
('was', 552),
('in', 541),
('that', 443)]
As we can see, the most popular words are actually quite boring to look at. They are often re-
ferred to as stop words109 and thus filtered out. That said, they still carry meaning and we will use
them nonetheless. However, one thing that is quite clear is that the word frequency decays rather
rapidly. The 10th most frequent word is less than 1/5 as common as the most popular one. To get
a better idea we plot the graph of the word frequency.
freqs = [freq for token, freq in vocab.token_freqs]

d2l.plot(freqs, xlabel='token: x', ylabel='frequency: n(x)',
xscale='log', yscale='log')
We are on to something quite fundamental here: the word frequency decays rapidly in a well
defined way. After dealing with the first four words as exceptions (ʻtheʼ, ʻiʼ, ʻandʼ, ʻofʼ), all remaining
words follow a straight line on a log-log plot. This means that words satisfy Zipfʼs law110 which
states that the item frequency is given by
n(x) ∝ (x + c)−α and hence log n(x) = −α log(x + c) + const. (8.3.7)
This should already give us pause if we want to model words by count statistics and smoothing.
After all, we will significantly overestimate the frequency of the tail, also known as the infrequent
109
https://en.wikipedia.org/wiki/Stop_words
110
https://en.wikipedia.org/wiki/Zipf%27s_law

words. But what about the other word combinations (such as bigrams, trigrams, and beyond)? Let
us see whether the bigram frequency behaves in the same manner as the unigram frequency.
bigram_tokens = [[pair for pair in zip(

line[:-1], line[1:])] for line in tokens]
bigram_vocab = d2l.Vocab(bigram_tokens)
print(bigram_vocab.token_freqs[:10])
[(('of', 'the'), 297), (('in', 'the'), 161), (('i', 'had'), 126), (('and', 'the'), 104), (('i
,→', 'was'), 104), (('the', 'time'), 97), (('it', 'was'), 94), (('to', 'the'), 81), (('as',
,→'i'), 75), (('of', 'a'), 69)]
One thing is notable here. Out of the 10 most frequent word pairs, 9 are composed of stop words
and only one is relevant to the actual book—“the time”. Furthermore, let us see whether the tri-
gram frequency behaves in the same manner.
trigram_tokens = [[triple for triple in zip(line[:-2], line[1:-1], line[2:])]

for line in tokens]
trigram_vocab = d2l.Vocab(trigram_tokens)
print(trigram_vocab.token_freqs[:10])
[(('the', 'time', 'traveller'), 53), (('the', 'time', 'machine'), 24), (('the', 'medical',
,→'man'), 22), (('it', 'seemed', 'to'), 14), (('it', 'was', 'a'), 14), (('i', 'began', 'to'),
,→ 13), (('i', 'did', 'not'), 13), (('i', 'saw', 'the'), 13), (('here', 'and', 'there'), 12),
,→ (('i', 'could', 'see'), 12)]
Last, let us visualize the token frequency among these three gram models: unigrams, bigrams,
and trigrams.
bigram_freqs = [freq for token, freq in bigram_vocab.token_freqs]

trigram_freqs = [freq for token, freq in trigram_vocab.token_freqs]
d2l.plot([freqs, bigram_freqs, trigram_freqs], xlabel='token',
ylabel='frequency', xscale='log', yscale='log',
legend=['unigram', 'bigram', 'trigram'])
The graph is quite exciting for a number of reasons. First, beyond unigram words, also sequences
of words appear to be following Zipfʼs law, albeit with a lower exponent, depending on sequence

length. Second, the number of distinct n-grams is not that large. This gives us hope that there is
quite a lot of structure in language. Third, many n-grams occur very rarely, which makes Laplace
smoothing rather unsuitable for language modeling. Instead, we will use deep learning based
models.
8.3.4 Training Data Preparation
Before introducing the model, let us assume we will use a neural network to train a language
model. Now the question is how to read minibatches of examples and labels at random. Since
sequence data is by its very nature sequential, we need to address the issue of processing it. We
did so in a rather ad-hoc manner when we introduced in Section 8.1. Let us formalize this a bit.
In Fig. 8.3.1, we visualized several possible ways to obtain 5-grams in a sentence, here a token is a
character. Note that we have quite some freedom since we could pick an arbitrary offset.
Fig. 8.3.1: Different offsets lead to different subsequences when splitting up text.
Hence, which one should we pick? In fact, all of them are equally good. But if we pick all off-
sets we end up with rather redundant data due to overlap, particularly if the sequences are long.
Picking just a random set of initial positions is no good either since it does not guarantee uniform
coverage of the array. For instance, if we pick n elements at random out of a set of n with random
replacement, the probability for a particular element not being picked is (1 − 1/n)n → e−1 . This
means that we cannot expect uniform coverage this way. Even randomly permuting a set of all
offsets does not offer good guarantees. Instead we can use a simple trick to get both coverage and
randomness: use a random offset, after which one uses the terms sequentially. We describe how
to accomplish this for both random sampling and sequential partitioning strategies below.
Random Sampling
The following code randomly generates a minibatch from the data each time. Here, the batch size
batch_size indicates the number of examples in each minibatch and num_steps is the length of the
sequence (or timesteps if we have a time series) included in each example. In random sampling,
each example is a sequence arbitrarily captured on the original sequence. The positions of two
adjacent random minibatches on the original sequence are not necessarily adjacent. The target is
to predict the next character based on what we have seen so far, hence the labels are the original
sequence, shifted by one character.

def seq_data_iter_random(corpus, batch_size, num_steps): #@save
# Offset the iterator over the data for uniform starts
corpus = corpus[random.randint(0, num_steps):]
# Subtract 1 extra since we need to account for label
num_examples = ((len(corpus) - 1) // num_steps)
example_indices = list(range(0, num_examples * num_steps, num_steps))
random.shuffle(example_indices)
def data(pos):
# This returns a sequence of length `num_steps` starting from `pos`
return corpus[pos: pos + num_steps]
# Discard half empty batches

num_batches = num_examples // batch_size
for i in range(0, batch_size * num_batches, batch_size):
# `batch_size` indicates the random examples read each time
batch_indices = example_indices[i:(i+batch_size)]
X = [data(j) for j in batch_indices]
Y = [data(j + 1) for j in batch_indices]
yield np.array(X), np.array(Y)
Let us generate an artificial sequence from 0 to 30. We assume that the batch size and numbers
of timesteps are 2 and 6 respectively. This means that depending on the offset we can generate
between 4 and 5 (x, y) pairs. With a minibatch size of 2, we only get 2 minibatches.
my_seq = list(range(30))
for X, Y in seq_data_iter_random(my_seq, batch_size=2, num_steps=6):
print('X: ', X, '\nY:', Y)
X: [[ 7. 8. 9. 10. 11. 12.]

[13. 14. 15. 16. 17. 18.]]
Y: [[ 8. 9. 10. 11. 12. 13.]
[14. 15. 16. 17. 18. 19.]]
X: [[ 1. 2. 3. 4. 5. 6.]
[19. 20. 21. 22. 23. 24.]]
Y: [[ 2. 3. 4. 5. 6. 7.]
[20. 21. 22. 23. 24. 25.]]
Sequential Partitioning
In addition to random sampling of the original sequence, we can also make the positions of two
adjacent random minibatches adjacent in the original sequence.
def seq_data_iter_consecutive(corpus, batch_size, num_steps): #@save

# Offset for the iterator over the data for uniform starts
offset = random.randint(0, num_steps)
# Slice out data: ignore `num_steps` and just wrap around
num_indices = ((len(corpus) - offset - 1) // batch_size) * batch_size
Xs = np.array(corpus[offset:offset+num_indices])
Ys = np.array(corpus[offset+1:offset+1+num_indices])
Xs, Ys = Xs.reshape(batch_size, -1), Ys.reshape(batch_size, -1)
num_batches = Xs.shape[1] // num_steps

for i in range(0, num_batches * num_steps, num_steps):
X = Xs[:, i:(i+num_steps)]
Y = Ys[:, i:(i+num_steps)]
yield X, Y
Using the same settings, print input X and label Y for each minibatch of examples read by sequen-
tial partitioning. The positions of two adjacent minibatches on the original sequence are adjacent.
for X, Y in seq_data_iter_consecutive(my_seq, batch_size=2, num_steps=6):

print('X: ', X, '\nY:', Y)
X: [[ 3. 4. 5. 6. 7. 8.]
[16. 17. 18. 19. 20. 21.]]
Y: [[ 4. 5. 6. 7. 8. 9.]
[17. 18. 19. 20. 21. 22.]]
X: [[ 9. 10. 11. 12. 13. 14.]
[22. 23. 24. 25. 26. 27.]]
Y: [[10. 11. 12. 13. 14. 15.]
[23. 24. 25. 26. 27. 28.]]
Now we wrap the above two sampling functions to a class so that we can use it as a data iterator
later.
class SeqDataLoader: #@save

"""An iterator to load sequence data."""
def __init__(self, batch_size, num_steps, use_random_iter, max_tokens):
if use_random_iter:
self.data_iter_fn = d2l.seq_data_iter_random
else:
self.data_iter_fn = d2l.seq_data_iter_consecutive
self.corpus, self.vocab = d2l.load_corpus_time_machine(max_tokens)
self.batch_size, self.num_steps = batch_size, num_steps
def __iter__(self):
return self.data_iter_fn(self.corpus, self.batch_size, self.num_steps)
Last, we define a function load_data_time_machine that returns both the data iterator and the
vocabulary, so we can use it similarly as other functions with load_data prefix.
def load_data_time_machine(batch_size, num_steps, #@save

use_random_iter=False, max_tokens=10000):
data_iter = SeqDataLoader(
batch_size, num_steps, use_random_iter, max_tokens)
return data_iter, data_iter.vocab

Summary
• Language models are an important technology for natural language processing.

• n-grams provide a convenient model for dealing with long sequences by truncating the de-
pendence.
• Long sequences suffer from the problem that they occur very rarely or never.
• Zipfʼs law governs the word distribution for not only unigrams but also the other n-grams.
• There is a lot of structure but not enough frequency to deal with infrequent word combina-
tions efficiently via Laplace smoothing.
• The main choices for sequence partitioning are picking between consecutive and random
sequences.
• Given the overall document length, it is usually acceptable to be slightly wasteful with the
documents and discard half-empty minibatches.
Exercises
1. Suppose there are 100, 000 words in the training dataset. How much word frequency and
multi-word adjacent frequency does a four-gram need to store?
2. Review the smoothed probability estimates. Why are they not accurate? Hint: we are dealing
with a contiguous sequence rather than singletons.
3. How would you model a dialogue?
4. Estimate the exponent of Zipfʼs law for unigrams, bigrams, and trigrams.
5. What other minibatch data sampling methods can you think of?
6. Why is it a good idea to have a random offset?
• Does it really lead to a perfectly uniform distribution over the sequences on the docu-
ment?
• What would you have to do to make things even more uniform?
7. If we want a sequence example to be a complete sentence, what kinds of problems does this
introduce in minibatch sampling? Why would we want to do this anyway?
Discussions111
8.4 Recurrent Neural Networks
In Section 8.3 we introduced n-gram models, where the conditional probability of word xt at po-
sition t only depends on the n − 1 previous words. If we want to check the possible effect of words
earlier than t − (n − 1) on xt , we need to increase n. However, the number of model parameters
would also increase exponentially with it, as we need to store |V |n numbers for a vocabulary V .
111

Hence, rather than modeling p(xt | xt−1 , . . . , xt−n+1 ) it is preferable to use a latent variable model
in which we have
p(xt | xt−1 , . . . , x1 ) ≈ p(xt | xt−1 , ht ). (8.4.1)
Here ht is a latent variable that stores the sequence information. A latent variable is also called as
hidden variable, hidden state or hidden state variable. The hidden state at time t could be computed
based on both input xt and hidden state ht−1 , that is
ht = f (xt , ht−1 ). (8.4.2)
For a sufficiently powerful function f , the latent variable model is not an approximation. After all,
ht could simply store all the data it observed so far. We discussed this in Section 8.1. But it could
potentially makes both computation and storage expensive.
Note that we also use h to denote the number of hidden units of a hidden layer. Hidden layers
and hidden states refer to two very different concepts. Hidden layers are, as explained, layers that
are hidden from view on the path from input to output. Hidden states are technically speaking
inputs to whatever we do at a given step. Instead, they can only be computed by looking at data
at previous iterations. In this sense they have much in common with latent variable models in
statistics, such as clustering or topic models where the clusters affect the output but cannot be
directly observed.
Recurrent neural networks are neural networks with hidden states. Before introducing this model,
let us first revisit the multi-layer perceptron introduced in Section 4.1.
8.4.1 Recurrent Networks Without Hidden States
Let us take a look at a multilayer perceptron with a single hidden layer. Given a minibatch of the
instances X ∈ Rn×d with sample size n and d inputs. Let the hidden layerʼs activation function be
ϕ. Hence, the hidden layerʼs output H ∈ Rn×h is calculated as
H = ϕ(XWxh + bh ). (8.4.3)
Here, we have the weight parameter Wxh ∈ Rd×h , bias parameter bh ∈ R1×h , and the number of
hidden units h, for the hidden layer.
The hidden variable H is used as the input of the output layer. The output layer is given by
O = HWhq + bq . (8.4.4)
Here, O ∈ Rn×q is the output variable, Whq ∈ Rh×q is the weight parameter, and bq ∈ R1×q is
the bias parameter of the output layer. If it is a classification problem, we can use softmax(O) to
compute the probability distribution of the output category.
This is entirely analogous to the regression problem we solved previously in Section 8.1, hence we
omit details. Suffice it to say that we can pick (xt , xt−1 ) pairs at random and estimate the param-
eters W and b of our network via autograd and stochastic gradient descent.
8.4. Recurrent Neural Networks 319

8.4.2 Recurrent Networks with Hidden States
Matters are entirely different when we have hidden states. Let us look at the structure in some
more detail. Remember that we often call iteration t as time t in an optimization algorithm, time
in a recurrent neural network refers to steps within an iteration. Assume that we have Xt ∈ Rn×d ,
t = 1, . . . , T , in an iteration. And Ht ∈ Rn×h is the hidden variable of timestep t from the se-
quence. Unlike the multilayer perceptron, here we save the hidden variable Ht−1 from the pre-
vious timestep and introduce a new weight parameter Whh ∈ Rh×h , to describe how to use the
hidden variable of the previous timestep in the current timestep. Specifically, the calculation of
the hidden variable of the current timestep is determined by the input of the current timestep
together with the hidden variable of the previous timestep:
Ht = ϕ(Xt Wxh + Ht−1 Whh + bh ). (8.4.5)
Compared with (8.4.3), we added one more Ht−1 Whh here. From the relationship between hidden
variables Ht and Ht−1 of adjacent timesteps, we know that those variables captured and retained
the sequenceʼs historical information up to the current timestep, just like the state or memory of
the neural networkʼs current timestep. Therefore, such a hidden variable is called a hidden state.
Since the hidden state uses the same definition of the previous timestep in the current timestep,
the computation of the equation above is recurrent, hence the name recurrent neural network
(RNN).
There are many different RNN construction methods. RNNs with a hidden state defined by the
equation above are very common. For timestep t, the output of the output layer is similar to the
computation in the multilayer perceptron:
Ot = Ht Whq + bq . (8.4.6)
RNN parameters include the weight Wxh ∈ Rd×h , Whh ∈ Rh×h of the hidden layer with the bias
bh ∈ R1×h , and the weight Whq ∈ Rh×q of the output layer with the bias bq ∈ R1×q . It is worth men-
tioning that RNNs always use these model parameters, even for different timesteps. Therefore,
the number of RNN model parameters does not grow as the number of timesteps increases.
Fig. 8.4.1 shows the computational logic of an RNN at three adjacent timesteps. In timestep t,
the computation of the hidden state can be treated as an entry of a fully connected layer with the
activation function ϕ after concatenating the input Xt with the hidden state Ht−1 of the previous
timestep. The output of the fully connected layer is the hidden state of the current timestep Ht .
Its model parameter is the concatenation of Wxh and Whh , with a bias of bh . The hidden state of
the current timestep t, Ht , will participate in computing the hidden state Ht+1 of the next timestep
t+1. What is more, Ht will become the input for Ot , the fully connected output layer of the current
timestep.

Fig. 8.4.1: An RNN with a hidden state.
8.4.3 Steps in a Language Model
Now we illustrate how RNNs can be used to build a language model. For simplicity of illustration
we use words rather than characters as the inputs, since the former are easier to comprehend. Let
the minibatch size be 1, and the sequence of the text be the beginning of our dataset, i.e., “the time
machine by H. G. Wells”. Fig. 8.4.2 illustrates how to estimate the next word based on the present
and previous words. During the training process, we run a softmax operation on the output from
the output layer for each timestep, and then use the cross-entropy loss function to compute the
error between the result and the label. Due to the recurrent computation of the hidden state in
the hidden layer, the output of timestep 3, O3 , is determined by the text sequence “the”, “time”,
and “machine” respectively. Since the next word of the sequence in the training data is “by”, the
loss of timestep 3 will depend on the probability distribution of the next word generated based on
the feature sequence “the”, “time”, “machine” and the label “by” of this timestep.
Fig. 8.4.2: Word-level RNN language model. The input and label sequences are the time machine
by H. and time machine by H. G. respectively.
In practice, each word is presented by a d dimensional vector, and we use a batch size n > 1.
Therefore, the input Xt at timestep t will be a n × d matrix, which is identical to what we discussed
before.
8.4. Recurrent Neural Networks 321

8.4.4 Perplexity
Last, let us discuss about how to measure the sequence model quality. One way is to check how
surprising the text is. A good language model is able to predict with high accuracy tokens that what
we will see next. Consider the following continuations of the phrase “It is raining”, as proposed
by different language models:
1. “It is raining outside”
2. “It is raining banana tree”
3. “It is raining piouw;kcj pwepoiut”
In terms of quality, example 1 is clearly the best. The words are sensible and logically coherent.
While it might not quite accurately reflect which word follows semantically (“in San Francisco” and
“in winter” would have been perfectly reasonable extensions), the model is able to capture which
kind of word follows. Example 2 is considerably worse by producing a nonsensical extension.
Nonetheless, at least the model has learned how to spell words and some degree of correlation
between words. Last, example 3 indicates a poorly trained model that does not fit data properly.
We might measure the quality of the model by computing p(w), i.e., the likelihood of the sequence.
Unfortunately this is a number that is hard to understand and difficult to compare. After all,
shorter sequences are much more likely to occur than the longer ones, hence evaluating the model
on Tolstoyʼs magnum opus “War and Peace”112 will inevitably produce a much smaller likelihood
than, say, on Saint-Exuperyʼs novella “The Little Prince”113 . What is missing is the equivalent of
an average.
Information theory comes handy here and we will introduce more in Section 18.11. If we want to
compress text, we can ask about estimating the next symbol given the current set of symbols. A
lower bound on the number of bits is given by − log2 p(xt | xt−1 , . . . , x1 ). A good language model
should allow us to predict the next word quite accurately. Thus, it should allow us to spend very
few bits on compressing the sequence. So we can measure it by the average number of bits that
we need to spend.
1∑
n
− log p(xt | xt−1 , . . . , x1 ). (8.4.7)
n
t=1
This makes the performance on documents of different lengths comparable. For historical rea-
sons, scientists in natural language processing prefer to use a quantity called perplexity rather than
bitrate. In a nutshell, it is the exponential of the above:
( )
1∑
n
PPL := exp − log p(xt | xt−1 , . . . , x1 ) . (8.4.8)
n
t=1
It can be best understood as the harmonic mean of the number of real choices that we have when
deciding which word to pick next. Note that perplexity naturally generalizes the notion of the
cross-entropy loss defined when we introduced the softmax regression (Section 3.4). That is, for
a single symbol both definitions are identical bar the fact that one is the exponential of the other.
Let us look at a number of cases:
• In the best case scenario, the model always estimates the probability of the next symbol as
1. In this case the perplexity of the model is 1.
112
https://www.gutenberg.org/files/2600/2600-h/2600-h.htm
113
https://en.wikipedia.org/wiki/The_Little_Prince

• In the worst case scenario, the model always predicts the probability of the label category
as 0. In this situation, the perplexity is infinite.
• At the baseline, the model predicts a uniform distribution over all tokens. In this case, the
perplexity equals the size of the dictionary len(vocab). In fact, if we were to store the se-
quence without any compression, this would be the best we could do to encode it. Hence,
this provides a nontrivial upper bound that any model must satisfy.
Summary
• A network that uses recurrent computation is called a recurrent neural network (RNN).
• The hidden state of the RNN can capture historical information of the sequence up to the
current timestep.
• The number of RNN model parameters does not grow as the number of timesteps increases.
• We can create language models using a character-level RNN.
Exercises
1. If we use an RNN to predict the next character in a text sequence, how many output dimen-
sions do we need?
2. Can you design a mapping for which an RNN with hidden states is exact? Hint: what about
a finite number of words?
3. What happens to the gradient if you backpropagate through a long sequence?
4. What are some of the problems associated with the simple sequence model described above?
Discussions114
8.5 Implementation of Recurrent Neural Networks from Scratch
In this section we implement a language model introduced in Chapter 8 from scratch. It is based
on a character-level recurrent neural network trained on H. G. Wellsʼ The Time Machine. As before,
we start by reading the dataset first, which is introduced in Section 8.3.
%matplotlib inline
import math
from mxnet import autograd, np, npx, gluon
npx.set_np()
batch_size, num_steps = 32, 35

train_iter, vocab = d2l.load_data_time_machine(batch_size, num_steps)
114
8.5. Implementation of Recurrent Neural Networks from Scratch 323

8.5.1 One-hot Encoding
Remember that each token is presented as a numerical index in train_iter. Feeding these indices
directly to the neural network might make it hard to learn. We often present each token as a more
expressive feature vector. The easiest representation is called one-hot encoding.
In a nutshell, we map each index to a different unit vector: assume that the number of different
tokens in the vocabulary is N (the len(vocab)) and the token indices range from 0 to N − 1. If
the index of a token is the integer i, then we create a vector ei of all 0s with a length of N and set
the element at position i to 1. This vector is the one-hot vector of the original token. The one-hot
vectors with indices 0 and 2 are shown below.
npx.one_hot(np.array([0, 2]), len(vocab))
array([[1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
[0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.]])
The shape of the minibatch we sample each time is (batch size, timestep). The one_hot function
transforms such a minibatch into a 3-D tensor with the last dimension equals to the vocabulary
size. We often transpose the input so that we will obtain a (timestep, batch size, vocabulary size)
output that fits into a sequence model easier.
npx.one_hot(X.T, 28).shape
(5, 2, 28)
8.5.2 Initializing the Model Parameters
Next, we initialize the model parameters for a RNN model. The number of hidden units
num_hiddens is a tunable parameter.
def get_params(vocab_size, num_hiddens, device):

num_inputs = num_outputs = vocab_size
def normal(shape):
return np.random.normal(scale=0.01, size=shape, ctx=device)
# Hidden layer parameters
W_xh = normal((num_inputs, num_hiddens))
W_hh = normal((num_hiddens, num_hiddens))
b_h = np.zeros(num_hiddens, ctx=device)
# Output layer parameters
W_hq = normal((num_hiddens, num_outputs))
b_q = np.zeros(num_outputs, ctx=device)
# Attach gradients
params = [W_xh, W_hh, b_h, W_hq, b_q]
param.attach_grad()
return params

8.5.3 RNN Model
First, we need an init_rnn_state function to return the hidden state at initialization. It returns a
tensor filled with 0 and with a shape of (batch size, number of hidden units). Using tuples makes
it easier to handle situations where the hidden state contains multiple variables (e.g., when com-
bining multiple layers in an RNN where each layer requires initializing).
def init_rnn_state(batch_size, num_hiddens, device):

return (np.zeros((batch_size, num_hiddens), ctx=device), )
The following rnn function defines how to compute the hidden state and output in a timestep. The
activation function here uses the tanh function. As described in :numref:sec_mlp, the mean value
of the tanh function is 0, when the elements are evenly distributed over the real numbers.
def rnn(inputs, state, params):

# Inputs shape: (num_steps, batch_size, vocab_size)
W_xh, W_hh, b_h, W_hq, b_q = params
H, = state
outputs = []
for X in inputs:
H = np.tanh(np.dot(X, W_xh) + np.dot(H, W_hh) + b_h)
Y = np.dot(H, W_hq) + b_q
outputs.append(Y)
return np.concatenate(outputs, axis=0), (H,)
Now we have all functions defined, next we create a class to wrap these functions and store pa-
rameters.
class RNNModelScratch: #@save

"""A RNN Model based on scratch implementations."""
def __init__(self, vocab_size, num_hiddens, device,
get_params, init_state, forward):
self.vocab_size, self.num_hiddens = vocab_size, num_hiddens
self.params = get_params(vocab_size, num_hiddens, device)
self.init_state, self.forward_fn = init_state, forward
def __call__(self, X, state):

X = npx.one_hot(X.T, self.vocab_size)
return self.forward_fn(X, state, self.params)
def begin_state(self, batch_size, ctx):

return self.init_state(batch_size, self.num_hiddens, ctx)
Let us do a sanity check whether inputs and outputs have the correct dimensions, e.g., to ensure
that the dimensionality of the hidden state has not changed.
num_hiddens = 512
model = RNNModelScratch(len(vocab), num_hiddens, d2l.try_gpu(), get_params,
init_rnn_state, rnn)
state = model.begin_state(X.shape[0], d2l.try_gpu())
Y, new_state = model(X.as_in_context(d2l.try_gpu()), state)
Y.shape, len(new_state), new_state[0].shape

((10, 28), 1, (2, 512))
We can see that the output shape is (number steps × batch size, vocabulary size), while the hidden
state shape remains the same, i.e., (batch size, number of hidden units).
8.5.4 Prediction
We first explain the predicting function so we can regularly check the prediction during training.
This function predicts the next num_predicts characters based on the prefix (a string containing
several characters). For the beginning of the sequence, we only update the hidden state. After
that we begin generating new characters and emitting them.
def predict_ch8(prefix, num_predicts, model, vocab, device): #@save

state = model.begin_state(batch_size=1, ctx=device)
outputs = [vocab[prefix[0]]]
get_input = lambda: np.array([outputs[-1]], ctx=device).reshape(1, 1)
for y in prefix[1:]: # Warmup state with prefix
_, state = model(get_input(), state)
outputs.append(vocab[y])
for _ in range(num_predicts): # Predict num_predicts steps
Y, state = model(get_input(), state)
outputs.append(int(Y.argmax(axis=1).reshape(1)))
return ''.join([vocab.idx_to_token[i] for i in outputs])
We test the predict_ch8 function first. Given that we did not train the network, it will generate
nonsensical predictions. We initialize it with the sequence traveller and have it generate 10 ad-
ditional characters.
predict_ch8('time traveller ', 10, model, vocab, d2l.try_gpu())
'time traveller iiiiiiiiii'
8.5.5 Gradient Clipping
For a sequence of length T , we compute the gradients over these T timesteps in an iteration, which
results in a chain of matrix-products with length O(T ) during backpropagating. As mentioned
in Section 4.8, it might result in numerical instability, e.g., the gradients may either explode or
vanish, when T is large. Therefore, RNN models often need extra help to stabilize the training.
Recall that when solving an optimization problem, we take update steps for the weights w in the
general direction of the negative gradient gt on a minibatch, say w − η · gt . Let us further assume
that the objective is well behaved, i.e., it is Lipschitz continuous with constant L, i.e.,
|l(w) − l(w′ )| ≤ L∥w − w′ ∥. (8.5.1)
In this case we can safely assume that if we update the weight vector by η · gt , we will not observe a
change by more than Lη∥gt ∥. This is both a curse and a blessing. A curse since it limits the speed
of making progress, whereas a blessing since it limits the extent to which things can go wrong if
we move in the wrong direction.

Sometimes the gradients can be quite large and the optimization algorithm may fail to converge.
We could address this by reducing the learning rate η or by some other higher order trick. But
what if we only rarely get large gradients? In this case such an approach may appear entirely
unwarranted. One alternative is to clip the gradients by projecting them back to a ball of a given
radius, say θ via
( )
θ
g ← min 1, g. (8.5.2)
∥g∥
By doing so we know that the gradient norm never exceeds θ and that the updated gradient is
entirely aligned with the original direction g. It also has the desirable side-effect of limiting the
influence any given minibatch (and within it any given sample) can exert on the weight vectors.
This bestows a certain degree of robustness to the model. Gradient clipping provides a quick
fix to the gradient exploding. While it does not entirely solve the problem, it is one of the many
techniques to alleviate it.
Below we define a function to clip the gradients of a model that is either a building from scratch
instance or a model constructed by the high-level APIs. Also note that we compute the gradient
norm over all parameters.
def grad_clipping(model, theta): #@save

if isinstance(model, gluon.Block):
params = [p.data() for p in model.collect_params().values()]
else:
params = model.params
norm = math.sqrt(sum((p.grad ** 2).sum() for p in params))
if norm > theta:
param.grad[:] *= theta / norm
8.5.6 Training
Let us first define the function to train the model on one data epoch. It differs from the models
training of Section 3.6 in three places:
1. Different sampling methods for sequential data (random sampling and sequential partition-
ing) will result in differences in the initialization of hidden states.
2. We clip the gradients before updating the model parameters. This ensures that the model
does not diverge even when gradients blow up at some point during the training process,
and it effectively reduces the step size automatically.
3. We use perplexity to evaluate the model. This ensures that sequences of different length are
comparable.
When the sequential partitioning is used, we initialize the hidden state at the beginning of each
epoch. Since the ith example in the next minibatch is adjacent to the current ith example, so the
next minibatch can use the current hidden state directly, we only detach the gradient so that we
compute the gradients within a minibatch. When using the random sampling, we need to re-
initialize the hidden state for each iteration since each example is sampled with a random position.
Same as the train_epoch_ch3 function in Section 3.6, we use generalized updater, which could be
either a high-level API trainer or a scratched implementation.

def train_epoch_ch8(model, train_iter, loss, updater, device, #@save
use_random_iter):
state, timer = None, d2l.Timer()
metric = d2l.Accumulator(2) # loss_sum, num_examples
for X, Y in train_iter:
if state is None or use_random_iter:
# Initialize state when either it is the first iteration or
# using random sampling.
state = model.begin_state(batch_size=X.shape[0], ctx=device)
else:
for s in state:
s.detach()
y = Y.T.reshape(-1)
py, state = model(X, state)
l = loss(py, y).mean()
l.backward()
grad_clipping(model, 1)
updater(batch_size=1) # Since used mean already
metric.add(l * y.size, y.size)
return math.exp(metric[0]/metric[1]), metric[1]/timer.stop()
The training function again supports either we implement the model from scratch or using high-
level APIs.
def train_ch8(model, train_iter, vocab, lr, num_epochs, device, #@save

use_random_iter=False):
# Initialize
animator = d2l.Animator(xlabel='epoch', ylabel='perplexity',
legend=['train'], xlim=[1, num_epochs])
if isinstance(model, gluon.Block):
model.initialize(ctx=device, force_reinit=True,
init=init.Normal(0.01))
trainer = gluon.Trainer(model.collect_params(),
'sgd', {'learning_rate': lr})
updater = lambda batch_size: trainer.step(batch_size)
else:
updater = lambda batch_size: d2l.sgd(model.params, lr, batch_size)
predict = lambda prefix: predict_ch8(prefix, 50, model, vocab, device)
# Train and check the progress.
ppl, speed = train_epoch_ch8(
model, train_iter, loss, updater, device, use_random_iter)
if epoch % 10 == 0:
print(predict('time traveller'))
animator.add(epoch+1, [ppl])
print(f'perplexity {ppl:.1f}, {speed:.1f} tokens/sec on {str(device)}')
print(predict('time traveller'))
print(predict('traveller'))
Now we can train a model. Since we only use 10, 000 tokens in the dataset, the model needs more
epochs to converge.

num_epochs, lr = 500, 1
train_ch8(model, train_iter, vocab, lr, num_epochs, d2l.try_gpu())
perplexity 1.0, 35037.7 tokens/sec on gpu(0)

time traveller it s against reason said filby what reason said
traveller it s against reason said filby what reason said
Finally let us check the results to use a random sampling iterator.
train_ch8(model, train_iter, vocab, lr, num_epochs, d2l.try_gpu(),

use_random_iter=True)

time traveller after the pauserequired for the proper assimilati
While implementing the above RNN model from scratch is instructive, it is not convenient. In the
next section we will see how to improve significantly on the current model and how to make it
faster and easier to implement.

Summary
• Sequence models need state initialization for training.

• Between sequential models you need to ensure to detach the gradients, to ensure that the
automatic differentiation does not propagate effects beyond the current sample.
• A simple RNN language model consists of an encoder, an RNN model, and a decoder.
• Gradient clipping prevents gradient explosion (but it cannot fix vanishing gradients).
• Perplexity calibrates model performance across different sequence length. It is the expo-
nentiated average of the cross-entropy loss.
• Sequential partitioning typically leads to better models.
Exercises
1. Show that one-hot encoding is equivalent to picking a different embedding for each object.
2. Adjust the hyperparameters to improve the perplexity.
• How low can you go? Adjust embeddings, hidden units, learning rate, etc.
• How well will it work on other books by H. G. Wells, e.g., The War of the Worlds115 .
3. Modify the predict function such as to use sampling rather than picking the most likely next
character.
• What happens?
• Bias the model towards more likely outputs, e.g., by sampling from q(wt |
wt−1 , . . . , w1 ) ∝ pα (wt | wt−1 , . . . , w1 ) for α > 1.
4. Run the code in this section without clipping the gradient. What happens?
5. Change sequential partitioning so that it does not separate hidden states from the computa-
tional graph. Does the running time change? How about the accuracy?
6. Replace the activation function used in this section with ReLU and repeat the experiments
in this section.
7. Prove that the perplexity is the inverse of the harmonic mean of the conditional word prob-
abilities.
Discussions116
115
http://www.gutenberg.org/ebooks/36
116

8.6 Concise Implementation of Recurrent Neural Networks
While Section 8.5 was instructive to see how recurrent neural networks (RNNs) are implemented,
this is not convenient or fast. This section will show how to implement the same language model
more efficiently using functions provided by Gluon. We begin as before by reading the “Time
Machine” corpus.

npx.set_np()

Gluonʼs rnn module provides a recurrent neural network implementation (beyond many other se-
quence models). We construct the recurrent neural network layer rnn_layer with a single hidden
layer and 256 hidden units, and initialize the weights.
num_hiddens = 256
rnn_layer = rnn.RNN(num_hiddens)
rnn_layer.initialize()
Initializing the state is straightforward. We invoke the member function rnn_layer.

begin_state(batch_size). This returns an initial state for each element in the minibatch. That
is, it returns an object of size (hidden layers, batch size, number of hidden units). The number of
hidden layers defaults to be 1. In fact, we have not even discussed yet what it means to have mul-
tiple layers—this will happen in Section 9.3. For now, suffice it to say that multiple layers simply
amount to the output of one RNN being used as the input for the next RNN.
batch_size = 1
state = rnn_layer.begin_state(batch_size=batch_size)
len(state), state[0].shape
(1, (1, 1, 256))
With a state variable and an input, we can compute the output with the updated state.
num_steps = 1
X = np.random.uniform(size=(num_steps, batch_size, len(vocab)))
Y, state_new = rnn_layer(X, state)
Y.shape, len(state_new), state_new[0].shape
((1, 1, 256), 1, (1, 1, 256))
Similar to Section 8.5, we define an RNNModel block by subclassing the Block class for a complete
recurrent neural network. Note that rnn_layer only contains the hidden recurrent layers, we need
8.6. Concise Implementation of Recurrent Neural Networks 331

to create a separate output layer. While in the previous section, we have the output layer within
the rnn block.
#@save
class RNNModel(nn.Block):
def __init__(self, rnn_layer, vocab_size, **kwargs):
super(RNNModel, self).__init__(**kwargs)
self.rnn = rnn_layer
self.vocab_size = vocab_size
self.dense = nn.Dense(vocab_size)
def forward(self, inputs, state):

X = npx.one_hot(inputs.T, self.vocab_size)
Y, state = self.rnn(X, state)
# The fully connected layer will first change the shape of `Y` to
# (`num_steps` * `batch_size`, `num_hiddens`). Its output shape is
# (`num_steps` * `batch_size`, `vocab_size`).
output = self.dense(Y.reshape(-1, Y.shape[-1]))
return output, state
def begin_state(self, *args, **kwargs):

return self.rnn.begin_state(*args, **kwargs)
8.6.2 Training and Predicting
Before training the model, let us make a prediction with the a model that has random weights.
device = d2l.try_gpu()
model = RNNModel(rnn_layer, len(vocab))
model.initialize(force_reinit=True, ctx=device)
d2l.predict_ch8('time traveller', 10, model, vocab, device)
'time travellervmjznnngii'
As is quite obvious, this model does not work at all. Next, we call train_ch8 with the same hyper-
parameters defined in Section 8.5 and train our model with Gluon.
d2l.train_ch8(model, train_iter, vocab, lr, num_epochs, device)

time traveller smiled round at us then still they could contit y
traveller you can show bamplyshoolatarlit theories betal t

Compared with the last section, this model achieves comparable perplexity, albeit within a shorter
period of time, due to the code being more optimized.
Summary
• Gluonʼs rnn module provides an implementation at the recurrent neural network layer.
• Gluonʼs nn.RNN instance returns the output and hidden state after forward computation. This
forward computation does not involve output layer computation.
• As before, the computational graph needs to be detached from previous steps for reasons of
efficiency.
Exercises
1. Compare the implementation with the previous section.

• Why does Gluonʼs implementation run faster?
• If you observe a significant difference beyond speed, try to find the reason.
2. Can you make the model overfit?
• Increase the number of hidden units.
• Increase the number of iterations.
• What happens if you adjust the clipping parameter?
3. Implement the autoregressive model of the introduction to the current chapter using an
RNN.
4. What happens if you increase the number of hidden layers in the RNN model? Can you make
the model work?
5. How well can you compress the text using this model?
• How many bits do you need?
• Why does not everyone use this model for text compression? Hint: what about the
compressor itself?
8.6. Concise Implementation of Recurrent Neural Networks 333

Discussions117
8.7 Backpropagation Through Time
So far we repeatedly alluded to things like exploding gradients, vanishing gradients, truncating back-
prop, and the need to detach the computational graph. For instance, in the previous section we in-
voked s.detach() on the sequence. None of this was really fully explained, in the interest of being
able to build a model quickly and to see how it works. In this section we will delve a bit more deeply
into the details of backpropagation for sequence models and why (and how) the math works. For a
more detailed discussion about randomization and backpropagation also see the paper by (Tallec
& Ollivier, 2017).
We encountered some of the effects of gradient explosion when we first implemented recurrent
neural networks (Section 8.5). In particular, if you solved the problems in the problem set, you
would have seen that gradient clipping is vital to ensure proper convergence. To provide a bet-
ter understanding of this issue, this section will review how gradients are computed for sequence
models. Note that there is nothing conceptually new in how it works. After all, we are still merely
applying the chain rule to compute gradients. Nonetheless, it is worth while reviewing backprop-
agation (Section 4.7) again.
Forward propagation in a recurrent neural network is relatively straightforward. Backpropagation
through time is actually a specific application of back propagation in recurrent neural networks. It
requires us to expand the recurrent neural network one timestep at a time to obtain the dependen-
cies between model variables and parameters. Then, based on the chain rule, we apply backprop-
agation to compute and store gradients. Since sequences can be rather long, the dependency can
be rather lengthy. For instance, for a sequence of 1000 characters, the first symbol could poten-
tially have significant influence on the symbol at position 1000. This is not really computationally
feasible (it takes too long and requires too much memory) and it requires over 1000 matrix-vector
products before we would arrive at that very elusive gradient. This is a process fraught with com-
putational and statistical uncertainty. In the following we will elucidate what happens and how
to address this in practice.
8.7.1 A Simplified Recurrent Network
We start with a simplified model of how an RNN works. This model ignores details about the
specifics of the hidden state and how it is updated. These details are immaterial to the analysis
and would only serve to clutter the notation, but make it look more intimidating. In this simplified
model, we denote ht as the hidden state, xt as the input, and ot as the output at timestep t. In
addition, wh and wo indicate the weights of hidden states and the output layer, respectively. As a
result, the hidden states and outputs at each timesteps can be explained as
ht = f (xt , ht−1 , wh ) and ot = g(ht , wo ). (8.7.1)
Hence, we have a chain of values {. . . , (ht−1 , xt−1 , ot−1 ), (ht , xt , ot ), . . .} that depend on each other
via recursive computation. The forward propagation is fairly straightforward. All we need is to
loop through the (xt , ht , ot ) triples one step at a time. The discrepancy between outputs ot and the
desired targets yt is then evaluated by an objective function as
∑
T
L(x, y, wh , wo ) = l(yt , ot ). (8.7.2)
t=1
117

For backpropagation, matters are a bit more tricky, especially when we compute the gradients
with regard to the parameters wh of the objective function L. To be specific, by the chain rule,
∑
T
∂wh L = ∂wh l(yt , ot )
t=1
(8.7.3)
∑T
= ∂ot l(yt , ot )∂ht g(ht , wh ) [∂wh ht ] .
t=1
The first and the second part of the derivative is easy to compute. The third part ∂wh ht is where
things get tricky, since we need to compute the effect of the parameters on ht .
To derive the above gradient, assume that we have three sequences {at }, {bt }, {ct } satisfying a0 =
0, a1 = b1 , and at = bt + ct at−1 for t = 1, 2, . . .. Then for t ≥ 1, it is easy to show
 
∑t−1 ∏t
at = bt +  cj  bi . (8.7.4)
i=1 j=i+1
Now let us apply (8.7.4) with
at = ∂wh ht , (8.7.5)
bt = ∂wh f (xt , ht−1 , wh ), (8.7.6)
ct = ∂ht−1 f (xt , ht−1 , wh ). (8.7.7)

Therefore, at = bt + ct at−1 becomes the following recursion
∂wh ht = ∂wh f (xt , ht−1 , w) + ∂h f (xt , ht−1 , wh )∂wh ht−1 . (8.7.8)
By (8.7.4), the third part will be

 
∑
t−1 ∏
t
∂wh ht = ∂wh f (xt , ht−1 , wh ) +  ∂hj−1 f (xj , hj−1 , wh ) ∂wh f (xi , hi−1 , wh ). (8.7.9)
i=1 j=i+1
While we can use the chain rule to compute ∂w ht recursively, this chain can get very long whenever
t is large. Let us discuss a number of strategies for dealing with this problem.
• Compute the full sum. This is very slow and gradients can blow up, since subtle changes in
the initial conditions can potentially affect the outcome a lot. That is, we could see things
similar to the butterfly effect where minimal changes in the initial conditions lead to dispro-
portionate changes in the outcome. This is actually quite undesirable in terms of the model
that we want to estimate. After all, we are looking for robust estimators that generalize well.
Hence this strategy is almost never used in practice.
• Truncate the sum after τ steps. This is what we have been discussing so far. This leads to
an approximation of the true gradient, simply by terminating the sum above at ∂w ht−τ . The
approximation error is thus given by ∂h f (xt , ht−1 , w)∂w ht−1 (multiplied by a product of gra-
dients involving ∂h f ). In practice this works quite well. It is what is commonly referred to as
truncated BPTT (backpropgation through time). One of the consequences of this is that the
model focuses primarily on short-term influence rather than long-term consequences. This
is actually desirable, since it biases the estimate towards simpler and more stable models.
8.7. Backpropagation Through Time 335

• Randomized Truncation. Last we can replace ∂wh ht by a random variable which is correct
in expectation but which truncates the sequence. This is achieved by using a sequence of ξt
where E[ξt ] = 1 and P (ξt = 0) = 1 − π and furthermore P (ξt = π −1 ) = π. We use this to
replace the gradient:
zt = ∂w f (xt , ht−1 , w) + ξt ∂h f (xt , ht−1 , w)∂w ht−1 . (8.7.10)
It follows from the definition of ξt that E[zt ] = ∂w ht . Whenever ξt = 0 the expansion terminates at
that point. This leads to a weighted sum of sequences of varying lengths where long sequences are
rare but appropriately overweighted. (Tallec & Ollivier, 2017) proposed this in their paper. Unfor-
tunately, while appealing in theory, the model does not work much better than simple truncation,
most likely due to a number of factors. First, the effect of an observation after a number of back-
propagation steps into the past is quite sufficient to capture dependencies in practice. Second,
the increased variance counteracts the fact that the gradient is more accurate. Third, we actually
want models that have only a short range of interaction. Hence, BPTT has a slight regularizing
effect which can be desirable.
Fig. 8.7.1: From top to bottom: randomized BPTT, regularly truncated BPTT and full BPTT
Fig. 8.7.1 illustrates the three cases when analyzing the first few words of The Time Machine: * The
first row is the randomized truncation which partitions the text into segments of varying length. *
The second row is the regular truncated BPTT which breaks it into sequences of the same length.
* The third row is the full BPTT that leads to a computationally infeasible expression.
8.7.2 The Computational Graph
In order to visualize the dependencies between model variables and parameters during computa-
tion in a recurrent neural network, we can draw a computational graph for the model, as shown
in Fig. 8.7.2. For example, the computation of the hidden states of timestep 3, h3 , depends on
the model parameters Whx and Whh , the hidden state of the last timestep h2 , and the input of the
current timestep x3 .

Fig. 8.7.2: Computational dependencies for a recurrent neural network model with three
timesteps. Boxes represent variables (not shaded) or parameters (shaded) and circles represent
operators.
8.7.3 BPTT in Detail
After discussing the general principle, let us discuss BPTT in detail. By decomposing W into differ-
ent sets of weight matrices (Whx , Whh and Woh ), we will get a simple linear latent variable model:
ht = Whx xt + Whh ht−1 and ot = Woh ht . (8.7.11)

∂L ∂L ∂L
Following the discussion in Section 4.7, we compute the gradients ∂Whx , ∂Whh , ∂Woh for
∑
T
L(x, y, W) = l(ot , yt ), (8.7.12)
t=1
where l(·) denotes the chosen loss function. Taking the derivatives with respect to Woh is fairly
straightforward and we obtain
∑
T
∂Woh L = prod (∂ot l(ot , yt ), ht ) , (8.7.13)
t=1
where prod(·) indicates the product of two or more matrices.

The dependency on Whx and Whh is a bit more tricky since it involves a chain of derivatives. We
begin with
∑
T
∂Whh L = prod (∂ot l(ot , yt ), Woh , ∂Whh ht ) ,
t=1
(8.7.14)
∑
T
∂Whx L = prod (∂ot l(ot , yt ), Woh , ∂Whx ht ) .
t=1
After all, hidden states depend on each other and on past inputs. The key quantity is how past
hidden states affect future hidden states.
( )T −t
∂ht ht+1 = W⊤ ⊤
hh and thus ∂ht hT = Whh . (8.7.15)
8.7. Backpropagation Through Time 337

Chaining terms together yields
t (
∑ )t−j
∂Whh ht = W⊤
hh hj
j=1
t (
(8.7.16)
∑ )t−j
∂Whx ht = W⊤
hh xj .
j=1
A number of things follow from this potentially very intimidating expression. First, it pays to store
intermediate results, i.e., powers of Whh as we work our way through the terms of the loss func-
tion L. Second, this simple linear example already exhibits some key problems of long sequence
models: it involves potentially very large powers Wjhh . In it, eigenvalues smaller than 1 vanish
for large j and eigenvalues larger than 1 diverge. This is numerically unstable and gives undue
importance to potentially irrelevant past detail. One way to address this is to truncate the sum
at a computationally convenient size. Later on in Chapter 9 we will see how more sophisticated
sequence models such as LSTMs can alleviate this further. In practice, this truncation is effected
by detaching the gradient after a given number of steps.
Summary
• Backpropagation through time is merely an application of backpropagation to sequence

models with a hidden state.
• Truncation is needed for computational convenience and numerical stability.
• High powers of matrices can lead to divergent and vanishing eigenvalues. This manifests
itself in the form of exploding or vanishing gradients.
• For efficient computation, intermediate values are cached.
Exercises
1. Assume that we have a symmetric matrix M ∈ Rn×n with eigenvalues λi . Without loss of
generality, assume that they are ordered in ascending order λi ≤ λi+1 . Show that Mk has
eigenvalues λki .
2. Prove that for a random vector x ∈ Rn , with high probability Mk x will be very much aligned
with the largest eigenvector vn of M. Formalize this statement.
3. What does the above result mean for gradients in a recurrent neural network?
4. Besides gradient clipping, can you think of any other methods to cope with gradient explo-
sion in recurrent neural networks?
Discussions118
118

9 | Modern Recurrent Neural Networks
Although we have learned the basics of recurrent neural networks, they are not sufficient for a
practitioner to solve todayʼs sequence learning problems. For instance, given the numerical un-
stability during gradient calculation, gated recurrent neural networks are much more common
in practice. We will begin by introducing two of such widely-used networks, namely gated recur-
rent units (GRUs) and long short term memory (LSTM), with illustrations using the same language
modeling problem as introduced in Chapter 8.
Furthermore, we will modify recurrent neural networks with a single undirectional hidden layer.
We will describe deep architectures, and discuss the bidirectional design with both forward and
backward recursion. They are frequently adopted in modern recurrent networks.
In fact, a large portion of sequence learning problems such as automatic speech recognition, text
to speech, and machine translation, consider both inputs and outputs to be sequences of arbi-
trary length. Finally, we will take machine translation as an example, and introduce the encoder-
decoder architecture based on recurrent neural networks and modern practices for such sequence
to sequence learning problems.
9.1 Gated Recurrent Units (GRU)
In the previous section, we discussed how gradients are calculated in a recurrent neural network.
In particular we found that long products of matrices can lead to vanishing or divergent gradients.
Let us briefly think about what such gradient anomalies mean in practice:
• We might encounter a situation where an early observation is highly significant for predict-
ing all future observations. Consider the somewhat contrived case where the first observa-
tion contains a checksum and the goal is to discern whether the checksum is correct at the
end of the sequence. In this case, the influence of the first token is vital. We would like to
have some mechanisms for storing vital early information in a memory cell. Without such a
mechanism, we will have to assign a very large gradient to this observation, since it affects
all subsequent observations.
• We might encounter situations where some symbols carry no pertinent observation. For
instance, when parsing a web page there might be auxiliary HTML code that is irrelevant
for the purpose of assessing the sentiment conveyed on the page. We would like to have
some mechanism for skipping such symbols in the latent state representation.
• We might encounter situations where there is a logical break between parts of a sequence.
For instance, there might be a transition between chapters in a book, or a transition between
a bear and a bull market for securities. In this case it would be nice to have a means of
resetting our internal state representation.
339
A number of methods have been proposed to address this. One of the earliest is Long Short Term
Memory (LSTM) (Hochreiter & Schmidhuber, 1997) which we will discuss in Section 9.2. Gated
Recurrent Unit (GRU) (Cho et al., 2014) is a slightly more streamlined variant that often offers
comparable performance and is significantly faster to compute. See also (Chung et al., 2014) for
more details. Due to its simplicity, let us start with the GRU.
9.1.1 Gating the Hidden State
The key distinction between regular RNNs and GRUs is that the latter support gating of the hidden
state. This means that we have dedicated mechanisms for when a hidden state should be updated
and also when it should be reset. These mechanisms are learned and they address the concerns
listed above. For instance, if the first symbol is of great importance we will learn not to update
the hidden state after the first observation. Likewise, we will learn to skip irrelevant temporary
observations. Last, we will learn to reset the latent state whenever needed. We discuss this in
detail below.
Reset Gates and Update Gates
The first thing we need to introduce are reset and update gates. We engineer them to be vectors
with entries in (0, 1) such that we can perform convex combinations. For instance, a reset vari-
able would allow us to control how much of the previous state we might still want to remember.
Likewise, an update variable would allow us to control how much of the new state is just a copy of
the old state.
We begin by engineering gates to generate these variables. Fig. 9.1.1 illustrates the inputs for
both reset and update gates in a GRU, given the current timestep input Xt and the hidden state of
the previous timestep Ht−1 . The output is given by a fully connected layer with a sigmoid as its
activation function.
Fig. 9.1.1: Reset and update gate in a GRU.
For a given timestep t, the minibatch input is Xt ∈ Rn×d (number of examples: n, number of
inputs: d) and the hidden state of the last timestep is Ht−1 ∈ Rn×h (number of hidden states: h).
340 Chapter 9. Modern Recurrent Neural Networks

Then, the reset gate Rt ∈ Rn×h and update gate Zt ∈ Rn×h are computed as follows:
Rt = σ(Xt Wxr + Ht−1 Whr + br ),

(9.1.1)
Zt = σ(Xt Wxz + Ht−1 Whz + bz ).
Here, Wxr , Wxz ∈ Rd×h and Whr , Whz ∈ Rh×h are weight parameters and br , bz ∈ R1×h are biases.
We use a sigmoid function (as introduced in Section 4.1) to transform input values to the interval
(0, 1).
Reset Gates in Action
We begin by integrating the reset gate with a regular latent state updating mechanism. In a con-
ventional RNN, we would have an hidden state update of the form
Ht = tanh(Xt Wxh + Ht−1 Whh + bh ). (9.1.2)
This is essentially identical to the discussion of the previous section, albeit with a nonlinearity
in the form of tanh to ensure that the values of the hidden states remain in the interval (−1, 1).
If we want to be able to reduce the influence of the previous states we can multiply Ht−1 with Rt
elementwise. Whenever the entries in the reset gate Rt are close to 1, we recover a conventional
RNN. For all entries of the reset gate Rt that are close to 0, the hidden state is the result of an MLP
with Xt as input. Any pre-existing hidden state is thus reset to defaults. This leads to the following
candidate hidden state (it is a candidate since we still need to incorporate the action of the update
gate).
H̃t = tanh(Xt Wxh + (Rt ⊙ Ht−1 ) Whh + bh ). (9.1.3)
Fig. 9.1.2 illustrates the computational flow after applying the reset gate. The symbol ⊙ indicates
pointwise multiplication between tensors.
Fig. 9.1.2: Candidate hidden state computation in a GRU. The multiplication is carried out ele-
mentwise.
9.1. Gated Recurrent Units (GRU) 341

Update Gates in Action
Next we need to incorporate the effect of the update gate Zt , as shown in :numref:fig_gru_3. This
determines the extent to which the new state Ht is just the old state Ht−1 and by how much the
new candidate state H̃t is used. The gating variable Zt can be used for this purpose, simply by
taking elementwise convex combinations between both candidates. This leads to the final update
equation for the GRU.
Ht = Zt ⊙ Ht−1 + (1 − Zt ) ⊙ H̃t . (9.1.4)
Fig. 9.1.3: Hidden state computation in a GRU. As before, the multiplication is carried out elemen-
twise.
Whenever the update gate Zt is close to 1, we simply retain the old state. In this case the infor-
mation from Xt is essentially ignored, effectively skipping timestep t in the dependency chain. In
contrast, whenever Zt is close to 0, the new latent state Ht approaches the candidate latent state H̃t .
These designs can help us cope with the vanishing gradient problem in RNNs and better capture
dependencies for time series with large timestep distances. In summary, GRUs have the following
two distinguishing features:
• Reset gates help capture short-term dependencies in time series.
• Update gates help capture long-term dependencies in time series.
To gain a better understanding of the model, let us implement a GRU from scratch.

Reading the Dataset
We begin by reading The Time Machine corpus that we used in Section 8.5. The code for reading
the dataset is given below:

from mxnet.gluon import rnn
npx.set_np()

The next step is to initialize the model parameters. We draw the weights from a Gaussian with
variance to be 0.01 and set the bias to 0. The hyperparameter num_hiddens defines the number of
hidden units. We instantiate all weights and biases relating to the update gate, the reset gate, and
the candidate hidden state itself. Subsequently, we attach gradients to all the parameters.
def get_params(vocab_size, num_hiddens, device):

def normal(shape):
def three():
return (normal((num_inputs, num_hiddens)),
normal((num_hiddens, num_hiddens)),
np.zeros(num_hiddens, ctx=device))
W_xz, W_hz, b_z = three() # Update gate parameter

W_xr, W_hr, b_r = three() # Reset gate parameter
W_xh, W_hh, b_h = three() # Candidate hidden state parameter
# Attach gradients
params = [W_xz, W_hz, b_z, W_xr, W_hr, b_r, W_xh, W_hh, b_h, W_hq, b_q]
param.attach_grad()
return params

Defining the Model
Now we will define the hidden state initialization function init_gru_state. Just like the
init_rnn_state function defined in Section 8.5, this function returns a tensor with a shape (batch
size, number of hidden units) whose values are all zeros.
def init_gru_state(batch_size, num_hiddens, device):

return (np.zeros(shape=(batch_size, num_hiddens), ctx=device), )
Now we are ready to define the GRU model. Its structure is the same as the basic RNN cell, except
that the update equations are more complex.
def gru(inputs, state, params):

W_xz, W_hz, b_z, W_xr, W_hr, b_r, W_xh, W_hh, b_h, W_hq, b_q = params
H, = state
outputs = []
for X in inputs:
Z = npx.sigmoid(np.dot(X, W_xz) + np.dot(H, W_hz) + b_z)
R = npx.sigmoid(np.dot(X, W_xr) + np.dot(H, W_hr) + b_r)
H_tilda = np.tanh(np.dot(X, W_xh) + np.dot(R * H, W_hh) + b_h)
H = Z * H + (1 - Z) * H_tilda
outputs.append(Y)
return np.concatenate(outputs, axis=0), (H,)
Training and Prediction
Training and prediction work in exactly the same manner as before. After training for one epoch,
the perplexity and the output sentence will be like the following.
vocab_size, num_hiddens, device = len(vocab), 256, d2l.try_gpu()

model = d2l.RNNModelScratch(len(vocab), num_hiddens, device, get_params,
init_gru_state, gru)


In Gluon, we can directly call the GRU class in the rnn module. This encapsulates all the config-
uration detail that we made explicit above. The code is significantly faster as it uses compiled
operators rather than Python for many details that we spelled out in detail before.
gru_layer = rnn.GRU(num_hiddens)
model = d2l.RNNModel(gru_layer, len(vocab))

traveller comint bee on it furedent a vegions of space and

Summary
• Gated recurrent neural networks are better at capturing dependencies for time series with
large timestep distances.
• Reset gates help capture short-term dependencies in time series.
• Update gates help capture long-term dependencies in time series.
• GRUs contain basic RNNs as their extreme case whenever the reset gate is switched on. They
can ignore sequences as needed.
Exercises
1. Compare runtime, perplexity, and the output strings for rnn.RNN and rnn.GRU implementa-
tions with each other.
2. Assume that we only want to use the input for timestep t′ to predict the output at timestep
t > t′ . What are the best values for the reset and update gates for each timestep?
3. Adjust the hyperparameters and observe and analyze the impact on running time, perplex-
ity, and the written lyrics.
4. What happens if you implement only parts of a GRU? That is, implement a recurrent cell
that only has a reset gate. Likewise, implement a recurrent cell only with an update gate.
Discussions119
9.2 Long Short Term Memory (LSTM)
The challenge to address long-term information preservation and short-term input skipping in
latent variable models has existed for a long time. One of the earliest approaches to address this
was the LSTM (Hochreiter & Schmidhuber, 1997). It shares many of the properties of the Gated Re-
current Unit (GRU). Interestingly, LSTMʼs design is slightly more complex than GRU but predates
GRU by almost two decades.
Arguably it is inspired by logic gates of a computer. To control a memory cell we need a number
of gates. One gate is needed to read out the entries from the cell (as opposed to reading any other
cell). We will refer to this as the output gate. A second gate is needed to decide when to read data
into the cell. We refer to this as the input gate. Last, we need a mechanism to reset the contents of
the cell, governed by a forget gate. The motivation for such a design is the same as before, namely
to be able to decide when to remember and when to ignore inputs in the latent state via a dedicated
mechanism. Let us see how this works in practice.
119

9.2.1 Gated Memory Cells
Three gates are introduced in LSTMs: the input gate, the forget gate, and the output gate. In addi-
tion to that we will introduce the memory cell that has the same shape as the hidden state. Strictly
speaking this is just a fancy version of a hidden state, engineered to record additional information.
Input Gates, Forget Gates, and Output Gates
Just like with GRUs, the data feeding into the LSTM gates is the input at the current timestep Xt and
the hidden state of the previous timestep Ht−1 . These inputs are processed by a fully connected
layer and a sigmoid activation function to compute the values of input, forget and output gates.
As a result, the three gatesʼ all output values are in the range of [0, 1]. Fig. 9.2.1 illustrates the data
flow for the input, forget, and output gates.
Fig. 9.2.1: Calculation of input, forget, and output gates in an LSTM.
We assume that there are h hidden units, the minibatch is of size n, and number of inputs is d.
Thus, the input is Xt ∈ Rn×d and the hidden state of the last timestep is Ht−1 ∈ Rn×h . Correspond-
ingly, the gates are defined as follows: the input gate is It ∈ Rn×h , the forget gate is Ft ∈ Rn×h ,
and the output gate is Ot ∈ Rn×h . They are calculated as follows:
It = σ(Xt Wxi + Ht−1 Whi + bi ),

Ft = σ(Xt Wxf + Ht−1 Whf + bf ), (9.2.1)
Ot = σ(Xt Wxo + Ht−1 Who + bo ),
where Wxi , Wxf , Wxo ∈ Rd×h and Whi , Whf , Who ∈ Rh×h are weight parameters and bi , bf , bo ∈
R1×h are bias parameters.
9.2. Long Short Term Memory (LSTM) 347

Candidate Memory Cell
Next we design the memory cell. Since we have not specified the action of the various gates yet,
we first introduce the candidate memory cell C̃t ∈ Rn×h . Its computation is similar to the three
gates described above, but using a tanh function with a value range for [−1, 1] as the activation
function. This leads to the following equation at timestep t.
C̃t = tanh(Xt Wxc + Ht−1 Whc + bc ). (9.2.2)
Here Wxc ∈ Rd×h and Whc ∈ Rh×h are weight parameters and bc ∈ R1×h is a bias parameter.
A quick illustration of the candidate memory cell is shown in Fig. 9.2.2.
Fig. 9.2.2: Computation of candidate memory cells in LSTM.
Memory Cell
In GRUs, we had a single mechanism to govern input and forgetting. Here in LSTMs we have two
parameters, It which governs how much we take new data into account via C̃t and the forget pa-
rameter Ft which addresses how much of the old memory cell content Ct−1 ∈ Rn×h we retain.
Using the same pointwise multiplication trick as before, we arrive at the following update equa-
tion.
Ct = Ft ⊙ Ct−1 + It ⊙ C̃t . (9.2.3)
If the forget gate is always approximately 1 and the input gate is always approximately 0, the past
memory cells Ct−1 will be saved over time and passed to the current timestep. This design was
introduced to alleviate the vanishing gradient problem and to better capture dependencies for
time series with long range dependencies. We thus arrive at the flow diagram in Fig. 9.2.3.

Fig. 9.2.3: Computation of memory cells in an LSTM. Here, the multiplication is carried out ele-
mentwise.
Hidden States
Last, we need to define how to compute the hidden state Ht ∈ Rn×h . This is where the output
gate comes into play. In LSTM it is simply a gated version of the tanh of the memory cell. This
ensures that the values of Ht are always in the interval (−1, 1). Whenever the output gate is 1 we
effectively pass all memory information through to the predictor, whereas for output 0 we retain
all the information only within the memory cell and perform no further processing. Fig. 9.2.4 has
a graphical illustration of the data flow.
Ht = Ot ⊙ tanh(Ct ). (9.2.4)
Fig. 9.2.4: Computation of the hidden state. Multiplication is elementwise.

Now let us implement an LSTM from scratch. As same as the experiments in the previous sections,
we first load data of The Time Machine.

npx.set_np()

Next we need to define and initialize the model parameters. As previously, the hyperparameter
num_hiddens defines the number of hidden units. We initialize weights following a Gaussian dis-
tribution with 0.01 standard deviation, and we set the biases to 0.
def get_lstm_params(vocab_size, num_hiddens, device):

def normal(shape):
def three():
return (normal((num_inputs, num_hiddens)),
normal((num_hiddens, num_hiddens)),
np.zeros(num_hiddens, ctx=device))
W_xi, W_hi, b_i = three() # Input gate parameters

W_xf, W_hf, b_f = three() # Forget gate parameters
W_xo, W_ho, b_o = three() # Output gate parameters
W_xc, W_hc, b_c = three() # Candidate cell parameters
# Attach gradients
params = [W_xi, W_hi, b_i, W_xf, W_hf, b_f, W_xo, W_ho, b_o, W_xc, W_hc,
b_c, W_hq, b_q]
param.attach_grad()
return params

Defining the Model
In the initialization function, the hidden state of the LSTM needs to return an additional mem-
ory cell with a value of 0 and a shape of (batch size, number of hidden units). Hence we get the
following state initialization.
def init_lstm_state(batch_size, num_hiddens, device):

return (np.zeros(shape=(batch_size, num_hiddens), ctx=device),
np.zeros(shape=(batch_size, num_hiddens), ctx=device))
The actual model is defined just like what we discussed before: providing three gates and an auxil-
iary memory cell. Note that only the hidden state is passed to the output layer. The memory cells
Ct do not participate in the output computation directly.
def lstm(inputs, state, params):

[W_xi, W_hi, b_i, W_xf, W_hf, b_f, W_xo, W_ho, b_o, W_xc, W_hc, b_c,
W_hq, b_q] = params
(H, C) = state
outputs = []
for X in inputs:
I = npx.sigmoid(np.dot(X, W_xi) + np.dot(H, W_hi) + b_i)
F = npx.sigmoid(np.dot(X, W_xf) + np.dot(H, W_hf) + b_f)
O = npx.sigmoid(np.dot(X, W_xo) + np.dot(H, W_ho) + b_o)
C_tilda = np.tanh(np.dot(X, W_xc) + np.dot(H, W_hc) + b_c)
C = F * C + I * C_tilda
H = O * np.tanh(C)
outputs.append(Y)
return np.concatenate(outputs, axis=0), (H, C)
Let us train an LSTM as same as what we did in Section 9.1, by calling the RNNModelScratch function
as introduced in Section 8.5.
vocab_size, num_hiddens, device = len(vocab), 256, d2l.try_gpu()

model = d2l.RNNModelScratch(len(vocab), num_hiddens, device, get_lstm_params,
init_lstm_state, lstm)

traveller but now you begin to seethe object of my investig

In Gluon, we can directly call the LSTM class in the rnn module. This encapsulates all the config-
uration details that we made explicit above. The code is significantly faster as it uses compiled
operators rather than Python for many details that we spelled out in detail before.
lstm_layer = rnn.LSTM(num_hiddens)
model = d2l.RNNModel(lstm_layer, len(vocab))

time traveller for so it will be convenient to speak of him was
traveller the parsporiccallime co bat exenp bas arount you
In many cases, LSTMs perform slightly better than GRUs but they are more costly to train and
execute due to the larger latent state size. LSTMs are the prototypical latent variable autoregressive
model with nontrivial state control. Many variants thereof have been proposed over the years, e.g.,
multiple layers, residual connections, different types of regularization. However, training LSTMs
and other sequence models (such as GRU) are quite costly due to the long range dependency of
the sequence. Later we will encounter alternative models such as Transformers that can be used
in some cases.

Summary
• LSTMs have three types of gates: input gates, forget gates, and output gates which control
the flow of information.
• The hidden layer output of LSTM includes hidden states and memory cells. Only hidden
states are passed into the output layer. Memory cells are entirely internal.
• LSTMs can cope with vanishing and exploding gradients.
Exercises
1. Adjust the hyperparameters. Observe and analyze the impact on runtime, perplexity, and
the generated output.
2. How would you need to change the model to generate proper words as opposed to sequences
of characters?
3. Compare the computational cost for GRUs, LSTMs, and regular RNNs for a given hidden
dimension. Pay special attention to the training and inference cost.
4. Since the candidate memory cells ensure that the value range is between −1 and 1 by using
the tanh function, why does the hidden state need to use the tanh function again to ensure
that the output value range is between −1 and 1?
5. Implement an LSTM for time series prediction rather than character sequence prediction.
Discussions120
9.3 Deep Recurrent Neural Networks
Up to now, we only discussed recurrent neural networks with a single unidirectional hidden layer.
In it the specific functional form of how latent variables and observations interact was rather ar-
bitrary. This is not a big problem as long as we have enough flexibility to model different types
of interactions. With a single layer, however, this can be quite challenging. In the case of the
perceptron, we fixed this problem by adding more layers. Within RNNs this is a bit trickier, since
we first need to decide how and where to add extra nonlinearity. Our discussion below focuses
primarily on LSTMs, but it applies to other sequence models, too.
• We could add extra nonlinearity to the gating mechanisms. That is, instead of using a single
perceptron we could use multiple layers. This leaves the mechanism of the LSTM unchanged.
Instead it makes it more sophisticated. This would make sense if we were led to believe
that the LSTM mechanism describes some form of universal truth of how latent variable
autoregressive models work.
• We could stack multiple layers of LSTMs on top of each other. This results in a mechanism
that is more flexible, due to the combination of several simple layers. In particular, data
might be relevant at different levels of the stack. For instance, we might want to keep high-
level data about financial market conditions (bear or bull market) available, whereas at a
lower level we only record shorter-term temporal dynamics.
120
9.3. Deep Recurrent Neural Networks 353

Beyond all this abstract discussion it is probably easiest to understand the family of models we are
interested in by reviewing Fig. 9.3.1. It describes a deep recurrent neural network with L hidden
layers. Each hidden state is continuously passed to both the next timestep of the current layer and
the current timestep of the next layer.
Fig. 9.3.1: Architecture of a deep recurrent neural network.
9.3.1 Functional Dependencies
At timestep t we assume that we have a minibatch Xt ∈ Rn×d (number of examples: n, number of

(ℓ)
inputs: d). The hidden state of hidden layer ℓ (ℓ = 1, . . . , L) is Ht ∈ Rn×h (number of hidden units:
h), the output layer variable is Ot ∈ R n×q (number of outputs: q) and a hidden layer activation
function fl for layer l. We compute the hidden state of layer 1 as before, using Xt as input. For all
subsequent layers, the hidden state of the previous layer is used in its place.
( )
(1) (1)
Ht = f1 Xt , Ht−1 ,
( ) (9.3.1)
(l) (l−1) (l)
Ht = fl Ht , Ht−1 .
Finally, the output layer is only based on the hidden state of hidden layer L. We use the output
function g to address this:
( )
(L)
Ot = g Ht . (9.3.2)
Just as with multilayer perceptrons, the number of hidden layers L and number of hidden units h
are hyper parameters. In particular, we can pick a regular RNN, a GRU, or an LSTM to implement
the model.

Fortunately many of the logistical details required to implement multiple layers of an RNN are
readily available in Gluon. To keep things simple we only illustrate the implementation using
such built-in functionality. The code is very similar to the one we used previously for LSTMs. In
fact, the only difference is that we specify the number of layers explicitly rather than picking the
default of a single layer. Let us begin by importing the appropriate modules and loading data.

from mxnet import npx
npx.set_np()

The architectural decisions (such as choosing parameters) are very similar to those of previous
sections. We pick the same number of inputs and outputs as we have distinct tokens, i.e., vo-
cab_size. The number of hidden units is still 256. The only difference is that we now select a
nontrivial number of layers num_layers = 2.
vocab_size, num_hiddens, num_layers = len(vocab), 256, 2

lstm_layer = rnn.LSTM(num_hiddens, num_layers)
9.3.3 Training
The actual invocation logic is identical to before. The only difference is that we now instantiate
two layers with LSTMs. This rather more complex architecture and the large number of epochs
slow down training considerably.

9.3. Deep Recurrent Neural Networks 355

Summary
• In deep recurrent neural networks, hidden state information is passed to the next timestep
of the current layer and the current timestep of the next layer.
• There exist many different flavors of deep RNNs, such as LSTMs, GRUs, or regular RNNs.
Conveniently these models are all available as parts of the rnn module in Gluon.
• Initialization of the models requires care. Overall, deep RNNs require considerable amount
of work (such as learning rate and clipping) to ensure proper convergence.
Exercises
1. Try to implement a two-layer RNN from scratch using the single layer implementation we
discussed in Section 8.5.
2. Replace the LSTM by a GRU and compare the accuracy.
3. Increase the training data to include multiple books. How low can you go on the perplexity
scale?
4. Would you want to combine sources of different authors when modeling text? Why is this a
good idea? What could go wrong?
Discussions121
9.4 Bidirectional Recurrent Neural Networks
So far we assumed that our goal is to model the next word given what we have seen so far, e.g., in
the context of a time series or in the context of a language model. While this is a typical scenario,
it is not the only one we might encounter. To illustrate the issue, consider the following three tasks
of filling in the blanks in a text:
1. I am _____
2. I am _____ very hungry.
121

3. I am _____ very hungry, I could eat half a pig.
Depending on the amount of information available, we might fill the blanks with very different
words such as “happy”, “not”, and “very”. Clearly the end of the phrase (if available) conveys sig-
nificant information about which word to pick. A sequence model that is incapable of taking
advantage of this will perform poorly on related tasks. For instance, to do well in named entity
recognition (e.g., to recognize whether “Green” refers to “Mr. Green” or to the color) longer-range
context is equally vital. To get some inspiration for addressing the problem let us take a detour to
graphical models.
9.4.1 Dynamic Programming
This section serves to illustrate the dynamic programming problem. The specific technical details
do not matter for understanding the deep learning counterpart but they help in motivating why
one might use deep learning and why one might pick specific architectures.
If we want to solve the problem using graphical models we could for instance design a latent vari-
able model as follows. We assume that there exists some latent variable ht which governs the
emissions xt that we observe via p(xt | ht ). Moreover, the transitions ht → ht+1 are given by some
state transition probability p(ht + 1 | ht ). The graphical model is then a Hidden Markov Model
(HMM) as in Fig. 9.4.1.
Fig. 9.4.1: Hidden Markov Model.
Thus, for a sequence of T observations we have the following joint probability distribution over
observed and hidden states:
∏
T
p(x, h) = p(h1 )p(x1 | h1 ) p(ht | ht−1 )p(xt | ht ). (9.4.1)
t=2
Now assume that we observe all xi with the exception of some xj and it is our goal to compute
p(xj | x−j ), where x−j = (x1 , x2 , . . . , xj−1 ). To accomplish this we need to sum over all possible
choices of h = (h1 , . . . , hT ). In case hi can take on k distinct values, this means that we need to
sum over k T terms—mission impossible! Fortunately there is an elegant solution for this: dynamic
programming. To see how it works, consider summing over the first two hidden variable h1 and
9.4. Bidirectional Recurrent Neural Networks 357

h2 . This yields:
∑
p(x) = p(x1 , . . . , xT ; h1 , . . . , hT )
h1 ,...,hT
∑ ∏
T
= p(h1 )p(x1 | h1 ) p(ht | ht−1 )p(xt | ht )
h1 ,...,hT t=2
 
∑ ∑ ∏
T
=  p(h1 )p(x1 | h1 )p(h2 | h1 ) p(x2 | h2 ) p(ht | ht−1 )p(xt | ht )
h2 ,...,hT h1 t=3
| {z }
=:π2 (h2 ) (9.4.2)
 
∑ ∑ ∏
T
=  
π2 (h2 )p(x2 | h2 )p(h3 | h2 ) p(x3 | h3 ) p(ht | ht−1 )p(xt | ht )
h3 ,...,hT h2 t=4
| {z }
=:π3 (h3 )
= ...
∑
= πT (hT )p(xT | hT ).
hT
In general we have the forward recursion as

∑
πt+1 (ht+1 ) = πt (ht )p(xt | ht )p(ht+1 | ht ). (9.4.3)
ht
The recursion is initialized as π1 (h1 ) = p(h1 ). In abstract terms this can be written as πt+1 =
f (πt , xt ), where f is some learnable function. This looks very much like the update equation in
the hidden variable models we discussed so far in the context of RNNs. Entirely analogously to
the forward recursion, we can also start a backward recursion. This yields:
∑
p(x) = p(x1 , . . . , xT ; h1 , . . . , hT )
h1 ,...,hT
∑ −1
T∏
= p(ht | ht−1 )p(xt | ht ) · p(hT | hT −1 )p(xT | hT )
h1 ,...,hT t=1
 
∑ −1
T∏ ∑
= p(ht | ht−1 )p(xt | ht ) ·  p(hT | hT −1 )p(xT | hT )
h1 ,...,hT −1 t=1 hT
| {z }
=:ρT −1 (hT −1 )
 
∑ −2
T∏ ∑
= p(ht | ht−1 )p(xt | ht ) ·  p(hT −1 | hT −2 )p(xT −1 | hT −1 )ρT −1 (hT −1 )
h1 ,...,hT −2 t=1 hT −1
| {z }
=:ρT −2 (hT −2 )
= ...,
∑
= p(h1 )p(x1 | h1 )ρ1 (h1 ).
h1
(9.4.4)
We can thus write the backward recursion as

∑
ρt−1 (ht−1 ) = p(ht | ht−1 )p(xt | ht )ρt (ht ), (9.4.5)
ht

with initialization ρT (hT ) = 1. These two recursions allow us to sum over T variables in O(kT )
(linear) time over all values of (h1 , . . . , hT ) rather than in exponential time. This is one of the
great benefits of the probabilistic inference with graphical models. It is a very special instance
of the (Aji & McEliece, 2000) proposed in 2000 by Aji and McEliece. Combining both forward and
backward pass, we are able to compute
∑
p(xj | x−j ) ∝ πj (hj )ρj (hj )p(xj | hj ). (9.4.6)
hj
Note that in abstract terms the backward recursion can be written as ρt−1 = g(ρt , xt ), where g is a
learnable function. Again, this looks very much like an update equation, just running backwards
unlike what we have seen so far in RNNs. Indeed, HMMs benefit from knowing future data when
it is available. Signal processing scientists distinguish between the two cases of knowing and not
knowing future observations as interpolation v.s. extrapolation. See the introductory chapter of
the book by (Doucet et al., 2001) on sequential Monte Carlo algorithms for more details.
9.4.2 Bidirectional Model
If we want to have a mechanism in RNNs that offers comparable look-ahead ability as in HMMs,
we need to modify the recurrent net design that we have seen so far. Fortunately, this is easy con-
ceptually. Instead of running an RNN only in the forward mode starting from the first symbol, we
start another one from the last symbol running from back to front. Bidirectional recurrent neural
networks add a hidden layer that passes information in a backward direction to more flexibly pro-
cess such information. Fig. 9.4.2 illustrates the architecture of a bidirectional recurrent neural
network with a single hidden layer.
Fig. 9.4.2: Architecture of a bidirectional recurrent neural network.
In fact, this is not too dissimilar to the forward and backward recursion we encountered above.
The main distinction is that in the previous case these equations had a specific statistical meaning.
Now they are devoid of such easily accessible interpretation and we can just treat them as generic
functions. This transition epitomizes many of the principles guiding the design of modern deep
networks: first, use the type of functional dependencies of classical statistical models, and then
use the models in a generic form.

Definition
Bidirectional RNNs were introduced by (Schuster & Paliwal, 1997). For a detailed discussion of
the various architectures see also the paper by (Graves & Schmidhuber, 2005). Let us look at the
specifics of such a network.
For a given timestep t, the minibatch input is Xt ∈ Rn×d (number of examples: n, number of in-
puts: d) and the hidden layer activation function is ϕ. In the bidirectional architecture, we assume
→
− ←
−
that the forward and backward hidden states for this timestep are H t ∈ Rn×h and H t ∈ Rn×h re-
spectively. Here h indicates the number of hidden units. We compute the forward and backward
hidden state updates as follows:
−
→ (f ) −
→ (f ) (f )
H t = ϕ(Xt Wxh + H t−1 Whh + bh ),
←
− (b) ←
− (b) (b)
(9.4.7)
H t = ϕ(Xt Wxh + H t+1 Whh + bh ).
(f ) (f ) (b) (b)
Here, the weight parameters Wxh ∈ Rd×h , Whh ∈ Rh×h , Wxh ∈ Rd×h , and Whh ∈ Rh×h , and bias
(f ) (b)
parameters bh ∈ R1×h and bh ∈ R1×h are all model parameters.
→
− ←−
Then we concatenate the forward and backward hidden states H t and H t to obtain the hidden
state Ht ∈ Rn×2h and feed it to the output layer. In deep bidirectional RNNs, the information
is passed on as input to the next bidirectional layer. Last, the output layer computes the output
Ot ∈ Rn×q (number of outputs: q):
Ot = Ht Whq + bq . (9.4.8)
Here, the weight parameter Whq ∈ R2h×q and the bias parameter bq ∈ R1×q are the model param-
eters of the output layer. The two directions can have different numbers of hidden units.
Computational Cost and Applications
One of the key features of a bidirectional RNN is that information from both ends of the sequence
is used to estimate the output. That is, we use information from both future and past observations
to predict the current one (a smoothing scenario). In the case of language models this is not quite
what we want. After all, we do not have the luxury of knowing the next to next symbol when
predicting the next one. Hence, if we were to use a bidirectional RNN naively we would not get a
very good accuracy: during training we have past and future data to estimate the present. During
test time we only have past data and thus poor accuracy (we will illustrate this in an experiment
below).
To add insult to injury, bidirectional RNNs are also exceedingly slow. The main reasons for this
are that they require both a forward and a backward pass and that the backward pass is dependent
on the outcomes of the forward propagation. Hence, gradients will have a very long dependency
chain.
In practice bidirectional layers are used very sparingly and only for a narrow set of applications,
such as filling in missing words, annotating tokens (e.g., for named entity recognition), or encod-
ing sequences wholesale as a step in a sequence processing pipeline (e.g., for machine transla-
tion). In short, handle with care!

Training a Bidirectional RNN for the Wrong Application
If we were to ignore all advice regarding the fact that bidirectional LSTMs use past and future
data and simply apply it to language models, we will get estimates with acceptable perplexity.
Nonetheless, the ability of the model to predict future symbols is severely compromised as the
example below illustrates. Despite reasonable perplexity, it only generates gibberish even after
many iterations. We include the code below as a cautionary example against using them in the
wrong context.

from mxnet import npx
npx.set_np()
# Load data
batch_size, num_steps, device = 32, 35, d2l.try_gpu()
# Define the model
vocab_size, num_hiddens, num_layers = len(vocab), 256, 2
lstm_layer = rnn.LSTM(num_hiddens, num_layers, bidirectional=True)
# Train the model

time travellerererererererererererererererererererererererererer
travellerererererererererererererererererererererererererer
The output is clearly unsatisfactory for the reasons described above. For a discussion of more
effective uses of bidirectional models, please see the sentiment classification in Section 15.2.

Summary
• In bidirectional recurrent neural networks, the hidden state for each timestep is simultane-
ously determined by the data prior to and after the current timestep.
• Bidirectional RNNs bear a striking resemblance with the forward-backward algorithm in
graphical models.
• Bidirectional RNNs are mostly useful for sequence embedding and the estimation of obser-
vations given bidirectional context.
• Bidirectional RNNs are very costly to train due to long gradient chains.
Exercises
1. If the different directions use a different number of hidden units, how will the shape of Ht
change?
2. Design a bidirectional recurrent neural network with multiple hidden layers.
3. Implement a sequence classification algorithm using bidirectional RNNs. Hint: use the RNN
to embed each word and then aggregate (average) all embedded outputs before sending the
output into an MLP for classification. For instance, if we have (o1 , o2 , o3 ), we compute ō =
1∑
3 i oi first and then use the latter for sentiment classification.
Discussions122
9.5 Machine Translation and the Dataset
So far we see how to use recurrent neural networks for language models, in which we predict the
next token given all previous tokens in an article. Now let us have a look at a different application,
machine translation, whose predict output is no longer a single token, but a list of tokens.
Machine translation (MT) refers to the automatic translation of a segment of text from one lan-
guage to another. Solving this problem with neural networks is often called neural machine trans-
lation (NMT). Compared to language models (Section 8.3), in which the corpus only contains a
single language, machine translation dataset has at least two languages, the source language and
the target language. In addition, each sentence in the source language is mapped to the accord-
ing translation in the target language. Therefore, the data preprocessing for machine translation
data is different to the one for language models. This section is dedicated to demonstrate how to
pre-process such a dataset and then load into a set of minibatches.

from mxnet import np, npx, gluon
import os
npx.set_np()
122

9.5.1 Reading and Preprocessing the Dataset
We first download a dataset that contains a set of English sentences with the corresponding French
translations. As can be seen that each line contains an English sentence with its French transla-
tion, which are separated by a TAB.
#@save
d2l.DATA_HUB['fra-eng'] = (d2l.DATA_URL + 'fra-eng.zip',
'94646ad1522d915e7b0f9296181140edcf86a4f5')
#@save
def read_data_nmt():
data_dir = d2l.download_extract('fra-eng')
with open(os.path.join(data_dir, 'fra.txt'), 'r') as f:
return f.read()
raw_text = read_data_nmt()
print(raw_text[0:106])
Go. Va !
Hi. Salut !
Run! Cours�!
Run! Courez�!
Who? Qui ?
Wow! Ça alors�!
Fire! Au feu !
Help! À l'aide�!
We perform several preprocessing steps on the raw text data, including ignoring cases, replacing
UTF-8 non-breaking space with space, and adding space between words and punctuation marks.
#@save
def preprocess_nmt(text):
def no_space(char, prev_char):
return char in set(',.!') and prev_char != ' '
text = text.replace('\u202f', ' ').replace('\xa0', ' ').lower()

out = [' ' + char if i > 0 and no_space(char, text[i-1]) else char
for i, char in enumerate(text)]
return ''.join(out)
text = preprocess_nmt(raw_text)
print(text[0:95])
go . va !
hi . salut !
run ! cours !
run ! courez !
who? qui ?
wow ! ça alors !
fire ! au feu !
9.5. Machine Translation and the Dataset 363

9.5.2 Tokenization
Different to using character tokens in Section 8.3, here a token is either a word or a punctuation
mark. The following function tokenizes the text data to return source and target. Each one is a list
of token list, with source[i] is the ith sentence in the source language and target[i] is the ith sen-
tence in the target language. To make the latter training faster, we sample the first num_examples
sentences pairs.
#@save
def tokenize_nmt(text, num_examples=None):
source, target = [], []
for i, line in enumerate(text.split('\n')):
if num_examples and i > num_examples:
break
parts = line.split('\t')
if len(parts) == 2:
source.append(parts[0].split(' '))
target.append(parts[1].split(' '))
return source, target
source, target = tokenize_nmt(text)

source[0:3], target[0:3]
([['go', '.'], ['hi', '.'], ['run', '!']],

[['va', '!'], ['salut', '!'], ['cours', '!']])
We visualize the histogram of the number of tokens per sentence in the following figure. As can be
seen, a sentence in average contains 5 tokens, and most of the sentences have less than 10 tokens.
d2l.set_figsize()
d2l.plt.hist([[len(l) for l in source], [len(l) for l in target]],
label=['source', 'target'])
d2l.plt.legend(loc='upper right');

9.5.3 Vocabulary
Since the tokens in the source language could be different to the ones in the target language, we
need to build a vocabulary for each of them. Since we are using words instead of characters as
tokens, it makes the vocabulary size significantly large. Here we map every token that appears
less than 3 times into the <unk> token Section 8.2. In addition, we need other special tokens such
as padding and sentence beginnings.
src_vocab = d2l.Vocab(source, min_freq=3,

reserved_tokens=['<pad>', '<bos>', '<eos>'])
len(src_vocab)
9140
9.5.4 Loading the Dataset
In language models, each example is a num_steps length sequence from the corpus, which may
be a segment of a sentence, or span over multiple sentences. In machine translation, an exam-
ple should contain a pair of source sentence and target sentence. These sentences might have
different lengths, while we need same length examples to form a minibatch.
One way to solve this problem is that if a sentence is longer than num_steps, we trim its length,
otherwise pad with a special <pad> token to meet the length. Therefore we could transform any
sentence to a fixed length.
#@save
def truncate_pad(line, num_steps, padding_token):
if len(line) > num_steps:
return line[:num_steps] # Trim
return line + [padding_token] * (num_steps - len(line)) # Pad
truncate_pad(src_vocab[source[0]], 10, src_vocab['<pad>'])
[47, 4, 1, 1, 1, 1, 1, 1, 1, 1]
Now we can convert a list of sentences into an (num_example, num_steps) index array. We also
record the length of each sentence without the padding tokens, called valid length, which might
be used by some models. In addition, we add the special “<bos>” and “<eos>” tokens to the target
sentences so that our model will know the signals for starting and ending predicting.
#@save
def build_array(lines, vocab, num_steps, is_source):
lines = [vocab[l] for l in lines]
if not is_source:
lines = [[vocab['<bos>']] + l + [vocab['<eos>']] for l in lines]
array = np.array([truncate_pad(
l, num_steps, vocab['<pad>']) for l in lines])
valid_len = (array != vocab['<pad>']).sum(axis=1)
return array, valid_len
Then we can construct minibatches based on these arrays.
9.5. Machine Translation and the Dataset 365

Finally, we define the function load_data_nmt to return the data iterator with the vocabularies for
source language and target language.
#@save
def load_data_nmt(batch_size, num_steps, num_examples=1000):
text = preprocess_nmt(read_data_nmt())
source, target = tokenize_nmt(text, num_examples)
src_vocab = d2l.Vocab(source, min_freq=3,
tgt_vocab = d2l.Vocab(target, min_freq=3,
src_array, src_valid_len = build_array(
source, src_vocab, num_steps, True)
tgt_array, tgt_valid_len = build_array(
target, tgt_vocab, num_steps, False)
data_arrays = (src_array, src_valid_len, tgt_array, tgt_valid_len)
data_iter = d2l.load_array(data_arrays, batch_size)
return src_vocab, tgt_vocab, data_iter
Let us read the first batch.
src_vocab, tgt_vocab, train_iter = load_data_nmt(batch_size=2, num_steps=8)

for X, X_vlen, Y, Y_vlen in train_iter:
print('X:', X.astype('int32'))
print('valid lengths for X:', X_vlen)
print('Y:', Y.astype('int32'))
print('valid lengths for Y:', Y_vlen)
break
X: [[ 5 45 8 4 1 1 1 1]
[10 70 4 1 1 1 1 1]]
valid lengths for X: [4 3]
Y: [[ 2 6 55 47 4 3 1 1]
[ 2 9 0 4 3 1 1 1]]
valid lengths for Y: [6 5]
Summary
• Machine translation (MT) refers to the automatic translation of a segment of text from one
language to another.
• We read, preprocess, and tokenize the datasets from both source language and target lan-
guage.

Exercises
1. Find a machine translation dataset online and process it.

Discussions123
9.6 Encoder-Decoder Architecture
The encoder-decoder architecture is a neural network design pattern. As shown in Fig. 9.6.1, the
architecture is partitioned into two parts, the encoder and the decoder. The encoderʼs role is to
encode the inputs into state, which often contains several tensors. Then the state is passed into
the decoder to generate the outputs. In machine translation, the encoder transforms a source
sentence, e.g., “Hello world.”, into state, e.g., a vector, that captures its semantic information. The
decoder then uses this state to generate the translated target sentence, e.g., “Bonjour le monde.”.
Fig. 9.6.1: The encoder-decoder architecture.
In this section, we will show an interface to implement this encoder-decoder architecture.
9.6.1 Encoder
The encoder is a normal neural network that takes inputs, e.g., a source sentence, to return out-
puts.
#@save
class Encoder(nn.Block):
"""The base encoder interface for the encoder-decoder architecture."""
super(Encoder, self).__init__(**kwargs)
def forward(self, X, *args):

raise NotImplementedError
9.6.2 Decoder
The decoder has an additional method init_state to parse the outputs of the encoder with pos-
sible additional information, e.g., the valid lengths of inputs, to return the state it needs. In the
forward method, the decoder takes both inputs, e.g., a target sentence and the state. It returns
outputs, with potentially modified state if the encoder contains RNN layers.
123
9.6. Encoder-Decoder Architecture 367

#@save
class Decoder(nn.Block):
"""The base decoder interface for the encoder-decoder architecture."""
super(Decoder, self).__init__(**kwargs)
def init_state(self, enc_outputs, *args):

def forward(self, X, state):

9.6.3 Model
The encoder-decoder model contains both an encoder and a decoder. We implement its forward
method for training. It takes both encoder inputs and decoder inputs, with optional additional
arguments. During computation, it first computes encoder outputs to initialize the decoder state,
and then returns the decoder outputs.
#@save
class EncoderDecoder(nn.Block):
"""The base class for the encoder-decoder architecture."""
def __init__(self, encoder, decoder, **kwargs):
super(EncoderDecoder, self).__init__(**kwargs)
self.encoder = encoder
self.decoder = decoder
def forward(self, enc_X, dec_X, *args):

enc_outputs = self.encoder(enc_X, *args)
dec_state = self.decoder.init_state(enc_outputs, *args)
return self.decoder(dec_X, dec_state)
Summary
• An encoder-decoder architecture is a neural network design pattern mainly in natural lan-

guage processing.
• An encoder is a network (FC, CNN, RNN, etc.) that takes the input, and outputs a feature
map, a vector or a tensor.
• A decoder is a network (usually the same network structure as encoder) that takes the feature
vector from the encoder, and gives the best closest match to the actual input or intended
output.

Exercises
1. Besides machine translation, can you think of another application scenarios where an
encoder-decoder architecture can fit?
2. Can you design a deep encoder-decoder architecture?
Discussions124
9.7 Sequence to Sequence
The sequence to sequence (seq2seq) model is based on the encoder-decoder architecture to gen-
erate a sequence output for a sequence input, as demonstrated in Fig. 9.7.1. Both the encoder and
the decoder use recurrent neural networks (RNNs) to handle sequence inputs of variable length.
The hidden state of the encoder is used directly to initialize the decoder hidden state to pass in-
formation from the encoder to the decoder.
Fig. 9.7.1: The sequence to sequence model architecture.
The layers in the encoder and the decoder are illustrated in Fig. 9.7.2.
Fig. 9.7.2: Layers in the encoder and the decoder.
In this section we will explain and implement the seq2seq model to train on the machine transla-
tion dataset.

from mxnet import np, npx, init, gluon, autograd
124
9.7. Sequence to Sequence 369

npx.set_np()
9.7.1 Encoder
Recall that the encoder of seq2seq can transform the inputs of variable length to a fixed-length
context vector c by encoding the sequence information into c. We usually use RNN layers within
the encoder. Suppose that we have an input sequence x1 , . . . , xT , where xt is the tth word. At
timestep t, the RNN will have two vectors as the input: the feature vector xt of xt and the hidden
state of the last timestep ht−1 . Let us denote the transformation of the RNNʼs hidden states by a
function f :
ht = f (xt , ht−1 ). (9.7.1)
Next, the encoder captures information of all the hidden states and encodes it into the context
vector c with a function q:
c = q(h1 , . . . , hT ). (9.7.2)
For example, if we choose q as q(h1 , . . . , hT ) = hT , then the context vector will be the final hidden
state hT .
So far what we describe above is a unidirectional RNN, where each timestepʼs hidden state depends
only on the previous timestepsʼ. We can also use other forms of RNNs such as GRUs, LSTMs, and
bidirectional RNNs to encode the sequential input.
Now let us implement the seq2seqʼs encoder. Here we use the word embedding layer to obtain the
feature vector according to the word index of the input language. Those feature vectors will be
fed to a multi-layer LSTM. The input for the encoder is a batch of sequences, which is 2-D tensor
with shape (batch size, sequence length). The encoder returns both the LSTM outputs, i.e., hidden
states of all the timesteps, as well as the hidden state and the memory cell of the final timestep.
#@save
class Seq2SeqEncoder(d2l.Encoder):
def __init__(self, vocab_size, embed_size, num_hiddens, num_layers,
dropout=0, **kwargs):
super(Seq2SeqEncoder, self).__init__(**kwargs)
self.embedding = nn.Embedding(vocab_size, embed_size)
self.rnn = rnn.LSTM(num_hiddens, num_layers, dropout=dropout)
def forward(self, X, *args):

# `X` shape: (`batch_size`, `seq_len`, èmbed_size`)
X = self.embedding(X)
# RNN needs first axes to be timestep, i.e., `seq_len`
X = X.swapaxes(0, 1)
state = self.rnn.begin_state(batch_size=X.shape[1], ctx=X.ctx)
out, state = self.rnn(X, state)
# òut` shape: (`seq_len`, `batch_size`, `num_hiddens`)
# `state` shape: (`num_layers`, `batch_size`, `num_hiddens`),
# where "state" contains the hidden state and the memory cell
return out, state

Next, we will create a minibatch sequence input with a batch size of 4 and 7 timesteps. We assume
the number of hidden layers of the LSTM unit is 2 and the number of hidden units is 16. The out-
put shape returned by the encoder after performing forward calculation on the input is (number
of timesteps, batch size, number of hidden units). The shape of the multi-layer hidden state of
the gated recurrent unit in the final timestep is (number of hidden layers, batch size, number of
hidden units). For the gated recurrent unit, the state list contains only one element, which is the
hidden state. If long short-term memory is used, the state list will also contain another element,
which is the memory cell.
encoder = Seq2SeqEncoder(vocab_size=10, embed_size=8, num_hiddens=16,

num_layers=2)
encoder.initialize()
X = np.zeros((4, 7))
output, state = encoder(X)
output.shape
(7, 4, 16)
Since an LSTM is used, the state list will contain both the hidden state and the memory cell with
same shape (number of hidden layers, batch size, number of hidden units). However, if a GRU
is used, the state list will contain only one element—the hidden state in the final timestep with
shape (number of hidden layers, batch size, number of hidden units).
len(state), state[0].shape, state[1].shape
(2, (2, 4, 16), (2, 4, 16))
9.7.2 Decoder
As we just introduced, the context vector c encodes the information from the whole input se-
quence x1 , . . . , xT . Suppose that the given outputs in the training set are y1 , . . . , yT ′ . At each
timestep t′ , the conditional probability of output yt′ will depend on the previous output sequence
y1 , . . . , yt′ −1 and the context vector c, i.e.,
P (yt′ | y1 , . . . , yt′ −1 , c). (9.7.3)
Hence, we can use another RNN as the decoder. At timestep t′ , the decoder will update its hidden
state st′ using three inputs: the feature vector yt′ −1 of yt′ −1 , the context vector c, and the hidden
state of the last timestep st′ −1 . Let us denote the transformation of the RNNʼs hidden states within
the decoder by a function g:
st′ = g(yt′ −1 , c, st′ −1 ). (9.7.4)
When implementing the decoder, we directly use the hidden state of the encoder in the final
timestep as the initial hidden state of the decoder. This requires that the encoder and decoder
RNNs have the same numbers of layers and hidden units. The LSTM forward calculation of the
decoder is similar to that of the encoder. The only difference is that we add a dense layer after
the LSTM layers, where the hidden size is the vocabulary size. The dense layer will predict the
confidence score for each word.

#@save
class Seq2SeqDecoder(d2l.Decoder):
super(Seq2SeqDecoder, self).__init__(**kwargs)
self.dense = nn.Dense(vocab_size, flatten=False)
def init_state(self, enc_outputs, *args):

return enc_outputs[1]

X = self.embedding(X).swapaxes(0, 1)
out, state = self.rnn(X, state)
# Make the batch to be the first dimension to simplify loss computation
out = self.dense(out).swapaxes(0, 1)
return out, state
We create a decoder with the same hyperparameters as the encoder. As we can see, the output
shape is changed to (batch size, the sequence length, vocabulary size).
decoder = Seq2SeqDecoder(vocab_size=10, embed_size=8,

num_hiddens=16, num_layers=2)
decoder.initialize()
state = decoder.init_state(encoder(X))
out, state = decoder(X, state)
out.shape, len(state), state[0].shape, state[1].shape
((4, 7, 10), 2, (2, 4, 16), (2, 4, 16))
9.7.3 The Loss Function
For each timestep, the decoder outputs a vocabulary-size confidence score vector to predict words.
Similar to language modeling, we can apply softmax to obtain the probabilities and then use cross-
entropy loss to calculate the loss. Note that we padded the target sentences to make them have
the same length, but we do not need to compute the loss on the padding symbols.
To implement the loss function that filters out some entries, we will use an operator called Se-
quenceMask. It can specify to mask the first dimension (axis=0) or the second one (axis=1). If the
second one is chosen, given a valid length vector len and 2-dim input X, this operator sets X[i,
len[i]:] = 0 for all iʼs.
X = np.array([[1, 2, 3], [4, 5, 6]])

npx.sequence_mask(X, np.array([1, 2]), True, axis=1)
array([[1., 0., 0.],

[4., 5., 0.]])
Apply to n-dim tensor X, it sets X[i, len[i]:, :, ..., :] = 0. In addition, we can specify the
filling value such as −1 as shown below.

X = np.ones((2, 3, 4))
npx.sequence_mask(X, np.array([1, 2]), True, value=-1, axis=1)
array([[[ 1., 1., 1., 1.],

[-1., -1., -1., -1.],
[-1., -1., -1., -1.]],
[[ 1., 1., 1., 1.],

[ 1., 1., 1., 1.],
[-1., -1., -1., -1.]]])
Now we can implement the masked version of the softmax cross-entropy loss. Note that each
Gluon loss function allows to specify per-example weights, in default they are 1s. Then we can
just use a zero weight for each example we would like to remove. So our customized loss function
accepts an additional valid_len argument to ignore some failing elements in each sequence.
#@save
class MaskedSoftmaxCELoss(gluon.loss.SoftmaxCELoss):
# `pred` shape: (`batch_size`, `seq_len`, `vocab_size`)
# `label` shape: (`batch_size`, `seq_len`)
# `valid_len` shape: (`batch_size`, )
def forward(self, pred, label, valid_len):
# weights shape: (batch_size, seq_len, 1)
weights = np.expand_dims(np.ones_like(label), axis=-1)
weights = npx.sequence_mask(weights, valid_len, True, axis=1)
return super(MaskedSoftmaxCELoss, self).forward(pred, label, weights)
For a sanity check, we create identical three sequences, keep 4 elements for the first sequence, 2
elements for the second sequence, and none for the last one. Then the first example loss should
be 2 times larger than the second one, and the last loss should be 0.
loss = MaskedSoftmaxCELoss()
loss(np.ones((3, 4, 10)), np.ones((3, 4)), np.array([4, 2, 0]))
array([2.3025851, 1.1512926, 0. ])
9.7.4 Training
During training, if the target sequence has length n, we feed the first n − 1 tokens into the decoder
as inputs, and the last n − 1 tokens are used as ground truth label.
#@save
def train_s2s_ch9(model, data_iter, lr, num_epochs, device):
model.initialize(init.Xavier(), force_reinit=True, ctx=device)
trainer = gluon.Trainer(model.collect_params(),
'adam', {'learning_rate': lr})
loss = MaskedSoftmaxCELoss()
animator = d2l.Animator(xlabel='epoch', ylabel='loss',
xlim=[1, num_epochs], ylim=[0, 0.25])
for epoch in range(1, num_epochs + 1):

timer = d2l.Timer()
metric = d2l.Accumulator(2) # loss_sum, num_tokens
for batch in data_iter:
X, X_vlen, Y, Y_vlen = [x.as_in_ctx(device) for x in batch]
Y_input, Y_label, Y_vlen = Y[:, :-1], Y[:, 1:], Y_vlen-1
Y_hat, _ = model(X, Y_input, X_vlen, Y_vlen)
l = loss(Y_hat, Y_label, Y_vlen)
l.backward()
d2l.grad_clipping(model, 1)
num_tokens = Y_vlen.sum()
trainer.step(num_tokens)
metric.add(l.sum(), num_tokens)
if epoch % 10 == 0:
animator.add(epoch, (metric[0]/metric[1],))
print(f'loss {metric[0] / metric[1]:.3f}, {metric[1] / timer.stop():.1f} '
f'tokens/sec on {str(device)}')
Next, we create a model instance and set hyperparameters. Then, we can train the model.
embed_size, num_hiddens, num_layers, dropout = 32, 32, 2, 0.0

lr, num_epochs, device = 0.005, 300, d2l.try_gpu()
src_vocab, tgt_vocab, train_iter = d2l.load_data_nmt(batch_size, num_steps)

encoder = Seq2SeqEncoder(
len(src_vocab), embed_size, num_hiddens, num_layers, dropout)
decoder = Seq2SeqDecoder(
len(tgt_vocab), embed_size, num_hiddens, num_layers, dropout)
model = d2l.EncoderDecoder(encoder, decoder)
train_s2s_ch9(model, train_iter, lr, num_epochs, device)
loss 0.026, 8491.3 tokens/sec on gpu(0)

9.7.5 Predicting
Here we implement the simplest method, greedy search, to generate an output sequence. As illus-
trated in Fig. 9.7.3, during predicting, we feed the same “<bos>” token to the decoder as training
at timestep 0. But the input token for a later timestep is the predicted token from the previous
timestep.
Fig. 9.7.3: Sequence to sequence model predicting with greedy search
#@save
def predict_s2s_ch9(model, src_sentence, src_vocab, tgt_vocab, num_steps,
device):
src_tokens = src_vocab[src_sentence.lower().split(' ')]
enc_valid_len = np.array([len(src_tokens)], ctx=device)
src_tokens = d2l.truncate_pad(src_tokens, num_steps, src_vocab['<pad>'])
enc_X = np.array(src_tokens, ctx=device)
# Add the batch size dimension
enc_outputs = model.encoder(np.expand_dims(enc_X, axis=0),
enc_valid_len)
dec_state = model.decoder.init_state(enc_outputs, enc_valid_len)
dec_X = np.expand_dims(np.array([tgt_vocab['<bos>']], ctx=device), axis=0)
predict_tokens = []
for _ in range(num_steps):
Y, dec_state = model.decoder(dec_X, dec_state)
# The token with highest score is used as the next timestep input
dec_X = Y.argmax(axis=2)
py = dec_X.squeeze(axis=0).astype('int32').item()
if py == tgt_vocab['<eos>']:
break
predict_tokens.append(py)
return ' '.join(tgt_vocab.to_tokens(predict_tokens))
Try several examples:
for sentence in ['Go .', 'Wow !', "I'm OK .", 'I won !']:
print(sentence + ' => ' + predict_s2s_ch9(
model, sentence, src_vocab, tgt_vocab, num_steps, device))
Go . => va !
Wow ! => <unk> !
I'm OK . => je vais bien .
I won ! => je l'ai emporté !

Summary
• The sequence to sequence (seq2seq) model is based on the encoder-decoder architecture to

generate a sequence output from a sequence input.
• We use multiple LSTM layers for both the encoder and the decoder.
Exercises
1. Can you think of other use cases of seq2seq besides neural machine translation?
2. What if the input sequence in the example of this section is longer?
3. If we do not use the SequenceMask in the loss function, what may happen?
Discussions125
9.8 Beam Search
In Section 9.7, we discussed how to train an encoder-decoder with input and output sequences
that are both of variable length. In this section, we are going to introduce how to use the encoder-
decoder to predict sequences of variable length.
As in Section 9.5, when preparing to train the dataset, we normally attach a special symbol “<eos>”
after each sentence to indicate the termination of the sequence. We will continue to use this math-
ematical symbol in the discussion below. For ease of discussion, we assume that the output of the
decoder is a sequence of text. Let the size of output text dictionary Y (contains special symbol
′
“<eos>”) be |Y|, and the maximum length of the output sequence be T ′ . There are a total O(|Y|T )
types of possible output sequences. All the subsequences after the special symbol “<eos>” in these
output sequences will be discarded. Besides, we still denote the context vector as c, which encodes
information of all the hidden states from the input.
9.8.1 Greedy Search
First, we will take a look at a simple solution: greedy search. For any timestep t′ of the output
sequence, we are going to search for the word with the highest conditional probability from |Y|
numbers of words, with
yt′ = argmax P (y | y1 , . . . , yt′ −1 , c) (9.8.1)

y∈Y
as the output. Once the “<eos>” symbol is detected, or the output sequence has reached its maxi-
mum length T ′ , the output is completed.
As we mentioned in our discussion of the decoder, the conditional probability of generating an
∏ ′
output sequence based on the input sequence is Tt′ =1 P (yt′ | y1 , . . . , yt′ −1 , c). We will take the out-
put sequence with the highest conditional probability as the optimal sequence. The main problem
with greedy search is that there is no guarantee that the optimal sequence will be obtained.
125

Take a look at the example below. We assume that there are four words “A”, “B”, “C”, and “<eos>”
in the output dictionary. The four numbers under each timestep in Fig. 9.8.1 represent the condi-
tional probabilities of generating “A”, “B”, “C”, and “<eos>” at that timestep, respectively. At each
timestep, greedy search selects the word with the highest conditional probability. Therefore, the
output sequence “A”, “B”, “C”, and “<eos>” will be generated in Fig. 9.8.1. The conditional proba-
bility of this output sequence is 0.5 × 0.4 × 0.4 × 0.6 = 0.048.
Fig. 9.8.1: The four numbers under each timestep represent the conditional probabilities of gen-
erating “A”, “B”, “C”, and “<eos>” at that timestep, respectively. At each timestep, greedy search
selects the word with the highest conditional probability.
Now, we will look at another example shown in Fig. 9.8.2. Unlike in Fig. 9.8.1, the following figure
Fig. 9.8.2 selects the word “C”, which has the second highest conditional probability at timestep
2. Since the output subsequences of timesteps 1 and 2, on which timestep 3 is based, are changed
from “A” and “B” in Fig. 9.8.1 to “A” and “C” in Fig. 9.8.2, the conditional probability of each word
generated at timestep 3 has also changed in Fig. 9.8.2. We choose the word “B”, which has the high-
est conditional probability. Now, the output subsequences of timestep 4 based on the first three
timesteps are “A”, “C”, and “B”, which are different from “A”, “B”, and “C” in Fig. 9.8.1. Therefore,
the conditional probability of generating each word in timestep 4 in Fig. 9.8.2 is also different
from that in Fig. 9.8.1. We find that the conditional probability of the output sequence “A”, “C”,
“B”, and “<eos>” at the current timestep is 0.5 × 0.3 × 0.6 × 0.6 = 0.054, which is higher than the
conditional probability of the output sequence obtained by greedy search. Therefore, the output
sequence “A”, “B”, “C”, and “<eos>” obtained by the greedy search is not an optimal sequence.
Fig. 9.8.2: The four numbers under each timestep represent the conditional probabilities of gener-
ating “A”, “B”, “C”, and “<eos>” at that timestep. At timestep 2, the word “C”, which has the second
highest conditional probability, is selected.
9.8. Beam Search 377

9.8.2 Exhaustive Search
If the goal is to obtain the optimal sequence, we may consider using exhaustive search: an exhaus-
tive examination of all possible output sequences, which outputs the sequence with the highest
conditional probability.
Although we can use an exhaustive search to obtain the optimal sequence, its computational over-
′
head O(|Y|T ) is likely to be excessively high. For example, when |Y| = 10000 and T ′ = 10, we
will need to evaluate 1000010 = 1040 sequences. This is next to impossible to complete. The
computational overhead of greedy search is O(|Y| T ′ ), which is usually significantly less than the
computational overhead of an exhaustive search. For example, when |Y| = 10000 and T ′ = 10,
we only need to evaluate 10000 × 10 = 1 × 105 sequences.
9.8.3 Beam Search
Beam search is an improved algorithm based on greedy search. It has a hyperparameter named
beam size, k. At timestep 1, we select k words with the highest conditional probability to be the first
word of the k candidate output sequences. For each subsequent timestep, we are going to select
the k output sequences with the highest conditional probability from the total of k |Y| possible
output sequences based on the k candidate output sequences from the previous timestep. These
will be the candidate output sequences for that timestep. Finally, we will filter out the sequences
containing the special symbol “<eos>” from the candidate output sequences of each timestep and
discard all the subsequences after it to obtain a set of final candidate output sequences.
Fig. 9.8.3: The beam search process. The beam size is 2 and the maximum length of the output
sequence is 3. The candidate output sequences are A, C, AB, CE, ABD, and CED.
Fig. 9.8.3 demonstrates the process of beam search with an example. Suppose that the vocabulary
of the output sequence contains only five elements: Y = {A, B, C, D, E} where one of them is
a special symbol “<eos>”. Set beam size to 2, the maximum length of the output sequence to 3.
At timestep 1 of the output sequence, suppose the words with the highest conditional probability

P (y1 | c) are A and C. At timestep 2, for all y2 ∈ Y, we compute
P (A, y2 | c) = P (A | c)P (y2 | A, c) (9.8.2)
and
P (C, y2 | c) = P (C | c)P (y2 | C, c), (9.8.3)
and pick the largest two among these 10 values, say
P (A, B | c) and P (C, E | c). (9.8.4)
Then at timestep 3, for all y3 ∈ Y, we compute
P (A, B, y3 | c) = P (A, B | c)P (y3 | A, B, c) (9.8.5)
and
P (C, E, y3 | c) = P (C, E | c)P (y3 | C, E, c), (9.8.6)
and pick the largest two among these 10 values, say
P (A, B, D | c) and P (C, E, D | c). (9.8.7)
As a result, we obtain 6 candidates output sequences: (1) A; (2) C; (3) A, B; (4) C, E; (5) A, B, D;
and (6) C, E, D. In the end, we will get the set of final candidate output sequences based on these
6 sequences.
In the set of final candidate output sequences, we will take the sequence with the highest score as
the output sequence from those below:
1 ∑
L
1
log P (y1 , . . . , yL ) = α log P (yt′ | y1 , . . . , yt′ −1 , c), (9.8.8)
Lα L ′
t =1
Here, L is the length of the final candidate sequence and the selection for α is generally 0.75. The
Lα on the denominator is a penalty on the logarithmic addition scores for the longer sequences
above. The computational overhead O(k |Y| T ′ ) of the beam search can be obtained through anal-
ysis. The result is between the computational overhead of greedy search and exhaustive search.
In addition, greedy search can be treated as a beam search with a beam size of 1. Beam search
strikes a balance between computational overhead and search quality using a flexible beam size
of k.
Summary
• Methods for predicting variable-length sequences include greedy search, exhaustive search,
and beam search.
• Beam search strikes a balance between computational overhead and search quality using a
flexible beam size.
9.8. Beam Search 379

Exercises
1. Can we treat an exhaustive search as a beam search with a special beam size? Why?
2. We used language models to generate sentences in Section 8.5. Which kind of search does
this output use? Can you improve it?
Discussions126
126

10 | Attention Mechanisms
As a bit of a historical digression, attention research is an enormous field with a long history in
cognitive neuroscience. Focalization, concentration of consciousness are of the essence of atten-
tion, which enable the human to prioritize the perception in order to deal effectively with others.
As a result, we do not process all the information that is available in the sensory input. At any
time, we are aware of only a small fraction of the information in the environment. In cognitive
neuroscience, there are several types of attention such as selective attention, covert attention, and
spatial attention. The theory ignites the spark in recent deep learning is the feature integration the-
ory of the selective attention, which was developed by Anne Treisman and Garry Gelade through
the paper (Treisman & Gelade, 1980) in 1980. This paper declares that when perceiving a stimulus,
features are registered early, automatically, and in parallel, while objects are identified separately
and at a later stage in processing. The theory has been one of the most influential psychological
models of human visual attention.
However, we will not indulge in too much theory of attention in neuroscience, but rather focus on
applying the attention idea in deep learning, where attention can be seen as a generalized pooling
method with bias alignment over inputs. In this chapter, we will provide you with some intuition
about how to transform the attention idea to the concrete mathematics models, and make them
work.
10.1 Attention Mechanisms
In Section 9.7, we encode the source sequence input information in the recurrent unit state and
then pass it to the decoder to generate the target sequence. A token in the target sequence may
closely relate to one or more tokens in the source sequence, instead of the whole source sequence.
For example, when translating “Hello world.” to “Bonjour le monde.”, “Bonjour” maps to “Hello”
and “monde” maps to “world”. In the seq2seq model, the decoder may implicitly select the corre-
sponding information from the state passed by the encoder. The attention mechanism, however,
makes this selection explicit.
Attention is a generalized pooling method with bias alignment over inputs. The core component
in the attention mechanism is the attention layer, or called attention for simplicity. An input of the
attention layer is called a query. For a query, attention returns an output based on the memory—a
set of key-value pairs encoded in the attention layer. To be more specific, assume that the memory
contains n key-value pairs, (k1 , v1 ), . . . , (kn , vn ), with ki ∈ Rdk , vi ∈ Rdv . Given a query q ∈ Rdq ,
the attention layer returns an output o ∈ Rdv with the same shape as the value.
381
Fig. 10.1.1: The attention layer returns an output based on the input query and its memory.
The full process of attention mechanism is expressed in Fig. 10.1.2. To compute the output of
attention, we first use a score function α that measures the similarity between the query and the
key. So for each key k1 , . . . , kn , we compute the scores a1 , . . . , an by
ai = α(q, ki ). (10.1.1)
Next we use softmax to obtain the attention weights, i.e.,

exp(ai )
b = softmax(a) , where bi = ∑ , b = [b1 , . . . , bn ]T . (10.1.2)
j exp(a j )
Finally, the output is a weighted sum of the values:

∑
n
o= bi vi . (10.1.3)
i=1
Fig. 10.1.2: The attention output is a weighted sum of the values.
Different choices of the score function lead to different attention layers. Below, we introduce two
commonly used attention layers. Before diving into the implementation, we first express two op-
erators to get you up and running: a masked version of the softmax operator masked_softmax and
a specialized dot operator batch_dot.
382 Chapter 10. Attention Mechanisms

import math
npx.set_np()
The masked softmax takes a 3-dimensional input and enables us to filter out some elements by
specifying a valid length for the last dimension. (Refer to Section 9.5 for the definition of a valid
length.) As a result, any value outside the valid length will be masked as 0. Let us implement the
masked_softmax function.
#@save
def masked_softmax(X, valid_len):
"""Perform softmax by filtering out some elements."""
# X: 3-D tensor, valid_len: 1-D or 2-D tensor
if valid_len is None:
return npx.softmax(X)
else:
shape = X.shape
if valid_len.ndim == 1:
valid_len = valid_len.repeat(shape[1], axis=0)
else:
valid_len = valid_len.reshape(-1)
# Fill masked elements with a large negative, whose exp is 0
X = npx.sequence_mask(X.reshape(-1, shape[-1]), valid_len, True,
axis=1, value=-1e6)
return npx.softmax(X).reshape(shape)
To illustrate how this function works, we construct two 2 × 4 matrices as the input. In addition, we
specify that the valid length equals to 2 for the first example, and 3 for the second example. Then,
as we can see from the following outputs, the values outside valid lengths are masked as zero.
masked_softmax(np.random.uniform(size=(2, 2, 4)), np.array([2, 3]))
array([[[0.488994 , 0.511006 , 0. , 0. ],
[0.43654838, 0.56345165, 0. , 0. ]],
[[0.28817102, 0.3519408 , 0.3598882 , 0. ],

[0.29034293, 0.25239873, 0.45725834, 0. ]]])
Moreover, the second operator batch_dot takes two inputs X and Y with shapes (b, n, m) and
(b, m, k), respectively, and returns an output with shape (b, n, k). To be specific, it computes b
dot products for i = {1, . . . , b}, i.e.,
Z[i, :, :] = X[i, :, :]Y [i, :, :]. (10.1.4)
npx.batch_dot(np.ones((2, 1, 3)), np.ones((2, 3, 2)))
array([[[3., 3.]],
[[3., 3.]]])
10.1. Attention Mechanisms 383

10.1.1 Dot Product Attention
Equipped with the above two operators: masked_softmax and batch_dot, let us dive into the details
of two widely used attention layers. The first one is the dot product attention: it assumes that the
query has the same dimension as the keys, namely q, ki ∈ Rd for all i. The dot product attention √
computes the scores by a dot product between the query and a key, which is then divided by d
to minimize the unrelated influence of the dimension d on the scores. In other words,
√
α(q, k) = ⟨q, k⟩/ d. (10.1.5)
Beyond the single-dimensional queries and keys, we can always generalize them to multi-
dimensional queries and keys. Assume that Q ∈ Rm×d contains m queries and K ∈ Rn×d has
all the n keys. We can compute all mn scores by
√
α(Q, K) = QK⊤ / d. (10.1.6)
With (10.1.6), we can implement the dot product attention layer DotProductAttention that sup-
ports a batch of queries and key-value pairs. In addition, for regularization we also use a dropout
layer.
#@save
class DotProductAttention(nn.Block):
def __init__(self, dropout, **kwargs):
super(DotProductAttention, self).__init__(**kwargs)
self.dropout = nn.Dropout(dropout)
# `query`: (`batch_size`, #queries, `d`)

# `key`: (`batch_size`, #kv_pairs, `d`)
# `value`: (`batch_size`, #kv_pairs, `dim_v`)
# `valid_len`: either (`batch_size`, ) or (`batch_size`, xx)
def forward(self, query, key, value, valid_len=None):
d = query.shape[-1]
# Set transpose_b=True to swap the last two dimensions of key
scores = npx.batch_dot(query, key, transpose_b=True) / math.sqrt(d)
attention_weights = self.dropout(masked_softmax(scores, valid_len))
return npx.batch_dot(attention_weights, value)
Let us test the class DotProductAttention in a toy example. First, create two batches, where each
batch has one query and 10 key-value pairs. Via the valid_len argument, we specify that we will
check the first 2 key-value pairs for the first batch and 6 for the second one. Therefore, even though
both batches have the same query and key-value pairs, we obtain different outputs.
atten = DotProductAttention(dropout=0.5)
atten.initialize()
keys = np.ones((2, 10, 2))
values = np.arange(40).reshape(1, 10, 4).repeat(2, axis=0)
atten(np.ones((2, 1, 2)), keys, values, np.array([2, 6]))
array([[[ 2. , 3. , 4. , 5. ]],
[[10. , 11. , 12.000001, 13. ]]])
As we can see above, dot product attention simply multiplies the query and key together, and
hopes to derive their similarities from there. Whereas, the query and key may not be of the same
dimension. To address such an issue, we may resort to the MLP attention.

10.1.2 MLP Attention
In MLP attention, we project both query and keys into Rh by learnable weights parameters. Assume
that the learnable weights are Wk ∈ Rh×dk , Wq ∈ Rh×dq , and v ∈ Rh . Then the score function is
defined by
α(k, q) = v⊤ tanh(Wk k + Wq q). (10.1.7)
Intuitively, you can imagine Wk k + Wq q as concatenating the key and value in the feature dimen-
sion and feeding them to a single hidden layer perceptron with hidden layer size h and output
layer size 1. In this hidden layer, the activation function is tanh and no bias is applied. Now let us
implement the MLP attention.
#@save
class MLPAttention(nn.Block):
def __init__(self, units, dropout, **kwargs):
super(MLPAttention, self).__init__(**kwargs)
# Use flatten=False to keep query's and key's 3-D shapes
self.W_k = nn.Dense(units, use_bias=False, flatten=False)
self.W_q = nn.Dense(units, use_bias=False, flatten=False)
self.v = nn.Dense(1, use_bias=False, flatten=False)
def forward(self, query, key, value, valid_len):

query, key = self.W_q(query), self.W_k(key)
# Expand query to (`batch_size`, #queries, 1, units), and key to
# (`batch_size`, 1, #kv_pairs, units). Then plus them with broadcast
features = np.expand_dims(query, axis=2) + np.expand_dims(key, axis=1)
features = np.tanh(features)
scores = np.squeeze(self.v(features), axis=-1)
attention_weights = self.dropout(masked_softmax(scores, valid_len))
return npx.batch_dot(attention_weights, value)
To test the above MLPAttention class, we use the same inputs as in the previous toy example. As
we can see below, despite MLPAttention containing an additional MLP model, we obtain the same
outputs as for DotProductAttention.
atten = MLPAttention(units=8, dropout=0.1)

atten.initialize()
atten(np.ones((2, 1, 2)), keys, values, np.array([2, 6]))
array([[[ 2. , 3. , 4. , 5. ]],
[[10. , 11. , 12.000001, 13. ]]])
10.1. Attention Mechanisms 385

Summary
• An attention layer explicitly selects related information.

• An attention layerʼs memory consists of key-value pairs, so its output is close to the values
whose keys are similar to the queries.
• Two commonly used attention models are dot product attention and MLP attention.
Exercises
1. What are the advantages and disadvantages for dot product attention and MLP attention,
respectively?
Discussions127
10.2 Sequence to Sequence with Attention Mechanisms
In this section, we add the attention mechanism to the sequence to sequence (seq2seq) model as
introduced in Section 9.7 to explicitly aggregate states with weights. Fig. 10.2.1 shows the model
architecture for encoding and decoding at the timestep t. Here, the memory of the attention layer
consists of all the information that the encoder has seen—the encoder output at each timestep.
During the decoding, the decoder output from the previous timestep t − 1 is used as the query.
The output of the attention model is viewed as the context information, and such context is con-
catenated with the decoder input Dt . Finally, we feed the concatenation into the decoder.
Fig. 10.2.1: The second timestep in decoding for the sequence to sequence model with attention
mechanism.
To illustrate the overall architecture of seq2seq with attention model, the layer structure of its
encoder and decoder is shown in Fig. 10.2.2.
127

Fig. 10.2.2: The layers in the sequence to sequence model with attention mechanism.

from mxnet.gluon import rnn, nn
npx.set_np()
10.2.1 Decoder
Since the encoder of seq2seq with attention mechanisms is the same as Seq2SeqEncoder in Section
9.7, we will just focus on the decoder. We add an MLP attention layer (MLPAttention) which has
the same hidden size as the LSTM layer in the decoder. Then we initialize the state of the decoder
by passing three items from the encoder:
• the encoder outputs of all timesteps: they are used as the attention layerʼs memory with
identical keys and values;
• the hidden state of the encoderʼs final timestep: it is used as the initial decoderʼs hidden
state;
• the encoder valid length: so the attention layer will not consider the padding tokens within
the encoder outputs.
At each timestep of the decoding, we use the hidden state of the decoderʼs last RNN layer as the
query for the attention layer. The attention modelʼs output is then concatenated with the input em-
bedding vector to feed into the RNN layer. Although the RNN layer hidden state also contains his-
tory information from decoder, the attention output explicitly selects the encoder outputs based
on enc_valid_len, so that the attention output suspends other irrelevant information.
Let us implement the Seq2SeqAttentionDecoder, and see how it differs from the decoder in
seq2seq from Section 9.7.2.
class Seq2SeqAttentionDecoder(d2l.Decoder):
super(Seq2SeqAttentionDecoder, self).__init__(**kwargs)
self.attention_cell = d2l.MLPAttention(num_hiddens, dropout)
10.2. Sequence to Sequence with Attention Mechanisms 387

def init_state(self, enc_outputs, enc_valid_len, *args):

outputs, hidden_state = enc_outputs
# Transpose òutputs` to (`batch_size`, `seq_len`, `num_hiddens`)
return (outputs.swapaxes(0, 1), hidden_state, enc_valid_len)

enc_outputs, hidden_state, enc_valid_len = state
X = self.embedding(X).swapaxes(0, 1)
outputs = []
for x in X:
# `query` shape: (`batch_size`, 1, `num_hiddens`)
query = np.expand_dims(hidden_state[0][-1], axis=1)
# `context` has same shape as `query`
context = self.attention_cell(
query, enc_outputs, enc_outputs, enc_valid_len)
# Concatenate on the feature dimension
x = np.concatenate((context, np.expand_dims(x, axis=1)), axis=-1)
# Reshape `x` to (1, `batch_size`, èmbed_size` + `num_hiddens`)
out, hidden_state = self.rnn(x.swapaxes(0, 1), hidden_state)
outputs.append(out)
outputs = self.dense(np.concatenate(outputs, axis=0))
return outputs.swapaxes(0, 1), [enc_outputs, hidden_state,
enc_valid_len]
Now we can test the seq2seq with attention model. To be consistent with the model without atten-
tion in Section 9.7, we use the same hyperparameters for vocab_size, embed_size, num_hiddens,
and num_layers. As a result, we get the same decoder output shape, but the state structure is
changed.
encoder = d2l.Seq2SeqEncoder(vocab_size=10, embed_size=8,

decoder = Seq2SeqAttentionDecoder(vocab_size=10, embed_size=8,
decoder.initialize()
X = np.zeros((4, 7))
state = decoder.init_state(encoder(X), None)
out, state = decoder(X, state)
out.shape, len(state), state[0].shape, len(state[1]), state[1][0].shape
((4, 7, 10), 3, (4, 7, 16), 2, (2, 4, 16))

10.2.2 Training
Similar to Section 9.7.4, we try a toy model by applying the same training hyperparameters and
the same training loss. As we can see from the result, since the sequences in the training dataset
are relatively short, the additional attention layer does not lead to a significant improvement. Due
to the computational overhead of both the encoderʼs and the decoderʼs attention layers, this model
is much slower than the seq2seq model without attention.
embed_size, num_hiddens, num_layers, dropout = 32, 32, 2, 0.0


encoder = d2l.Seq2SeqEncoder(
len(src_vocab), embed_size, num_hiddens, num_layers, dropout)
decoder = Seq2SeqAttentionDecoder(
len(tgt_vocab), embed_size, num_hiddens, num_layers, dropout)
d2l.train_s2s_ch9(model, train_iter, lr, num_epochs, device)
Last, we predict several sample examples.
print(sentence + ' => ' + d2l.predict_s2s_ch9(
Go . => <unk> !
Wow ! => <unk> !
I'm OK . => je vais bien .
I won ! => j'ai gagné !
10.2. Sequence to Sequence with Attention Mechanisms 389

Summary
• The seq2seq model with attention adds an additional attention layer to the model without
attention.
• The decoder of the seq2seq with attention model passes three items from the encoder: the
encoder outputs of all timesteps, the hidden state of the encoderʼs final timestep, and the
encoder valid length.
Exercises
1. Compare Seq2SeqAttentionDecoder and Seq2seqDecoder by using the same parameters and

checking their losses.
2. Can you think of any use cases where Seq2SeqAttentionDecoder will outperform
Seq2seqDecoder?
Discussions128
10.3 Transformer
In previous chapters, we have covered major neural network architectures such as convolution
neural networks (CNNs) and recurrent neural networks (RNNs). Let us recap their pros and cons:
• CNNs are easy to parallelize at a layer but cannot capture the variable-length sequential de-
pendency very well.
• RNNs are able to capture the long-range, variable-length sequential information, but suffer
from inability to parallelize within a sequence.
To combine the advantages from both CNNs and RNNs, (Vaswani et al., 2017) designed a novel
architecture using the attention mechanism. This architecture, which is called as Transformer,
achieves parallelization by capturing recurrence sequence with attention and at the same time
encodes each itemʼs position in the sequence. As a result, Transformer leads to a compatible model
with significantly shorter training time.
Similar to the seq2seq model in Section 9.7, Transformer is also based on the encoder-decoder
architecture. However, Transformer differs to the former by replacing the recurrent layers in
seq2seq with multi-head attention layers, incorporating the position-wise information through po-
sition encoding, and applying layer normalization. We compare Transformer and seq2seq side-by-
side in Fig. 10.3.1.
Overall, these two models are similar to each other: the source sequence embeddings are fed
into n repeated blocks. The outputs of the last block are then used as attention memory for the
decoder. The target sequence embeddings are similarly fed into n repeated blocks in the decoder,
and the final outputs are obtained by applying a dense layer with vocabulary size to the last blockʼs
outputs.
128

Dense
Transformer Add & norm
Position-
State wise FFN
Add & norm Add & norm
Position- Muti-head xn
wise FFN attention
Seq2seq with Attention nx

Add & norm Add & norm
Encoder Decoder Multi-head Masked

attention multi-head
Attention Dense attention
nx Recurrent layer Recurrent layer xn Positional Positional

+ +
encoding encoding
Embedding Embedding Embedding Embedding
Sources Targets Sources Targets
Fig. 10.3.1: The Transformer architecture.
On the flip side, Transformer differs from the seq2seq with attention model in the following:
1. Transformer block: a recurrent layer in seq2seq is replaced by a Transformer block. This
block contains a multi-head attention layer and a position-wise feed-forward network with two
layers for the encoder. For the decoder, another multi-head attention layer is used to take
the encoder state.
2. Add and norm: the inputs and outputs of both the multi-head attention layer or the position-
wise feed-forward network, are processed by two “add and norm” layer that contains a resid-
ual structure and a layer normalization layer.
3. Position encoding: since the self-attention layer does not distinguish the item order in a se-
quence, a positional encoding layer is used to add sequential information into each sequence
item.
In the rest of this section, we will equip you with each new component introduced by Transformer,
and get you up and running to construct a machine translation model.

import math
npx.set_np()
10.3. Transformer 391

10.3.1 Multi-Head Attention
Before the discussion of the multi-head attention layer, let us quick express the self-attention ar-
chitecture. The self-attention model is a normal attention model, with its query, its key, and its
value being copied exactly the same from each item of the sequential inputs. As we illustrate in
Fig. 10.3.2, self-attention outputs a same-length sequential output for each input item. Compared
with a recurrent layer, output items of a self-attention layer can be computed in parallel and,
therefore, it is easy to obtain a highly-efficient implementation.
Fig. 10.3.2: Self-attention architecture.
The multi-head attention layer consists of h parallel self-attention layers, each one is called a head.
For each head, before feeding into the attention layer, we project the queries, keys, and values with
three dense layers with hidden sizes pq , pk , and pv , respectively. The outputs of these h attention
heads are concatenated and then processed by a final dense layer.
Fig. 10.3.3: Multi-head attention
Assume that the dimension for a query, a key, and a value are dq , dk , and dv , respectively. Then,
for each head i = 1, . . . , h, we can train learnable parameters Wq ∈ Rpq ×dq , Wk ∈ Rpk ×dk , and
(i) (i)
Wv ∈ Rpv ×dv . Therefore, the output for each head is

(i)
(i)
o(i) = attention(W(i) (i)
q q, Wk k, Wv v), (10.3.1)
where attention can be any attention layer, such as the DotProductAttention and MLPAttention as
we introduced in Section 10.1.
After that, the output with length pv from each of the h attention heads are concatenated to be an
output of length hpv , which is then passed the final dense layer with do hidden units. The weights

of this dense layer can be denoted by Wo ∈ Rdo ×hpv . As a result, the multi-head attention output
will be
 (1) 
o
 .. 
o = Wo  .  . (10.3.2)
o(h)
Now we can implement the multi-head attention. Assume that the multi-head attention contain
the number heads num_heads = h, the hidden size num_hiddens = pq = pk = pv are the same for
the query, key, and value dense layers. In addition, since the multi-head attention keeps the same
dimensionality between its input and its output, we have the output feature size do = num_hiddens
as well.
#@save
class MultiHeadAttention(nn.Block):
def __init__(self, num_hiddens, num_heads, dropout, use_bias=False, **kwargs):
super(MultiHeadAttention, self).__init__(**kwargs)
self.num_heads = num_heads
self.attention = d2l.DotProductAttention(dropout)
self.W_q = nn.Dense(num_hiddens, use_bias=use_bias, flatten=False)
self.W_k = nn.Dense(num_hiddens, use_bias=use_bias, flatten=False)
self.W_v = nn.Dense(num_hiddens, use_bias=use_bias, flatten=False)
self.W_o = nn.Dense(num_hiddens, use_bias=use_bias, flatten=False)
def forward(self, query, key, value, valid_len):

# For self-attention, `query`, `key`, and `value` shape:
# (`batch_size`, `seq_len`, `dim`), where `seq_len` is the length of
# input sequence. `valid_len` shape is either (`batch_size`, ) or
# (`batch_size`, `seq_len`).
# Project and transpose `query`, `key`, and `value` from

# (`batch_size`, `seq_len`, `num_hiddens`) to
# (`batch_size` * `num_heads`, `seq_len`, `num_hiddens` / `num_heads`)
query = transpose_qkv(self.W_q(query), self.num_heads)
key = transpose_qkv(self.W_k(key), self.num_heads)
value = transpose_qkv(self.W_v(value), self.num_heads)
if valid_len is not None:

# Copy `valid_len` by `num_heads` times
if valid_len.ndim == 1:
valid_len = np.tile(valid_len, self.num_heads)
else:
valid_len = np.tile(valid_len, (self.num_heads, 1))
# For self-attention, òutput` shape:

output = self.attention(query, key, value, valid_len)
# òutput_concat` shape: (`batch_size`, `seq_len`, `num_hiddens`)

output_concat = transpose_output(output, self.num_heads)
return self.W_o(output_concat)
Here are the definitions of the transpose functions transpose_qkv and transpose_output, which
are the inverse of each other.

#@save
def transpose_qkv(X, num_heads):
# Input `X` shape: (`batch_size`, `seq_len`, `num_hiddens`).
# Output `X` shape:
# (`batch_size`, `seq_len`, `num_heads`, `num_hiddens` / `num_heads`)
X = X.reshape(X.shape[0], X.shape[1], num_heads, -1)
# `X` shape:
# (`batch_size`, `num_heads`, `seq_len`, `num_hiddens` / `num_heads`)
X = X.transpose(0, 2, 1, 3)
# òutput` shape:
output = X.reshape(-1, X.shape[2], X.shape[3])
return output
#@save
def transpose_output(X, num_heads):
# A reversed version of `transpose_qkv`
X = X.reshape(-1, num_heads, X.shape[1], X.shape[2])
X = X.transpose(0, 2, 1, 3)
return X.reshape(X.shape[0], X.shape[1], -1)
Let us test the MultiHeadAttention model in the a toy example. Create a multi-head attention
with the hidden size do = 100, the output will share the same batch size and sequence length as
the input, but the last dimension will be equal to the num_hiddens = 100.
cell = MultiHeadAttention(90, 9, 0.5)

cell.initialize()
X = np.ones((2, 4, 5))
valid_len = np.array([2, 3])
cell(X, X, X, valid_len).shape
(2, 4, 90)
10.3.2 Position-wise Feed-Forward Networks
Another key component in the Transformer block is called position-wise feed-forward network
(FFN). It accepts a 3-dimensional input with shape (batch size, sequence length, feature size). The
position-wise FFN consists of two dense layers that applies to the last dimension. Since the same
two dense layers are used for each position item in the sequence, we referred to it as position-wise.
Indeed, it is equivalent to applying two 1 × 1 convolution layers.
Below, the PositionWiseFFN shows how to implement a position-wise FFN with two dense layers
of hidden size ffn_num_hiddens and pw_num_outputs, respectively.
#@save
class PositionWiseFFN(nn.Block):
def __init__(self, ffn_num_hiddens, pw_num_outputs, **kwargs):
super(PositionWiseFFN, self).__init__(**kwargs)
self.dense1 = nn.Dense(ffn_num_hiddens, flatten=False,

activation='relu')
self.dense2 = nn.Dense(pw_num_outputs, flatten=False)

return self.dense2(self.dense1(X))
Similar to the multi-head attention, the position-wise feed-forward network will only change the
last dimension size of the input—the feature dimension. In addition, if two items in the input
sequence are identical, the according outputs will be identical as well.
ffn = PositionWiseFFN(4, 8)
ffn.initialize()
ffn(np.ones((2, 3, 4)))[0]
array([[ 9.15348239e-04, -7.27669394e-04, 1.14063594e-04,

-8.76279722e-04, -1.02867256e-03, 8.02748313e-04,
-4.53725770e-05, 2.15598906e-04],
[ 9.15348239e-04, -7.27669394e-04, 1.14063594e-04,
-8.76279722e-04, -1.02867256e-03, 8.02748313e-04,
-4.53725770e-05, 2.15598906e-04],
[ 9.15348239e-04, -7.27669394e-04, 1.14063594e-04,
-8.76279722e-04, -1.02867256e-03, 8.02748313e-04,
-4.53725770e-05, 2.15598906e-04]])
10.3.3 Add and Norm
Besides the above two components in the Transformer block, the “add and norm” within the block
also plays a key role to connect the inputs and outputs of other layers smoothly. To explain, we
add a layer that contains a residual structure and a layer normalization after both the multi-head
attention layer and the position-wise FFN network. Layer normalization is similar to batch normal-
ization in Section 7.5. One difference is that the mean and variances for the layer normalization
are calculated along the last dimension, e.g X.mean(axis=-1) instead of the first batch dimension,
e.g., X.mean(axis=0). Layer normalization prevents the range of values in the layers from chang-
ing too much, which allows faster training and better generalization ability.
MXNet has both LayerNorm and BatchNorm implemented within the nn block. Let us call both of
them and see the difference in the example below.
layer = nn.LayerNorm()
layer.initialize()
batch = nn.BatchNorm()
batch.initialize()
X = np.array([[1, 2], [2, 3]])
# Compute mean and variance from `X` in the training mode
print('layer norm:', layer(X), '\nbatch norm:', batch(X))
layer norm: [[-0.99998 0.99998]

[-0.99998 0.99998]]

batch norm: [[-0.99998 -0.99998]
[ 0.99998 0.99998]]
Now let us implement the connection block AddNorm together. AddNorm accepts two inputs X and
Y . We can deem X as the original input in the residual network, and Y as the outputs from either
the multi-head attention layer or the position-wise FFN network. In addition, we apply dropout
on Y for regularization.
#@save
class AddNorm(nn.Block):
def __init__(self, dropout, **kwargs):
super(AddNorm, self).__init__(**kwargs)
self.ln = nn.LayerNorm()
def forward(self, X, Y):

return self.ln(self.dropout(Y) + X)
Due to the residual connection, X and Y should have the same shape.
add_norm = AddNorm(0.5)
add_norm.initialize()
add_norm(np.ones((2, 3, 4)), np.ones((2, 3, 4))).shape
(2, 3, 4)
10.3.4 Positional Encoding
Unlike the recurrent layer, both the multi-head attention layer and the position-wise feed-forward
network compute the output of each item in the sequence independently. This feature enables
us to parallelize the computation, but it fails to model the sequential information for a given se-
quence. To better capture the sequential information, the Transformer model uses the positional
encoding to maintain the positional information of the input sequence.
To explain, assume that X ∈ Rl×d is the embedding of an example, where l is the sequence length
and d is the embedding size. This positional encoding layer encodes Xʼs position P ∈ Rl×d and
outputs P + X.
The position P is a 2-D matrix, where i refers to the order in the sentence, and j refers to the
position along the embedding vector dimension. In this way, each value in the origin sequence is
then maintained using the equations below:
Pi,2j = sin(i/100002j/d ), (10.3.3)
Pi,2j+1 = cos(i/100002j/d ), (10.3.4)

for i = 0, . . . , l − 1 and j = 0, . . . , ⌊(d − 1)/2⌋.
Fig. 10.3.4 illustrates the positional encoding.

Fig. 10.3.4: Positional encoding.
#@save
class PositionalEncoding(nn.Block):
def __init__(self, num_hiddens, dropout, max_len=1000):
super(PositionalEncoding, self).__init__()
# Create a long enough `P`
self.P = np.zeros((1, max_len, num_hiddens))
X = np.arange(0, max_len).reshape(-1, 1) / np.power(
10000, np.arange(0, num_hiddens, 2) / num_hiddens)
self.P[:, :, 0::2] = np.sin(X)
self.P[:, :, 1::2] = np.cos(X)

X = X + self.P[:, :X.shape[1], :].as_in_ctx(X.ctx)
return self.dropout(X)
Now we test the PositionalEncoding class with a toy model for 4 dimensions. As we can see, the
4th dimension has the same frequency as the 5th but with different offset (i.e. phase) because one
is produced by a sine function and the other is produced by a cosine function. The 6th and 7th
dimensions have lower frequency.
pe = PositionalEncoding(20, 0)
pe.initialize()
Y = pe(np.zeros((1, 100, 20)))
d2l.plot(np.arange(100), Y[0, :, 4:8].T, figsize=(6, 2.5),
legend=["dim %d" % p for p in [4, 5, 6, 7]])

10.3.5 Encoder
Armed with all the essential components of Transformer, let us first build a Transformer en-
coder block. This encoder contains a multi-head attention layer, a position-wise feed-forward
network, and two “add and norm” connection blocks. As shown in the code, for both of the atten-
tion model and the positional FFN model in the EncoderBlock, their outputsʼ dimension are equal
to the num_hiddens. This is due to the nature of the residual block, as we need to add these outputs
back to the original value during “add and norm”.
#@save
class EncoderBlock(nn.Block):
def __init__(self, num_hiddens, ffn_num_hiddens, num_heads, dropout,
use_bias=False, **kwargs):
super(EncoderBlock, self).__init__(**kwargs)
self.attention = MultiHeadAttention(num_hiddens, num_heads, dropout,
use_bias)
self.addnorm1 = AddNorm(dropout)
self.ffn = PositionWiseFFN(ffn_num_hiddens, num_hiddens)
def forward(self, X, valid_len):

Y = self.addnorm1(X, self.attention(X, X, X, valid_len))
return self.addnorm2(Y, self.ffn(Y))
Due to the residual connections, this block will not change the input shape. It means that the
num_hiddens argument should be equal to the input size of the last dimension. In our toy example
below, num_hiddens = 24, ffn_num_hiddens = 48, num_heads = 8, and dropout = 0.5.
X = np.ones((2, 100, 24))

encoder_blk = EncoderBlock(24, 48, 8, 0.5)
encoder_blk.initialize()
encoder_blk(X, valid_len).shape
(2, 100, 24)
Now it comes to the implementation of the entire Transformer encoder. With the Transformer
encoder, n blocks of EncoderBlock stack up one after another. Because of the residual connection,
the embedding layer size d√
is same as the Transformer block output size. Also note that we multiply
the embedding output by d to prevent its values from being too small.
#@save
class TransformerEncoder(d2l.Encoder):
def __init__(self, vocab_size, num_hiddens, ffn_num_hiddens,
num_heads, num_layers, dropout, use_bias=False, **kwargs):
super(TransformerEncoder, self).__init__(**kwargs)
self.num_hiddens = num_hiddens
self.embedding = nn.Embedding(vocab_size, num_hiddens)
self.pos_encoding = PositionalEncoding(num_hiddens, dropout)
self.blks = nn.Sequential()
for _ in range(num_layers):
self.blks.add(
EncoderBlock(num_hiddens, ffn_num_hiddens, num_heads, dropout,
use_bias))

def forward(self, X, valid_len, *args):

X = self.pos_encoding(self.embedding(X) * math.sqrt(self.num_hiddens))
for blk in self.blks:
X = blk(X, valid_len)
return X
Let us create an encoder with two stacked Transformer encoder blocks, whose hyperparameters
are the same as before. Similar to the previous toy exampleʼs parameters, we add two more pa-
rameters vocab_size to be 200 and num_layers to be 2 here.
encoder = TransformerEncoder(200, 24, 48, 8, 2, 0.5)

encoder(np.ones((2, 100)), valid_len).shape
(2, 100, 24)
10.3.6 Decoder
The Transformer decoder block looks similar to the Transformer encoder block. However, be-
sides the two sub-layers (the multi-head attention layer and the positional encoding network), the
decoder Transformer block contains a third sub-layer, which applies multi-head attention on the
output of the encoder stack. Similar to the Transformer encoder block, the Transformer decoder
block employs “add and norm”, i.e., the residual connections and the layer normalization to con-
nect each of the sub-layers.
To be specific, at timestep t, assume that xt is the current input, i.e., the query. As illustrated in
Fig. 10.3.5, the keys and values of the self-attention layer consist of the current query with all the
past queries x1 , . . . , xt−1 .
Fig. 10.3.5: Predict at timestep t for a self-attention layer.
During training, the output for the t-query could observe all the previous key-value pairs. It results
in an different behavior from prediction. Thus, during prediction we can eliminate the unneces-
sary information by specifying the valid length to be t for the tth query.
class DecoderBlock(nn.Block):
# ì` means it is the i-th block in the decoder
def __init__(self, num_hiddens, ffn_num_hiddens, num_heads,

dropout, i, **kwargs):
super(DecoderBlock, self).__init__(**kwargs)
self.i = i
self.attention1 = MultiHeadAttention(num_hiddens, num_heads, dropout)
self.attention2 = MultiHeadAttention(num_hiddens, num_heads, dropout)
self.ffn = PositionWiseFFN(ffn_num_hiddens, num_hiddens)

enc_outputs, enc_valid_len = state[0], state[1]
# `state[2][i]` contains the past queries for this block
if state[2][self.i] is None:
key_values = X
else:
key_values = np.concatenate((state[2][self.i], X), axis=1)
state[2][self.i] = key_values
batch_size, seq_len, _ = X.shape
# Shape: (batch_size, seq_len), the values in the j-th column
# are j+1
valid_len = np.tile(np.arange(1, seq_len + 1, ctx=X.ctx),
(batch_size, 1))
else:
valid_len = None
X2 = self.attention1(X, key_values, key_values, valid_len)

Y = self.addnorm1(X, X2)
Y2 = self.attention2(Y, enc_outputs, enc_outputs, enc_valid_len)
Z = self.addnorm2(Y, Y2)
return self.addnorm3(Z, self.ffn(Z)), state
Similar to the Transformer encoder block, num_hiddens should be equal to the last dimension size
of X.
decoder_blk = DecoderBlock(24, 48, 8, 0.5, 0)

decoder_blk.initialize()
X = np.ones((2, 100, 24))
state = [encoder_blk(X, valid_len), valid_len, [None]]
decoder_blk(X, state)[0].shape
(2, 100, 24)
The construction of the entire Transformer decoder is identical to the Transformer encoder, ex-
cept for the additional dense layer to obtain the output confidence scores.
Let us implement the Transformer decoder TransformerDecoder. Besides the regular hyperpa-
rameters such as the vocab_size and num_hiddens, the Transformer decoder also needs the Trans-
former encoderʼs outputs enc_outputs and env_valid_len.
class TransformerDecoder(d2l.Decoder):
def __init__(self, vocab_size, num_hiddens, ffn_num_hiddens,

num_heads, num_layers, dropout, **kwargs):
super(TransformerDecoder, self).__init__(**kwargs)
self.num_hiddens = num_hiddens
self.num_layers = num_layers
self.embedding = nn.Embedding(vocab_size, num_hiddens)
self.pos_encoding = PositionalEncoding(num_hiddens, dropout)
for i in range(num_layers):
self.blks.add(
DecoderBlock(num_hiddens, ffn_num_hiddens, num_heads,
dropout, i))
def init_state(self, enc_outputs, env_valid_len, *args):

return [enc_outputs, env_valid_len, [None]*self.num_layers]

X = self.pos_encoding(self.embedding(X) * math.sqrt(self.num_hiddens))
X, state = blk(X, state)
return self.dense(X), state
10.3.7 Training
Finally, we can build an encoder-decoder model with the Transformer architecture. Similar to
the seq2seq with attention model in Section 10.2, we use the following hyperparameters: two
Transformer blocks with both the embedding size and the block output size to be 32. In addition,
we use 4 heads, and set the hidden size to be twice larger than the output size.
num_hiddens, num_layers, dropout, batch_size, num_steps = 32, 2, 0.0, 64, 10

ffn_num_hiddens, num_heads = 64, 4
encoder = TransformerEncoder(
len(src_vocab), num_hiddens, ffn_num_hiddens, num_heads, num_layers,
dropout)
decoder = TransformerDecoder(
len(src_vocab), num_hiddens, ffn_num_hiddens, num_heads, num_layers,
dropout)
d2l.train_s2s_ch9(model, train_iter, lr, num_epochs, device)

As we can see from the training time and accuracy, compared with the seq2seq model with atten-
tion model, Transformer runs faster per epoch, and converges faster at the beginning.
We can use the trained Transformer to translate some simple sentences.
print(sentence + ' => ' + d2l.predict_s2s_ch9(
Go . => va <unk> !
Wow ! => <unk> !
I'm OK . => ça va <unk> va <unk> va <unk> tom .
I won ! => je l'ai emporté gagné !
Summary
• The Transformer model is based on the encoder-decoder architecture.

• Multi-head attention layer contains h parallel attention layers.
• Position-wise feed-forward network consists of two dense layers that apply to the last dimen-
sion.
• Layer normalization differs from batch normalization by normalizing along the last dimen-
sion (the feature dimension) instead of the first (batch size) dimension.
• Positional encoding is the only place that adds positional information to the Transformer
model.

Exercises
1. Try a larger size of epochs and compare the loss between seq2seq model and Transformer
model.
2. Can you think of any other benefit of positional encoding?
3. Compare layer normalization and batch normalization, when shall we apply which?
Discussions129
129

11 | Optimization Algorithms
If you read the book in sequence up to this point you already used a number of advanced optimiza-
tion algorithms to train deep learning models. They were the tools that allowed us to continue
updating model parameters and to minimize the value of the loss function, as evaluated on the
training set. Indeed, anyone content with treating optimization as a black box device to minimize
objective functions in a simple setting might well content oneself with the knowledge that there
exists an array of incantations of such a procedure (with names such as “Adam”, “NAG”, or “SGD”).
To do well, however, some deeper knowledge is required. Optimization algorithms are important
for deep learning. On one hand, training a complex deep learning model can take hours, days, or
even weeks. The performance of the optimization algorithm directly affects the modelʼs training
efficiency. On the other hand, understanding the principles of different optimization algorithms
and the role of their parameters will enable us to tune the hyperparameters in a targeted manner
to improve the performance of deep learning models.
In this chapter, we explore common deep learning optimization algorithms in depth. Almost all
optimization problems arising in deep learning are nonconvex. Nonetheless, the design and anal-
ysis of algorithms in the context of convex problems has proven to be very instructive. It is for
that reason that this section includes a primer on convex optimization and the proof for a very
simple stochastic gradient descent algorithm on a convex objective function.
11.1 Optimization and Deep Learning
In this section, we will discuss the relationship between optimization and deep learning as well
as the challenges of using optimization in deep learning. For a deep learning problem, we will
usually define a loss function first. Once we have the loss function, we can use an optimization
algorithm in attempt to minimize the loss. In optimization, a loss function is often referred to as
the objective function of the optimization problem. By tradition and convention most optimiza-
tion algorithms are concerned with minimization. If we ever need to maximize an objective there
is a simple solution: just flip the sign on the objective.
405
11.1.1 Optimization and Estimation
Although optimization provides a way to minimize the loss function for deep learning, in essence,
the goals of optimization and deep learning are fundamentally different. The former is primarily
concerned with minimizing an objective whereas the latter is concerned with finding a suitable
model, given a finite amount of data. In Section 4.4, we discussed the difference between these
two goals in detail. For instance, training error and generalization error generally differ: since the
objective function of the optimization algorithm is usually a loss function based on the training
dataset, the goal of optimization is to reduce the training error. However, the goal of statistical
inference (and thus of deep learning) is to reduce the generalization error. To accomplish the
latter we need to pay attention to overfitting in addition to using the optimization algorithm to
reduce the training error. We begin by importing a few libraries for this chapter.
%matplotlib inline
from mpl_toolkits import mplot3d
npx.set_np()
Next we define two functions, the expected function f and the empirical function g, to illustrate
this issue. Here the g is less smooth than f since we have only a finite amount of data.
def f(x): return x * np.cos(np.pi * x)

def g(x): return f(x) + 0.2 * np.cos(5 * np.pi * x)
The graph below illustrates that the minimum of the training error may be at a different location
than the minimum of the expected error (or of the test error).
def annotate(text, xy, xytext): #@save

d2l.plt.gca().annotate(text, xy=xy, xytext=xytext,
arrowprops=dict(arrowstyle='->'))
x = np.arange(0.5, 1.5, 0.01)

d2l.set_figsize((4.5, 2.5))
d2l.plot(x, [f(x), g(x)], 'x', 'risk')
annotate('empirical risk', (1.0, -1.2), (0.5, -1.1))
annotate('expected risk', (1.1, -1.05), (0.95, -0.5))
406 Chapter 11. Optimization Algorithms

11.1.2 Optimization Challenges in Deep Learning
In this chapter, we are going to focus specifically on the performance of the optimization algorithm
in minimizing the objective function, rather than a modelʼs generalization error. In Section 3.1
we distinguished between analytical solutions and numerical solutions in optimization problems.
In deep learning, most objective functions are complicated and do not have analytical solutions.
Instead, we must use numerical optimization algorithms. The optimization algorithms below all
fall into this category.
There are many challenges in deep learning optimization. Some of the most vexing ones are local
minima, saddle points and vanishing gradients. Let us have a look at a few of them.
Local Minima
For the objective function f (x), if the value of f (x) at x is smaller than the values of f (x) at any
other points in the vicinity of x, then f (x) could be a local minimum. If the value of f (x) at x is
the minimum of the objective function over the entire domain, then f (x) is the global minimum.
For example, given the function
f (x) = x · cos(πx) for − 1.0 ≤ x ≤ 2.0, (11.1.1)
we can approximate the local minimum and global minimum of this function.
x = np.arange(-1.0, 2.0, 0.01)

d2l.plot(x, [f(x), ], 'x', 'f(x)')
annotate('local minimum', (-0.3, -0.25), (-0.77, -1.0))
annotate('global minimum', (1.1, -0.95), (0.6, 0.8))
The objective function of deep learning models usually has many local optima. When the nu-
merical solution of an optimization problem is near the local optimum, the numerical solution
obtained by the final iteration may only minimize the objective function locally, rather than glob-
ally, as the gradient of the objective functionʼs solutions approaches or becomes zero. Only some
degree of noise might knock the parameter out of the local minimum. In fact, this is one of the
beneficial properties of stochastic gradient descent where the natural variation of gradients over
minibatches is able to dislodge the parameters from local minima.
11.1. Optimization and Deep Learning 407

Saddle Points
Besides local minima, saddle points are another reason for gradients to vanish. A saddle point130
is any location where all gradients of a function vanish but which is neither a global nor a local
minimum. Consider the function f (x) = x3 . Its first and second derivative vanish for x = 0.
Optimization might stall at the point, even though it is not a minimum.
x = np.arange(-2.0, 2.0, 0.01)

d2l.plot(x, [x**3], 'x', 'f(x)')
annotate('saddle point', (0, -0.2), (-0.52, -5.0))
Saddle points in higher dimensions are even more insidious, as the example below shows. Con-
sider the function f (x, y) = x2 − y 2 . It has its saddle point at (0, 0). This is a maximum with
respect to y and a minimum with respect to x. Moreover, it looks like a saddle, which is where this
mathematical property got its name.
x, y = np.meshgrid(np.linspace(-1.0, 1.0, 101), np.linspace(-1.0, 1.0, 101))

z = x**2 - y**2
ax = d2l.plt.figure().add_subplot(111, projection='3d')
ax.plot_wireframe(x, y, z, **{'rstride': 10, 'cstride': 10})
ax.plot([0], [0], [0], 'rx')
ticks = [-1, 0, 1]
d2l.plt.xticks(ticks)
d2l.plt.yticks(ticks)
ax.set_zticks(ticks)
d2l.plt.xlabel('x')
d2l.plt.ylabel('y');
130
https://en.wikipedia.org/wiki/Saddle_point

We assume that the input of a function is a k-dimensional vector and its output is a scalar, so its
Hessian matrix will have k eigenvalues (refer to Section 18.1). The solution of the function could
be a local minimum, a local maximum, or a saddle point at a position where the function gradient
is zero:
• When the eigenvalues of the functionʼs Hessian matrix at the zero-gradient position are all
positive, we have a local minimum for the function.
• When the eigenvalues of the functionʼs Hessian matrix at the zero-gradient position are all
negative, we have a local maximum for the function.
• When the eigenvalues of the functionʼs Hessian matrix at the zero-gradient position are neg-
ative and positive, we have a saddle point for the function.
For high-dimensional problems the likelihood that at least some of the eigenvalues are negative
is quite high. This makes saddle points more likely than local minima. We will discuss some ex-
ceptions to this situation in the next section when introducing convexity. In short, convex func-
tions are those where the eigenvalues of the Hessian are never negative. Sadly, though, most deep
learning problems do not fall into this category. Nonetheless it is a great tool to study optimization
algorithms.
Vanishing Gradients
Probably the most insidious problem to encounter are vanishing gradients. For instance, assume
that we want to minimize the function f (x) = tanh(x) and we happen to get started at x = 4.
As we can see, the gradient of f is close to nil. More specifically f ′ (x) = 1 − tanh2 (x) and thus
f ′ (4) = 0.0013. Consequently optimization will get stuck for a long time before we make progress.
This turns out to be one of the reasons that training deep learning models was quite tricky prior
to the introduction of the ReLU activation function.
x = np.arange(-2.0, 5.0, 0.01)

d2l.plot(x, [np.tanh(x)], 'x', 'f(x)')
annotate('vanishing gradient', (4, 1), (2, 0.0))
11.1. Optimization and Deep Learning 409

As we saw, optimization for deep learning is full of challenges. Fortunately there exists a robust
range of algorithms that perform well and that are easy to use even for beginners. Furthermore,
it is not really necessary to find the best solution. Local optima or even approximate solutions
thereof are still very useful.
Summary
• Minimizing the training error does not guarantee that we find the best set of parameters to
minimize the expected error.
• The optimization problems may have many local minima.
• The problem may have even more saddle points, as generally the problems are not convex.
• Vanishing gradients can cause optimization to stall. Often a reparameterization of the prob-
lem helps. Good initialization of the parameters can be beneficial, too.
Exercises
1. Consider a simple multilayer perceptron with a single hidden layer of, say, d dimensions in
the hidden layer and a single output. Show that for any local minimum there are at least d!
equivalent solutions that behave identically.
2. Assume that we have a symmetric random matrix M where the entries Mij = Mji are each
drawn from some probability distribution pij . Furthermore assume that pij (x) = pij (−x),
i.e., that the distribution is symmetric (see e.g., (Wigner, 1958) for details).
• Prove that the distribution over eigenvalues is also symmetric. That is, for any eigen-
vector v the probability that the associated eigenvalue λ satisfies P (λ > 0) = P (λ < 0).
• Why does the above not imply P (λ > 0) = 0.5?
3. What other challenges involved in deep learning optimization can you think of?
4. Assume that you want to balance a (real) ball on a (real) saddle.
• Why is this hard?
• Can you exploit this effect also for optimization algorithms?

Discussions131
11.2 Convexity
Convexity plays a vital role in the design of optimization algorithms. This is largely due to the fact
that it is much easier to analyze and test algorithms in this context. In other words, if the algorithm
performs poorly even in the convex setting we should not hope to see great results otherwise.
Furthermore, even though the optimization problems in deep learning are generally nonconvex,
they often exhibit some properties of convex ones near local minima. This can lead to exciting
new optimization variants such as (Izmailov et al., 2018).
11.2.1 Basics
Let us begin with the basics.
Sets
Sets are the basis of convexity. Simply put, a set X in a vector space is convex if for any a, b ∈ X
the line segment connecting a and b is also in X. In mathematical terms this means that for all
λ ∈ [0, 1] we have
λ · a + (1 − λ) · b ∈ X whenever a, b ∈ X. (11.2.1)
This sounds a bit abstract. Consider the picture Fig. 11.2.1. The first set is not convex since there
are line segments that are not contained in it. The other two sets suffer no such problem.
Fig. 11.2.1: Three shapes, the left one is nonconvex, the others are convex
Definitions on their own are not particularly useful unless you can do something with them. In
this case we can look at unions and intersections as shown in Fig. 11.2.2. Assume that X and Y
are convex sets. Then X ∩ Y is also convex. To see this, consider any a, b ∈ X ∩ Y . Since X and Y
are convex, the line segments connecting a and b are contained in both X and Y . Given that, they
also need to be contained in X ∩ Y , thus proving our first theorem.
131
11.2. Convexity 411

Fig. 11.2.2: The intersection between two convex sets is convex
We can strengthen this result with little effort: given convex sets Xi , their intersection ∩i Xi is
convex. To see that the converse is not true, consider two disjoint sets X ∩ Y = ∅. Now pick a ∈ X
and b ∈ Y . The line segment in :numref:fig_nonconvex connecting a and b needs to contain some
part that is neither in X nor Y , since we assumed that X ∩ Y = ∅. Hence the line segment is not
in X ∪ Y either, thus proving that in general unions of convex sets need not be convex.
Fig. 11.2.3: The union of two convex sets need not be convex
Typically the problems in deep learning are defined on convex domains. For instance Rd is a
convex set (after all, the line between any two points in Rd remains in Rd ). In some cases we work
with variables of bounded length, such as balls of radius r as defined by {x|x ∈ Rd and ∥x∥2 ≤ r}.
Functions
Now that we have convex sets we can introduce convex functions f . Given a convex set X a func-
tion defined on it f : X → R is convex if for all x, x′ ∈ X and for all λ ∈ [0, 1] we have
λf (x) + (1 − λ)f (x′ ) ≥ f (λx + (1 − λ)x′ ). (11.2.2)
To illustrate this let us plot a few functions and check which ones satisfy the requirement. We
need to import a few libraries.
%matplotlib inline
npx.set_np()
Let us define a few functions, both convex and nonconvex.

f = lambda x: 0.5 * x**2 # Convex
g = lambda x: np.cos(np.pi * x) # Nonconvex
h = lambda x: np.exp(0.5 * x) # Convex
x, segment = np.arange(-2, 2, 0.01), np.array([-1.5, 1])

_, axes = d2l.plt.subplots(1, 3, figsize=(9, 3))
for ax, func in zip(axes, [f, g, h]):
d2l.plot([x, segment], [func(x), func(segment)], axes=ax)
As expected, the cosine function is nonconvex, whereas the parabola and the exponential function
are. Note that the requirement that X is a convex set is necessary for the condition to make sense.
Otherwise the outcome of f (λx + (1 − λ)x′ ) might not be well defined. Convex functions have a
number of desirable properties.
Jensen’s Inequality
One of the most useful tools is Jensenʼs inequality. It amounts to a generalization of the definition
of convexity:
( )
∑ ∑
αi f (xi ) ≥ f αi xi and Ex [f (x)] ≥ f (Ex [x]) , (11.2.3)
i i
∑
where αi are nonnegative real numbers such that i αi = 1. In other words, the expectation of a
convex function is larger than the convex function of an expectation. To prove the first inequality
we repeatedly apply the definition of convexity to one term in the sum at a time. The expectation
can be proven by taking the limit over finite segments.
One of the common applications of Jensenʼs inequality is with regard to the log-likelihood of par-
tially observed random variables. That is, we use
Ey∼P (y) [− log P (x | y)] ≥ − log P (x). (11.2.4)

∫
This follows since P (y)P (x | y)dy = P (x). This is used in variational methods. Here y is typically
the unobserved random variable, P (y) is the best guess of how it might be distributed and P (x) is
the distribution with y integrated out. For instance, in clustering y might be the cluster labels and
P (x | y) is the generative model when applying cluster labels.
11.2. Convexity 413

11.2.2 Properties
Convex functions have a few useful properties. We describe them as follows.
No Local Minima
In particular, convex functions do not have local minima. Let us assume the contrary and prove
it wrong. If x ∈ X is a local minimum there exists some neighborhood of x for which f (x) is
the smallest value. Since x is only a local minimum there has to be another x′ ∈ X for which
f (x′ ) < f (x). However, by convexity the function values on the entire line λx + (1 − λ)x′ have to
be less than f (x) since for λ ∈ [0, 1)
f (x) > λf (x) + (1 − λ)f (x′ ) ≥ f (λx + (1 − λ)x′ ). (11.2.5)
This contradicts the assumption that f (x) is a local minimum. For instance, the function f (x) =
(x + 1)(x − 1)2 has a local minimum for x = 1. However, it is not a global minimum.
f = lambda x: (x-1)**2 * (x+1)

d2l.set_figsize()
d2l.plot([x, segment], [f(x), f(segment)], 'x', 'f(x)')
The fact that convex functions have no local minima is very convenient. It means that if we min-
imize functions we cannot “get stuck”. Note, though, that this does not mean that there cannot
be more than one global minimum or that there might even exist one. For instance, the function
f (x) = max(|x|−1, 0) attains its minimum value over the interval [−1, 1]. Conversely, the function
f (x) = exp(x) does not attain a minimum value on R. For x → −∞ it asymptotes to 0, however
there is no x for which f (x) = 0.

Convex Functions and Sets
Convex functions define convex sets as below-sets. They are defined as
Sb := {x|x ∈ X and f (x) ≤ b}. (11.2.6)
Such sets are convex. Let us prove this quickly. Remember that for any x, x′ ∈ Sb we need to
show that λx + (1 − λ)x′ ∈ Sb as long as λ ∈ [0, 1]. But this follows directly from the definition of
convexity since f (λx + (1 − λ)x′ ) ≤ λf (x) + (1 − λ)f (x′ ) ≤ b.
Have a look at the function f (x, y) = 0.5x2 + cos(2πy) below. It is clearly nonconvex. The level
sets are correspondingly nonconvex. In fact, they are typically composed of disjoint sets.
x, y = np.meshgrid(np.linspace(-1.0, 1.0, 101), np.linspace(-1.0, 1.0, 101))

z = x**2 + 0.5 * np.cos(2 * np.pi * y)
# Plot the 3D surface
d2l.set_figsize((6, 4))
ax.contour(x, y, z, offset=-1)
ax.set_zlim(-1, 1.5)
# Adjust labels
for func in [d2l.plt.xticks, d2l.plt.yticks, ax.set_zticks]:
func([-1, 0, 1])
/var/lib/jenkins/miniconda3/envs/d2l-en-release-0/lib/python3.7/site-packages/numpy/core/
,→_asarray.py:136: VisibleDeprecationWarning: Creating an ndarray from ragged nested␣
,→sequences (which is a list-or-tuple of lists-or-tuples-or ndarrays with different lengths␣

,→or shapes) is deprecated. If you meant to do this, you must specify 'dtype=object' when␣
,→creating the ndarray

return array(a, dtype, copy=False, order=order, subok=True)
11.2. Convexity 415

Derivatives and Convexity
Whenever the second derivative of a function exists it is very easy to check for convexity. All we
need to do is check whether ∂x2 f (x) ⪰ 0, i.e., whether all of its eigenvalues are nonnegative. For
instance, the function f (x) = 21 ∥x∥22 is convex since ∂x2 f = 1, i.e., its derivative is the identity
matrix.
The first thing to realize is that we only need to prove this property for one-dimensional functions.
After all, in general we can always define some function g(z) = f (x + z · v). This function has the
first and second derivatives g ′ = (∂x f )⊤ v and g ′′ = v⊤ (∂x2 f )v respectively. In particular, g ′′ ≥ 0 for
all v whenever the Hessian of f is positive semidefinite, i.e., whenever all of its eigenvalues are
greater equal than zero. Hence back to the scalar case.
To see that f ′′ (x) ≥ 0 for convex functions we use the fact that
( )
1 1 x+ϵ x−ϵ
f (x + ϵ) + f (x − ϵ) ≥ f + = f (x). (11.2.7)
2 2 2 2
Since the second derivative is given by the limit over finite differences it follows that
f (x + ϵ) + f (x − ϵ) − 2f (x)
f ′′ (x) = lim ≥ 0. (11.2.8)
ϵ→0 ϵ2
To see that the converse is true we use the fact that f ′′ ≥ 0 implies that f ′ is a monotonically
increasing function. Let a < x < b be three points in R. We use the mean value theorem to
express
f (x) − f (a) = (x − a)f ′ (α) for some α ∈ [a, x] and

(11.2.9)
f (b) − f (x) = (b − x)f ′ (β) for some β ∈ [x, b].
By monotonicity f ′ (β) ≥ f ′ (α), hence
f (b) − f (a) = f (b) − f (x) + f (x) − f (a)

= (b − x)f ′ (β) + (x − a)f ′ (α) (11.2.10)
′
≥ (b − a)f (α).
By geometry it follows that f (x) is below the line connecting f (a) and f (b), thus proving convexity.
We omit a more formal derivation in favor of a graph below.
f = lambda x: 0.5 * x**2

x = np.arange(-2, 2, 0.01)
axb, ab = np.array([-1.5, -0.5, 1]), np.array([-1.5, 1])
d2l.set_figsize()
d2l.plot([x, axb, ab], [f(x) for x in [x, axb, ab]], 'x', 'f(x)')
d2l.annotate('a', (-1.5, f(-1.5)), (-1.5, 1.5))
d2l.annotate('b', (1, f(1)), (1, 1.5))
d2l.annotate('x', (-0.5, f(-0.5)), (-1.5, f(-0.5)))

11.2.3 Constraints
One of the nice properties of convex optimization is that it allows us to handle constraints effi-
ciently. That is, it allows us to solve problems of the form:
minimize f (x)
x (11.2.11)
subject to ci (x) ≤ 0 for all i ∈ {1, . . . , N }.
Here f is the objective and the functions ci are constraint functions. To see what this does consider
the case where c1 (x) = ∥x∥2 − 1. In this case the parameters x are constrained to the unit ball. If a
second constraint is c2 (x) = v⊤ x + b, then this corresponds to all x lying on a halfspace. Satisfying
both constraints simultaneously amounts to selecting a slice of a ball as the constraint set.
Lagrange Function
In general, solving a constrained optimization problem is difficult. One way of addressing it stems
from physics with a rather simple intuition. Imagine a ball inside a box. The ball will roll to the
place that is lowest and the forces of gravity will be balanced out with the forces that the sides of
the box can impose on the ball. In short, the gradient of the objective function (i.e., gravity) will
be offset by the gradient of the constraint function (need to remain inside the box by virtue of the
walls “pushing back”). Note that any constraint that is not active (i.e., the ball does not touch the
wall) will not be able to exert any force on the ball.
Skipping over the derivation of the Lagrange function L (see e.g., the book by Boyd and Vanden-
berghe for details (Boyd & Vandenberghe, 2004)) the above reasoning can be expressed via the
following saddlepoint optimization problem:
∑
L(x, α) = f (x) + αi ci (x) where αi ≥ 0. (11.2.12)
i
Here the variables αi are the so-called Lagrange Multipliers that ensure that a constraint is properly
enforced. They are chosen just large enough to ensure that ci (x) ≤ 0 for all i. For instance, for
any x for which ci (x) < 0 naturally, weʼd end up picking αi = 0. Moreover, this is a saddlepoint
optimization problem where one wants to maximize L with respect to α and simultaneously mini-
mize it with respect to x. There is a rich body of literature explaining how to arrive at the function
L(x, α). For our purposes it is sufficient to know that the saddlepoint of L is where the original
constrained optimization problem is solved optimally.
11.2. Convexity 417

Penalties
One way of satisfying constrained optimization problems at least approximately is to adapt the La-
grange function L. Rather than satisfying ci (x) ≤ 0 we simply add αi ci (x) to the objective function
f (x). This ensures that the constraints will not be violated too badly.
In fact, we have been using this trick all along. Consider weight decay in Section 4.5. In it we add
2 ∥w∥ to the objective function to ensure that w does not grow too large. Using the constrained
λ 2
optimization point of view we can see that this will ensure that ∥w∥2 − r2 ≤ 0 for some radius r.
Adjusting the value of λ allows us to vary the size of w.
In general, adding penalties is a good way of ensuring approximate constraint satisfaction. In
practice this turns out to be much more robust than exact satisfaction. Furthermore, for noncon-
vex problems many of the properties that make the exact approach so appealing in the convex
case (e.g., optimality) no longer hold.
Projections
An alternative strategy for satisfying constraints are projections. Again, we encountered them
before, e.g., when dealing with gradient clipping in Section 8.5. There we ensured that a gradient
has length bounded by c via
g ← g · min(1, c/∥g∥). (11.2.13)
This turns out to be a projection of g onto the ball of radius c. More generally, a projection on a
(convex) set X is defined as
ProjX (x) = argmin ∥x − x′ ∥2 . (11.2.14)

x′ ∈X
It is thus the closest point in X to x. This sounds a bit abstract. Fig. 11.2.4 explains it somewhat
more clearly. In it we have two convex sets, a circle and a diamond. Points inside the set (yellow)
remain unchanged. Points outside the set (black) are mapped to the closest point inside the set
(red). While for L2 balls this leaves the direction unchanged, this need not be the case in general,
as can be seen in the case of the diamond.
Fig. 11.2.4: Convex Projections
One of the uses for convex projections is to compute sparse weight vectors. In this case we project
w onto an L1 ball (the latter is a generalized version of the diamond in the picture above).

Summary
In the context of deep learning the main purpose of convex functions is to motivate optimiza-
tion algorithms and help us understand them in detail. In the following we will see how gradient
descent and stochastic gradient descent can be derived accordingly.
• Intersections of convex sets are convex. Unions are not.
• The expectation of a convex function is larger than the convex function of an expectation
(Jensenʼs inequality).
• A twice-differentiable function is convex if and only if its second derivative has only non-
negative eigenvalues throughout.
• Convex constraints can be added via the Lagrange function. In practice simply add them
with a penalty to the objective function.
• Projections map to points in the (convex) set closest to the original point.
Exercises
1. Assume that we want to verify convexity of a set by drawing all lines between points within
the set and checking whether the lines are contained.
• Prove that it is sufficient to check only the points on the boundary.
• Prove that it is sufficient to check only the vertices of the set.
2. Denote by Bp [r] := {x|x ∈ Rd and ∥x∥p ≤ r} the ball of radius r using the p-norm. Prove that
Bp [r] is convex for all p ≥ 1.
3. Given convex functions f and g show that max(f, g) is convex, too. Prove that min(f, g) is
not convex.
4. Prove that the normalization
∑ of the softmax function is convex. More specifically prove the
convexity of f (x) = log i exp(xi ).
5. Prove that linear subspaces are convex sets, i.e., X = {x|Wx = b}.
6. Prove that in the case of linear subspaces with b = 0 the projection ProjX can be written as
Mx for some matrix M.
7. Show that for convex twice differentiable functions f we can write f (x + ϵ) = f (x) + ϵf ′ (x) +
1 2 ′′
2 ϵ f (x + ξ) for some ξ ∈ [0, ϵ].
8. Given a vector w ∈ Rd with ∥w∥1 > 1 compute the projection on the ℓ1 unit ball.
• As intermediate step write out the penalized objective ∥w − w′ ∥22 + λ∥w′ ∥1 and compute
the solution for a given λ > 0.
• Can you find the ʻrightʼ value of λ without a lot of trial and error?
9. Given a convex set X and two vectors x and y prove that projections never increase distances,
i.e., ∥x − y∥ ≥ ∥ProjX (x) − ProjX (y)∥.
Discussions132
132
11.2. Convexity 419

11.3 Gradient Descent
In this section we are going to introduce the basic concepts underlying gradient descent. This is
brief by necessity. See e.g., (Boyd & Vandenberghe, 2004) for an in-depth introduction to convex
optimization. Although the latter is rarely used directly in deep learning, an understanding of
gradient descent is key to understanding stochastic gradient descent algorithms. For instance,
the optimization problem might diverge due to an overly large learning rate. This phenomenon
can already be seen in gradient descent. Likewise, preconditioning is a common technique in
gradient descent and carries over to more advanced algorithms. Let us start with a simple special
case.
11.3.1 Gradient Descent in One Dimension
Gradient descent in one dimension is an excellent example to explain why the gradient descent
algorithm may reduce the value of the objective function. Consider some continuously differen-
tiable real-valued function f : R → R. Using a Taylor expansion (Section 18.3) we obtain that
f (x + ϵ) = f (x) + ϵf ′ (x) + O(ϵ2 ). (11.3.1)
That is, in first approximation f (x + ϵ) is given by the function value f (x) and the first derivative
f ′ (x) at x. It is not unreasonable to assume that for small ϵ moving in the direction of the negative
gradient will decrease f . To keep things simple we pick a fixed step size η > 0 and choose ϵ =
−ηf ′ (x). Plugging this into the Taylor expansion above we get
f (x − ηf ′ (x)) = f (x) − ηf ′2 (x) + O(η 2 f ′2 (x)). (11.3.2)
If the derivative f ′ (x) ̸= 0 does not vanish we make progress since ηf ′2 (x) > 0. Moreover, we can
always choose η small enough for the higher order terms to become irrelevant. Hence we arrive
at
f (x − ηf ′ (x)) ⪅ f (x). (11.3.3)
This means that, if we use
x ← x − ηf ′ (x) (11.3.4)
to iterate x, the value of function f (x) might decline. Therefore, in gradient descent we first choose
an initial value x and a constant η > 0 and then use them to continuously iterate x until the stop
condition is reached, for example, when the magnitude of the gradient |f ′ (x)| is small enough or
the number of iterations has reached a certain value.
For simplicity we choose the objective function f (x) = x2 to illustrate how to implement gradient
descent. Although we know that x = 0 is the solution to minimize f (x), we still use this simple
function to observe how x changes. As always, we begin by importing all required modules.
%matplotlib inline
npx.set_np()

f = lambda x: x**2 # Objective function
gradf = lambda x: 2 * x # Its derivative
Next, we use x = 10 as the initial value and assume η = 0.2. Using gradient descent to iterate x for
10 times we can see that, eventually, the value of x approaches the optimal solution.
def gd(eta):
x = 10.0
results = [x]
for i in range(10):
x -= eta * gradf(x)
results.append(float(x))
print('epoch 10, x:', x)
return results
res = gd(0.2)
epoch 10, x: 0.06046617599999997
The progress of optimizing over x can be plotted as follows.
def show_trace(res):
n = max(abs(min(res)), abs(max(res)))
f_line = np.arange(-n, n, 0.01)
d2l.set_figsize()
d2l.plot([f_line, res], [[f(x) for x in f_line], [f(x) for x in res]],
'x', 'f(x)', fmts=['-', '-o'])
show_trace(res)
11.3. Gradient Descent 421

Learning Rate
The learning rate η can be set by the algorithm designer. If we use a learning rate that is too small,
it will cause x to update very slowly, requiring more iterations to get a better solution. To show
what happens in such a case, consider the progress in the same optimization problem for η = 0.05.
As we can see, even after 10 steps we are still very far from the optimal solution.
show_trace(gd(0.05))
epoch 10, x: 3.4867844009999995
Conversely, if we use an excessively high learning rate, |ηf ′ (x)| might be too large for the first-
order Taylor expansion formula. That is, the term O(η 2 f ′2 (x)) in :eqref:gd-taylor might become
significant. In this case, we cannot guarantee that the iteration of x will be able to lower the value
of f (x). For example, when we set the learning rate to η = 1.1, x overshoots the optimal solution
x = 0 and gradually diverges.
show_trace(gd(1.1))
epoch 10, x: 61.917364224000096

Local Minima
To illustrate what happens for nonconvex functions consider the case of f (x) = x · cos cx. This
function has infinitely many local minima. Depending on our choice of learning rate and de-
pending on how well conditioned the problem is, we may end up with one of many solutions.
The example below illustrates how an (unrealistically) high learning rate will lead to a poor local
minimum.
c = np.array(0.15 * np.pi)
f = lambda x: x * np.cos(c * x)
gradf = lambda x: np.cos(c * x) - c * x * np.sin(c * x)
show_trace(gd(2))
epoch 10, x: -1.5281651
11.3.2 Multivariate Gradient Descent
Now that we have a better intuition of the univariate case, let us consider the situation where
x ∈ Rd . That is, the objective function f : Rd → R maps vectors into scalars. Correspondingly its
gradient is multivariate, too. It is a vector consisting of d partial derivatives:
[ ]⊤
∂f (x) ∂f (x) ∂f (x)
∇f (x) = , ,..., . (11.3.5)
∂x1 ∂x2 ∂xd
Each partial derivative element ∂f (x)/∂xi in the gradient indicates the rate of change of f at x
with respect to the input xi . As before in the univariate case we can use the corresponding Taylor
approximation for multivariate functions to get some idea of what we should do. In particular, we
have that
f (x + ϵ) = f (x) + ϵ⊤ ∇f (x) + O(∥ϵ∥2 ). (11.3.6)
In other words, up to second order terms in ϵ the direction of steepest descent is given by the
negative gradient −∇f (x). Choosing a suitable learning rate η > 0 yields the prototypical gradient
descent algorithm:
x ← x − η∇f (x). (11.3.7)

To see how the algorithm behaves in practice let us construct an objective function f (x) = x21 +2x22
with a two-dimensional vector x = [x1 , x2 ]⊤ as input and a scalar as output. The gradient is given
by ∇f (x) = [2x1 , 4x2 ]⊤ . We will observe the trajectory of x by gradient descent from the initial
position [−5, −2]. We need two more helper functions. The first uses an update function and
applies it 20 times to the initial value. The second helper visualizes the trajectory of x.
def train_2d(trainer, steps=20): #@save

"""Optimize a 2-dim objective function with a customized trainer."""
# s1 and s2 are internal state variables and will
# be used later in the chapter
x1, x2, s1, s2 = -5, -2, 0, 0
results = [(x1, x2)]
for i in range(steps):
x1, x2, s1, s2 = trainer(x1, x2, s1, s2)
results.append((x1, x2))
return results
def show_trace_2d(f, results): #@save

"""Show the trace of 2D variables during optimization."""
d2l.set_figsize()
d2l.plt.plot(*zip(*results), '-o', color='#ff7f0e')
x1, x2 = np.meshgrid(np.arange(-5.5, 1.0, 0.1),
np.arange(-3.0, 1.0, 0.1))
d2l.plt.contour(x1, x2, f(x1, x2), colors='#1f77b4')
d2l.plt.xlabel('x1')
d2l.plt.ylabel('x2')
Next, we observe the trajectory of the optimization variable x for learning rate η = 0.1. We can see
that after 20 steps the value of x approaches its minimum at [0, 0]. Progress is fairly well-behaved
albeit rather slow.
f = lambda x1, x2: x1 ** 2 + 2 * x2 ** 2 # Objective

gradf = lambda x1, x2: (2 * x1, 4 * x2) # Gradient
def gd(x1, x2, s1, s2):

(g1, g2) = gradf(x1, x2) # Compute gradient
return (x1 - eta * g1, x2 - eta * g2, 0, 0) # Update variables
eta = 0.1
show_trace_2d(f, train_2d(gd))

11.3.3 Adaptive Methods
As we could see in Section 11.3.1, getting the learning rate η “just right” is tricky. If we pick it too
small, we make no progress. If we pick it too large, the solution oscillates and in the worst case it
might even diverge. What if we could determine η automatically or get rid of having to select a step
size at all? Second order methods that look not only at the value and gradient of the objective but
also at its curvature can help in this case. While these methods cannot be applied to deep learning
directly due to the computational cost, they provide useful intuition into how to design advanced
optimization algorithms that mimic many of the desirable properties of the algorithms outlined
below.
Newton’s Method
Reviewing the Taylor expansion of f there is no need to stop after the first term. In fact, we can
write it as
1
f (x + ϵ) = f (x) + ϵ⊤ ∇f (x) + ϵ⊤ ∇∇⊤ f (x)ϵ + O(∥ϵ∥3 ). (11.3.8)
2
To avoid cumbersome notation we define Hf := ∇∇⊤ f (x) to be the Hessian of f . This is a d × d
matrix. For small d and simple problems Hf is easy to compute. For deep networks, on the other
hand, Hf may be prohibitively large, due to the cost of storing O(d2 ) entries. Furthermore it may
be too expensive to compute via backpropagation as we would need to apply backpropagation
to the backpropagation call graph. For now let us ignore such considerations and look at what
algorithm weʼd get.
After all, the minimum of f satisfies ∇f (x) = 0. Taking derivatives of (11.3.8) with regard to ϵ and
ignoring higher order terms we arrive at
∇f (x) + Hf ϵ = 0 and hence ϵ = −Hf−1 ∇f (x). (11.3.9)
That is, we need to invert the Hessian Hf as part of the optimization problem.
For f (x) = 12 x2 we have ∇f (x) = x and Hf = 1. Hence for any x we obtain ϵ = −x. In other
words, a single step is sufficient to converge perfectly without the need for any adjustment! Alas,
we got a bit lucky here since the Taylor expansion was exact. Let us see what happens in other
problems.
c = np.array(0.5)
f = lambda x: np.cosh(c * x) # Objective
gradf = lambda x: c * np.sinh(c * x) # Derivative
hessf = lambda x: c**2 * np.cosh(c * x) # Hessian
def newton(eta=1):
x = 10.0
results = [x]
for i in range(10):
x -= eta * gradf(x) / hessf(x)
results.append(float(x))
print('epoch 10, x:', x)
return results
show_trace(newton())

epoch 10, x: 0.0
Now let us see what happens when we have a nonconvex function, such as f (x) = x cos(cx). After
all, note that in Newtonʼs method we end up dividing by the Hessian. This means that if the second
derivative is negative we would walk into the direction of increasing f . That is a fatal flaw of the
algorithm. Let us see what happens in practice.
c = np.array(0.15 * np.pi)
f = lambda x: x * np.cos(c * x)
gradf = lambda x: np.cos(c * x) - c * x * np.sin(c * x)
hessf = lambda x: - 2 * c * np.sin(c * x) - x * c**2 * np.cos(c * x)
show_trace(newton())
epoch 10, x: 26.834133
This went spectacularly wrong. How can we fix it? One way would be to “fix” the Hessian by taking
its absolute value instead. Another strategy is to bring back the learning rate. This seems to defeat
the purpose, but not quite. Having second order information allows us to be cautious whenever
the curvature is large and to take longer steps whenever the objective is flat. Let us see how this

works with a slightly smaller learning rate, say η = 0.5. As we can see, we have quite an efficient
algorithm.
show_trace(newton(0.5))
epoch 10, x: 7.26986
Convergence Analysis
We only analyze the convergence rate for convex and three times differentiable f , where at its
minimum x∗ the second derivative is nonzero, i.e., where f ′′ (x∗ ) > 0. The multivariate proof is
a straightforward extension of the argument below and omitted since it doesnʼt help us much in
terms of intuition.
Denote by xk the value of x at the k-th iteration and let ek := xk −x∗ be the distance from optimality.
By Taylor series expansion we have that the condition f ′ (x∗ ) = 0 can be written as
1
0 = f ′ (xk − ek ) = f ′ (xk ) − ek f ′′ (xk ) + e2k f ′′′ (ξk ). (11.3.10)
2
This holds for some ξk ∈ [xk − ek , xk ]. Recall that we have the update xk+1 = xk − f ′ (xk )/f ′′ (xk ).
Dividing the above expansion by f ′′ (xk ) yields
1
ek − f ′ (xk )/f ′′ (xk ) = e2k f ′′′ (ξk )/f ′′ (xk ). (11.3.11)
2
Plugging in the update equations leads to the following bound ek+1 ≤ e2k f ′′′ (ξk )/f ′ (xk ). Conse-
quently, whenever we are in a region of bounded f ′′′ (ξk )/f ′′ (xk ) ≤ c, we have a quadratically
decreasing error ek+1 ≤ ce2k .
As an aside, optimization researchers call this linear convergence, whereas a condition such as
ek+1 ≤ αek would be called a constant rate of convergence. Note that this analysis comes with
a number of caveats: We do not really have much of a guarantee when we will reach the region
of rapid convergence. Instead, we only know that once we reach it, convergence will be very
quick. Second, this requires that f is well-behaved up to higher order derivatives. It comes down
to ensuring that f does not have any “surprising” properties in terms of how it might change its
values.

Preconditioning
Quite unsurprisingly computing and storing the full Hessian is very expensive. It is thus desir-
able to find alternatives. One way to improve matters is by avoiding to compute the Hessian in
its entirety but only compute the diagonal entries. While this is not quite as good as the full New-
ton method, it is still much better than not using it. Moreover, estimates for the main diagonal
elements are what drives some of the innovation in stochastic gradient descent optimization al-
gorithms. This leads to update algorithms of the form
x ← x − ηdiag(Hf )−1 ∇f (x). (11.3.12)
To see why this might be a good idea consider a situation where one variable denotes height in mil-
limeters and the other one denotes height in kilometers. Assuming that for both the natural scale
is in meters we have a terrible mismatch in parameterizations. Using preconditioning removes
this. Effectively preconditioning with gradient descent amounts to selecting a different learning
rate for each coordinate.
Gradient Descent with Line Search
One of the key problems in gradient descent was that we might overshoot the goal or make insuf-
ficient progress. A simple fix for the problem is to use line search in conjunction with gradient
descent. That is, we use the direction given by ∇f (x) and then perform binary search as to which
step length η minimizes f (x − η∇f (x)).
This algorithm converges rapidly (for an analysis and proof see e.g., (Boyd & Vandenberghe,
2004)). However, for the purpose of deep learning this is not quite so feasible, since each step
of the line search would require us to evaluate the objective function on the entire dataset. This is
way too costly to accomplish.
Summary
• Learning rates matter. Too large and we diverge, too small and we do not make progress.
• Gradient descent can get stuck in local minima.
• In high dimensions adjusting the learning rate is complicated.
• Preconditioning can help with scale adjustment.
• Newtonʼs method is a lot faster once it has started working properly in convex problems.
• Beware of using Newtonʼs method without any adjustments for nonconvex problems.
Exercises
1. Experiment with different learning rates and objective functions for gradient descent.
2. Implement line search to minimize a convex function in the interval [a, b].
• Do you need derivatives for binary search, i.e., to decide whether to pick [a, (a + b)/2]
or [(a + b)/2, b].
• How rapid is the rate of convergence for the algorithm?

• Implement the algorithm and apply it to minimizing log(exp(x) + exp(−2 ∗ x − 3)).
3. Design an objective function defined on R2 where gradient descent is exceedingly slow. Hint:
scale different coordinates differently.
4. Implement the lightweight version of Newtonʼs method using preconditioning:
• Use diagonal Hessian as preconditioner.
• Use the absolute values of that rather than the actual (possibly signed) values.
• Apply this to the problem above.
5. Apply the algorithm above to a number of objective functions (convex or not). What happens
if you rotate coordinates by 45 degrees?
Discussions133
11.4 Stochastic Gradient Descent
In this section, we are going to introduce the basic principles of stochastic gradient descent.
%matplotlib inline
import math
npx.set_np()
11.4.1 Stochastic Gradient Updates
In deep learning, the objective function is usually the average of the loss functions for each exam-
ple in the training dataset. We assume that fi (x) is the loss function of the training dataset with n
examples, an index of i, and parameter vector of x, then we have the objective function
1∑
n
f (x) = fi (x). (11.4.1)
n
i=1
The gradient of the objective function at x is computed as
1∑
n
∇f (x) = ∇fi (x). (11.4.2)
n
i=1
If gradient descent is used, the computing cost for each independent variable iteration is O(n),
which grows linearly with n. Therefore, when the model training dataset is large, the cost of gra-
dient descent for each iteration will be very high.
Stochastic gradient descent (SGD) reduces computational cost at each iteration. At each iteration
of stochastic gradient descent, we uniformly sample an index i ∈ {1, . . . , n} for data examples at
random, and compute the gradient ∇fi (x) to update x:
x ← x − η∇fi (x). (11.4.3)

133
11.4. Stochastic Gradient Descent 429

Here, η is the learning rate. We can see that the computing cost for each iteration drops from
O(n) of the gradient descent to the constant O(1). We should mention that the stochastic gradient
∇fi (x) is the unbiased estimate of gradient ∇f (x).
1∑
n
Ei ∇fi (x) = ∇fi (x) = ∇f (x). (11.4.4)
n
i=1
This means that, on average, the stochastic gradient is a good estimate of the gradient.
Now, we will compare it to gradient descent by adding random noise with a mean of 0 and a vari-
ance of 1 to the gradient to simulate a SGD.
f = lambda x1, x2: x1 ** 2 + 2 * x2 ** 2 # Objective

gradf = lambda x1, x2: (2 * x1, 4 * x2) # Gradient
def sgd(x1, x2, s1, s2):

global lr # Learning rate scheduler
(g1, g2) = gradf(x1, x2)
# Simulate noisy gradient
g1 += np.random.normal(0.0, 1, (1,))
g2 += np.random.normal(0.0, 1, (1,))
eta_t = eta * lr() # Learning rate at time t
return (x1 - eta_t * g1, x2 - eta_t * g2, 0, 0) # Update variables
eta = 0.1
lr = (lambda: 1) # Constant learning rate
d2l.show_trace_2d(f, d2l.train_2d(sgd, steps=50))
/var/lib/jenkins/miniconda3/envs/d2l-en-release-0/lib/python3.7/site-packages/numpy/core/
,→_asarray.py:136: VisibleDeprecationWarning: Creating an ndarray from ragged nested␣
,→sequences (which is a list-or-tuple of lists-or-tuples-or ndarrays with different lengths␣
,→or shapes) is deprecated. If you meant to do this, you must specify 'dtype=object' when␣
,→creating the ndarray
return array(a, dtype, copy=False, order=order, subok=True)
As we can see, the trajectory of the variables in the SGD is much more noisy than the one we
observed in gradient descent in the previous section. This is due to the stochastic nature of the
gradient. That is, even when we arrive near the minimum, we are still subject to the uncertainty

injected by the instantaneous gradient via η∇fi (x). Even after 50 steps the quality is still not so
good. Even worse, it will not improve after additional steps (we encourage the reader to exper-
iment with a larger number of steps to confirm this on his own). This leaves us with the only
alternative—change the learning rate η. However, if we pick this too small, we will not make any
meaningful progress initially. On the other hand, if we pick it too large, we will not get a good
solution, as seen above. The only way to resolve these conflicting goals is to reduce the learning
rate dynamically as optimization progresses.
This is also the reason for adding a learning rate function lr into the sgd step function. In the
example above any functionality for learning rate scheduling lies dormant as we set the associated
lr function to be constant, i.e., lr = (lambda: 1).
11.4.2 Dynamic Learning Rate
Replacing η with a time-dependent learning rate η(t) adds to the complexity of controlling conver-
gence of an optimization algorithm. In particular, need to figure out how rapidly η should decay.
If it is too quick, we will stop optimizing prematurely. If we decrease it too slowly, we waste too
much time on optimization. There are a few basic strategies that are used in adjusting η over time
(we will discuss more advanced strategies in a later chapter):
η(t) = ηi if ti ≤ t ≤ ti+1 piecewise constant

η(t) = η0 · e−λt exponential (11.4.5)
−α
η(t) = η0 · (βt + 1) polynomial
In the first scenario we decrease the learning rate, e.g., whenever progress in optimization has
stalled. This is a common strategy for training deep networks. Alternatively we could decrease it
much more aggressively by an exponential decay. Unfortunately this leads to premature stopping
before the algorithm has converged. A popular choice is polynomial decay with α = 0.5. In the
case of convex optimization there are a number of proofs which show that this rate is well behaved.
Let us see what this looks like in practice.
def exponential():
global ctr
ctr += 1
return math.exp(-0.1 * ctr)
ctr = 1
lr = exponential # Set up learning rate

As expected, the variance in the parameters is significantly reduced. However, this comes at the
expense of failing to converge to the optimal solution x = (0, 0). Even after 1000 steps are we are
still very far away from the optimal solution. Indeed, the algorithm fails to converge at all. On the
other hand, if we use a polynomial decay where the learning rate decays with the inverse square
root of the number of steps convergence is good.
def polynomial():
global ctr
ctr += 1
return (1 + 0.1 * ctr)**(-0.5)
ctr = 1
lr = polynomial # Set up learning rate
There exist many more choices for how to set the learning rate. For instance, we could start with a
small rate, then rapidly ramp up and then decrease it again, albeit more slowly. We could even al-
ternate between smaller and larger learning rates. There exists a large variety of such schedules.
For now let us focus on learning rate schedules for which a comprehensive theoretical analysis
is possible, i.e., on learning rates in a convex setting. For general nonconvex problems it is very
difficult to obtain meaningful convergence guarantees, since in general minimizing nonlinear

nonconvex problems is NP hard. For a survey see e.g., the excellent lecture notes134 of Tibshirani
2015.
11.4.3 Convergence Analysis for Convex Objectives
The following is optional and primarily serves to convey more intuition about the problem. We
limit ourselves to one of the simplest proofs, as described by (Nesterov & Vial, 2000). Significantly
more advanced proof techniques exist, e.g., whenever the objective function is particularly well
behaved. (Hazan et al., 2008) show that for strongly convex functions, i.e., for functions that can
be bounded from below by x⊤ Qx, it is possible to minimize them in a small number of steps while
decreasing the learning rate like η(t) = η0 /(βt + 1). Unfortunately this case never really occurs in
deep learning and we are left with a much more slowly decreasing rate in practice.
Consider the case where
wt+1 = wt − ηt ∂w l(xt , w). (11.4.6)
In particular, assume that xt is drawn from some distribution P (x) and that l(x, w) is a convex
function in w for all x. Last denote by
R(w) = Ex∼P [l(x, w)] (11.4.7)
the expected risk and by R∗ its minimum with regard to w. Last let w∗ be the minimizer (we
assume that it exists within the domain which w is defined). In this case we can track the distance
between the current parameter wt and the risk minimizer w∗ and see whether it improves over
time:
∥wt+1 − w∗ ∥2 = ∥wt − ηt ∂w l(xt , w) − w∗ ∥2
(11.4.8)
= ∥wt − w∗ ∥2 + ηt2 ∥∂w l(xt , w)∥2 − 2ηt ⟨wt − w∗ , ∂w l(xt , w)⟩ .
The gradient ∂w l(xt , w) can be bounded from above by some Lipschitz constant L, hence we have
that
ηt2 ∥∂w l(xt , w)∥2 ≤ ηt2 L2 . (11.4.9)
We are mostly interested in how the distance between wt and w∗ changes in expectation. In fact,
for any specific sequence of steps the distance might well increase, depending on whichever xt
we encounter. Hence we need to bound the inner product. By convexity we have that
l(xt , w∗ ) ≥ l(xt , wt ) + ⟨w∗ − wt , ∂w l(xt , wt )⟩ . (11.4.10)
Using both inequalities and plugging it into the above we obtain a bound on the distance between
parameters at time t + 1 as follows:
∥wt − w∗ ∥2 − ∥wt+1 − w∗ ∥2 ≥ 2ηt (l(xt , wt ) − l(xt , w∗ )) − ηt2 L2 . (11.4.11)
This means that we make progress as long as the expected difference between current loss and
the optimal loss outweighs ηt L2 . Since the former is bound to converge to 0 it follows that the
learning rate ηt also needs to vanish.
Next we take expectations over this expression. This yields
[ ] [ ]
Ewt ∥wt − w∗ ∥2 − Ewt+1 |wt ∥wt+1 − w∗ ∥2 ≥ 2ηt [E[R[wt ]] − R∗ ] − ηt2 L2 . (11.4.12)
134
https://www.stat.cmu.edu/~ryantibs/convexopt-F15/lectures/26-nonconvex.pdf

The last step involves summing over the inequalities for t ∈ {t, . . . , T }. Since the sum telescopes
and by dropping the lower term we obtain
∑
T ∑
T
∗ 2 ∗
∥w0 − w ∥ ≥ 2 ηt [E[R[wt ]] − R ] − L 2
ηt2 . (11.4.13)
t=1 t=1
Note that we exploited that w0 is given and thus the expectation can be dropped. Last define
∑T
t=1 ηt wt
w̄ := ∑ T
. (11.4.14)
t=1 η t
Then by convexity it follows that

∑ ∑
ηt E[R[wt ]] ≥ ηt · [E[w̄]] . (11.4.15)
t
Plugging this into the above inequality yields the bound

∑
r2 + L2 Tt=1 ηt2
∗
[E[w̄]] − R ≤ ∑ . (11.4.16)
2 Tt=1 ηt
Here r2 := ∥w0 − w∗ ∥2 is a bound on the distance between the initial choice of parameters and
the final outcome. In short, the speed of convergence depends on how rapidly the loss function
changes via the Lipschitz constant L and how far away from optimality the initial value is r. Note
that the bound is in terms of w̄ rather than wT . This is the case since w̄ is a smoothed version of
the optimization path. Now let us analyze some choices for ηt .
√
• Known Time Horizon. Whenever √ r, L and
√ T are known we can pick η = r/L T . This
√ yields
as upper bound rL(1 + 1/T )/2 T < rL/ T . That is, we converge with rate O(1/ T ) to the
optimal solution.
• Unknown Time √ Horizon. Whenever we want to have a good solution for any time T we can
pick η = O(1/ T ). This
√ costs us an extra logarithmic factor and it leads to an upper bound
of the form O(log T / T ).
Note that for strongly convex losses l(x, w′ ) ≥ l(x, w) + ⟨w′ − w, ∂w l(x, w)⟩ + λ2 ∥w − w′ ∥2 we can
design even more rapidly converging optimization schedules. In fact, an exponential decay in η
leads to a bound of the form O(log T /T ).
11.4.4 Stochastic Gradients and Finite Samples
So far we have played a bit fast and loose when it comes to talking about stochastic gradient de-
scent. We posited that we draw instances xi , typically with labels yi from some distribution p(x, y)
and that we use this to update the weights w in some manner. In particular, for a finite sample size
1 ∑n
we simply argued that the discrete distribution p(x, y) = n i=1 δxi (x)δyi (y) allows us to perform
SGD over it.
However, this is not really what we did. In the toy examples in the current section we simply
added noise to an otherwise non-stochastic gradient, i.e., we pretended to have pairs (xi , yi ). It
turns out that this is justified here (see the exercises for a detailed discussion). More troubling is
that in all previous discussions we clearly did not do this. Instead we iterated over all instances
exactly once. To see why this is preferable consider the converse, namely that we are sampling

n observations from the discrete distribution with replacement. The probability of choosing an
element i at random is N −1 . Thus to choose it at least once is
P (choose i) = 1 − P (omit i) = 1 − (1 − N −1 )N ≈ 1 − e−1 ≈ 0.63. (11.4.17)
A
(Nsimilar
) −1 reasoning shows that the probability of picking a sample exactly once is given by
N (1 − N −1 )N −1 = N −1 (1 − N −1 )N ≈ e−1 ≈ 0.37. This leads to an increased variance
1 N
and decreased data efficiency relative to sampling without replacement. Hence, in practice we
perform the latter (and this is the default choice throughout this book). Last note that repeated
passes through the dataset traverse it in a different random order.
Summary
• For convex problems we can prove that for a wide choice of learning rates Stochastic Gradi-
ent Descent will converge to the optimal solution.
• For deep learning this is generally not the case. However, the analysis of convex problems
gives us useful insight into how to approach optimization, namely to reduce the learning
rate progressively, albeit not too quickly.
• Problems occur when the learning rate is too small or too large. In practice a suitable learn-
ing rate is often found only after multiple experiments.
• When there are more examples in the training dataset, it costs more to compute each itera-
tion for gradient descent, so SGD is preferred in these cases.
• Optimality guarantees for SGD are in general not available in nonconvex cases since the
number of local minima that require checking might well be exponential.
Exercises
1. Experiment with different learning rate schedules for SGD and with different numbers of
iterations. In particular, plot the distance from the optimal solution (0, 0) as a function of
the number of iterations.
2. Prove that for the function f (x1 , x2 ) = x21 + 2x22 adding normal noise to the gradient is equiv-
alent to minimizing a loss function l(x, w) = (x1 − w1 )2 + 2(x2 − w2 )2 where x is drawn from
a normal distribution.
• Derive mean and variance of the distribution for x.
• Show that this property holds in general for objective functions f (x) = 12 (x − µ)⊤ Q(x −
µ) for Q ⪰ 0.
3. Compare convergence of SGD when you sample from {(x1 , y1 ), . . . , (xm , ym )} with replace-
ment and when you sample without replacement.
4. How would you change the SGD solver if some gradient (or rather some coordinate associ-
ated with it) was consistently larger than all other gradients?
5. Assume that f (x) = x2 (1 + sin x). How many local minima does f have? Can you change f
in such a way that to minimize it one needs to evaluate all local minima?
Discussions135
135

11.5 Minibatch Stochastic Gradient Descent
So far we encountered two extremes in the approach to gradient based learning: Section 11.3 uses
the full dataset to compute gradients and to update parameters, one pass at a time. Conversely
Section 11.4 processes one observation at a time to make progress. Each of them has its own draw-
backs. Gradient Descent is not particularly data efficient whenever data is very similar. Stochastic
Gradient Descent is not particularly computationally efficient since CPUs and GPUs cannot exploit
the full power of vectorization. This suggests that there might be a happy medium, and in fact,
thatʼs what we have been using so far in the examples we discussed.
11.5.1 Vectorization and Caches
At the heart of the decision to use minibatches is computational efficiency. This is most easily
understood when considering parallelization to multiple GPUs and multiple servers. In this case
we need to send at least one image to each GPU. With 8 GPUs per server and 16 servers we already
arrive at a minibatch size of 128.
Things are a bit more subtle when it comes to single GPUs or even CPUs. These devices have mul-
tiple types of memory, often multiple type of compute units and different bandwidth constraints
between them. For instance, a CPU has a small number of registers and then L1, L2 and in some
cases even L3 cache (which is shared between the different processor cores). These caches are of
increasing size and latency (and at the same time they are of decreasing bandwidth). Suffice it to
say, the processor is capable of performing many more operations than what the main memory
interface is able to provide.
• A 2GHz CPU with 16 cores and AVX-512 vectorization can process up to 2 · 109 · 16 · 32 = 1012
bytes per second. The capability of GPUs easily exceeds this number by a factor of 100.
On the other hand, a midrange server processor might not have much more than 100 GB/s
bandwidth, i.e., less than one tenth of what would be required to keep the processor fed. To
make matters worse, not all memory access is created equal: first, memory interfaces are
typically 64 bit wide or wider (e.g., on GPUs up to 384 bit), hence reading a single byte incurs
the cost of a much wider access.
• There is significant overhead for the first access whereas sequential access is relatively cheap
(this is often called a burst read). There are many more things to keep in mind, such as
caching when we have multiple sockets, chiplets and other structures. A detailed discussion
of this is beyond the scope of this section. See e.g., this Wikipedia article136 for a more in-
depth discussion.
The way to alleviate these constraints is to use a hierarchy of CPU caches which are actually fast
enough to supply the processor with data. This is the driving force behind batching in deep learn-
ing. To keep matters simple, consider matrix-matrix multiplication, say A = BC. We have a num-
ber of options for calculating A. For instance we could try the following:
1. We could compute Aij = Bi,: C⊤
:,j , i.e., we could compute it element-wise by means of dot
products.
2. We could compute A:,j = BC⊤
:,j , i.e., we could compute it one column at a time. Likewise we
could compute A one row Ai,: at a time.
3. We could simply compute A = BC.
136
https://en.wikipedia.org/wiki/Cache_hierarchy

4. We could break B and C into smaller block matrices and compute A one block at a time.
If we follow the first option, we will need to copy one row and one column vector into the CPU
each time we want to compute an element Aij . Even worse, due to the fact that matrix elements
are aligned sequentially we are thus required to access many disjoint locations for one of the two
vectors as we read them from memory. The second option is much more favorable. In it, we are
able to keep the column vector C:,j in the CPU cache while we keep on traversing through B. This
halves the memory bandwidth requirement with correspondingly faster access. Of course, option
3 is most desirable. Unfortunately, most matrices might not entirely fit into cache (this is what we
are discussing after all). However, option 4 offers a practically useful alternative: we can move
blocks of the matrix into cache and multiply them locally. Optimized libraries take care of this for
us. Let us have a look at how efficient these operations are in practice.
Beyond computational efficiency, the overhead introduced by Python and by the deep learning
framework itself is considerable. Recall that each time we execute a command the Python in-
terpreter sends a command to the MXNet engine which needs to insert it into the computational
graph and deal with it during scheduling. Such overhead can be quite detrimental. In short, it is
highly advisable to use vectorization (and matrices) whenever possible.
%matplotlib inline
npx.set_np()
timer = d2l.Timer()
A = np.zeros((256, 256))
B = np.random.normal(0, 1, (256, 256))
C = np.random.normal(0, 1, (256, 256))
Element-wise assignment simply iterates over all rows and columns of B and C respectively to
assign the value to A.
# Compute A = BC one element at a time

timer.start()
for i in range(256):
for j in range(256):
A[i, j] = np.dot(B[i, :], C[:, j])
A.wait_to_read()
timer.stop()
60.47386956214905
A faster strategy is to perform column-wise assignment.
# Compute A = BC one column at a time

timer.start()
for j in range(256):
A[:, j] = np.dot(B, C[:, j])
A.wait_to_read()
timer.stop()
11.5. Minibatch Stochastic Gradient Descent 437

0.18439102172851562
Last, the most effective manner is to perform the entire operation in one block. Let us see what
the respective speed of the operations is.
# Compute A = BC in one go
timer.start()
A = np.dot(B, C)
A.wait_to_read()
timer.stop()
# Multiply and add count as separate operations (fused in practice)

gigaflops = [2/i for i in timer.times]
print(f'performance in Gigaflops: element {gigaflops[0]:.3f}, '
f'column {gigaflops[1]:.3f}, full {gigaflops[2]:.3f}')
performance in Gigaflops: element 0.033, column 10.847, full 2337.311
11.5.2 Minibatches
In the past we took it for granted that we would read minibatches of data rather than single observa-
tions to update parameters. We now give a brief justification for it. Processing single observations
requires us to perform many single matrix-vector (or even vector-vector) multiplications, which is
quite expensive and which incurs a significant overhead on behalf of the underlying deep learning
framework. This applies both to evaluating a network when applied to data (often referred to as
inference) and when computing gradients to update parameters. That is, this applies whenever
we perform w ← w − ηt gt where
gt = ∂w f (xt , w) (11.5.1)
We can increase the computational efficiency of this operation by applying it to a minibatch of

observations at a time. That is, we replace the gradient gt over a single observation by one over a
small batch
1 ∑
gt = ∂w f (xi , w) (11.5.2)
|B |t
i∈Bt
Let us see what this does to the statistical properties of gt : since both xt and also all elements
of the minibatch Bt are drawn uniformly at random from the training set, the expectation of the
gradient remains unchanged. The variance, on the other hand, is reduced significantly. Since the
minibatch gradient is composed of b := |Bt | independent gradients which are being averaged, its
standard deviation is reduced by a factor of b− 2 . This, by itself, is a good thing, since it means that
1
the updates are more reliably aligned with the full gradient.
Naively this would indicate that choosing a large minibatch Bt would be universally desirable.
Alas, after some point, the additional reduction in standard deviation is minimal when compared
to the linear increase in computational cost. In practice we pick a minibatch that is large enough
to offer good computational efficiency while still fitting into the memory of a GPU. To illustrate the
savings let us have a look at some code. In it we perform the same matrix-matrix multiplication,
but this time broken up into “minibatches” of 64 columns at a time.

timer.start()
for j in range(0, 256, 64):
A[:, j:j+64] = np.dot(B, C[:, j:j+64])
timer.stop()
print(f'performance in Gigaflops: block {2 / timer.times[3]:.3f}')
performance in Gigaflops: block 548.490
As we can see, the computation on the minibatch is essentially as efficient as on the full matrix. A
word of caution is in order. In Section 7.5 we used a type of regularization that was heavily depen-
dent on the amount of variance in a minibatch. As we increase the latter, the variance decreases
and with it the benefit of the noise-injection due to batch normalization. See e.g., (Ioffe, 2017) for
details on how to rescale and compute the appropriate terms.
Let us have a look at how minibatches are efficiently generated from data. In the following we use
a dataset developed by NASA to test the wing noise from different aircraft137 to compare these opti-
mization algorithms. For convenience we only use the first 1, 500 examples. The data is whitened
for preprocessing, i.e., we remove the mean and rescale the variance to 1 per coordinate.
#@save
d2l.DATA_HUB['airfoil'] = (d2l.DATA_URL + 'airfoil_self_noise.dat',
'76e5be1548fd8222e5074cf0faae75edff8cf93f')
#@save
def get_data_ch11(batch_size=10, n=1500):
data = np.genfromtxt(d2l.download('airfoil'),
dtype=np.float32, delimiter='\t')
data = (data - data.mean(axis=0)) / data.std(axis=0)
data_iter = d2l.load_array(
(data[:n, :-1], data[:n, -1]), batch_size, is_train=True)
return data_iter, data.shape[1]-1
Recall the minibatch SGD implementation from Section 3.2. In the following we provide a slightly
more general implementation. For convenience it has the same call signature as the other opti-
mization algorithms introduced later in this chapter. Specifically, we add the status input states
and place the hyperparameter in dictionary hyperparams. In addition, we will average the loss of
each minibatch example in the training function, so the gradient in the optimization algorithm
does not need to be divided by the batch size.
def sgd(params, states, hyperparams):

for p in params:
p[:] -= hyperparams['lr'] * p.grad
137
https://archive.ics.uci.edu/ml/datasets/Airfoil+Self-Noise

Next, we implement a generic training function to facilitate the use of the other optimization al-
gorithms introduced later in this chapter. It initializes a linear regression model and can be used
to train the model with minibatch SGD and other algorithms introduced subsequently.
#@save
def train_ch11(trainer_fn, states, hyperparams, data_iter,
feature_dim, num_epochs=2):
# Initialization
w = np.random.normal(scale=0.01, size=(feature_dim, 1))
b = np.zeros(1)
w.attach_grad()
b.attach_grad()
net, loss = lambda X: d2l.linreg(X, w, b), d2l.squared_loss
# Train
xlim=[0, num_epochs], ylim=[0.22, 0.35])
n, timer = 0, d2l.Timer()
for _ in range(num_epochs):
l = loss(net(X), y).mean()
l.backward()
trainer_fn([w, b], states, hyperparams)
n += X.shape[0]
if n % 200 == 0:
timer.stop()
animator.add(n/X.shape[0]/len(data_iter),
(d2l.evaluate_loss(net, data_iter, loss),))
timer.start()
print(f'loss: {animator.Y[0][-1]:.3f}, {timer.avg():.3f} sec/epoch')
return timer.cumsum(), animator.Y[0]
Let us see how optimization proceeds for batch gradient descent. This can be achieved by setting
the minibatch size to 1500 (i.e., to the total number of examples). As a result the model parameters
are updated only once per epoch. There is little progress. In fact, after 6 steps progress stalls.
def train_sgd(lr, batch_size, num_epochs=2):

data_iter, feature_dim = get_data_ch11(batch_size)
return train_ch11(
sgd, None, {'lr': lr}, data_iter, feature_dim, num_epochs)
gd_res = train_sgd(1, 1500, 10)
loss: 0.254, 0.071 sec/epoch

When the batch size equals 1, we use SGD for optimization. For simplicity of implementation we
picked a constant (albeit small) learning rate. In SGD, the model parameters are updated when-
ever an example is processed. In our case this amounts to 1500 updates per epoch. As we can see,
the decline in the value of the objective function slows down after one epoch. Although both the
procedures processed 1500 examples within one epoch, SGD consumes more time than gradient
descent in our experiment. This is because SGD updated the parameters more frequently and
since it is less efficient to process single observations one at a time.
sgd_res = train_sgd(0.005, 1)
Last, when the batch size equals 100, we use minibatch SGD for optimization. The time required
per epoch is shorter than the time needed for SGD and the time for batch gradient descent.
mini1_res = train_sgd(.4, 100)

Reducing the batch size to 10, the time for each epoch increases because the workload for each
batch is less efficient to execute.
mini2_res = train_sgd(.05, 10)
Finally, we compare the time vs. loss for the preview four experiments. As can be seen, despite
SGD converges faster than GD in terms of number of examples processed, it uses more time to
reach the same loss than GD because that computing gradient example by example is not efficient.
Minibatch SGD is able to trade-off the convergence speed and computation efficiency. A minibatch
size 10 is more efficient than SGD; a minibatch size 100 even outperforms GD in terms of runtime.
d2l.set_figsize([6, 3])
d2l.plot(*list(map(list, zip(gd_res, sgd_res, mini1_res, mini2_res))),
'time (sec)', 'loss', xlim=[1e-2, 10],
legend=['gd', 'sgd', 'batch size=100', 'batch size=10'])
d2l.plt.gca().set_xscale('log')

In Gluon, we can use the Trainer class to call optimization algorithms. This is used to implement
a generic training function. We will use this throughout the current chapter.
#@save
def train_concise_ch11(tr_name, hyperparams, data_iter, num_epochs=2):
# Initialization
trainer = gluon.Trainer(net.collect_params(), tr_name, hyperparams)
xlim=[0, num_epochs], ylim=[0.22, 0.35])
n, timer = 0, d2l.Timer()
for _ in range(num_epochs):
l = loss(net(X), y)
l.backward()
n += X.shape[0]
if n % 200 == 0:
timer.stop()
animator.add(n/X.shape[0]/len(data_iter),
(d2l.evaluate_loss(net, data_iter, loss),))
timer.start()
print(f'loss: {animator.Y[0][-1]:.3f}, {timer.avg():.3f} sec/epoch')
Using Gluon to repeat the last experiment shows identical behavior.
data_iter, _ = get_data_ch11(10)
train_concise_ch11('sgd', {'learning_rate': 0.05}, data_iter)

Summary
• Vectorization makes code more efficient due to reduced overhead arising from the deep
learning framework and due to better memory locality and caching on CPUs and GPUs.
• There is a trade-off between statistical efficiency arising from SGD and computational effi-
ciency arising from processing large batches of data at a time.
• Minibatch stochastic gradient descent offers the best of both worlds: computational and
statistical efficiency.
• In minibatch SGD we process batches of data obtained by a random permutation of the train-
ing data (i.e., each observation is processed only once per epoch, albeit in random order).
• It is advisable to decay the learning rates during training.
• In general, minibatch SGD is faster than SGD and gradient descent for convergence to a
smaller risk, when measured in terms of clock time.
Exercises
1. Modify the batch size and learning rate and observe the rate of decline for the value of the
objective function and the time consumed in each epoch.
2. Read the MXNet documentation and use the Trainer class set_learning_rate function to
reduce the learning rate of the minibatch SGD to 1/10 of its previous value after each epoch.
3. Compare minibatch SGD with a variant that actually samples with replacement from the train-
ing set. What happens?
4. An evil genie replicates your dataset without telling you (i.e., each observation occurs twice
and your dataset grows to twice its original size, but nobody told you). How does the behavior
of SGD, minibatch SGD and that of gradient descent change?
Discussions138
138

11.6 Momentum
In Section 11.4 we reviewed what happens when performing stochastic gradient descent, i.e.,
when performing optimization where only a noisy variant of the gradient is available. In partic-
ular, we noticed that for noisy gradients we need to be extra cautious when it comes to choosing
the learning rate in the face of noise. If we decrease it too rapidly, convergence stalls. If we are
too lenient, we fail to converge to a good enough solution since noise keeps on driving us away
from optimality.
11.6.1 Basics
In this section, we will explore more effective optimization algorithms, especially for certain types
of optimization problems that are common in practice.
Leaky Averages
The previous section saw us discussing minibatch SGD as a means for accelerating computation.
It also had the nice side-effect that averaging gradients reduced the amount of variance. The mini-
batch SGD can be calculated by:
1 ∑ 1 ∑
gt,t−1 = ∂w f (xi , wt−1 ) = hi,t−1 . (11.6.1)
|Bt | |Bt |
i∈Bt i∈Bt
To keep the notation simple, here we used hi,t−1 = ∂w f (xi , wt−1 ) as the SGD for sample i using
the weights updated at time t − 1. It would be nice if we could benefit from the effect of variance
reduction even beyond averaging gradients on a mini-batch. One option to accomplish this task
is to replace the gradient computation by a “leaky average”:
vt = βvt−1 + gt,t−1 (11.6.2)
for some β ∈ (0, 1). This effectively replaces the instantaneous gradient by one thatʼs been aver-
aged over multiple past gradients. v is called momentum. It accumulates past gradients similar to
how a heavy ball rolling down the objective function landscape integrates over past forces. To see
what is happening in more detail let us expand vt recursively into
∑
t−1
2
vt = β vt−2 + βgt−1,t−2 + gt,t−1 = . . . , = β τ gt−τ,t−τ −1 . (11.6.3)
τ =0
Large β amounts to a long-range average, whereas small β amounts to only a slight correction
relative to a gradient method. The new gradient replacement no longer points into the direction
of steepest descent on a particular instance any longer but rather in the direction of a weighted
average of past gradients. This allows us to realize most of the benefits of averaging over a batch
without the cost of actually computing the gradients on it. We will revisit this averaging procedure
in more detail later.
The above reasoning formed the basis for what is now known as accelerated gradient methods,
such as gradients with momentum. They enjoy the additional benefit of being much more effec-
tive in cases where the optimization problem is ill-conditioned (i.e., where there are some direc-
tions where progress is much slower than in others, resembling a narrow canyon). Furthermore,
they allow us to average over subsequent gradients to obtain more stable directions of descent.
11.6. Momentum 445

Indeed, the aspect of acceleration even for noise-free convex problems is one of the key reasons
why momentum works and why it works so well.
As one would expect, due to its efficacy momentum is a well-studied subject in optimization for
deep learning and beyond. See e.g., the beautiful expository article139 by (Goh, 2017) for an in-
depth analysis and interactive animation. It was proposed by (Polyak, 1964). (Nesterov, 2018)
has a detailed theoretical discussion in the context of convex optimization. Momentum in deep
learning has been known to be beneficial for a long time. See e.g., the discussion by (Sutskever et
al., 2013) for details.
An Ill-conditioned Problem
To get a better understanding of the geometric properties of the momentum method we revisit
gradient descent, albeit with a significantly less pleasant objective function. Recall that in Section
11.3 we used f (x) = x21 + 2x22 , i.e., a moderately distorted ellipsoid objective. We distort this
function further by stretching it out in the x1 direction via
f (x) = 0.1x21 + 2x22 . (11.6.4)
As before f has its minimum at (0, 0). This function is very flat in the direction of x1 . Let us
see what happens when we perform gradient descent as before on this new function. We pick a
learning rate of 0.4.
%matplotlib inline
npx.set_np()
eta = 0.4
def f_2d(x1, x2):
return 0.1 * x1 ** 2 + 2 * x2 ** 2
def gd_2d(x1, x2, s1, s2):
return (x1 - eta * 0.2 * x1, x2 - eta * 4 * x2, 0, 0)
d2l.show_trace_2d(f_2d, d2l.train_2d(gd_2d))
139
https://distill.pub/2017/momentum/

By construction, the gradient in the x2 direction is much higher and changes much more rapidly
than in the horizontal x1 direction. Thus we are stuck between two undesirable choices: if we
pick a small learning rate we ensure that the solution does not diverge in the x2 direction but we
are saddled with slow convergence in the x1 direction. Conversely, with a large learning rate we
progress rapidly in the x1 direction but diverge in x2 . The example below illustrates what hap-
pens even after a slight increase in learning rate from 0.4 to 0.6. Convergence in the x1 direction
improves but the overall solution quality is much worse.
eta = 0.6
d2l.show_trace_2d(f_2d, d2l.train_2d(gd_2d))
The Momentum Method
The momentum method allows us to solve the gradient descent problem described above. Look-
ing at the optimization trace above we might intuit that averaging gradients over the past would
work well. After all, in the x1 direction this will aggregate well-aligned gradients, thus increasing
the distance we cover with every step. Conversely, in the x2 direction where gradients oscillate,
an aggregate gradient will reduce step size due to oscillations that cancel each other out. Using vt
instead of the gradient gt yields the following update equations:
vt ← βvt−1 + gt,t−1 ,
(11.6.5)
xt ← xt−1 − ηt vt .
Note that for β = 0 we recover regular gradient descent. Before delving deeper into the mathe-
matical properties let us have a quick look at how the algorithm behaves in practice.
def momentum_2d(x1, x2, v1, v2):

v1 = beta * v1 + 0.2 * x1
v2 = beta * v2 + 4 * x2
return x1 - eta * v1, x2 - eta * v2, v1, v2
eta, beta = 0.6, 0.5

d2l.show_trace_2d(f_2d, d2l.train_2d(momentum_2d))
11.6. Momentum 447

As we can see, even with the same learning rate that we used before, momentum still converges
well. Let us see what happens when we decrease the momentum parameter. Halving it to β =
0.25 leads to a trajectory that barely converges at all. Nonetheless, it is a lot better than without
momentum (when the solution diverges).
eta, beta = 0.6, 0.25

d2l.show_trace_2d(f_2d, d2l.train_2d(momentum_2d))
Note that we can combine momentum with SGD and in particular, minibatch-SGD. The only
change is that in that case we replace the gradients gt,t−1 with gt . Last, for convenience we initial-
ize v0 = 0 at time t = 0. Let us look at what leaky averaging actually does to the updates.

Effective Sample Weight
∑t−1 τ ∑∞ τ 1
Recall that vt = τ =0 β gt−τ,t−τ −1 . In the limit the terms add up to τ =0 β = 1−β . In other
η
words, rather than taking a step of size η in GD or SGD we take a step of size 1−β while at the same
time, dealing with a potentially much better behaved descent direction. These are two benefits in
one. To illustrate how weighting behaves for different choices of β consider the diagram below.
d2l.set_figsize()
betas = [0.95, 0.9, 0.6, 0]
for beta in betas:
x = d2l.numpy(np.arange(40))
d2l.plt.plot(x, beta ** x, label=f'beta = {beta:.2f}')
d2l.plt.xlabel('time')
d2l.plt.legend();
11.6.2 Practical Experiments
Let us see how momentum works in practice, i.e., when used within the context of a proper opti-
mizer. For this we need a somewhat more scalable implementation.
Implementation from Scratch
Compared with (minibatch) SGD the momentum method needs to maintain a set of auxiliary vari-
ables, i.e., velocity. It has the same shape as the gradients (and variables of the optimization prob-
lem). In the implementation below we call these variables states.
def init_momentum_states(feature_dim):
v_w = np.zeros((feature_dim, 1))
v_b = np.zeros(1)
return (v_w, v_b)
def sgd_momentum(params, states, hyperparams):

for p, v in zip(params, states):
v[:] = hyperparams['momentum'] * v + p.grad
p[:] -= hyperparams['lr'] * v
11.6. Momentum 449

Let us see how this works in practice.
def train_momentum(lr, momentum, num_epochs=2):

d2l.train_ch11(sgd_momentum, init_momentum_states(feature_dim),
{'lr': lr, 'momentum': momentum}, data_iter,
feature_dim, num_epochs)
data_iter, feature_dim = d2l.get_data_ch11(batch_size=10)

train_momentum(0.02, 0.5)
When we increase the momentum hyperparameter momentum to 0.9, it amounts to a significantly

1
larger effective sample size of 1−0.9 = 10. We reduce the learning rate slightly to 0.01 to keep
matters under control.

Reducing the learning rate further addresses any issue of non-smooth optimization problems.
Setting it to 0.005 yields good convergence properties.
Concise Implementation
There is very little to do in Gluon since the standard sgd solver already had momentum built in.
Setting matching parameters yields a very similar trajectory.
d2l.train_concise_ch11('sgd', {'learning_rate': 0.005, 'momentum': 0.9},

data_iter)
11.6. Momentum 451

11.6.3 Theoretical Analysis
So far the 2D example of f (x) = 0.1x21 + 2x22 seemed rather contrived. We will now see that this is
actually quite representative of the types of problem one might encounter, at least in the case of
minimizing convex quadratic objective functions.
Quadratic Convex Functions
Consider the function

1
h(x) = x⊤ Qx + x⊤ c + b. (11.6.6)
2
This is a general quadratic function. For positive definite matrices Q ≻ 0, i.e., for matrices with
positive eigenvalues this has a minimizer at x∗ = −Q−1 c with minimum value b− 12 c⊤ Q−1 c. Hence
we can rewrite h as
1 1
h(x) = (x − Q−1 c)⊤ Q(x − Q−1 c) + b − c⊤ Q−1 c. (11.6.7)
2 2
The gradient is given by ∂x f (x) = Q(x − Q−1 c). That is, it is given by the distance between x
and the minimizer, multiplied by Q. Consequently also the momentum is a linear combination of
terms Q(xt − Q−1 c).
Since Q is positive definite it can be decomposed into its eigensystem via Q = O⊤ ΛO for an or-
thogonal (rotation) matrix O and a diagonal matrix Λ of positive eigenvalues. This allows us to
perform a change of variables from x to z := O(x − Q−1 c) to obtain a much simplified expression:
1
h(z) = z⊤ Λz + b′ . (11.6.8)
2
Here c′ = b − 12 c⊤ Q−1 c. Since O is only an orthogonal matrix this does not perturb the gradients
in a meaningful way. Expressed in terms of z gradient descent becomes
zt = zt−1 − Λzt−1 = (I − Λ)zt−1 . (11.6.9)
The important fact in this expression is that gradient descent does not mix between different
eigenspaces. That is, when expressed in terms of the eigensystem of Q the optimization prob-
lem proceeds in a coordinate-wise manner. This also holds for momentum.
vt = βvt−1 + Λzt−1
zt = zt−1 − η (βvt−1 + Λzt−1 ) (11.6.10)
= (I − ηΛ)zt−1 − ηβvt−1 .
In doing this we just proved the following theorem: Gradient Descent with and without momen-
tum for a convex quadratic function decomposes into coordinate-wise optimization in the direc-
tion of the eigenvectors of the quadratic matrix.

Scalar Functions
λ 2
Given the above result let us see what happens when we minimize the function f (x) = 2x . For
gradient descent we have
xt+1 = xt − ηλxt = (1 − ηλ)xt . (11.6.11)
Whenever |1 − ηλ| < 1 this optimization converges at an exponential rate since after t steps we
have xt = (1 − ηλ)t x0 . This shows how the rate of convergence improves initially as we increase
the learning rate η until ηλ = 1. Beyond that things diverge and for ηλ > 2 the optimization
problem diverges.
lambdas = [0.1, 1, 10, 19]

eta = 0.1
for lam in lambdas:
t = d2l.numpy(np.arange(20))
d2l.plt.plot(t, (1 - eta * lam) ** t, label=f'lambda = {lam:.2f}')
d2l.plt.xlabel('time')
d2l.plt.legend();
To analyze convergence in the case of momentum we begin by rewriting the update equations in
terms of two scalars: one for x and one for the momentum v. This yields:
[ ] [ ][ ] [ ]
vt+1 β λ vt v
= = R(β, η, λ) t . (11.6.12)
xt+1 −ηβ (1 − ηλ) xt xt
We used R to denote the 2 × 2 governing convergence behavior. After t steps the initial choice
[v0 , x0 ] becomes R(β, η, λ)t [v0 , x0 ]. Hence, it is up to the eigenvalues of R to determine the speed
of convergence. See the Distill post140 of (Goh, 2017) for a great animation and (Flammarion &
140
https://distill.pub/2017/momentum/
11.6. Momentum 453

Bach, 2015) for a detailed analysis. One can show that 0 < ηλ < 2 + 2β momentum converges.
This is a larger range of feasible parameters when compared to 0 < ηλ < 2 for gradient descent. It
also suggests that in general large values of β are desirable. Further details require a fair amount
of technical detail and we suggest that the interested reader consult the original publications.
Summary
• Momentum replaces gradients with a leaky average over past gradients. This accelerates
convergence significantly.
• It is desirable for both noise-free gradient descent and (noisy) stochastic gradient descent.
• Momentum prevents stalling of the optimization process that is much more likely to occur
for stochastic gradient descent.
1
• The effective number of gradients is given by 1−β due to exponentiated downweighting of
past data.
• In the case of convex quadratic problems this can be analyzed explicitly in detail.
• Implementation is quite straightforward but it requires us to store an additional state vector
(momentum v).
Exercises
1. Use other combinations of momentum hyperparameters and learning rates and observe and
analyze the different experimental results.
2. Try out GD and momentum for a quadratic problem where you have multiple eigenvalues,
1∑
i.e., f (x) = 2 i λi x2i , e.g., λi = 2−i . Plot how the values of x decrease for the initialization
xi = 1.
3. Derive minimum value and minimizer for h(x) = 12 x⊤ Qx + x⊤ c + b.
4. What changes when we perform SGD with momentum? What happens when we use mini-
batch SGD with momentum? Experiment with the parameters?
Discussions141
11.7 Adagrad
Let us begin by considering learning problems with features that occur infrequently.
141

11.7.1 Sparse Features and Learning Rates
Imagine that we are training a language model. To get good accuracy we typically want to decrease
the learning rate as we keep on training, usually at a rate of O(t− 2 ) or slower. Now consider a
1
model training on sparse features, i.e., features that occur only infrequently. This is common for
natural language, e.g., it is a lot less likely that we will see the word preconditioning than learning.
However, it is also common in other areas such as computational advertising and personalized
collaborative filtering. After all, there are many things that are of interest only for a small number
of people.
Parameters associated with infrequent features only receive meaningful updates whenever these
features occur. Given a decreasing learning rate we might end up in a situation where the parame-
ters for common features converge rather quickly to their optimal values, whereas for infrequent
features we are still short of observing them sufficiently frequently before their optimal values can
be determined. In other words, the learning rate either decreases too slowly for frequent features
or too quickly for infrequent ones.
A possible hack to redress this issue would be to count the number of times we see a particular
feature and to use this as a clock for adjusting learning rates. That is, rather than choosing a
learning rate of the form η = √ηt+c
0
we could use ηi = √ η0 . Here s(i, t) counts the number of
s(i,t)+c
nonzeros for feature i that we have observed up to time t. This is actually quite easy to implement
at no meaningful overhead. However, it fails whenever we do not quite have sparsity but rather
just data where the gradients are often very small and only rarely large. After all, it is unclear
where one would draw the line between something that qualifies as an observed feature or not.
Adagrad by (Duchi et al., 2011) addresses this by replacing the rather crude counter s(i, t) by an
aggregate of the squares of previously observed gradients. In particular, it uses s(i, t + 1) =
s(i, t) + (∂i f (x))2 as a means to adjust the learning rate. This has two benefits: first, we no longer
need to decide just when a gradient is large enough. Second, it scales automatically with the mag-
nitude of the gradients. Coordinates that routinely correspond to large gradients are scaled down
significantly, whereas others with small gradients receive a much more gentle treatment. In prac-
tice this leads to a very effective optimization procedure for computational advertising and related
problems. But this hides some of the additional benefits inherent in Adagrad that are best under-
stood in the context of preconditioning.
11.7.2 Preconditioning
Convex optimization problems are good for analyzing the characteristics of algorithms. After all,
for most nonconvex problems it is difficult to derive meaningful theoretical guarantees, but intu-
ition and insight often carry over. Let us look at the problem of minimizing f (x) = 12 x⊤ Qx+c⊤ x+b.
As we saw in Section 11.6, it is possible to rewrite this problem in terms of its eigendecompo-
sition Q = U⊤ ΛU to arrive at a much simplified problem where each coordinate can be solved
individually:
1
f (x) = f¯(x̄) = x̄⊤ Λx̄ + c̄⊤ x̄ + b. (11.7.1)
2
Here we used x = Ux and consequently c = Uc. The modified problem has as its minimizer
x̄ = −Λ−1 c̄ and minimum value − 21 c̄⊤ Λ−1 c̄ + b. This is much easier to compute since Λ is a
diagonal matrix containing the eigenvalues of Q.
If we perturb c slightly we would hope to find only slight changes in the minimizer of f . Unfortu-
nately this is not the case. While slight changes in c lead to equally slight changes in c̄, this is not
11.7. Adagrad 455

the case for the minimizer of f (and of f¯ respectively). Whenever the eigenvalues Λi are large we
will see only small changes in x̄i and in the minimum of f¯. Conversely, for small Λi changes in x̄i
can be dramatic. The ratio between the largest and the smallest eigenvalue is called the condition
number of an optimization problem.
Λ1
κ= . (11.7.2)
Λd
If the condition number κ is large, it is difficult to solve the optimization problem accurately. We
need to ensure that we are careful in getting a large dynamic range of values right. Our analysis
leads to an obvious, albeit somewhat naive question: couldnʼt we simply “fix” the problem by dis-
torting the space such that all eigenvalues are 1. In theory this is quite easy: we only need the
1
eigenvalues and eigenvectors of Q to rescale the problem from x to one in z := Λ 2 Ux. In the
new coordinate system x⊤ Qx could be simplified to ∥z∥2 . Alas, this is a rather impractical sugges-
tion. Computing eigenvalues and eigenvectors is in general much more expensive than solving the
actual problem.
While computing eigenvalues exactly might be expensive, guessing them and computing them
even somewhat approximately may already be a lot better than not doing anything at all. In par-
ticular, we could use the diagonal entries of Q and rescale it accordingly. This is much cheaper
than computing eigenvalues.
Q̃ = diag− 2 (Q)Qdiag− 2 (Q).

1 1
(11.7.3)
√
In this case we have Q̃ij = Qij / Qii Qjj and specifically Q̃ii = 1 for all i. In most cases this
simplifies the condition number considerably. For instance, the cases we discussed previously,
this would entirely eliminate the problem at hand since the problem is axis aligned.
Unfortunately we face yet another problem: in deep learning we typically do not even have access
to the second derivative of the objective function: for x ∈ Rd the second derivative even on a
minibatch may require O(d2 ) space and work to compute, thus making it practically infeasible.
The ingenious idea of Adagrad is to use a proxy for that elusive diagonal of the Hessian that is both
relatively cheap to compute and effective—the magnitude of the gradient itself.
In order to see why this works, let us look at f¯(x̄). We have that
∂x̄ f¯(x̄) = Λx̄ + c̄ = Λ (x̄ − x̄0 ) , (11.7.4)
where x̄0 is the minimizer of f¯. Hence the magnitude of the gradient depends both on Λ and the
distance from optimality. If x̄ − x̄0 didnʼt change, this would be all thatʼs needed. After all, in this
case the magnitude of the gradient ∂x̄ f¯(x̄) suffices. Since AdaGrad is a stochastic gradient descent
algorithm, we will see gradients with nonzero variance even at optimality. As a result we can safely
use the variance of the gradients as a cheap proxy for the scale of the Hessian. A thorough analysis
is beyond the scope of this section (it would be several pages). We refer the reader to (Duchi et al.,
2011) for details.

11.7.3 The Algorithm
Let us formalize the discussion from above. We use the variable st to accumulate past gradient
variance as follows.
gt = ∂w l(yt , f (xt , w)),
st = st−1 + g2t , (11.7.5)
η
wt = wt−1 − √ · gt .
st + ϵ
Here the operation are applied coordinate wise. That is, v2 has entries vi2 . Likewise √1 has entries
v
√1
viand u · v has entries ui vi . As before η is the learning rate and ϵ is an additive constant that
ensures that we do not divide by 0. Last, we initialize s0 = 0.
Just like in the case of momentum we need to keep track of an auxiliary variable, in this case to
allow for an individual learning rate per coordinate. This does not increase the cost of Adagrad
significantly relative to SGD, simply since the main cost is typically to compute l(yt , f (xt , w)) and
its derivative.
Note that accumulating squared gradients in st means that st grows essentially at linear rate (some-
what slower than linearly in practice, since the gradients initially diminish). This leads to an
O(t− 2 ) learning rate, albeit adjusted on a per coordinate basis. For convex problems this is per-
1
fectly adequate. In deep learning, though, we might want to decrease the learning rate rather
more slowly. This led to a number of Adagrad variants that we will discuss in the subsequent
chapters. For now let us see how it behaves in a quadratic convex problem. We use the same
problem as before:
f (x) = 0.1x21 + 2x22 . (11.7.6)
We are going to implement Adagrad using the same learning rate previously, i.e., η = 0.4. As
we can see, the iterative trajectory of the independent variable is smoother. However, due to the
cumulative effect of st , the learning rate continuously decays, so the independent variable does
not move as much during later stages of iteration.
%matplotlib inline
import math
npx.set_np()
def adagrad_2d(x1, x2, s1, s2):

eps = 1e-6
g1, g2 = 0.2 * x1, 4 * x2
s1 += g1 ** 2
s2 += g2 ** 2
x1 -= eta / math.sqrt(s1 + eps) * g1
return x1, x2, s1, s2
def f_2d(x1, x2):

return 0.1 * x1 ** 2 + 2 * x2 ** 2
eta = 0.4
d2l.show_trace_2d(f_2d, d2l.train_2d(adagrad_2d))
11.7. Adagrad 457

As we increase the learning rate to 2 we see much better behavior. This already indicates that the
decrease in learning rate might be rather aggressive, even in the noise-free case and we need to
ensure that parameters converge appropriately.
eta = 2
d2l.show_trace_2d(f_2d, d2l.train_2d(adagrad_2d))
Just like the momentum method, Adagrad needs to maintain a state variable of the same shape as
the parameters.
def init_adagrad_states(feature_dim):
s_w = np.zeros((feature_dim, 1))
s_b = np.zeros(1)
return (s_w, s_b)
def adagrad(params, states, hyperparams):

eps = 1e-6
for p, s in zip(params, states):

s[:] += np.square(p.grad)
p[:] -= hyperparams['lr'] * p.grad / np.sqrt(s + eps)
Compared to the experiment in Section 11.5 we use a larger learning rate to train the model.

d2l.train_ch11(adagrad, init_adagrad_states(feature_dim),
{'lr': 0.1}, data_iter, feature_dim);
Using the Trainer instance of the algorithm adagrad, we can invoke the Adagrad algorithm in
Gluon.
d2l.train_concise_ch11('adagrad', {'learning_rate': 0.1}, data_iter)
11.7. Adagrad 459

Summary
• Adagrad decreases the learning rate dynamically on a per-coordinate basis.

• It uses the magnitude of the gradient as a means of adjusting how quickly progress is
achieved - coordinates with large gradients are compensated with a smaller learning rate.
• Computing the exact second derivative is typically infeasible in deep learning problems due
to memory and computational constraints. The gradient can be a useful proxy.
• If the optimization problem has a rather uneven uneven structure Adagrad can help mitigate
the distortion.
• Adagrad is particularly effective for sparse features where the learning rate needs to decrease
more slowly for infrequently occurring terms.
• On deep learning problems Adagrad can sometimes be too aggressive in reducing learning
rates. We will discuss strategies for mitigating this in the context of Section 11.10.
Exercises
1. Prove that for an orthogonal matrix U and a vector c the following holds: ∥c − δ∥2 = ∥Uc −
Uδ∥2 . Why does this mean that the magnitude of perturbations does not change after an
orthogonal change of variables?
2. Try out Adagrad for f (x) = 0.1x21 + 2x22 and also for the objective function was rotated by 45
degrees, i.e., f (x) = 0.1(x1 + x2 )2 + 2(x1 − x2 )2 . Does it behave differently?
3. Prove Gerschgorinʼs
∑ circle theorem142 which states that eigenvalues λi of a matrix M satisfy
|λi − Mjj | ≤ k̸=j |Mjk | for at least one choice of j.
4. What does Gerschgorinʼs theorem tell us about the eigenvalues of the diagonally precondi-
tioned matrix diag− 2 (M)Mdiag− 2 (M)?
1 1
5. Try out Adagrad for a proper deep network, such as Section 6.6 when applied to Fashion
MNIST.
6. How would you need to modify Adagrad to achieve a less aggressive decay in learning rate?
142
https://en.wikipedia.org/wiki/Gershgorin_circle_theorem

Discussions143
11.8 RMSProp
One of the key issues in Section 11.7 is that the learning rate decreases at a predefined schedule
of effectively O(t− 2 ). While this is generally appropriate for convex problems, it might not be
1
ideal for nonconvex ones, such as those encountered in deep learning. Yet, the coordinate-wise
adaptivity of Adagrad is highly desirable as a preconditioner.
(Tieleman & Hinton, 2012) proposed the RMSProp algorithm as a simple fix to decouple rate
scheduling from coordinate-adaptive learning rates. The issue is that Adagrad accumulates the
squares of the gradient gt into a state vector st = st−1 + g2t . As a result st keeps on growing without
bound due to the lack of normalization, essentially linearly as the algorithm converges.
One way of fixing this problem would be to use st /t. For reasonable distributions of gt this will
converge. Unfortunately it might take a very long time until the limit behavior starts to matter
since the procedure remembers the full trajectory of values. An alternative is to use a leaky av-
erage in the same way we used in the momentum method, i.e., st ← γst−1 + (1 − γ)g2t for some
parameter γ > 0. Keeping all other parts unchanged yields RMSProp.
Let us write out the equations in detail.
st ← γst−1 + (1 − γ)g2t ,
η (11.8.1)
xt ← xt−1 − √ ⊙ gt .
st + ϵ
The constant ϵ > 0 is typically set to 10−6 to ensure that we do not suffer from division by zero
or overly large step sizes. Given this expansion we are now free to control the learning rate η
independently of the scaling that is applied on a per-coordinate basis. In terms of leaky averages
we can apply the same reasoning as previously applied in the case of the momentum method.
Expanding the definition of st yields
st = (1 − γ)g2t + γst−1
( ) (11.8.2)
= (1 − γ) g2t + γg2t−1 + γ 2 gt−2 + . . . , .
1
As before in Section 11.6 we use 1+γ +γ 2 +. . . , = 1−γ . Hence the sum of weights is normalized to 1
−1
with a half-life time of an observation of γ . Let us visualize the weights for the past 40 timesteps
for various choices of γ.
%matplotlib inline
import math
npx.set_np()
143
11.8. RMSProp 461

d2l.set_figsize()
gammas = [0.95, 0.9, 0.8, 0.7]
for gamma in gammas:
x = d2l.numpy(np.arange(40))
d2l.plt.plot(x, (1-gamma) * gamma ** x, label=f'gamma = {gamma:.2f}')
d2l.plt.xlabel('time');
As before we use the quadratic function f (x) = 0.1x21 + 2x22 to observe the trajectory of RMSProp.
Recall that in Section 11.7, when we used Adagrad with a learning rate of 0.4, the variables moved
only very slowly in the later stages of the algorithm since the learning rate decreased too quickly.
Since η is controlled separately this does not happen with RMSProp.
def rmsprop_2d(x1, x2, s1, s2):

g1, g2, eps = 0.2 * x1, 4 * x2, 1e-6
s1 = gamma * s1 + (1 - gamma) * g1 ** 2
s2 = gamma * s2 + (1 - gamma) * g2 ** 2
return x1, x2, s1, s2
def f_2d(x1, x2):

return 0.1 * x1 ** 2 + 2 * x2 ** 2
eta, gamma = 0.4, 0.9

d2l.show_trace_2d(f_2d, d2l.train_2d(rmsprop_2d))

Next, we implement RMSProp to be used in a deep network. This is equally straightforward.
def init_rmsprop_states(feature_dim):
s_w = np.zeros((feature_dim, 1))
s_b = np.zeros(1)
return (s_w, s_b)
def rmsprop(params, states, hyperparams):

gamma, eps = hyperparams['gamma'], 1e-6
for p, s in zip(params, states):
s[:] = gamma * s + (1 - gamma) * np.square(p.grad)
p[:] -= hyperparams['lr'] * p.grad / np.sqrt(s + eps)
We set the initial learning rate to 0.01 and the weighting term γ to 0.9. That is, s aggregates on
average over the past 1/(1 − γ) = 10 observations of the square gradient.

d2l.train_ch11(rmsprop, init_rmsprop_states(feature_dim),
{'lr': 0.01, 'gamma': 0.9}, data_iter, feature_dim);
11.8. RMSProp 463

Since RMSProp is a rather popular algorithm it is also available in the Trainer instance. All we
need to do is instantiate it using an algorithm named rmsprop, assigning γ to the parameter gamma1.
d2l.train_concise_ch11('rmsprop', {'learning_rate': 0.01, 'gamma1': 0.9},

data_iter)
Summary
• RMSProp is very similar to Adagrad insofar as both use the square of the gradient to scale
coefficients.
• RMSProp shares with momentum the leaky averaging. However, RMSProp uses the tech-
nique to adjust the coefficient-wise preconditioner.
• The learning rate needs to be scheduled by the experimenter in practice.
• The coefficient γ determines how long the history is when adjusting the per-coordinate
scale.
Exercises
1. What happens experimentally if we set γ = 1? Why?

2. Rotate the optimization problem to minimize f (x) = 0.1(x1 + x2 )2 + 2(x1 − x2 )2 . What
happens to the convergence?
3. Try out what happens to RMSProp on a real machine learning problem, such as training on
Fashion-MNIST. Experiment with different choices for adjusting the learning rate.
4. Would you want to adjust γ as optimization progresses? How sensitive is RMSProp to this?
Discussions144
144

11.9 Adadelta
Adadelta is yet another variant of AdaGrad (Section 11.7). The main difference lies in the fact that
it decreases the amount by which the learning rate is adaptive to coordinates. Moreover, tradi-
tionally it referred to as not having a learning rate since it uses the amount of change itself as
calibration for future change. The algorithm was proposed in (Zeiler, 2012). It is fairly straight-
forward, given the discussion of previous algorithms so far.
In a nutshell, Adadelta uses two state variables, st to store a leaky average of the second moment
of the gradient and ∆xt to store a leaky average of the second moment of the change of parameters
in the model itself. Note that we use the original notation and naming of the authors for compati-
bility with other publications and implementations (there is no other real reason why one should
use different Greek variables to indicate a parameter serving the same purpose in momentum,
Adagrad, RMSProp, and Adadelta).
Here are the technical details of Adadelta. Given the parameter du jour is ρ, we obtain the follow-
ing leaky updates similarly to Section 11.8:
st = ρst−1 + (1 − ρ)g2t . (11.9.1)
The difference to Section 11.8 is that we perform updates with the rescaled gradient g′t , i.e.,
xt = xt−1 − g′t . (11.9.2)
So what is the rescaled gradient g′t ? We can calculate it as follows:

√
′ ∆xt−1 + ϵ
gt = √ ⊙ gt , (11.9.3)
st + ϵ
where ∆xt−1 is the leaky average of the squared rescaled gradients g′t . We initialize ∆x0 to be 0
and update it at each step with g′t , i.e.,
∆xt = ρ∆xt−1 + (1 − ρ)g′t ,

2
(11.9.4)
and ϵ (a small value such as 10−5 ) is added to maintain numerical stability.
11.9.2 Implementation
Adadelta needs to maintain two state variables for each variable, st and ∆xt . This yields the fol-
lowing implementation.
%matplotlib inline
npx.set_np()
def init_adadelta_states(feature_dim):
s_w, s_b = np.zeros((feature_dim, 1)), np.zeros(1)
delta_w, delta_b = np.zeros((feature_dim, 1)), np.zeros(1)
11.9. Adadelta 465

return ((s_w, delta_w), (s_b, delta_b))
def adadelta(params, states, hyperparams):

rho, eps = hyperparams['rho'], 1e-5
for p, (s, delta) in zip(params, states):
# In-place updates via [:]
s[:] = rho * s + (1 - rho) * np.square(p.grad)
g = (np.sqrt(delta + eps) / np.sqrt(s + eps)) * p.grad
p[:] -= g
delta[:] = rho * delta + (1 - rho) * g * g
Choosing ρ = 0.9 amounts to a half-life time of 10 for each parameter update. This tends to work
quite well. We get the following behavior.

d2l.train_ch11(adadelta, init_adadelta_states(feature_dim),
{'rho': 0.9}, data_iter, feature_dim);
For a concise implementation we simply use the adadelta algorithm from the Trainer class. This
yields the following one-liner for a much more compact invocation.
d2l.train_concise_ch11('adadelta', {'rho': 0.9}, data_iter)

Summary
• Adadelta has no learning rate parameter. Instead, it uses the rate of change in the parame-
ters itself to adapt the learning rate.
• Adadelta requires two state variables to store the second moments of gradient and the
change in parameters.
• Adadelta uses leaky averages to keep a running estimate of the appropriate statistics.
Exercises
1. Adjust the value of ρ. What happens?

2. Show how to implement the algorithm without the use of g′t . Why might this be a good idea?
3. Is Adadelta really learning rate free? Could you find optimization problems that break
Adadelta?
4. Compare Adadelta to Adagrad and RMS prop to discuss their convergence behavior.
Discussions145
11.10 Adam
In the discussions leading up to this section we encountered a number of techniques for efficient
optimization. Let us recap them in detail here:
• We saw that Section 11.4 is more effective than Gradient Descent when solving optimization
problems, e.g., due to its inherent resilience to redundant data.
• We saw that Section 11.5 affords significant additional efficiency arising from vectorization,
using larger sets of observations in one minibatch. This is the key to efficient multi-machine,
multi-GPU and overall parallel processing.
145
11.10. Adam 467

• Section 11.6 added a mechanism for aggregating a history of past gradients to accelerate
convergence.
• Section 11.7 used per-coordinate scaling to allow for a computationally efficient precondi-
tioner.
• Section 11.8 decoupled per-coordinate scaling from a learning rate adjustment.
Adam (Kingma & Ba, 2014) combines all these techniques into one efficient learning algorithm.
As expected, this is an algorithm that has become rather popular as one of the more robust and
effective optimization algorithms to use in deep learning. It is not without issues, though. In
particular, (Reddi et al., 2019) show that there are situations where Adam can diverge due to poor
variance control. In a follow-up work (Zaheer et al., 2018) proposed a hotfix to Adam, called Yogi
which addresses these issues. More on this later. For now let us review the Adam algorithm.
One of the key components of Adam is that it uses exponential weighted moving averages (also
known as leaky averaging) to obtain an estimate of both the momentum and also the second mo-
ment of the gradient. That is, it uses the state variables
vt ← β1 vt−1 + (1 − β1 )gt ,
(11.10.1)
st ← β2 st−1 + (1 − β2 )g2t .
Here β1 and β2 are nonnegative weighting parameters. Common choices for them are β1 = 0.9
and β2 = 0.999. That is, the variance estimate moves much more slowly than the momentum term.
Note that if we initialize v0 = s0 = 0 we have a significant amount of bias initially towards smaller
∑ t
values. This can be addressed by using the fact that ti=0 β i = 1−β
1−β to re-normalize terms. Cor-
respondingly the normalized state variables are given by
vt st
v̂t = and ŝt = . (11.10.2)
1 − β1t 1 − β2t
Armed with the proper estimates we can now write out the update equations. First, we rescale the
gradient in a manner very much akin to that of RMSProp to obtain
η v̂t
g′t = √ . (11.10.3)
ŝt + ϵ
Unlike RMSProp our update uses the momentum v̂t rather than the gradient itself. Moreover,
there is a slight cosmetic difference as the rescaling happens using √ŝ1+ϵ instead of √ŝ1+ϵ . The
t t
former works arguably slightly better in practice, hence the deviation from RMSProp. Typically
we pick ϵ = 10−6 for a good trade-off between numerical stability and fidelity.
Now we have all the pieces in place to compute updates. This is slightly anticlimactic and we have
a simple update of the form
xt ← xt−1 − g′t . (11.10.4)
Reviewing the design of Adam its inspiration is clear. Momentum and scale are clearly visible in
the state variables. Their rather peculiar definition forces us to debias terms (this could be fixed
by a slightly different initialization and update condition). Second, the combination of both terms
is pretty straightforward, given RMSProp. Last, the explicit learning rate η allows us to control the
step length to address issues of convergence.

11.10.2 Implementation
Implementing Adam from scratch is not very daunting. For convenience we store the timestep
counter t in the hyperparams dictionary. Beyond that all is straightforward.
%matplotlib inline
npx.set_np()
def init_adam_states(feature_dim):
v_w, v_b = np.zeros((feature_dim, 1)), np.zeros(1)
s_w, s_b = np.zeros((feature_dim, 1)), np.zeros(1)
return ((v_w, s_w), (v_b, s_b))
def adam(params, states, hyperparams):

beta1, beta2, eps = 0.9, 0.999, 1e-6
for p, (v, s) in zip(params, states):
v[:] = beta1 * v + (1 - beta1) * p.grad
s[:] = beta2 * s + (1 - beta2) * np.square(p.grad)
v_bias_corr = v / (1 - beta1 ** hyperparams['t'])
s_bias_corr = s / (1 - beta2 ** hyperparams['t'])
p[:] -= hyperparams['lr'] * v_bias_corr / (np.sqrt(s_bias_corr) + eps)
hyperparams['t'] += 1
We are ready to use Adam to train the model. We use a learning rate of η = 0.01.

d2l.train_ch11(adam, init_adam_states(feature_dim),
{'lr': 0.01, 't': 1}, data_iter, feature_dim);
A more concise implementation is straightforward since adam is one of the algorithms provided
as part of the Gluon trainer optimization library. Hence we only need to pass configuration pa-
rameters for an implementation in Gluon.
11.10. Adam 469

d2l.train_concise_ch11('adam', {'learning_rate': 0.01}, data_iter)
11.10.3 Yogi
One of the problems of Adam is that it can fail to converge even in convex settings when the second
moment estimate in st blows up. As a fix (Zaheer et al., 2018) proposed a refined update (and
initialization) for st . To understand whatʼs going on, let us rewrite the Adam update as follows:
( )
st ← st−1 + (1 − β2 ) g2t − st−1 . (11.10.5)
Whenever g2t has high variance or updates are sparse, st might forget past values too quickly. A
possible fix for this is to replace g2t − st−1 by g2t ⊙ sgn(g2t − st−1 ). Now the magnitude of the update
no longer depends on the amount of deviation. This yields the Yogi updates
st ← st−1 + (1 − β2 )g2t ⊙ sgn(g2t − st−1 ). (11.10.6)
The authors furthermore advise to initialize the momentum on a larger initial batch rather than
just initial pointwise estimate. We omit the details since they are not material to the discussion
and since even without this convergence remains pretty good.
def yogi(params, states, hyperparams):

beta1, beta2, eps = 0.9, 0.999, 1e-3
for p, (v, s) in zip(params, states):
v[:] = beta1 * v + (1 - beta1) * p.grad
s[:] = s + (1 - beta2) * np.sign(
np.square(p.grad) - s) * np.square(p.grad)
v_bias_corr = v / (1 - beta1 ** hyperparams['t'])
s_bias_corr = s / (1 - beta2 ** hyperparams['t'])
p[:] -= hyperparams['lr'] * v_bias_corr / (np.sqrt(s_bias_corr) + eps)
hyperparams['t'] += 1

d2l.train_ch11(yogi, init_adam_states(feature_dim),
{'lr': 0.01, 't': 1}, data_iter, feature_dim);

Summary
• Adam combines features of many optimization algorithms into a fairly robust update rule.
• Created on the basis of RMSProp, Adam also uses EWMA on the minibatch stochastic gradi-
ent
• Adam uses bias correction to adjust for a slow startup when estimating momentum and a
second moment.
• For gradients with significant variance we may encounter issues with convergence. They
can be amended by using larger minibatches or by switching to an improved estimate for st .
Yogi offers such an alternative.
Exercises
1. Adjust the learning rate and observe and analyze the experimental results.
2. Can you rewrite momentum and second moment updates such that it does not require bias
correction?
3. Why do you need to reduce the learning rate η as we converge?
4. Try to construct a case for which Adam diverges and Yogi converges?
Discussions146
146
11.10. Adam 471

11.11 Learning Rate Scheduling
So far we primarily focused on optimization algorithms for how to update the weight vectors rather
than on the rate at which they are being updated. Nonetheless, adjusting the learning rate is often
just as important as the actual algorithm. There are a number of aspects to consider:
• Most obviously the magnitude of the learning rate matters. If it is too large, optimization
diverges, if it is too small, it takes too long to train or we end up with a suboptimal result.
We saw previously that the condition number of the problem matters (see e.g., Section 11.6
for details). Intuitively it is the ratio of the amount of change in the least sensitive direction
vs. the most sensitive one.
• Secondly, the rate of decay is just as important. If the learning rate remains large we may
simply end up bouncing around the minimum and thus not reach optimality. Section 11.5
discussed this in some detail and we analyzed performance guarantees in Section 11.4. In
short, we want the rate to decay, but probably more slowly than O(t− 2 ) which would be a
1
good choice for convex problems.

• Another aspect that is equally important is initialization. This pertains both to how the pa-
rameters are set initially (review Section 4.8 for details) and also how they evolve initially.
This goes under the moniker of warmup, i.e., how rapidly we start moving towards the so-
lution initially. Large steps in the beginning might not be beneficial, in particular since the
initial set of parameters is random. The initial update directions might be quite meaning-
less, too.
• Lastly, there are a number of optimization variants that perform cyclical learning rate ad-
justment. This is beyond the scope of the current chapter. We recommend the reader to
review details in (Izmailov et al., 2018), e.g., how to obtain better solutions by averaging
over an entire path of parameters.
Given the fact that there is a lot of detail needed to manage learning rates, most deep learning
frameworks have tools to deal with this automatically. In the current chapter we will review the
effects that different schedules have on accuracy and also show how this can be managed effi-
ciently via a learning rate scheduler.
11.11.1 Toy Problem
We begin with a toy problem that is cheap enough to compute easily, yet sufficiently nontrivial to
illustrate some of the key aspects. For that we pick a slightly modernized version of LeNet (relu
instead of sigmoid activation, MaxPooling rather than AveragePooling), as applied to Fashion-
MNIST. Moreover, we hybridize the network for performance. Since most of the code is standard
we just introduce the basics without further detailed discussion. See Chapter 6 for a refresher as
needed.
%matplotlib inline
from mxnet import autograd, gluon, init, lr_scheduler, np, npx
npx.set_np()
net = nn.HybridSequential()
net.add(nn.Conv2D(channels=6, kernel_size=5, padding=2, activation='relu'),

nn.Conv2D(channels=16, kernel_size=5, activation='relu'),
nn.Dense(120, activation='relu'),
nn.Dense(10))
net.hybridize()
batch_size = 256
train_iter, test_iter = d2l.load_data_fashion_mnist(batch_size=batch_size)
# The code is almost identical to `d2l.train_ch6` that defined in the lenet

# section of chapter convolutional neural networks
def train(net, train_iter, test_iter, num_epochs, loss, trainer, device):
metric = d2l.Accumulator(3) # train_loss, train_acc, num_examples
for i, (X, y) in enumerate(train_iter):
y_hat = net(X)
l = loss(y_hat, y)
l.backward()
metric.add(l.sum(), d2l.accuracy(y_hat, y), X.shape[0])
train_loss = metric[0] / metric[2]
train_acc = metric[1] / metric[2]
if (i + 1) % 50 == 0:
animator.add(epoch + i / len(train_iter),
(train_loss, train_acc, None))
test_acc = d2l.evaluate_accuracy_gpu(net, test_iter)
print(f'train loss {train_loss:.3f}, train acc {train_acc:.3f}, '
f'test acc {test_acc:.3f}')
Let us have a look at what happens if we invoke this algorithm with default settings, such as a
learning rate of 0.3 and train for 30 iterations. Note how the training accuracy keeps on increas-
ing while progress in terms of test accuracy stalls beyond a point. The gap between both curves
indicates overfitting.

train(net, train_iter, test_iter, num_epochs, loss, trainer, device)
train loss 0.141, train acc 0.947, test acc 0.884
11.11. Learning Rate Scheduling 473

11.11.2 Schedulers
One way of adjusting the learning rate is to set it explicitly at each step. This is conveniently
achieved by the set_learning_rate method. We could adjust it downward after every epoch (or
even after every minibatch), e.g., in a dynamic manner in response to how optimization is pro-
gressing.
trainer.set_learning_rate(0.1)
print(f'learning rate is now {trainer.learning_rate:.2f}')
learning rate is now 0.10
More generally we want to define a scheduler. When invoked with the number of updates it returns
the appropriate value of the learning rate. Let us define a simple one that sets the learning rate to
η = η0 (t + 1)− 2 .
1
class SquareRootScheduler:
def __init__(self, lr=0.1):
self.lr = lr
def __call__(self, num_update):

return self.lr * pow(num_update + 1.0, -0.5)
Let us plot its behavior over a range of values.
scheduler = SquareRootScheduler(lr=1.0)
d2l.plot(np.arange(num_epochs), [scheduler(t) for t in range(num_epochs)])

Now let us see how this plays out for training on Fashion-MNIST. We simply provide the scheduler
as an additional argument to the training algorithm.

{'lr_scheduler': scheduler})
This worked quite a bit better than previously. Two things stand out: the curve was rather more
smooth than previously. Secondly, there was less overfitting. Unfortunately it is not a well-
resolved question as to why certain strategies lead to less overfitting in theory. There is some
argument that a smaller stepsize will lead to parameters that are closer to zero and thus simpler.
However, this does not explain the phenomenon entirely since we do not really stop early but
simply reduce the learning rate gently.

11.11.3 Policies
While we cannot possibly cover the entire variety of learning rate schedulers, we attempt to give
a brief overview of popular policies below. Common choices are polynomial decay and piecewise
constant schedules. Beyond that, cosine learning rate schedules have been found to work well
empirically on some problems. Lastly, on some problems it is beneficial to warm up the optimizer
prior to using large learning rates.
Factor Scheduler
One alternative to a polynomial decay would be a multiplicative one, that is ηt+1 ← ηt · α for
α ∈ (0, 1). To prevent the learning rate from decaying beyond a reasonable lower bound the
update equation is often modified to ηt+1 ← max(ηmin , ηt · α).
class FactorScheduler:
def __init__(self, factor=1, stop_factor_lr=1e-7, base_lr=0.1):
self.factor = factor
self.stop_factor_lr = stop_factor_lr
self.base_lr = base_lr
def __call__(self, num_update):

self.base_lr = max(self.stop_factor_lr, self.base_lr * self.factor)
return self.base_lr
scheduler = FactorScheduler(factor=0.9, stop_factor_lr=1e-2, base_lr=2.0)

d2l.plot(np.arange(50), [scheduler(t) for t in range(50)])
This can also be accomplished by a built-in scheduler in MXNet via the lr_scheduler.
FactorScheduler object. It takes a few more parameters, such as warmup period, warmup mode
(linear or constant), the maximum number of desired updates, etc.; Going forward we will use
the built-in schedulers as appropriate and only explain their functionality here. As illustrated, it
is fairly straightforward to build your own scheduler if needed.

Multi Factor Scheduler
A common strategy for training deep networks is to keep the learning rate piecewise constant and
to decrease it by a given amount every so often. That is, given a set of times when to decrease the
rate, such as s = {5, 10, 20} decrease ηt+1 ← ηt · α whenever t ∈ s. Assuming that the values are
halved at each step we can implement this as follows.
scheduler = lr_scheduler.MultiFactorScheduler(step=[15, 30], factor=0.5,

base_lr=0.5)
The intuition behind this piecewise constant learning rate schedule is that one lets optimization
proceed until a stationary point has been reached in terms of the distribution of weight vectors.
Then (and only then) do we decrease the rate such as to obtain a higher quality proxy to a good
local minimum. The example below shows how this can produce ever slightly better solutions.


Cosine Scheduler
A rather perplexing heuristic was proposed by (Loshchilov & Hutter, 2016). It relies on the obser-
vation that we might not want to decrease the learning rate too drastically in the beginning and
moreover, that we might want to “refine” the solution in the end using a very small learning rate.
This results in a cosine-like schedule with the following functional form for learning rates in the
range t ∈ [0, T ].
η0 − ηT
ηt = ηT + (1 + cos(πt/T )) (11.11.1)
2
Here η0 is the initial learning rate, ηT is the target rate at time T . Furthermore, for t > T we simply
pin the value to ηT without increasing it again. In the following example, we set the max update
step T = 20.
scheduler = lr_scheduler.CosineScheduler(max_update=20, base_lr=0.5,

final_lr=0.01)
In the context of computer vision this schedule can lead to improved results. Note, though, that
such improvements are not guaranteed (as can be seen below).


Warmup
In some cases initializing the parameters is not sufficient to guarantee a good solution. This par-
ticularly a problem for some advanced network designs that may lead to unstable optimization
problems. We could address this by choosing a sufficiently small learning rate to prevent diver-
gence in the beginning. Unfortunately this means that progress is slow. Conversely, a large learn-
ing rate initially leads to divergence.
A rather simple fix for this dilemma is to use a warmup period during which the learning rate
increases to its initial maximum and to cool down the rate until the end of the optimization process.
For simplicity one typically uses a linear increase for this purpose. This leads to a schedule of the
form indicated below.
scheduler = lr_scheduler.CosineScheduler(20, warmup_steps=5, base_lr=0.5,

final_lr=0.01)
Note that the network converges better initially (in particular observe the performance during the
first 5 epochs).

Warmup can be applied to any scheduler (not just cosine). For a more detailed discussion of learn-
ing rate schedules and many more experiments see also (Gotmare et al., 2018). In particular they
find that a warmup phase limits the amount of divergence of parameters in very deep networks.
This makes intuitively sense since we would expect significant divergence due to random initial-
ization in those parts of the network that take the most time to make progress in the beginning.
Summary
• Decreasing the learning rate during training can lead to improved accuracy and (most per-
plexingly) reduced overfitting of the model.
• A piecewise decrease of the learning rate whenever progress has plateaued is effective in
practice. Essentially this ensures that we converge efficiently to a suitable solution and only
then reduce the inherent variance of the parameters by reducing the learning rate.
• Cosine schedulers are popular for some computer vision problems. See e.g., GluonCV147 for
details of such a scheduler.
• A warmup period before optimization can prevent divergence.
• Optimization serves multiple purposes in deep learning. Besides minimizing the training
objective, different choices of optimization algorithms and learning rate scheduling can lead
to rather different amounts of generalization and overfitting on the test set (for the same
amount of training error).
147
http://gluon-cv.mxnet.io

Exercises
1. Experiment with the optimization behavior for a given fixed learning rate. What is the best
model you can obtain this way?
2. How does convergence change if you change the exponent of the decrease in the learning
rate? Use PolyScheduler for your convenience in the experiments.
3. Apply the cosine scheduler to large computer vision problems, e.g., training ImageNet. How
does it affect performance relative to other schedulers?
4. How long should warmup last?
5. Can you connect optimization and sampling? Start by using results from (Welling & Teh,
2011) on Stochastic Gradient Langevin Dynamics.
Discussions148
148

12 | Computational Performance
In deep learning, datasets are usually large and model computation is complex. Therefore, we
are always very concerned about computing performance. This chapter will focus on the impor-
tant factors that affect computing performance: imperative programming, symbolic program-
ming, asynchronous programing, automatic parallel computation, and multi-GPU computation.
By studying this chapter, you should be able to further improve the computing performance of
the models that have been implemented in the previous chapters, for example, by reducing the
model training time without affecting the accuracy of the model.
12.1 Compilers and Interpreters
So far, this book has focused on imperative programming, which makes use of statements such as
print, + or if to change a programʼs state. Consider the following example of a simple imperative
program.
def add(a, b):

return a + b
def fancy_func(a, b, c, d):

e = add(a, b)
f = add(c, d)
g = add(e, f)
return g
print(fancy_func(1, 2, 3, 4))
10
Python is an interpreted language. When evaluating fancy_func it performs the operations mak-
ing up the functionʼs body in sequence. That is, it will evaluate e = add(a, b) and it will store the
results as variable e, thereby changing the programʼs state. The next two statements f = add(c,
d) and g = add(e, f) will be excecuted similarly, performing additions and storing the results as
variables. Fig. 12.1.1 illustrates the flow of data.
483
Fig. 12.1.1: Data flow in an imperative program.
Although imperative programming is convenient, it may be inefficient. On one hand, even if the
add function is repeatedly called throughout fancy_func, Python will execute the three function
calls individually. If these are executed, say, on a GPU (or even on multiple GPUs), the overhead
arising from the Python interpreter can become overwhelming. Moreover, it will need to save
the variable values of e and f until all the statements in fancy_func have been executed. This is
because we do not know whether the variables e and f will be used by other parts of the program
after the statements e = add(a, b) and f = add(c, d) have been executed.
12.1.1 Symbolic Programming
Consider the alternative, symbolic programming where computation is usually performed only
once the process has been fully defined. This strategy is used by multiple deep learning frame-
works, including Theano, Keras and TensorFlow (the latter two have since acquired imperative
extensions). It usually involves the following steps:
1. Define the operations to be executed.
2. Compile the operations into an executable program.
3. Provide the required inputs and call the compiled program for execution.
This allows for a significant amount of optimization. First off, we can skip the Python interpreter
in many cases, thus removing a performance bottleneck that can become significant on multiple
fast GPUs paired with a single Python thread on a CPU. Secondly, a compiler might optimize and
rewrite the above code into print((1 + 2) + (3 + 4)) or even print(10). This is possible since a
compiler gets to see the full code before turning it into machine instructions. For instance, it can
release memory (or never allocate it) whenever a variable is no longer needed. Or it can transform
the code entirely into an equivalent piece. To get a better idea consider the following simulation
of imperative programming (it is Python after all) below.
def add_():
return '''
def add(a, b):
return a + b
'''
def fancy_func_():
return '''
e = add(a, b)
f = add(c, d)
484 Chapter 12. Computational Performance

g = add(e, f)
return g
'''
def evoke_():
return add_() + fancy_func_() + 'print(fancy_func(1, 2, 3, 4))'
prog = evoke_()
print(prog)
y = compile(prog, '', 'exec')
exec(y)
def add(a, b):

return a + b

e = add(a, b)
f = add(c, d)
g = add(e, f)
return g
print(fancy_func(1, 2, 3, 4))
10
The differences between imperative (interpreted) programming and symbolic programming are
as follows:
• Imperative programming is easier. When imperative programming is used in Python, the
majority of the code is straightforward and easy to write. It is also easier to debug imperative
programming code. This is because it is easier to obtain and print all relevant intermediate
variable values, or use Pythonʼs built-in debugging tools.
• Symbolic programming is more efficient and easier to port. It makes it easier to optimize
the code during compilation, while also having the ability to port the program into a format
independent of Python. This allows the program to be run in a non-Python environment,
thus avoiding any potential performance issues related to the Python interpreter.
12.1.2 Hybrid Programming
Historically most deep learning frameworks choose between an imperative or a symbolic ap-
proach. For example, Theano, TensorFlow (inspired by the latter), Keras and CNTK formulate
models symbolically. Conversely, Chainer and PyTorch take an imperative approach. An imper-
ative mode was added to TensorFlow 2.0 (via Eager) and Keras in later revisions. When designing
Gluon, developers considered whether it would be possible to combine the benefits of both pro-
gramming models. This led to a hybrid model that lets users develop and debug using pure imper-
ative programming, while having the ability to convert most programs into symbolic programs to
be run when product-level computing performance and deployment are required.
In practice this means that we build models using either the HybridBlock or the HybridSequential
and HybridConcurrent classes. By default, they are executed in the same way Block or Sequential
and Concurrent classes are executed in imperative programming. HybridSequential is a subclass
of HybridBlock ( just like Sequential subclasses Block). When the hybridize function is called,
12.1. Compilers and Interpreters 485

Gluon compiles the model into the form used in symbolic programming. This allows one to op-
timize the compute-intensive components without sacrifices in the way a model is implemented.
We will illustrate the benefits below, focusing on sequential models and blocks only (the concur-
rent composition works analogously).
12.1.3 HybridSequential
The easiest way to get a feel for how hybridization works is to consider deep networks with mul-
tiple layers. Conventionally the Python interpreter will need to execute the code for all layers to
generate an instruction that can then be forwarded to a CPU or a GPU. For a single (fast) compute
device this does not cause any major issues. On the other hand, if we use an advanced 8-GPU
server such as an AWS P3dn.24xlarge instance Python will struggle to keep all GPUs busy. The
single-threaded Python interpreter becomes the bottleneck here. Let us see how we can address
this for significant parts of the code by replacing Sequential by HybridSequential. We begin by
defining a simple MLP.

npx.set_np()
# factory for networks

def get_net():
nn.Dense(2))
net.initialize()
return net
x = np.random.normal(size=(1, 512))
net = get_net()
net(x)
array([[ 0.16526176, -0.14005631]])
By calling the hybridize function, we are able to compile and optimize the computation in the
MLP. The modelʼs computation result remains unchanged.
net.hybridize()
net(x)
array([[ 0.16526176, -0.14005631]])
This seems almost too good to be true: simply designate a block to be HybridSequential, write
the same code as before and invoke hybridize. Once this happens the network is optimized (we
will benchmark the performance below). Unfortunately this does not work magically for every
layer. That said, the blocks provided by Gluon are by default subclasses of HybridBlock and thus
hybridizable. A layer will not be optimized if it inherits from the Block instead.

Acceleration by Hybridization
To demonstrate the performance improvement gained by compilation we compare the time

needed to evaluate net(x) before and after hybridization. Let us define a function to measure
this time first. It will come handy throughout the chapter as we set out to measure (and improve)
performance.
#@save
class Benchmark:
def __init__(self, description='Done'):
self.description = description
def __enter__(self):
self.timer = d2l.Timer()
return self
def __exit__(self, *args):

print(f'{self.description}: {self.timer.stop():.4f} sec')
Now we can invoke the network twice, once with and once without hybridization.
net = get_net()
with Benchmark('Without hybridization'):
for i in range(1000): net(x)
npx.waitall()
net.hybridize()
with Benchmark('With hybridization'):
for i in range(1000): net(x)
npx.waitall()
Without hybridization: 0.8433 sec

With hybridization: 0.3059 sec
As is observed in the above results, after a HybridSequential instance calls the hybridize function,
computing performance is improved through the use of symbolic programming.
Serialization
One of the benefits of compiling the models is that we can serialize (save) the model and its pa-
rameters to disk. This allows us to store a model in a manner that is independent of the front-end
language of choice. This allows us to deploy trained models to other devices and easily use other
front-end programming languages. At the same time the code is often faster than what can be
achieved in imperative programming. Let us see the export method in action.
net.export('my_mlp')
!ls -lh my_mlp*
-rw-r--r-- 1 jenkins jenkins 643K Aug 9 21:34 my_mlp-0000.params

-rw-r--r-- 1 jenkins jenkins 3.0K Aug 9 21:34 my_mlp-symbol.json

The model is decomposed into a (large binary) parameter file and a JSON description of the pro-
gram required to execute to compute the model. The files can be read by other front-end languages
supported by Python or MXNet, such as C++, R, Scala, and Perl. Let us have a look at the model
description.
!head my_mlp-symbol.json
{
"nodes": [
{
"op": "null",
"name": "data",
"inputs": []
},
{
"op": "null",
"name": "dense3_weight",
Things are slightly more tricky when it comes to models that resemble code more closely. Basically
hybridization needs to deal with control flow and Python overhead in a much more immediate
manner. Moreover,
Contrary to the Block instance, which needs to use the forward function, for a HybridBlock in-
stance we need to use the hybrid_forward function.
Earlier, we demonstrated that, after calling the hybridize method, the model is able to achieve su-
perior computing performance and portability. Note, though that hybridization can affect model
flexibility, in particular in terms of control flow. We will illustrate how to design more general
models and also how compilation will remove spurious Python elements.
class HybridNet(nn.HybridBlock):
super(HybridNet, self).__init__(**kwargs)
self.hidden = nn.Dense(4)
def hybrid_forward(self, F, x):

print('module F: ', F)
print('value x: ', x)
x = F.npx.relu(self.hidden(x))
print('result : ', x)
return self.output(x)
The code above implements a simple network with 4 hidden units and 2 outputs. hybrid_forward
takes an additional argument - the module F. This is needed since, depending on whether the code
has been hybridized or not, it will use a slightly different library (ndarray or symbol) for processing.
Both classes perform very similar functions and MXNet automatically determines the argument.
To understand what is going on we print the arguments as part of the function invocation.
net = HybridNet()
net.initialize()
x = np.random.normal(size=(1, 3))
net(x)

module F: <module 'mxnet.ndarray' from '/var/lib/jenkins/miniconda3/envs/d2l-en-release-0/
,→lib/python3.7/site-packages/mxnet/ndarray/__init__.py'>
value x: [[-0.6338663 0.40156594 0.46456942]]

result : [[0.01641375 0. 0. 0. ]]
array([[0.00097611, 0.00019453]])
Repeating the forward computation will lead to the same output (we omit details). Now let us see
what happens if we invoke the hybridize method.
net.hybridize()
net(x)
module F: <module 'mxnet.symbol' from '/var/lib/jenkins/miniconda3/envs/d2l-en-release-0/

,→lib/python3.7/site-packages/mxnet/symbol/__init__.py'>
value x: <_Symbol data>
result : <_Symbol hybridnet0_relu0>
array([[0.00097611, 0.00019453]])
Instead of using ndarray we now use the symbol module for F. Moreover, even though the input is
of ndarray type, the data flowing through the network is now converted to symbol type as part of
the compilation process. Repeating the function call leads to a surprising outcome:
net(x)
array([[0.00097611, 0.00019453]])
This is quite different from what we saw previously. All print statements, as defined in hy-
brid_forward are omitted. Indeed, after hybridization the execution of net(x) does not involve
the Python interpreter any longer. This means that any spurious Python code is omitted (such
as print statements) in favor of a much more streamlined execution and better performance. In-
stead, MXNet directly calls the C++ backend. Also note that some functions are not supported in
the symbol module (like asnumpy) and operations in-place like a += b and a[:] = a + b must
be rewritten as a = a + b. Nonetheless, compilation of models is worth the effort whenever
speed matters. The benefit can range from small percentage points to more than twice the speed,
depending on the complexity of the model, the speed of the CPU and the speed and number of
GPUs.
Summary
• Imperative programming makes it easy to design new models since it is possible to write
code with control flow and the ability to use a large amount of the Python software ecosys-
tem.
• Symbolic programming requires that we specify the program and compile it before execut-
ing it. The benefit is improved performance.
• MXNet is able to combine the advantages of both approaches as needed.

• Models constructed by the HybridSequential and HybridBlock classes are able to convert
imperative programs into symbolic programs by calling the hybridize method.
Exercises
1. Design a network using the HybridConcurrent class. Alternatively look at Networks with Par-
allel Concatenations (GoogLeNet) (page 274) for a network to compose.
2. Add x.asnumpy() to the first line of the hybrid_forward function of the HybridNet class in
this section. Execute the code and observe the errors you encounter. Why do they happen?
3. What happens if we add control flow, i.e., the Python statements if and for in the hy-
brid_forward function?
4. Review the models that interest you in the previous chapters and use the HybridBlock class
or HybridSequential class to implement them.
Discussions149
12.2 Asynchronous Computation
Todayʼs computers are highly parallel systems, consisting of multiple CPU cores (often multiple
threads per core), multiple processing elements per GPU and often multiple GPUs per device. In
short, we can process many different things at the same time, often on different devices. Unfor-
tunately Python is not a great way of writing parallel and asynchronous code, at least not with
some extra help. After all, Python is single-threaded and this is unlikely to change in the future.
Deep learning frameworks such as MXNet and TensorFlow utilize an asynchronous programming
model to improve performance (PyTorch uses Pythonʼs own scheduler leading to a different per-
formance trade-off). Hence, understanding how asynchronous programming works helps us to
develop more efficient programs, by proactively reducing computational requirements and mu-
tual dependencies. This allows us to reduce memory overhead and increase processor utilization.
We begin by importing the necessary libraries.

import numpy, os, subprocess
npx.set_np()
12.2.1 Asynchrony via Backend
For a warmup consider the following toy problem - we want to generate a random matrix and
multiply it. Let us do that both in NumPy and in MXNet NP to see the difference.
with d2l.Benchmark('numpy'):
for _ in range(10):
a = numpy.random.normal(size=(1000, 1000))
b = numpy.dot(a, a)
149

with d2l.Benchmark('mxnet.np'):
for _ in range(10):
a = np.random.normal(size=(1000, 1000))
b = np.dot(a, a)
numpy: 0.9055 sec

mxnet.np: 0.0041 sec
This is orders of magnitude faster. At least it seems to be so. Since both are executed on the same
processor something else must be going on. Forcing MXNet to finish all computation prior to
returning shows what happened previously: computation is being executed by the backend while
the frontend returns control to Python.
with d2l.Benchmark():
for _ in range(10):
a = np.random.normal(size=(1000, 1000))
b = np.dot(a, a)
npx.waitall()
Done: 0.6968 sec
Broadly speaking, MXNet has a frontend for direct interaction with the users, e.g., via Python,
as well as a backend used by the system to perform the computation. As shown in Fig. 12.2.1,
users can write MXNet programs in various frontend languages, such as Python, R, Scala, and C++.
Regardless of the frontend programming language used, the execution of MXNet programs occurs
primarily in the backend of C++ implementations. Operations issued by the frontend language are
passed on to the backend for execution. The backend manages its own threads that continuously
collect and execute queued tasks. Note that for this to work the backend must be able to keep track
of the dependencies between various steps in the computational graph. Hence, it is not possible
to parallelize operations that depend on each other.
Fig. 12.2.1: Programming Frontends.
Let us look at another toy example to understand the dependency graph a bit better.
x = np.ones((1, 2))
y = np.ones((1, 2))
12.2. Asynchronous Computation 491

z = x * y + 2
z
array([[3., 3.]])
Fig. 12.2.2: Dependencies.
The code snippet above is also illustrated in Fig. 12.2.2. Whenever the Python frontend thread
executes one of the first three statements, it simply returns the task to the backend queue. When
the last statementʼs results need to be printed, the Python frontend thread will wait for the C++
backend thread to finish computing result of the variable z. One benefit of this design is that the
Python frontend thread does not need to perform actual computations. Thus, there is little impact
on the programʼs overall performance, regardless of Pythonʼs performance. Fig. 12.2.3 illustrates
how frontend and backend interact.
Fig. 12.2.3: Frontend and Backend.
12.2.2 Barriers and Blockers
There are a number of operations that will force Python to wait for completion: * Most obviously
npx.waitall() waits until all computation has completed, regardless of when the compute in-
structions were issued. In practice it is a bad idea to use this operator unless absolutely necessary
since it can lead to poor performance. * If we just want to wait until a specific variable is available
we can call z.wait_to_read(). In this case MXNet blocks return to Python until the variable z has
been computed. Other computation may well continue afterwards.
Let us see how this works in practice:

with d2l.Benchmark('waitall'):
b = np.dot(a, a)
npx.waitall()
with d2l.Benchmark('wait_to_read'):
b = np.dot(a, a)
b.wait_to_read()
waitall: 0.0043 sec

wait_to_read: 0.0043 sec
Both operations take approximately the same time to complete. Besides the obvious blocking op-
erations we recommend that the reader is aware of implicit blockers. Printing a variable clearly
requires the variable to be available and is thus a blocker. Lastly, conversions to NumPy via z.
asnumpy() and conversions to scalars via z.item() are blocking, since NumPy has no notion of
asynchrony. It needs access to the values just like the print function. Copying small amounts
of data frequently from MXNetʼs scope to NumPy and back can destroy performance of an oth-
erwise efficient code, since each such operation requires the computational graph to evaluate all
intermediate results needed to get the relevant term before anything else can be done.
with d2l.Benchmark('numpy conversion'):

b = np.dot(a, a)
b.asnumpy()
with d2l.Benchmark('scalar conversion'):

b = np.dot(a, a)
b.sum().item()
numpy conversion: 0.0046 sec

scalar conversion: 0.0155 sec
12.2.3 Improving Computation
On a heavily multithreaded system (even regular laptops have 4 threads or more and on multi-
socket servers this number can exceed 256) the overhead of scheduling operations can become
significant. This is why it is highly desirable to have computation and scheduling occur asyn-
chronously and in parallel. To illustrate the benefit of doing this let us see what happens if we
increment a variable by 1 multiple times, both in sequence or asynchronously. We simulate syn-
chronous execution by inserting a wait_to_read() barrier in between each addition.
with d2l.Benchmark('synchronous'):
for _ in range(1000):
y = x + 1
y.wait_to_read()
with d2l.Benchmark('asynchronous'):
for _ in range(1000):
y = x + 1
y.wait_to_read()

synchronous: 0.0624 sec
asynchronous: 0.0572 sec
A slightly simplified interaction between the Python frontend thread and the C++ backend thread
can be summarized as follows:
1. The frontend orders the backend to insert the calculation task y = x + 1 into the queue.
2. The backend then receives the computation tasks from the queue and performs the actual
computations.
3. The backend then returns the computation results to the frontend.
Assume that the durations of these three stages are t1 , t2 and t3 , respectively. If we do not use
asynchronous programming, the total time taken to perform 1000 computations is approximately
1000(t1 + t2 + t3 ). If asynchronous programming is used, the total time taken to perform 1000
computations can be reduced to t1 + 1000t2 + t3 (assuming 1000t2 > 999t1 ), since the frontend
does not have to wait for the backend to return computation results for each loop.
12.2.4 Improving Memory Footprint
Imagine a situation where we keep on inserting operations into the backend by executing Python
code on the frontend. For instance, the frontend might insert a large number of minibatch tasks
within a very short time. After all, if no meaningful computation happens in Python this can
be done quite quickly. If each of these tasks can be launched quickly at the same time this may
cause a spike in memory usage. Given a finite amount of memory available on GPUs (and even
on CPUs) this can lead to resource contention or even program crashes. Some readers might have
noticed that previous training routines made use of synchronization methods such as item or even
asnumpy.
We recommend to use these operations carefully, e.g., for each minibatch, such as to balance
computational efficiency and memory footprint. To illustrate what happens let us implement a
simple training loop for a deep network and measure its memory consumption and timing. Below
is the mock data generator and deep network.
def data_iter():
timer = d2l.Timer()
num_batches, batch_size = 150, 1024
for i in range(num_batches):
X = np.random.normal(size=(batch_size, 512))
y = np.ones((batch_size,))
yield X, y
if (i + 1) % 50 == 0:
print(f'batch {i + 1}, time {timer.stop():.4f} sec')
nn.Dense(512, activation='relu'), nn.Dense(1))
net.initialize()
trainer = gluon.Trainer(net.collect_params(), 'sgd')
Next we need a tool to measure the memory footprint of our code. We use a relatively primitive
ps call to accomplish this (note that the latter only works on Linux and MacOS). For a much more

detailed analysis of what is going on here use e.g., Nvidiaʼs Nsight150 or Intelʼs vTune151 .
def get_mem():
res = subprocess.check_output(['ps', 'u', '-p', str(os.getpid())])
return int(str(res).split()[15]) / 1e3
Before we can begin testing we need to initialize the parameters of the network and process one
batch. Otherwise it would be tricky to see what the additional memory consumption is. See Sec-
tion 5.3 for further details related to initialization.
for X, y in data_iter():
break
loss(y, net(X)).wait_to_read()
To ensure that we do not overflow the task buffer on the backend we insert a wait_to_read call for
the loss function at the end of each loop. This forces the forward propagation to complete before
a new forward propagation is commenced. Note that a (possibly more elegant) alternative would
have been to track the loss in a scalar variable and to force a barrier via the item call.
mem = get_mem()
with d2l.Benchmark('time per epoch'):
l = loss(y, net(X))
l.backward()
l.wait_to_read() # Barrier before a new batch
npx.waitall()
print(f'increased memory: {get_mem() - mem:f} MB')
batch 50, time 3.0298 sec

time per epoch: 9.1362 sec
increased memory: 5.200000 MB
As we see, the timing of the minibatches lines up quite nicely with the overall runtime of the opti-
mization code. Moreover, memory footprint only increases slightly. Now let us see what happens
if we drop the barrier at the end of each minibatch.
mem = get_mem()
with d2l.Benchmark('time per epoch'):
l = loss(y, net(X))
l.backward()
npx.waitall()
print(f'increased memory: {get_mem() - mem:f} MB')
150
https://developer.nvidia.com/nsight-compute-2019_5
151
https://software.intel.com/en-us/vtune

time per epoch: 9.1249 sec
increased memory: 2.040000 MB
Even though the time to issue instructions for the backend is an order of magnitude smaller, we
still need to perform computation. Consequently a large amount of intermediate results cannot
be released and may pile up in memory. While this didnʼt cause any issues in the toy example
above, it might well have resulted in out of memory situations when left unchecked in real world
scenarios.
Summary
• MXNet decouples the Python frontend from an execution backend. This allows for fast asyn-
chronous insertion of commands into the backend and associated parallelism.
• Asynchrony leads to a rather responsive frontend. However, use caution not to overfill the
task queue since it may lead to excessive memory consumption.
• It is recommended to synchronize for each minibatch to keep frontend and backend approx-
imately synchronized.
• Be aware of the fact that conversions from MXNetʼs memory management to Python will
force the backend to wait until the specific variable is ready. print, asnumpy and item all have
this effect. This can be desirable but a carless use of synchronization can ruin performance.
• Chip vendors offer sophisticated performance analysis tools to obtain a much more fine-
grained insight into the efficiency of deep learning.
Exercises
1. We mentioned above that using asynchronous computation can reduce the total amount of
time needed to perform 1000 computations to t1 + 1000t2 + t3 . Why do we have to assume
1000t2 > 999t1 here?
2. How would you need to modify the training loop if you wanted to have an overlap of one
minibatch each? I.e., if you wanted to ensure that batch bt finishes before batch bt+2 com-
mences?
3. What might happen if we want to execute code on CPUs and GPUs simultaneously? Should
you still insist on synchronizing after every minibatch has been issued?
4. Measure the difference between waitall and wait_to_read. Hint: perform a number of
instructions and synchronize for an intermediate result.
Discussions152
152

12.3 Automatic Parallelism
MXNet automatically constructs computational graphs at the backend. Using a computational

graph, the system is aware of all the dependencies, and can selectively execute multiple non-
interdependent tasks in parallel to improve speed. For instance, Fig. 12.2.2 in Section 12.2 ini-
tializes two variables independently. Consequently the system can choose to execute them in
parallel.
Typically, a single operator will use all the computational resources on all CPUs or on a single
GPU. For example, the dot operator will use all cores (and threads) on all CPUs, even if there
are multiple CPU processors on a single machine. The same applies to a single GPU. Hence par-
allelization is not quite so useful single-device computers. With multiple devices things matter
more. While parallelization is typically most relevant between multiple GPUs, adding the local
CPU will increase performance slightly. See e.g., (Hadjis et al., 2016) for a paper that focuses on
training computer vision models combining a GPU and a CPU. With the convenience of an auto-
matically parallelizing framework we can accomplish the same goal in a few lines of Python code.
More broadly, our discussion of automatic parallel computation focuses on parallel computation
using both CPUs and GPUs, as well as the parallelization of computation and communication. We
begin by importing the required packages and modules. Note that we need at least one GPU to
run the experiments in this section.

npx.set_np()
12.3.1 Parallel Computation on CPUs and GPUs
Let us start by defining a reference workload to test - the run function below performs 10 matrix-
matrix multiplications on the device of our choosing using data allocated into two variables, x_cpu
and x_gpu.
def run(x):
return [x.dot(x) for _ in range(10)]
x_cpu = np.random.uniform(size=(2000, 2000))

x_gpu = np.random.uniform(size=(6000, 6000), ctx=d2l.try_gpu())
Now we apply the function to the data. To ensure that caching does not play a role in the results
we warm up the devices by performing a single pass on each of them prior to measuring.
run(x_cpu) # Warm-up both devices

run(x_gpu)
npx.waitall()
with d2l.Benchmark('CPU time'):

run(x_cpu)
npx.waitall()
with d2l.Benchmark('GPU time'):

run(x_gpu)
npx.waitall()
12.3. Automatic Parallelism 497

CPU time: 0.2868 sec
GPU time: 0.3007 sec
If we remove the waitall() between both tasks the system is free to parallelize computation on
both devices automatically.
with d2l.Benchmark('CPU & GPU'):

run(x_cpu)
run(x_gpu)
npx.waitall()
CPU & GPU: 0.3005 sec
In the above case the total execution time is less than the sum of its parts, since MXNet automat-
ically schedules computation on both CPU and GPU devices without the need for sophisticated
code on behalf of the user.
12.3.2 Parallel Computation and Communication
In many cases we need to move data between different devices, say between CPU and GPU, or be-
tween different GPUs. This occurs e.g., when we want to perform distributed optimization where
we need to aggregate the gradients over multiple accelerator cards. Let us simulate this by com-
puting on the GPU and then copying the results back to the CPU.
def copy_to_cpu(x):
return [y.copyto(npx.cpu()) for y in x]
with d2l.Benchmark('Run on GPU'):

y = run(x_gpu)
npx.waitall()
with d2l.Benchmark('Copy to CPU'):

y_cpu = copy_to_cpu(y)
npx.waitall()
Run on GPU: 0.3038 sec

Copy to CPU: 1.0119 sec
This is somewhat inefficient. Note that we could already start copying parts of y to the CPU while
the remainder of the list is still being computed. This situation occurs, e.g., when we compute the
(backprop) gradient on a minibatch. The gradients of some of the parameters will be available
earlier than that of others. Hence it works to our advantage to start using PCI-Express bus band-
width while the GPU is still running. Removing waitall between both parts allows us to simulate
this scenario.
with d2l.Benchmark('Run on GPU and copy to CPU'):

y = run(x_gpu)
y_cpu = copy_to_cpu(y)
npx.waitall()

Run on GPU and copy to CPU: 1.0859 sec
The total time required for both operations is (as expected) significantly less than the sum of their
parts. Note that this task is different from parallel computation as it uses a different resource: the
bus between CPU and GPUs. In fact, we could compute on both devices and communicate, all at
the same time. As noted above, there is a dependency between computation and communication:
y[i] must be computed before it can be copied to the CPU. Fortunately, the system can copy y[i-1]
while computing y[i] to reduce the total running time.
We conclude with an illustration of the computational graph and its dependencies for a simple
two-layer MLP when training on a CPU and two GPUs, as depicted in Fig. 12.3.1. It would be quite
painful to schedule the parallel program resulting from this manually. This is where it is advan-
tageous to have a graph based compute backend for optimization.
Fig. 12.3.1: Two layer MLP on a CPU and 2 GPUs.
12.3. Automatic Parallelism 499

Summary
• Modern systems have a variety of devices, such as multiple GPUs and CPUs. They can be
used in parallel, asynchronously.
• Modern systems also have a variety of resources for communication, such as PCI Express,
storage (typically SSD or via network), and network bandwidth. They can be used in parallel
for peak efficiency.
• The backend can improve performance through through automatic parallel computation
and communication.
Exercises
1. 10 operations were performed in the run function defined in this section. There are no de-
pendencies between them. Design an experiment to see if MXNet will automatically execute
them in parallel.
2. When the workload of an individual operator is sufficiently small, parallelization can help
even on a single CPU or GPU. Design an experiment to verify this.
3. Design an experiment that uses parallel computation on CPU, GPU and communication be-
tween both devices.
4. Use a debugger such as NVIDIAʼs Nsight to verify that your code is efficient.
5. Designing computation tasks that include more complex data dependencies, and run exper-
iments to see if you can obtain the correct results while improving performance.
Discussions153
12.4 Hardware
Building systems with great performance requires a good understanding of the algorithms and
models to capture the statistical aspects of the problem. At the same time it is also indispensable
to have at least a modicum of knowledge of the underlying hardware. The current section is no
substitute for a proper course on hardware and systems design. Instead, it might serve as a starting
point for understanding why some algorithms are more efficient than others and how to achieve
good throughput. Good design can easily make a difference of an order of magnitude and, in turn,
this can make the difference between being able to train a network (e.g., in a week) or not at all (in
3 months, thus missing the deadline). We will start by looking at computers. Then we will zoom in
to look more carefully at CPUs and GPUs. Lastly we zoom out to review how multiple computers
are connected in a server center or in the cloud. This is not a GPU purchase guide. For this review
Section 19.5. An introduction to cloud computing with AWS can be found in Section 19.3.
Impatient readers may be able to get by with Fig. 12.4.1. It is taken from Colin Scottʼs interactive
post154 which gives a good overview of the progress over the past decade. The original numbers are
due to Jeff Deanʼs Stanford talk from 2010155 . The discussion below explains some of the rationale
for these numbers and how they can guide us in designing algorithms. The discussion below is
153
154
https://people.eecs.berkeley.edu/~rcs/research/interactive_latency.html
155
https://static.googleusercontent.com/media/research.google.com/en//people/jeff/Stanford-DL-Nov-2010.pdf

very high level and cursory. It is clearly no substitute for a proper course but rather just meant to
provide enough information for a statistical modeler to make suitable design decisions. For an
in-depth overview of computer architecture we refer the reader to (Hennessy & Patterson, 2011)
or a recent course on the subject, such as the one by Arste Asanovic156 .
Fig. 12.4.1: Latency Numbers every Programmer should know.
12.4.1 Computers
Most deep learning researchers have access to a computer with a fair amount of memory, com-
pute, some form of an accelerator such as a GPU, or multiples thereof. It consists of several key
components:
• A processor, also referred to as CPU which is able to execute the programs we give it (in
addition to running an operating system and many other things), typically consisting of 8 or
more cores.
• Memory (RAM) to store and retrieve the results from computation, such as weight vectors,
activations, often training data.
• An Ethernet network connection (sometimes multiple) with speeds ranging from 1Gbit/s to
100Gbit/s (on high end servers more advanced interconnects can be found).
• A high speed expansion bus (PCIe) to connect the system to one or more GPUs. Servers have
up to 8 accelerators, often connected in an advanced topology, desktop systems have 1-2,
depending on the budget of the user and the size of the power supply.
• Durable storage, such as a magnetic harddrive (HDD), solid state (SSD), in many cases con-
nected using the PCIe bus, provides efficient transfer of training data to the system and stor-
age of intermediate checkpoints as needed.
Fig. 12.4.2: Connectivity of components

156
http://inst.eecs.berkeley.edu/~cs152/sp19/
12.4. Hardware 501

As Fig. 12.4.2 indicates, most components (network, GPU, storage) are connected to the CPU across
the PCI Express bus. It consists of multiple lanes that are directly attached to the CPU. For instance
AMDʼs Threadripper 3 has 64 PCIe 4.0 lanes, each of which is capable 16 Gbit/s data transfer in both
directions. The memory is directly attached to the CPU with a total bandwidth of up to 100 GB/s.
When we run code on a computer we need to shuffle data to the processors (CPU or GPU), perform
computation and then move the results off the processor back to RAM and durable storage. Hence,
in order to get good performance we need to make sure that this works seamlessly without any
one of the systems becoming a major bottleneck. For instance, if we cannot load images quickly
enough the processor will not have any work to do. Likewise, if we cannot move matrices quickly
enough to the CPU (or GPU), its processing elements will starve. Finally, if we want to synchronize
multiple computers across the network, the latter should not slow down computation. One option
is to interleave communication and computation. Let us have a look at the various components in
more detail.
12.4.2 Memory
At its most basic memory is used to store data that needs to be readily accessible. At present CPU
RAM is typically of the DDR4157 variety, offering 20-25GB/s bandwidth per module. Each module
has a 64 bit wide bus. Typically pairs of memory modules are used to allow for multiple channels.
CPUs have between 2 and 4 memory channels, i.e., they have between 40GB/s and 100GB/s peak
memory bandwidth. Often there are two banks per channel. For instance AMDʼs Zen 3 Thread-
ripper has 8 slots.
While these numbers are impressive, indeed, they only tell part of the story. When we want to
read a portion from memory we first need to tell the memory module where the information can
be found. That is, we first need to send the address to RAM. Once this accomplished we can choose
to read just a single 64bit record or a long sequence of records. The latter is called burst read. In
a nutshell, sending an address to memory and setting up the transfer takes approximately 100ns
(details depend on the specific timing coefficients of the memory chips used), every subsequent
transfer takes only 0.2ns. In short, the first read is 500 times as expensive as subsequent ones! We
could perform up to 10, 000, 000 random reads per second. This suggests that we avoid random
memory access as far as possible and use burst reads (and writes) instead.
Matters are a bit more complex when we take into account that we have multiple banks. Each bank
can read memory largely independently. This means two things: the effective number of random
reads is up to 4x higher, provided that they are spread evenly across memory. It also means that
it is still a bad idea to perform random reads since burst reads are 4x faster, too. Secondly, due
to memory alignment to 64 bit boundaries it is a good idea to align any datastructures with the
same boundaries. Compilers do this pretty much automatically158 when the appropriate flags are
set. Curious readers are encouraged to review a lecture on DRAMs such as the one by Zeshan
Chishti159 .
GPU memory is subject to even higher bandwidth requirements since they have many more pro-
cessing elements than CPUs. By and large there are two options to address them. One is to make
the memory bus significantly wider. For instance NVIDIAʼs RTX 2080 Ti has a 352 bit wide bus.
This allows for much more information to be transferred at the same time. Secondly, GPUs use
specific high-performance memory. Consumer grade devices, such as NVIDIAʼs RTX and Titan
157
https://en.wikipedia.org/wiki/DDR4_SDRAM
158
https://en.wikipedia.org/wiki/Data_structure_alignment
159
http://web.cecs.pdx.edu/~zeshan/ece585_lec5.pdf

series typically use GDDR6160 chips with over 500 GB/s aggregate bandwidth. An alternative is to
use HBM (high bandwidth memory) modules. They use a very different interface and connect
directly with GPUs on a dedicated silicon wafer. This makes them very expensive and their use is
typically limited to high end server chips, such as the NVIDIA Volta V100 series of accelerators.
Quite unsurprisingly GPU memory is much smaller than CPU memory due to its higher cost. For
our purposes, by and large their performance characteristics are similar, just a lot faster. We can
safely ignore the details for the purpose of this book. They only matter when tuning GPU kernels
for high throughput.
12.4.3 Storage
We saw that some of the key characteristics of RAM were bandwidth and latency. The same is true
for storage devices, just that the differences can be even more extreme.
Hard Disks have been in use for over half a century. In a nutshell they contain a number of spin-
ning platters with heads that can be positioned to read / write at any given track. High end end
disks hold up to 16TB on 9 platters. One of the key benefits of HDDs is that they are relatively in-
expensive. One of their many downsides are their typically catastrophic failure modes and their
relatively high read latency.
To understand the latter, consider the fact that HDDs spin at around 7,200 RPM. If they were much
faster they would shatter due to the centrifugal force exerted on the platters. This has a major
downside when it comes to accessing a specific sector on the disk: we need to wait until the platter
has rotated in position (we can move the heads but not accelerate the actual disks). Hence it can
take over 8ms until the requested data is available. A common way this is expressed is to say that
HDDs can operate at approximately 100 IOPs. This number has essentially remained unchanged
for the past two decades. Worse still, it is equally difficult to increase bandwidth (it is in the order
of 100-200 MB/s). After all, each head reads a track of bits, hence the bit rate only scales with
the square root of the information density. As a result HDDs are quickly becoming relegated to
archival storage and low-grade storage for very large datasets.
Solid State Drives use Flash memory to store information persistently. This allows for much faster
access to stored records. Modern SSDs can operate at 100,000 to 500,000 IOPs, i.e., up to 3 orders
of magnitude faster than HDDs. Furthermore, their bandwidth can reach 1-3GB/s, i.e., one order
of magnitude faster than HDDs. These improvements sound almost too good to be true. Indeed,
they come with a number of caveats, due to the way SSDs are designed.
• SSDs store information in blocks (256 KB or larger). They can only be written as a whole,
which takes significant time. Consequently bit-wise random writes on SSD have very poor
performance. Likewise, writing data in general takes significant time since the block has
to be read, erased and then rewritten with new information. By now SSD controllers and
firmware have developed algorithms to mitigate this. Nonetheless writes can be much
slower, in particular for QLC (quad level cell) SSDs. The key for improved performance is to
maintain a queue of operations, to prefer reads and to write in large blocks if possible.
• The memory cells in SSDs wear out relatively quickly (often already after a few thousand
writes). Wear-level protection algorithms are able to spread the degradation over many cells.
That said, it is not recommended to use SSDs for swap files or for large aggregations of log-
files.
• Lastly, the massive increase in bandwidth has forced computer designers to attach SSDs di-
rectly to the PCIe bus. The drives capable of handling this, referred to as NVMe (Non Volatile
160
https://en.wikipedia.org/wiki/GDDR6_SDRAM
12.4. Hardware 503

Memory enhanced), can use up to 4 PCIe lanes. This amounts to up to 8GB/s on PCIe 4.0.
Cloud Storage provides a configurable range of performance. That is, the assignment of storage
to virtual machines is dynamic, both in terms of quantity and in terms speed, as chosen by the
user. We recommend that the user increase the provisioned number of IOPs whenever latency is
too high, e.g., during training with many small records.
12.4.4 CPUs
Central Processing Units (CPUs) are the centerpiece of any computer (as before we give a very
high level description focusing primarily on what matters for efficient deep learning models).
They consist of a number of key components: processor cores which are able to execute machine
code, a bus connecting them (the specific topology differs significantly between processor mod-
els, generations and vendors), and caches to allow for higher bandwidth and lower latency mem-
ory access than what is possible by reads from main memory. Lastly, almost all modern CPUs
contain vector processing units to aid with high performance linear algebra and convolutions, as
they are common in media processing and machine learning.
Fig. 12.4.3: Intel Skylake consumer quad-core CPU
Fig. 12.4.3 depicts an Intel Skylake consumer grade quad-core CPU. It has an integrated GPU,
caches, and a ringbus connecting the four cores. Peripherals (Ethernet, WiFi, Bluetooth, SSD
controller, USB, etc.) are either part of the chipset or directly attached (PCIe) to the CPU.

Microarchitecture
Each of the processor cores consists of a rather sophisticated set of components. While details
differ between generations and vendors, the basic functionality is pretty much standard. The
front end loads instructions and tries to predict which path will be taken (e.g., for control flow).
Instructions are then decoded from assembly code to microinstructions. Assembly code is often
not the lowest level code that a processor executes. Instead, complex instructions may be decoded
into a set of more lower level operations. These are then processed by the actual execution core.
Often the latter is capable of performing many operations simultaneously. For instance, the ARM
Cortex A77 core of Fig. 12.4.4 is able to perform up to 8 operations simultaneously.
Fig. 12.4.4: ARM Cortex A77 Microarchitecture Overview
This means that efficient programs might be able to perform more than one instruction per clock
cycle, provided that they can be carried out independently. Not all units are created equal. Some
specialize in integer instructions whereas others are optimized for floating point performance.
To increase throughput, the processor might also follow multiple codepaths simultaneously in a
branching instruction and then discard the results of the branches not taken. This is why branch
prediction units matter (on the frontend) such that only the most promising paths are pursued.
Vectorization
Deep learning is extremely compute hungry. Hence, to make CPUs suitable for machine learning,
one needs to perform many operations in one clock cycle. This is achieved via vector units. They
have different names: on ARM they are called NEON, on x86 the latest generation is referred to as
AVX2161 units. A common aspect is that they are able to perform SIMD (single instruction multiple
data) operations. Fig. 12.4.5 shows how 8 short integers can be added in one clock cycle on ARM.
161
https://en.wikipedia.org/wiki/Advanced_Vector_Extensions
12.4. Hardware 505

Fig. 12.4.5: 128 bit NEON vectorization
Depending on architecture choices, such registers are up to 512 bit long, allowing for the combina-
tion of up to 64 pairs of numbers. For instance, we might be multiplying two numbers and adding
them to a third, which is also known as a fused multiply-add. Intelʼs OpenVino162 uses these to
achieve respectable throughput for deep learning on server grade CPUs. Note, though, that this
number is entirely dwarved by what GPUs are capable of achieving. For instance, NVIDIAʼs RTX
2080 Ti has 4,352 CUDA cores, each of which is capable of processing such an operation at any
time.
Cache
Consider the following situation: we have a modest CPU core with 4 cores as depicted in Fig. 12.4.3
above, running at 2GHz frequency. Moreover, let us assume that we have an IPC (instructions per
clock) count of 1 and that the units have AVX2 with 256bit width enabled. Let us furthermore as-
sume that at least one of the registers used for AVX2 operations needs to be retrieved from mem-
ory. This means that the CPU consumes 4x256bit = 1kbit of data per clock cycle. Unless we are
able to transfer 2 · 109 · 128 = 256 · 109 bytes to the processor per second the processing elements
are going to starve. Unfortunately the memory interface of such a chip only supports 20-40 GB/s
data transfer, i.e., one order of magnitude less. The fix is to avoid loading new data from memory
as far as possible and rather to cache it locally on the CPU. This is where caches come in handy
(see this Wikipedia article163 for a primer). Commonly the following names / concepts are used:
• Registers are strictly speaking not part of the cache. They help stage instructions. That said,
CPU registers are memory locations that a CPU can access at clock speed without any delay
penalty. CPUs have tens of registers. It is up to the compiler (or programmer) to use registers
efficiently. For instance the C programming language has a register keyword.
• L1 caches are the first line of defense against high memory bandwidth requirements. L1
caches are tiny (typical sizes might be 32-64kB) and often split into data and instructions
caches. When data is found in the L1 cache access is very fast. If it cannot be found there,
the search progresses down the cache hierarchy.
• L2 caches are the next stop. Depending on architecture design and processor size they might
be exclusive. They might be accessible only by a given core or shared between multiple
cores. L2 caches are larger (typically 256-512kB per core) and slower than L1. Furthermore,
162
https://01.org/openvinotoolkit
163
https://en.wikipedia.org/wiki/Cache_hierarchy

to access something in L2 we first need to check to realize that the data is not in L1, which
adds a small amount of extra latency.
• L3 caches are shared between multiple cores and can be quite large. AMDʼs Epyc 3 server
CPUs have a whopping 256MB of cache spread across multiple chiplets. More typical num-
bers are in the 4-8MB range.
Predicting which memory elements will be needed next is one of the key optimization parame-
ters in chip design. For instance, it is advisable to traverse memory in a forward direction since
most caching algorithms will try to read ahead rather than backwards. Likewise, keeping memory
access patterns local is a good way of improving performance. Adding caches is a double-edge
sword. On one hand they ensure that the processor cores do not starve of data. At the same time
they increase chip size, using up area that otherwise could have been spent on increasing process-
ing power. Moreover, cache misses can be expensive. Consider the worst case scenario, depicted
in Fig. 12.4.6. A memory location is cached on processor 0 when a thread on processor 1 requests
the data. To obtain it, processor 0 needs to stop what it is doing, write the information back to
main memory and then let processor 1 read it from memory. During this operation both proces-
sors wait. Quite potentially such code runs more slowly on multiple processors when compared to
an efficient single-processor implementation. This is one more reason for why there is a practical
limit to cache sizes (besides their physical size).
Fig. 12.4.6: False sharing (image courtesy of Intel)
12.4.5 GPUs and other Accelerators
It is not an exaggeration to claim that deep learning would not have been successful without GPUs.
By the same token, it is quite reasonable to argue that GPU manufacturersʼ fortunes have been
increased significantly due to deep learning. This co-evolution of hardware and algorithms has
led to a situation where for better or worse deep learning is the preferable statistical modeling
paradigm. Hence it pays to understand the specific benefits that GPUs and related accelerators
such as the TPU (Jouppi et al., 2017) offer.
Of note is a distinction that is often made in practice: accelerators are optimized either for train-
ing or inference. For the latter we only need to compute the forward propagation in a network.
No storage of intermediate data is needed for backpropagation. Moreover, we may not need very
precise computation (FP16 or INT8 typically suffice). On the other hand, during training all inter-
mediate results need storing to compute gradients. Moreover, accumulating gradients requires
higher precision to avoid numerical underflow (or overflow). This means that FP16 (or mixed
12.4. Hardware 507

precision with FP32) is the minimum required. All of this necessitates faster and larger memory
(HBM2 vs. GDDR6) and more processing power. For instance, NVIDIAʼs Turing164 T4 GPUs are
optimized for inference whereas the V100 GPUs are preferable for training.
Recall Fig. 12.4.5. Adding vector units to a processor core allowed us to increase throughput sig-
nificantly (in the example in the figure we were able to perform 16 operations simultaneously).
What if we added operations that optimized not just operations between vectors but also between
matrices? This strategy led to Tensor Cores (more on this shortly). Secondly, what if we added
many more cores? In a nutshell, these two strategies summarize the design decisions in GPUs.
Fig. 12.4.7 gives an overview over a basic processing block. It contains 16 integer and 16 float-
ing point units. In addition to that, two Tensor Cores accelerate a narrow subset of additional
operations relevant for deep learning. Each Streaming Multiprocessor (SM) consists of four such
blocks.
Fig. 12.4.7: NVIDIA Turing Processing Block (image courtesy of NVIDIA)
12 streaming multiprocessors are then grouped into graphics processing clusters which make up
the high-end TU102 processors. Ample memory channels and an L2 cache complement the setup.
Fig. 12.4.8 has the relevant details. One of the reasons for designing such a device is that individual
blocks can be added or removed as needed to allow for more compact chips and to deal with yield
issues (faulty modules might not be activated). Fortunately programming such devices is well
hidden from the casual deep learning researcher beneath layers of CUDA and framework code.
In particular, more than one of the programs might well be executed simultaneously on the GPU,
provided that there are available resources. Nonetheless it pays to be aware of the limitations of
the devices to avoid picking models that do not fit into device memory.
164
https://devblogs.nvidia.com/nvidia-turing-architecture-in-depth/

Fig. 12.4.8: NVIDIA Turing Architecture (image courtesy of NVIDIA)
A last aspect that is worth mentioning in more detail are TensorCores. They are an example of
a recent trend of adding more optimized circuits that are specifically effective for deep learning.
For instance, the TPU added a systolic array (Kung, 1988) for fast matrix multiplication. There
the design was to support a very small number (one for the first generation of TPUs) of large op-
erations. TensorCores are at the other end. They are optimized for small operations involving
between 4x4 and 16x16 matrices, depending on their numerical precision. Fig. 12.4.9 gives an
overview of the optimizations.
Fig. 12.4.9: NVIDIA TensorCores in Turing (image courtesy of NVIDIA)
Obviously when optimizing for computation we end up making certain compromises. One of
them is that GPUs are not very good at handling interrupts and sparse data. While there are no-
table exceptions, such as Gunrock165 (Wang et al., 2016), the access pattern of sparse matrices and
vectors do not go well with the high bandwidth burst read operations where GPUs excel. Match-
ing both goals is an area of active research. See e.g., DGL166 , a library tuned for deep learning on
165
https://github.com/gunrock/gunrock
166
http://dgl.ai
12.4. Hardware 509

graphs.
12.4.6 Networks and Buses
Whenever a single device is insufficient for optimization we need to transfer data to and from it
to synchronize processing. This is where networks and buses come in handy. We have a number
of design parameters: bandwidth, cost, distance and flexibility. On one end we have WiFi which
has a pretty good range, is very easy to use (no wires, after all), cheap but it offers comparatively
mediocre bandwidth and latency. No machine learning researcher within their right mind would
use it to build a cluster of servers. In what follows we focus on interconnects that are suitable for
deep learning.
• PCIe is a dedicated bus for very high bandwidth point to point connections (up to 16 Gbs on
PCIe 4.0) per lane. Latency is in the order of single-digit microseconds (5 µs). PCIe links
are precious. Processors only have a limited number of them: AMDʼs EPYC 3 has 128 lanes,
Intelʼs Xeon has up to 48 lanes per chip; on desktop grade CPUs the numbers are 20 (Ryzen
9) and 16 (Core i9) respectively. Since GPUs have typically 16 lanes this limits the number of
GPUs that can connect to the CPU at full bandwidth. After all, they need to share the links
with other high bandwidth peripherals such as storage and Ethernet. Just like with RAM
access, large bulk transfers are preferable due to reduced packet overhead.
• Ethernet is the most commonly used way of connecting computers. While it is significantly
slower than PCIe, it is very cheap and resilient to install and covers much longer distances.
Typical bandwidth for low-grade servers is 1 GBit/s. Higher end devices (e.g., C5 instances167
in the cloud) offer between 10 and 100 GBit/s bandwidth. As in all previous cases data trans-
mission has significant overheads. Note that we almost never use raw Ethernet directly but
rather a protocol that is executed on top of the physical interconnect (such as UDP or TCP/IP).
This adds further overhead. Like PCIe, Ethernet is designed to connect two devices, e.g., a
computer and a switch.
• Switches allow us to connect multiple devices in a manner where any pair of them can
carry out a (typically full bandwidth) point to point connection simultaneously. For in-
stance, Ethernet switches might connect 40 servers at high cross-sectional bandwidth. Note
that switches are not unique to traditional computer networks. Even PCIe lanes can be
switched168 . This occurs e.g., to connect a large number of GPUs to a host processor, as
is the case for the P2 instances169 .
• NVLink is an alternative to PCIe when it comes to very high bandwidth interconnects. It
offers up to 300 Gbit/s data transfer rate per link. Server GPUs (Volta V100) have 6 links
whereas consumer grade GPUs (RTX 2080 Ti) have only one link, operating at a reduced 100
Gbit/s rate. We recommend to use NCCL170 to achieve high data transfer between GPUs.
167
https://aws.amazon.com/ec2/instance-types/c5/
168
https://www.broadcom.com/products/pcie-switches-bridges/pcie-switches
169
https://aws.amazon.com/ec2/instance-types/p2/
170
https://github.com/NVIDIA/nccl

Summary
• Devices have overheads for operations. Hence it is important to aim for a small number of
large transfers rather than many small ones. This applies to RAM, SSDs, Networks and GPUs.
• Vectorization is key for performance. Make sure you are aware of the specific abilities of your
accelerator. E.g., some Intel Xeon CPUs are particularly good for INT8 operations, NVIDIA
Volta GPUs excel at FP16 matrix-matrix operations and NVIDIA Turing shines at FP16, INT8
and INT4 operations.
• Numerical overflow due to small datatypes can be a problem during training (and to a lesser
extent during inference).
• Aliasing can significantly degrade performance. For instance, memory alignment on 64 bit
CPUs should be done with respect to 64 bit boundaries. On GPUs it is a good idea to keep
convolution sizes aligned e.g., to TensorCores.
• Match your algorithms to the hardware (memory footprint, bandwidth, etc.). Great speedup
(orders of magnitude) can be achieved when fitting the parameters into caches.
• We recommend that you sketch out the performance of a novel algorithm on paper before
verifying the experimental results. Discrepancies of an order-of-magnitude or more are rea-
sons for concern.
• Use profilers to debug performance bottlenecks.
• Training and inference hardware have different sweet spots in terms of price / performance.
12.4.7 More Latency Numbers
The summary in Table 12.4.1 and Table 12.4.2 are due to Eliot Eshelman171 who maintains an
updated version of the numbers as a GitHub Gist172 .
Table 12.4.1: Common Latency Numbers.

Action Time Notes
L1 cache reference/hit 1.5 ns 4 cycles
Floating-point add/mult/FMA 1.5 ns 4 cycles
L2 cache reference/hit 5 ns 12 ~ 17 cycles
Branch mispredict 6 ns 15 ~ 20 cycles
L3 cache hit (unshared cache) 16 ns 42 cycles
L3 cache hit (shared in another core) 25 ns 65 cycles
Mutex lock/unlock 25 ns
L3 cache hit (modified in another core) 29 ns 75 cycles
L3 cache hit (on a remote CPU socket) 40 ns 100 ~ 300 cycles (40 ~ 116 ns)
QPI hop to a another CPU (per hop) 40 ns
64MB memory ref. (local CPU) 46 ns TinyMemBench on Broadwell E5-2690v4
64MB memory ref. (remote CPU) 70 ns TinyMemBench on Broadwell E5-2690v4
256MB memory ref. (local CPU) 75 ns TinyMemBench on Broadwell E5-2690v4
Intel Optane random write 94 ns UCSD Non-Volatile Systems Lab
continues on next page
171
https://gist.github.com/eshelman
172
https://gist.github.com/eshelman/343a1c46cb3fba142c1afdcdeec17646
12.4. Hardware 511

Table 12.4.1 – continued from previous page
Action Time Notes
256MB memory ref. (remote CPU) 120 ns TinyMemBench on Broadwell E5-2690v4
Intel Optane random read 305 ns UCSD Non-Volatile Systems Lab
Send 4KB over 100 Gbps HPC fabric 1 µs MVAPICH2 over Intel Omni-Path
Compress 1KB with Google Snappy 3 µs
Send 4KB over 10 Gbps ethernet 10 µs
Write 4KB randomly to NVMe SSD 30 µs DC P3608 NVMe SSD (QOS 99% is 500µs)
Transfer 1MB to/from NVLink GPU 30 µs ~33GB/s on NVIDIA 40GB NVLink
Transfer 1MB to/from PCI-E GPU 80 µs ~12GB/s on PCIe 3.0 x16 link
Read 4KB randomly from NVMe SSD 120 µs DC P3608 NVMe SSD (QOS 99%)
Read 1MB sequentially from NVMe SSD 208 µs ~4.8GB/s DC P3608 NVMe SSD
Write 4KB randomly to SATA SSD 500 µs DC S3510 SATA SSD (QOS 99.9%)
Read 4KB randomly from SATA SSD 500 µs DC S3510 SATA SSD (QOS 99.9%)
Round trip within same datacenter 500 µs One-way ping is ~250µs
Read 1MB sequentially from SATA SSD 2 ms ~550MB/s DC S3510 SATA SSD
Read 1MB sequentially from disk 5 ms ~200MB/s server HDD
Random Disk Access (seek+rotation) 10 ms
Send packet CA->Netherlands->CA 150 ms
Table 12.4.2: Latency Numbers for NVIDIA Tesla GPUs.

Action Time Notes
GPU Shared Memory access 30 ns 30~90 cycles (bank conflicts add latency)
GPU Global Memory access 200 ns 200~800 cycles
Launch CUDA kernel on GPU 10 µs Host CPU instructs GPU to start kernel
Transfer 1MB to/from NVLink GPU 30 µs ~33GB/s on NVIDIA 40GB NVLink
Transfer 1MB to/from PCI-E GPU 80 µs ~12GB/s on PCI-Express x16 link
Exercises
1. Write C code to test whether there is any difference in speed between accessing memory
aligned or misaligned relative to the external memory interface. Hint: be careful of caching
effects.
2. Test the difference in speed between accessing memory in sequence or with a given stride.
3. How could you measure the cache sizes on a CPU?
4. How would you lay out data across multiple memory channels for maximum bandwidth?
How would you lay it out if you had many small threads?
5. An enterprise class HDD is spinning at 10,000 rpm. What is the absolutely minimum time
an HDD needs to spend worst case before it can read data (you can assume that heads move
almost instantaneously)? Why are 2.5” HDDs becoming popular for commercial servers (rel-
ative to 3.5” and 5.25” drives)?
6. Assume that an HDD manufacturer increases the storage density from 1 Tbit per square inch
to 5 Tbit per square inch. How much information can you store on a ring on a 2.5” HDD? Is
there a difference between the inner and outer tracks?

7. The AWS P2 instances have 16 K80 Kepler GPUs. Use lspci on a p2.16xlarge and a p2.8xlarge
instance to understand how the GPUs are connected to the CPUs. Hint: keep your eye out
for PCI PLX bridges.
8. Going from 8 bit to 16 bit datatypes increases the amount of silicon approximately by 4x.
Why? Why might NVIDIA have added INT4 operations to their Turing GPUs.
9. Given 6 high speed links between GPUs (such as for the Volta V100 GPUs), how would you
connect 8 of them? Look up the connectivity used in the P3.16xlarge servers.
10. How much faster is it to read forward through memory vs. reading backwards? Does this
number differ between different computers and CPU vendors? Why? Write C code and ex-
periment with it.
11. Can you measure the cache size of your disk? What is it for a typical HDD? Do SSDs need a
cache?
12. Measure the packet overhead when sending messages across the Ethernet. Look up the dif-
ference between UDP and TCP/IP connections.
13. Direct Memory Access allows devices other than the CPU to write (and read) directly to
(from) memory. Why is this a good idea?
14. Look at the performance numbers for the Turing T4 GPU. Why does the performance ʻonlyʼ
double as you go from FP16 to INT8 and INT4?
15. What is the shortest time it should take for a packet on a roundtrip between San Francisco
and Amsterdam? Hint: you can assume that the distance is 10,000km.
Discussions173
12.5 Training on Multiple GPUs
So far we discussed how to train models efficiently on CPUs and GPUs. We even showed how deep
learning frameworks such as MXNet (and TensorFlow) allow one to parallelize computation and
communication automatically between them in Section 12.3. Lastly, we showed in Section 5.6
how to list all available GPUs on a computer using nvidia-smi. What we did not discuss is how to
actually parallelize deep learning training (we omit any discussion of inference on multiple GPUs
here as it is a rather rarely used and advanced topic that goes beyond the scope of this book).
Instead, we implied in passing that one would somehow split the data across multiple devices and
make it work. The present section fills in the details and shows how to train a network in parallel
when starting from scratch. Details on how to take advantage of functionality in Gluon is relegated
to Section 12.6. We assume that the reader is familiar with minibatch SGD algorithms such as the
ones described in Section 11.5.
173
12.5. Training on Multiple GPUs 513

12.5.1 Splitting the Problem
Let us start with a simple computer vision problem and a slightly archaic network, e.g., with mul-
tiple layers of convolutions, pooling, and possibly a few dense layers in the end. That is, let us
start with a network that looks quite similar to LeNet (LeCun et al., 1998) or AlexNet (Krizhevsky
et al., 2012). Given multiple GPUs (2 if it is a desktop server, 4 on a g4dn.12xlarge, 8 on an AWS
p3.16xlarge, or 16 on a p2.16xlarge), we want to partition training in a manner as to achieve good
speedup while simultaneously benefitting from simple and reproducible design choices. Multi-
ple GPUs, after all, increase both memory and compute ability. In a nutshell, we have a number of
choices, given a minibatch of training data that we want to classify.
• We could partition the network layers across multiple GPUs. That is, each GPU takes as input
the data flowing into a particular layer, processes data across a number of subsequent layers
and then sends the data to the next GPU.
– This allows us to process data with larger networks when compared to what a single
GPU could handle.
– Memory footprint per GPU can be well controlled (it is a fraction of the total network
footprint)
– The interface between layers (and thus GPUs) requires tight synchronization. This can
be tricky, in particular if the computational workloads are not properly matched be-
tween layers. The problem is exacerbated for large numbers of GPUs.
– The interface between layers requires large amounts of data transfer (activations, gra-
dients). This may overwhelm the bandwidth of the GPU buses.
– Compute intensive, yet sequential operations are nontrivial to partition. See e.g.,
(Mirhoseini et al., 2017) for a best effort in this regard. It remains a difficult problem
and it is unclear whether it is possible to achieve good (linear) scaling on nontrivial
problems. We do not recommend it unless there is excellent framework / OS support
for chaining together multiple GPUs.
• We could split the work required by individual layers. For instance, rather than computing
64 channels on a single GPU we could split up the problem across 4 GPUs, each of which
generate data for 16 channels. Likewise, for a dense layer we could split the number of
output neurons. Fig. 12.5.1 illustrates this design. The figure is taken from (Krizhevsky et al.,
2012) where this strategy was used to deal with GPUs that had a very small memory footprint
(2GB at the time).
– This allows for good scaling in terms of computation, provided that the number of chan-
nels (or neurons) is not too small.
– Multiple GPUs can process increasingly larger networks since the memory available
scales linearly.
– We need a very large number of synchronization / barrier operations since each layer
depends on the results from all other layers.
– The amount of data that needs to be transferred is potentially even larger than when
distributing layers across GPUs. We do not recommend this approach due to its band-
width cost and complexity.

Fig. 12.5.1: Model parallelism in the original AlexNet design due to limited GPU memory.
• Lastly we could partition data across multiple GPUs. This way all GPUs perform the same
type of work, albeit on different observations. Gradients are aggregated between GPUs after
each minibatch.
– This is the simplest approach and it can be applied in any situation.
– Adding more GPUs does not allow us to train larger models.
– We only need to synchronize after each minibatch. That said, it is highly desirable to
start exchanging gradients parameters already while others are still being computed.
– Large numbers of GPUs lead to very large minibatch sizes, thus reducing training effi-
ciency.
By and large, data parallelism is the most convenient way to proceed, provided that we have ac-
cess to GPUs with sufficiently large memory. See also (Li et al., 2014) for a detailed description of
partitioning for distributed training. GPU memory used to be a problem in the early days of deep
learning. By now this issue has been resolved for all but the most unusual cases. We focus on data
parallelism in what follows.
12.5.2 Data Parallelism
Assume that there are k GPUs on a machine. Given the model to be trained, each GPU will maintain
a complete set of model parameters independently. Training proceeds as follows (see Fig. 12.5.2
for details on data parallel training on two GPUs).

Fig. 12.5.2: Calculation of minibatch stochastic gradient using data parallelism and two GPUs.
• In any iteration of training, given a random minibatch, we split the examples in the batch
into k portions and distribute them evenly across the GPUs.
• Each GPU calculates loss and gradient of the model parameters based on the minibatch sub-
set it was assigned and the model parameters it maintains.
• The local gradients of each of the k GPUs are aggregated to obtain the current minibatch
stochastic gradient.
• The aggregate gradient is re-distributed to each GPU.
• Each GPU uses this minibatch stochastic gradient to update the complete set of model pa-
rameters that it maintains.
A comparison of different ways of parallelization on multiple GPUs is depicted in Fig. 12.5.3. Note
that in practice we increase the minibatch size k-fold when training on k GPUs such that each GPU
has the same amount of work to do as if we were training on a single GPU only. On a 16 GPU server
this can increase the minibatch size considerably and we may have to increase the learning rate
accordingly. Also note that Section 7.5 needs to be adjusted (e.g., by keeping a separate batch
norm coefficient per GPU). In what follows we will use Section 6.6 as the toy network to illustrate
multi-GPU training. As always we begin by importing the relevant packages and modules.

Fig. 12.5.3: Parallelization on multiple GPUs. From left to right - original problem, network parti-
tioning, layer partitioning, data parallelism.
%matplotlib inline
npx.set_np()
12.5.3 A Toy Network
We use LeNet as introduced in Section 6.6. We define it from scratch to illustrate parameter ex-
change and synchronization in detail.
# Initialize model parameters

scale = 0.01
W1 = np.random.normal(scale=scale, size=(20, 1, 3, 3))
b1 = np.zeros(20)
W2 = np.random.normal(scale=scale, size=(50, 20, 5, 5))
b2 = np.zeros(50)
W3 = np.random.normal(scale=scale, size=(800, 128))
b3 = np.zeros(128)
W4 = np.random.normal(scale=scale, size=(128, 10))
b4 = np.zeros(10)
params = [W1, b1, W2, b2, W3, b3, W4, b4]
# Define the model

def lenet(X, params):
h1_conv = npx.convolution(data=X, weight=params[0], bias=params[1],
kernel=(3, 3), num_filter=20)
h1_activation = npx.relu(h1_conv)
h1 = npx.pooling(data=h1_activation, pool_type='avg', kernel=(2, 2),
stride=(2, 2))

h2_conv = npx.convolution(data=h1, weight=params[2], bias=params[3],
kernel=(5, 5), num_filter=50)
h2_activation = npx.relu(h2_conv)
h2 = npx.pooling(data=h2_activation, pool_type='avg', kernel=(2, 2),
stride=(2, 2))
h2 = h2.reshape(h2.shape[0], -1)
h3_linear = np.dot(h2, params[4]) + params[5]
h3 = npx.relu(h3_linear)
y_hat = np.dot(h3, params[6]) + params[7]
return y_hat
# Cross-entropy loss function

12.5.4 Data Synchronization
For efficient multi-GPU training we need two basic operations: firstly we need to have the ability
to distribute a list of parameters to multiple devices and to attach gradients (get_params). Without
parameters it is impossible to evaluate the network on a GPU. Secondly, we need the ability to sum
parameters across multiple devices, i.e., we need an allreduce function.
def get_params(params, device):

new_params = [p.copyto(device) for p in params]
for p in new_params:
p.attach_grad()
return new_params
Let us try it out by copying the model parameters of lenet to gpu(0).
new_params = get_params(params, d2l.try_gpu(0))

print('b1 weight:', new_params[1])
print('b1 grad:', new_params[1].grad)
b1 weight: [0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.] @gpu(0)

b1 grad: [0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.] @gpu(0)
Since we didnʼt perform any computation yet, the gradient with regard to the bias weights is still 0.
Now let us assume that we have a vector distributed across multiple GPUs. The following allre-
duce function adds up all vectors and broadcasts the result back to all GPUs. Note that for this to
work we need to copy the data to the device accumulating the results.
def allreduce(data):
for i in range(1, len(data)):
data[0][:] += data[i].copyto(data[0].ctx)
for i in range(1, len(data)):
data[0].copyto(data[i])
Let us test this by creating vectors with different values on different devices and aggregate them.

data = [np.ones((1, 2), ctx=d2l.try_gpu(i)) * (i + 1) for i in range(2)]
print('before allreduce:\n', data[0], '\n', data[1])
allreduce(data)
print('after allreduce:\n', data[0], '\n', data[1])
before allreduce:
[[1. 1.]] @gpu(0)
[[2. 2.]] @gpu(1)
after allreduce:
[[3. 3.]] @gpu(0)
[[3. 3.]] @gpu(1)
12.5.5 Distributing Data
We need a simple utility function to distribute a minibatch evenly across multiple GPUs. For in-
stance, on 2 GPUs weʼd like to have half of the data to be copied to each of the GPUs. Since it is
more convenient and more concise, we use the built-in split and load function in Gluon (to try it
out on a 4 × 5 matrix).
data = np.arange(20).reshape(4, 5)
devices = [npx.gpu(0), npx.gpu(1)]
split = gluon.utils.split_and_load(data, devices)
print('input :', data)
print('load into', devices)
print('output:', split)
input : [[ 0. 1. 2. 3. 4.]
[ 5. 6. 7. 8. 9.]
[10. 11. 12. 13. 14.]
[15. 16. 17. 18. 19.]]
load into [gpu(0), gpu(1)]
output: [array([[0., 1., 2., 3., 4.],
[5., 6., 7., 8., 9.]], ctx=gpu(0)), array([[10., 11., 12., 13., 14.],
[15., 16., 17., 18., 19.]], ctx=gpu(1))]
For later reuse we define a split_batch function which splits both data and labels.
#@save
def split_batch(X, y, devices):
"""Split `X` and `y` into multiple devices."""
assert X.shape[0] == y.shape[0]
return (gluon.utils.split_and_load(X, devices),
gluon.utils.split_and_load(y, devices))

12.5.6 Training
Now we can implement multi-GPU training on a single minibatch. Its implementation is primarily
based on the data parallelism approach described in this section. We will use the auxiliary func-
tions we just discussed, allreduce and split_and_load, to synchronize the data among multiple
GPUs. Note that we do not need to write any specific code to achieve parallelism. Since the com-
putational graph does not have any dependencies across devices within a minibatch, it is executed
in parallel automatically.
def train_batch(X, y, device_params, devices, lr):

X_shards, y_shards = split_batch(X, y, devices)
with autograd.record(): # Loss is calculated separately on each GPU
losses = [loss(lenet(X_shard, device_W), y_shard)
for X_shard, y_shard, device_W in zip(
X_shards, y_shards, device_params)]
for l in losses: # Back Propagation is performed separately on each GPU
l.backward()
# Sum all gradients from each GPU and broadcast them to all GPUs
for i in range(len(device_params[0])):
allreduce([device_params[c][i].grad for c in range(len(devices))])
# The model parameters are updated separately on each GPU
for param in device_params:
d2l.sgd(param, lr, X.shape[0]) # Here, we use a full-size batch
Now, we can define the training function. It is slightly different from the ones used in the previous
chapters: we need to allocate the GPUs and copy all the model parameters to all devices. Obvi-
ously each batch is processed using train_batch to deal with multiple GPUs. For convenience
(and conciseness of code) we compute the accuracy on a single GPU (this is inefficient since the
other GPUs are idle).
def train(num_gpus, batch_size, lr):

devices = [d2l.try_gpu(i) for i in range(num_gpus)]
# Copy model parameters to num_gpus GPUs
device_params = [get_params(params, d) for d in devices]
# num_epochs, times, acces = 10, [], []
num_epochs = 10
animator = d2l.Animator('epoch', 'test acc', xlim=[1, num_epochs])
timer = d2l.Timer()
timer.start()
# Perform multi-GPU training for a single minibatch
train_batch(X, y, device_params, devices, lr)
npx.waitall()
timer.stop()
# Verify the model on GPU 0
animator.add(epoch + 1, (d2l.evaluate_accuracy_gpu(
lambda x: lenet(x, device_params[0]), test_iter, devices[0]),))
print(f'test acc: {animator.Y[0][-1]:.2f}, {timer.avg():.1f} sec/epoch '
f'on {str(devices)}')

12.5.7 Experiment
Let us see how well this works on a single GPU. We use a batch size of 256 and a learning rate of
0.2.
train(num_gpus=1, batch_size=256, lr=0.2)
test acc: 0.85, 2.5 sec/epoch on [gpu(0)]
By keeping the batch size and learning rate unchanged and changing the number of GPUs to 2,
we can see that the improvement in test accuracy is roughly the same as in the results from the
previous experiment. In terms of the optimization algorithms, they are identical. Unfortunately
there is no meaningful speedup to be gained here: the model is simply too small; moreover we only
have a small dataset, where our slightly unsophisticated approach to implementing multi-GPU
training suffered from significant Python overhead. We will encounter more complex models and
more sophisticated ways of parallelization going forward. Let us see what happens nonetheless
for Fashion-MNIST.
test acc: 0.86, 4.2 sec/epoch on [gpu(0), gpu(1)]

Summary
• There are multiple ways to split deep network training over multiple GPUs. We could split
them between layers, across layers, or across data. The former two require tightly chore-
ographed data transfers. Data parallelism is the simplest strategy.
• Data parallel training is straightforward. However, it increases the effective minibatch size
to be efficient.
• Data is split across multiple GPUs, each GPU executes its own forward and backward opera-
tion and subsequently gradients are aggregated and results broadcast back to the GPUs.
• Large minibatches may require a slightly increased learning rate.
Exercises
1. When training on multiple GPUs, change the minibatch size from b to k · b, i.e., scale it up by
the number of GPUs.
2. Compare accuracy for different learning rates. How does it scale with the number of GPUs.
3. Implement a more efficient allreduce that aggregates different parameters on different GPUs
(why is this more efficient in the first place).
4. Implement multi-GPU test accuracy computation.
Discussions174
12.6 Concise Implementation for Multiple GPUs
Implementing parallelism from scratch for every new model is no fun. Moreover, there is signif-
icant benefit in optimizing synchronization tools for high performance. In the following we will
show how to do this using Gluon. The math and the algorithms are the same as in Section 12.5. As
before we begin by importing the required modules (quite unsurprisingly you will need at least
two GPUs to run this notebook).

npx.set_np()
174

12.6.1 A Toy Network
Let us use a slightly more meaningful network than LeNet from the previous section thatʼs still
sufficiently easy and quick to train. We pick a ResNet-18 variant (He et al., 2016a). Since the input
images are tiny we modify it slightly. In particular, the difference to Section 7.6 is that we use a
smaller convolution kernel, stride, and padding at the beginning. Moreover, we remove the max-
pooling layer.
#@save
def resnet18(num_classes):
"""A slightly modified ResNet-18 model."""
blk.add(d2l.Residual(
num_channels, use_1x1conv=True, strides=2))
else:
blk.add(d2l.Residual(num_channels))
return blk
# This model uses a smaller convolution kernel, stride, and padding and
# removes the maximum pooling layer
nn.BatchNorm(), nn.Activation('relu'))
net.add(nn.GlobalAvgPool2D(), nn.Dense(num_classes))
return net
12.6.2 Parameter Initialization and Logistics
The initialize method allows us to set initial defaults for parameters on a device of our choice.
For a refresher see Section 4.8. What is particularly convenient is that it also lets us initialize the
network on multiple devices simultaneously. Let us try how this works in practice.
net = resnet18(10)
# get a list of GPUs
devices = d2l.try_all_gpus()
# initialize the network on all of them
net.initialize(init=init.Normal(sigma=0.01), ctx=devices)
Using the split_and_load function introduced in the previous section we can divide a minibatch
of data and copy portions to the list of devices provided by the context variable. The network
object automatically uses the appropriate GPU to compute the value of the forward propagation.
As before we generate 4 observations and split them over the GPUs.
x = np.random.uniform(size=(4, 1, 28, 28))

x_shards = gluon.utils.split_and_load(x, devices)
net(x_shards[0]), net(x_shards[1])
12.6. Concise Implementation for Multiple GPUs 523

(array([[ 2.2610238e-06, 2.2045988e-06, -5.4046791e-06, 1.2869939e-06,
5.1373154e-06, -3.8298003e-06, 1.4338991e-07, 5.4683424e-06,
-2.8279187e-06, -3.9651113e-06],
[ 2.0698699e-06, 2.0084674e-06, -5.6382478e-06, 1.0498467e-06,
5.5506403e-06, -4.1065500e-06, 6.0830121e-07, 5.4521761e-06,
-3.7365019e-06, -4.1891644e-06]], ctx=gpu(0)),
array([[ 2.4629785e-06, 2.6015518e-06, -5.4362636e-06, 1.2938228e-06,
5.6387912e-06, -4.1360131e-06, 3.5758842e-07, 5.5125242e-06,
-3.1957325e-06, -4.2976335e-06],
[ 1.9431677e-06, 2.2600434e-06, -5.2698206e-06, 1.4807442e-06,
5.4830939e-06, -3.9678885e-06, 7.5754315e-08, 5.6764375e-06,
-3.2530229e-06, -4.0943960e-06]], ctx=gpu(1)))
Once data passes through the network, the corresponding parameters are initialized on the device
the data passed through. This means that initialization happens on a per-device basis. Since we
picked GPU 0 and GPU 1 for initialization, the network is initialized only there, and not on the
CPU. In fact, the parameters do not even exist on the device. We can verify this by printing out the
parameters and observing any errors that might arise.
weight = net[0].params.get('weight')
try:
weight.data()
except RuntimeError:
print('not initialized on cpu')
weight.data(devices[0])[0], weight.data(devices[1])[0]
not initialized on cpu
(array([[[ 0.01382882, -0.01183044, 0.01417866],

[-0.00319718, 0.00439528, 0.02562625],
[-0.00835081, 0.01387452, -0.01035946]]], ctx=gpu(0)),
array([[[ 0.01382882, -0.01183044, 0.01417866],
[-0.00319718, 0.00439528, 0.02562625],
[-0.00835081, 0.01387452, -0.01035946]]], ctx=gpu(1)))
Lastly let us replace the code to evaluate the accuracy by one that works in parallel across multiple
devices. This serves as a replacement of the evaluate_accuracy_gpu function from Section 6.6.
The main difference is that we split a batch before invoking the network. All else is essentially
identical.
#@save
def evaluate_accuracy_gpus(net, data_iter, split_f=d2l.split_batch):
# Query the list of devices
devices = list(net.collect_params().values())[0].list_ctx()
metric = d2l.Accumulator(2) # num_corrected_examples, num_examples
for features, labels in data_iter:
X_shards, y_shards = split_f(features, labels, devices)
# Run in parallel
pred_shards = [net(X_shard) for X_shard in X_shards]
metric.add(sum(float(d2l.accuracy(pred_shard, y_shard)) for
pred_shard, y_shard in zip(

pred_shards, y_shards)), labels.size)
12.6.3 Training
As before, the training code needs to perform a number of basic functions for efficient parallelism:
• Network parameters need to be initialized across all devices.
• While iterating over the dataset minibatches are to be divided across all devices.
• We compute the loss and its gradient in parallel across devices.
• Losses are aggregated (by the trainer method) and parameters are updated accordingly.
In the end we compute the accuracy (again in parallel) to report the final value of the network.
The training routine is quite similar to implementations in previous chapters, except that we need
to split and aggregate data.
def train(num_gpus, batch_size, lr):

ctx = [d2l.try_gpu(i) for i in range(num_gpus)]
net.initialize(init=init.Normal(sigma=0.01), ctx=ctx, force_reinit=True)
timer, num_epochs = d2l.Timer(), 10
animator = d2l.Animator('epoch', 'test acc', xlim=[1, num_epochs])
timer.start()
for features, labels in train_iter:
X_shards, y_shards = d2l.split_batch(features, labels, ctx)
losses = [loss(net(X_shard), y_shard) for X_shard, y_shard
in zip(X_shards, y_shards)]
for l in losses:
l.backward()
npx.waitall()
timer.stop()
animator.add(epoch + 1, (evaluate_accuracy_gpus(net, test_iter),))
print(f'test acc: {animator.Y[0][-1]:.2f}, {timer.avg():.1f} sec/epoch '
f'on {str(ctx)}')
12.6. Concise Implementation for Multiple GPUs 525

12.6.4 Experiments
Let us see how this works in practice. As a warmup we train the network on a single GPU.
test acc: 0.93, 13.1 sec/epoch on [gpu(0)]
Next we use 2 GPUs for training. Compared to LeNet the model for ResNet-18 is considerably
more complex. This is where parallelization shows its advantage. The time for computation is
meaningfully larger than the time for synchronizing parameters. This improves scalability since
the overhead for parallelization is less relevant.
test acc: 0.93, 6.8 sec/epoch on [gpu(0), gpu(1)]

Summary
• Gluon provides primitives for model initialization across multiple devices by providing a
context list.
• Data is automatically evaluated on the devices where the data can be found.
• Take care to initialize the networks on each device before trying to access the parameters on
that device. Otherwise you will encounter an error.
• The optimization algorithms automatically aggregate over multiple GPUs.
Exercises
1. This section uses ResNet-18. Try different epochs, batch sizes, and learning rates. Use more
GPUs for computation. What happens if you try this on a p2.16xlarge instance with 16 GPUs?
2. Sometimes, different devices provide different computing power. We could use the GPUs
and the CPU at the same time. How should we divide the work? Is it worth the effort? Why?
Why not?
3. What happens if we drop npx.waitall()? How would you modify training such that you have
an overlap of up to two steps for parallelism?
Discussions175
12.7 Parameter Servers
As we move from single GPUs to multiple GPUs and then to multiple servers containing multi-
ple GPUs, possibly all spread out across multiple racks and network switches our algorithms for
distributed and parallel training need to become much more sophisticated. Details matter since
different interconnects have very different bandwidth (e.g., NVLink can offer up to 100GB/s across
6 links in an appropriate setting, PCIe 3.0 16x lanes offer 16GB/s while even high speed 100 GbE
Ethernet only amounts to 10GB/s). At the same time it is unreasonable to expect that a statistical
modeler be an expert in networking and systems.
The core idea of the parameter server was introduced in (Smola & Narayanamurthy, 2010) in the
context of distributed latent variable models. A description of the push and pull semantics then
followed in (Ahmed et al., 2012) and a description of the system and an open source library fol-
lowed in (Li et al., 2014). In the following we will motivate the components needed for efficiency.
175
12.7. Parameter Servers 527

12.7.1 Data Parallel Training
Let us review the data parallel training approach to distributed training. We will use this to the
exclusion of all others in this section since it is significantly simpler to implement in practice.
There are virtually no use cases (besides deep learning on graphs) where any other strategy for
parallelism is preferred since GPUs have plenty of memory nowadays. Fig. 12.7.1 describes the
variant of data parallelism that we implemented in the previous section. The key aspect in it is
that the aggregation of gradients occurs on GPU0 before the updated parameters are rebroadcast
to all GPUs.
Fig. 12.7.1: Left: single GPU training; Right: a variant of multi-GPU training. It proceeds as fol-
lows. (1) we compute loss and gradient, (2) all gradients are aggregated on one GPU, (3) parameter
update happens and the parameters are re-distributed to all GPUs.
In retrospect, the decision to aggregate on GPU0 seems rather ad-hoc. After all, we might just as
well aggregate on the CPU. In fact, we could even decide to aggregate some of the parameters on
one GPU and some others on another. Provided that the optimization algorithm supports this,
there is no real reason for why we could not. For instance, if we have four parameter vectors
v1 , . . . , v4 with associated gradients g1 , . . . , g4 we could aggregate the gradients on one GPU each.
∑
gi = gij (12.7.1)
j∈GPUs

This reasoning seems arbitrary and frivolous. After all, the math is the same throughout. How-
ever, we are dealing with real physical hardware where different buses have different bandwidth
as discussed in Section 12.4. Consider a real 4-way GPU server as described in Fig. 12.7.2. If it
is particularly well connected, it might have a 100 GbE network card. More typical numbers are
in the 1-10 GbE range with an effective bandwidth of 100MB/s to 1GB/s. Since the CPUs have too
few PCIe lanes to connect to all GPUs directly (e.g., consumer grade Intel CPUs have 24 lanes) we
need a multiplexer176 . The bandwidth from the CPU on a 16x Gen3 link is 16GB/s. This is also the
speed at which each of the GPUs is connected to the switch. This means that it is more effective to
communicate between the devices.
Fig. 12.7.2: A 4-way GPU server.
For the sake of the argument let us assume that the gradients ʻweightʼ 160MB. In this case it takes
30ms to send the gradients from all 3 remaining GPUs to the fourth one (each transfer takes 10ms
= 160MB / 16 GB/s). Add another 30ms to transmit the weight vectors back we arrive at a total of
60ms. If we send all data to the CPU we incur a penalty of 40ms since each of the four GPUs needs
to send the data to the CPU, yielding a total of 80ms. Lastly assume that we are able to split the
gradients into 4 parts of 40MB each. Now we can aggregate each of the parts on a different GPU
simultaneously since the PCIe switch offers a full-bandwidth operation between all links. Instead
of 30ms this takes 7.5ms, yielding a total of 15ms for a synchronization operation. In short, de-
pending on how we synchronize parameters the same operation can take anywhere from 15ms to
80ms. Fig. 12.7.3 depicts the different strategies for exchanging parameters.
176
https://www.broadcom.com/products/pcie-switches-bridges/pcie-switches

Fig. 12.7.3: Synchronization strategies.
Note that we have yet another tool at our disposal when it comes to improving performance: in a
deep network it takes some time to compute all gradients from the top to the bottom. We can begin
synchronizing gradients for some parameter groups even while we are still busy computing them
for others (the technical details for that are somewhat involved). See e.g., (Sergeev & DelBalso,
2018) for details on how to do this in Horovod177 .
12.7.2 Ring Synchronization
When it comes to synchronization on modern deep learning hardware we often encounter sig-
nificantly bespoke network connectivity. For instance, the AWS P3.16xlarge and NVIDIA DGX-2
instances share the connectivity structure of Fig. 12.7.4. Each GPU connects to a host CPU via a
PCIe link which operates at best at 16 GB/s. Additionally each GPU also has 6 NVLink connections,
each of which is capable of transferring 300 Gbit/s bidirectionally. This amounts to around 18 GB/s
per link per direction. In short, the aggregate NVLink bandwidth is significantly higher than the
PCIe bandwidth. The question is how to use it most efficiently.
177
https://github.com/horovod/horovod

Fig. 12.7.4: NVLink connectivity on 8GPU V100 servers (image courtesy of NVIDIA).
It turns out (Wang et al., 2018) that the optimal synchronization strategy is to decompose the net-
work into two rings and to use them to synchronize data directly. Fig. 12.7.5 illustrates that the
network can be decomposed into one ring (1-2-3-4-5-6-7-8-1) with double NVLink bandwidth and
into one (1-4-6-3-5-8-2-7-1) with regular bandwidth. Designing an efficient synchronization proto-
col in this case is nontrivial.

Fig. 12.7.5: Decomposition of the NVLink network into two rings.
Consider the following thought experiment: given a ring of n compute nodes (or GPUs) we can
send gradients from the first to the second node. There it is added to the local gradient and sent
on to the third node, and so on. After n − 1 steps the aggregate gradient can be found in the last-
visited node. That is, the time to aggregate gradients grows linearly with the number of nodes.
But if we do this the algorithm is quite inefficient. After all, at any time there is only one of the
nodes communicating. What if we broke the gradients into n chunks and started synchronizing
chunk i starting at node i. Since each chunk is of size 1/n the total time is now (n − 1)/n ≈ 1. In
other words, the time spent to aggregate gradients does not grow as we increase the size of the ring.
This is quite an astonishing result. Fig. 12.7.6 illustrates the sequence of steps on n = 4 nodes.

Fig. 12.7.6: Ring synchronization across 4 nodes. Each node starts transmitting parts of gradients
to its left neighbor until the assembled gradient can be found in its right neighbor.
If we use the same example of synchronizing 160MB across 8 V100 GPUs we arrive at approximately
2 · 160MB/(3 · 18GB/s) ≈ 6ms This is quite a bit better than using the PCIe bus, even though we are
now using 8 GPUs. Note that in practice these numbers are quite a bit worse, since deep learning
frameworks often fail to assemble communication into large burst transfers. Moreover, timing is
critical. Note that there is a common misconception that ring synchronization is fundamentally
different from other synchronization algorithms. The only difference is that the synchronization
path is somewhat more elaborate when compared to a simple tree.
12.7.3 Multi-Machine Training
Distributed training on multiple machines adds a further challenge: we need to communicate

with servers that are only connected across a comparatively lower bandwidth fabric which can be
over an order of magnitude slower in some cases. Synchronization across devices is tricky. After
all, different machines running training code will have subtly different speed. Hence we need to
synchronize them if we want to use synchronous distributed optimization. Fig. 12.7.7 illustrates
how distributed parallel training occurs.
1. A (different) batch of data is read on each machine, split across multiple GPUs and trans-
ferred to GPU memory. There predictions and gradients are computed on each GPU batch
separately.
2. The gradients from all local GPUs are aggregated on one GPU (or alternatively parts of it are
aggregated over different GPUs.
3. The gradients are sent to the CPU.
4. The CPU sends the gradients to a central parameter server which aggregates all the gradi-
ents.

5. The aggregate gradients are then used to update the weight vectors and the updated weight
vectors are broadcast back to the individual CPUs.
6. The information is sent to one (or multiple) GPUs.
7. The updated weight vectors are spread across all GPUs.
Fig. 12.7.7: Multi-machine multi-GPU distributed parallel training.
Each of these operations seems rather straightforward. And, indeed, they can be carried out ef-
ficiently within a single machine. Once we look at multiple machines, though, we can see that
the central parameter server becomes the bottleneck. After all, the bandwidth per server is lim-
ited, hence for m workers the time it takes to send all gradients to the server is O(m). We can
break through this barrier by increasing the number of servers to n. At this point each server only
needs to store O(1/n) of the parameters, hence the total time for updates and optimization be-
comes O(m/n). Matching both numbers yields constant scaling regardless of how many workers
we are dealing with. In practice we use the same machines both as workers and as servers. Fig.
12.7.8 illustrates the design. See also (Li et al., 2014) for details. In particular, ensuring that mul-
tiple machines work without unreasonable delays is nontrivial. We omit details on barriers and
will only briefly touch on synchronous and asynchronous updates below.

Fig. 12.7.8: Top - a single parameter server is a bottleneck since its bandwidth is finite. Bottom -
multiple parameter servers store parts of the parameters with aggregate bandwidth.
12.7.4 (key,value) Stores
Implementing the steps required for distributed multi-GPU training in practice is nontrivial. In
particular, given the many different choices that we might encounter. This is why it pays to use a
common abstraction, namely that of a (key,value) store with redefined update semantics. Across
many servers and many GPUs the gradient computation can be defined as
∑ ∑
gi = gijk . (12.7.2)
k∈workers j∈GPUs
The key aspect in this operation is that it is a commutative reduction, that is, it turns many vectors
into one and the order in which the operation is applied does not matter. This is great for our
purposes since we do not (need to) have fine grained control over when which gradient is received.
Note that it is possible for us to perform the reduction stagewise. Furthermore, note that this
operation is independent between blocks i pertaining to different parameters (and gradients).
This allows us to define the following two operations: push, which accumulates gradients, and
pull, which retrieves aggregate gradients. Since we have many different sets of gradients (after
all, we have many layers), we need to index the gradients with a key i. This similarity to (key,value)
stores, such as the one introduced in Dynamo (DeCandia et al., 2007) is not by coincidence. They,

too, satisfy many similar characteristics, in particular when it comes to distributing the parame-
ters across multiple servers.
• push(key, value) sends a particular gradient (the value) from a worker to a common storage.
There the parameter is aggregated, e.g., by summing it up.
• pull(key, value) retrieves an aggregate parameter from common storage, e.g., after com-
bining the gradients from all workers.
By hiding all the complexity about synchronization behind a simple push and pull operation we
can decouple the concerns of the statistical modeler who wants to be able to express optimization
in simple terms and the systems engineer who needs to deal with the complexity inherent in dis-
tributed synchronization. In the next section we will experiment with such a (key,value) store in
practice.
Summary
• Synchronization needs to be highly adaptive to specific network infrastructure and connec-

tivity within a server. This can make a significant difference to the time it takes to synchro-
nize.
• Ring-synchronization can be optimal for P3 and DGX-2 servers. For others possibly not so
much.
• A hierarchical synchronization strategy works well when adding multiple parameter servers
for increased bandwidth.
• Asynchronous communication (while computation is still ongoing) can improve perfor-
mance.
Exercises
1. Can you increase the ring synchronization even further? Hint: you can send messages in
both directions.
2. Fully asynchronous. Some delays permitted?
3. Fault tolerance. How? What if we lose a server? Is this a problem?
4. Checkpointing
5. Tree aggregation. Can you do it faster?
6. Other reductions (commutative semiring).
Discussions178
178

13 | Computer Vision
Many applications in the area of computer vision are closely related to our daily lives, now and in
the future, whether medical diagnostics, driverless vehicles, camera monitoring, or smart filters.
In recent years, deep learning technology has greatly enhanced computer vision systemsʼ perfor-
mance. It can be said that the most advanced computer vision applications are nearly inseparable
from deep learning.
We have introduced deep learning models commonly used in the area of computer vision in the
chapter “Convolutional Neural Networks” and have practiced simple image classification tasks. In
this chapter, we will introduce image augmentation and fine tuning methods and apply them to
image classification. Then, we will explore various methods of object detection. After that, we
will learn how to use fully convolutional networks to perform semantic segmentation on images.
Then, we explain how to use style transfer technology to generate images that look like the cover of
this book. Finally, we will perform practice exercises on two important computer vision datasets
to review the content of this chapter and the previous chapters.
13.1 Image Augmentation
We mentioned that large-scale datasets are prerequisites for the successful application of deep
neural networks in Section 7.1. Image augmentation technology expands the scale of training
datasets by making a series of random changes to the training images to produce similar, but dif-
ferent, training examples. Another way to explain image augmentation is that randomly changing
training examples can reduce a modelʼs dependence on certain properties, thereby improving its
capability for generalization. For example, we can crop the images in different ways, so that the
objects of interest appear in different positions, reducing the modelʼs dependence on the posi-
tion where objects appear. We can also adjust the brightness, color, and other factors to reduce
modelʼs sensitivity to color. It can be said that image augmentation technology contributed greatly
to the success of AlexNet. In this section, we will discuss this technology, which is widely used in
computer vision.
First, import the packages or modules required for the experiment in this section.
%matplotlib inline
from mxnet import autograd, gluon, image, init, np, npx
npx.set_np()
537
13.1.1 Common Image Augmentation Method
In this experiment, we will use an image with a shape of 400 × 500 as an example.
d2l.set_figsize()
img = image.imread('../img/cat1.jpg')
d2l.plt.imshow(img.asnumpy());
Most image augmentation methods have a certain degree of randomness. To make it easier for
us to observe the effect of image augmentation, we next define the auxiliary function apply. This
function runs the image augmentation method aug multiple times on the input image img and
shows all results.
def apply(img, aug, num_rows=2, num_cols=4, scale=1.5):

Y = [aug(img) for _ in range(num_rows * num_cols)]
d2l.show_images(Y, num_rows, num_cols, scale=scale)
Flipping and Cropping
Flipping the image left and right usually does not change the category of the object. This is one
of the earliest and most widely used methods of image augmentation. Next, we use the trans-
forms module to create the RandomFlipLeftRight instance, which introduces a 50% chance that
the image is flipped left and right.
apply(img, gluon.data.vision.transforms.RandomFlipLeftRight())
538 Chapter 13. Computer Vision

Flipping up and down is not as commonly used as flipping left and right. However, at least for this
example image, flipping up and down does not hinder recognition. Next, we create a RandomFlip-
TopBottom instance for a 50% chance of flipping the image up and down.
apply(img, gluon.data.vision.transforms.RandomFlipTopBottom())
In the example image we used, the cat is in the middle of the image, but this may not be the case
for all images. In Section 6.5, we explained that the pooling layer can reduce the sensitivity of the
convolutional layer to the target location. In addition, we can make objects appear at different
positions in the image in different proportions by randomly cropping the image. This can also
reduce the sensitivity of the model to the target position.
In the following code, we randomly crop a region with an area of 10% to 100% of the original
area, and the ratio of width to height of the region is randomly selected from between 0.5 and 2.
Then, the width and height of the region are both scaled to 200 pixels. Unless otherwise stated, the
random number between a and b in this section refers to a continuous value obtained by uniform
sampling in the interval [a, b].
shape_aug = gluon.data.vision.transforms.RandomResizedCrop(
(200, 200), scale=(0.1, 1), ratio=(0.5, 2))
apply(img, shape_aug)
13.1. Image Augmentation 539

Changing the Color
Another augmentation method is changing colors. We can change four aspects of the image color:
brightness, contrast, saturation, and hue. In the example below, we randomly change the bright-
ness of the image to a value between 50% (1 − 0.5) and 150% (1 + 0.5) of the original image.
apply(img, gluon.data.vision.transforms.RandomBrightness(0.5))
Similarly, we can randomly change the hue of the image.
apply(img, gluon.data.vision.transforms.RandomHue(0.5))

We can also create a RandomColorJitter instance and set how to randomly change the brightness,
contrast, saturation, and hue of the image at the same time.
color_aug = gluon.data.vision.transforms.RandomColorJitter(
brightness=0.5, contrast=0.5, saturation=0.5, hue=0.5)
apply(img, color_aug)
Overlying Multiple Image Augmentation Methods
In practice, we will overlay multiple image augmentation methods. We can overlay the different
image augmentation methods defined above and apply them to each image by using a Compose
instance.
augs = gluon.data.vision.transforms.Compose([
gluon.data.vision.transforms.RandomFlipLeftRight(), color_aug, shape_aug])
apply(img, augs)

13.1.2 Using an Image Augmentation Training Model
Next, we will look at how to apply image augmentation in actual training. Here, we use the CIFAR-
10 dataset, instead of the Fashion-MNIST dataset we have been using. This is because the position
and size of the objects in the Fashion-MNIST dataset have been normalized, and the differences in
color and size of the objects in CIFAR-10 dataset are more significant. The first 32 training images
in the CIFAR-10 dataset are shown below.
d2l.show_images(gluon.data.vision.CIFAR10(
train=True)[0:32][0], 4, 8, scale=0.8);
In order to obtain definitive results during prediction, we usually only apply image augmentation
to the training example, and do not use image augmentation with random operations during pre-
diction. Here, we only use the simplest random left-right flipping method. In addition, we use
a ToTensor instance to convert minibatch images into the format required by MXNet, i.e., 32-bit
floating point numbers with the shape of (batch size, number of channels, height, width) and
value range between 0 and 1.

train_augs = gluon.data.vision.transforms.Compose([
gluon.data.vision.transforms.RandomFlipLeftRight(),
gluon.data.vision.transforms.ToTensor()])
test_augs = gluon.data.vision.transforms.Compose([
gluon.data.vision.transforms.ToTensor()])
Next, we define an auxiliary function to make it easier to read the image and apply image augmen-
tation. The transform_first function provided by Gluonʼs dataset applies image augmentation to
the first element of each training example (image and label), i.e., the element at the top of the
image. For detailed descriptions of DataLoader, refer to Section 3.5.
def load_cifar10(is_train, augs, batch_size):

return gluon.data.DataLoader(
gluon.data.vision.CIFAR10(train=is_train).transform_first(augs),
batch_size=batch_size, shuffle=is_train,
num_workers=d2l.get_dataloader_workers())
Using a Multi-GPU Training Model
We train the ResNet-18 model described in Section 7.6 on the CIFAR-10 dataset. We will also apply
the methods described in Section 12.6 and use a multi-GPU training model.
Next, we define the training function to train and evaluate the model using multiple GPUs.
#@save
def train_batch_ch13(net, features, labels, loss, trainer, devices,
split_f=d2l.split_batch):
X_shards, y_shards = split_f(features, labels, devices)
pred_shards = [net(X_shard) for X_shard in X_shards]
ls = [loss(pred_shard, y_shard) for pred_shard, y_shard
in zip(pred_shards, y_shards)]
for l in ls:
l.backward()
# The True flag allows parameters with stale gradients, which is useful
# later (e.g., in fine-tuning BERT)
trainer.step(labels.shape[0], ignore_stale_grad=True)
train_loss_sum = sum([float(l.sum()) for l in ls])
train_acc_sum = sum(d2l.accuracy(pred_shard, y_shard)
for pred_shard, y_shard in zip(pred_shards, y_shards))
return train_loss_sum, train_acc_sum
#@save
def train_ch13(net, train_iter, test_iter, loss, trainer, num_epochs,
devices=d2l.try_all_gpus(), split_f=d2l.split_batch):
num_batches, timer = len(train_iter), d2l.Timer()
animator = d2l.Animator(xlabel='epoch', xlim=[0, num_epochs], ylim=[0, 1],
# Store training_loss, training_accuracy, num_examples, num_features

for i, (features, labels) in enumerate(train_iter):
timer.start()
l, acc = train_batch_ch13(
net, features, labels, loss, trainer, devices, split_f)
metric.add(l, acc, labels.shape[0], labels.size)
timer.stop()
if (i + 1) % (num_batches // 5) == 0:
animator.add(epoch + i / num_batches,
(metric[0] / metric[2], metric[1] / metric[3],
None))
test_acc = d2l.evaluate_accuracy_gpus(net, test_iter, split_f)
print(f'loss {metric[0] / metric[2]:.3f}, train acc '
f'{metric[1] / metric[3]:.3f}, test acc {test_acc:.3f}')
print(f'{metric[2] * num_epochs / timer.sum():.1f} examples/sec on '
f'{str(devices)}')
Now, we can define the train_with_data_aug function to use image augmentation to train the
model. This function obtains all available GPUs and uses Adam as the optimization algorithm
for training. It then applies image augmentation to the training dataset, and finally calls the
train_ch13 function just defined to train and evaluate the model.
batch_size, devices, net = 256, d2l.try_all_gpus(), d2l.resnet18(10)

net.initialize(init=init.Xavier(), ctx=devices)
def train_with_data_aug(train_augs, test_augs, net, lr=0.001):

train_iter = load_cifar10(True, train_augs, batch_size)
test_iter = load_cifar10(False, test_augs, batch_size)
train_ch13(net, train_iter, test_iter, loss, trainer, 10, devices)
Now we train the model using image augmentation of random flipping left and right.
train_with_data_aug(train_augs, test_augs, net)

4811.5 examples/sec on [gpu(0), gpu(1)]

Summary
• Image augmentation generates random images based on existing training data to cope with
overfitting.
• In order to obtain definitive results during prediction, we usually only apply image augmen-
tation to the training example, and do not use image augmentation with random operations
during prediction.
• We can obtain classes related to image augmentation from Gluonʼs transforms module.
Exercises
1. Train the model without using image augmentation: train_with_data_aug(no_aug,

no_aug). Compare training and testing accuracy when using and not using image augmenta-
tion. Can this comparative experiment support the argument that image augmentation can
mitigate overfitting? Why?
2. Add different image augmentation methods in model training based on the CIFAR-10
dataset. Observe the implementation results.
3. With reference to the MXNet documentation, what other image augmentation methods are
provided in Gluonʼs transforms module?
Discussions179
179

13.2 Fine-Tuning
In earlier chapters, we discussed how to train models on the Fashion-MNIST training dataset,
which only has 60,000 images. We also described ImageNet, the most widely used large-scale
image dataset in the academic world, with more than 10 million images and objects of over 1000
categories. However, the size of datasets that we often deal with is usually larger than the first, but
smaller than the second.
Assume we want to identify different kinds of chairs in images and then push the purchase link
to the user. One possible method is to first find a hundred common chairs, take one thousand
different images with different angles for each chair, and then train a classification model on the
collected image dataset. Although this dataset may be larger than Fashion-MNIST, the number of
examples is still less than one tenth of ImageNet. This may result in the overfitting of the com-
plicated model applicable to ImageNet on this dataset. At the same time, because of the limited
amount of data, the accuracy of the final trained model may not meet the practical requirements.
In order to deal with the above problems, an obvious solution is to collect more data. However,
collecting and labeling data can consume a lot of time and money. For example, in order to collect
the ImageNet datasets, researchers have spent millions of dollars of research funding. Although,
recently, data collection costs have dropped significantly, the costs still cannot be ignored.
Another solution is to apply transfer learning to migrate the knowledge learned from the source
dataset to the target dataset. For example, although the images in ImageNet are mostly unrelated
to chairs, models trained on this dataset can extract more general image features that can help
identify edges, textures, shapes, and object composition. These similar features may be equally
effective for recognizing a chair.
In this section, we introduce a common technique in transfer learning: fine tuning. As shown in
Fig. 13.2.1, fine tuning consists of the following four steps:
1. Pre-train a neural network model, i.e., the source model, on a source dataset (e.g., the Ima-
geNet dataset).
2. Create a new neural network model, i.e., the target model. This replicates all model designs
and their parameters on the source model, except the output layer. We assume that these
model parameters contain the knowledge learned from the source dataset and this knowl-
edge will be equally applicable to the target dataset. We also assume that the output layer
of the source model is closely related to the labels of the source dataset and is therefore not
used in the target model.
3. Add an output layer whose output size is the number of target dataset categories to the target
model, and randomly initialize the model parameters of this layer.
4. Train the target model on a target dataset, such as a chair dataset. We will train the output
layer from scratch, while the parameters of all remaining layers are fine-tuned based on the
parameters of the source model.

Fig. 13.2.1: Fine tuning.
13.2.1 Hot Dog Recognition
Next, we will use a specific example for practice: hot dog recognition. We will fine-tune the ResNet
model trained on the ImageNet dataset based on a small dataset. This small dataset contains thou-
sands of images, some of which contain hot dogs. We will use the model obtained by fine tuning
to identify whether an image contains a hot dog.
First, import the packages and modules required for the experiment. Gluonʼs model_zoo package
provides a common pre-trained model. If you want to get more pre-trained models for computer
vision, you can use the GluonCV Toolkit180 .
%matplotlib inline
from mxnet import gluon, init, np, npx
import os
npx.set_np()
Obtaining the Dataset
The hot dog dataset we use was taken from online images and contains 1, 400 positive images con-
taining hot dogs and the same number of negative images containing other foods. 1, 000 images
of various classes are used for training and the rest are used for testing.
We first download the compressed dataset and get two folders hotdog/train and hotdog/test.
Both folders have hotdog and not-hotdog category subfolders, each of which has corresponding
image files.
180
https://gluon-cv.mxnet.io
13.2. Fine-Tuning 547

#@save
d2l.DATA_HUB['hotdog'] = (d2l.DATA_URL+'hotdog.zip',
'fba480ffa8aa7e0febbb511d181409f899b9baa5')
data_dir = d2l.download_extract('hotdog')
Downloading ../data/hotdog.zip from http://d2l-data.s3-accelerate.amazonaws.com/hotdog.zip...
We create two ImageFolderDataset instances to read all the image files in the training dataset and
testing dataset, respectively.
train_imgs = gluon.data.vision.ImageFolderDataset(
os.path.join(data_dir, 'train'))
test_imgs = gluon.data.vision.ImageFolderDataset(
os.path.join(data_dir, 'test'))
The first 8 positive examples and the last 8 negative images are shown below. As you can see, the
images vary in size and aspect ratio.
hotdogs = [train_imgs[i][0] for i in range(8)]

not_hotdogs = [train_imgs[-i - 1][0] for i in range(8)]
d2l.show_images(hotdogs + not_hotdogs, 2, 8, scale=1.4);
During training, we first crop a random area with random size and random aspect ratio from the
image and then scale the area to an input with a height and width of 224 pixels. During testing,
we scale the height and width of images to 256 pixels, and then crop the center area with height
and width of 224 pixels to use as the input. In addition, we normalize the values of the three RGB
(red, green, and blue) color channels. The average of all values of the channel is subtracted from
each value and then the result is divided by the standard deviation of all values of the channel to
produce the output.
# We specify the mean and variance of the three RGB channels to normalize the
# image channel
normalize = gluon.data.vision.transforms.Normalize(
[0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
train_augs = gluon.data.vision.transforms.Compose([
gluon.data.vision.transforms.RandomResizedCrop(224),
gluon.data.vision.transforms.ToTensor(),

normalize])
test_augs = gluon.data.vision.transforms.Compose([
gluon.data.vision.transforms.Resize(256),
gluon.data.vision.transforms.CenterCrop(224),
normalize])
Defining and Initializing the Model
We use ResNet-18, which was pre-trained on the ImageNet dataset, as the source model. Here, we
specify pretrained=True to automatically download and load the pre-trained model parameters.
The first time they are used, the model parameters need to be downloaded from the Internet.
pretrained_net = gluon.model_zoo.vision.resnet18_v2(pretrained=True)
The pre-trained source model instance contains two member variables: features and output.
The former contains all layers of the model, except the output layer, and the latter is the output
layer of the model. The main purpose of this division is to facilitate the fine tuning of the model
parameters of all layers except the output layer. The member variable output of source model is
given below. As a fully connected layer, it transforms ResNetʼs final global average pooling layer
output into 1000 class output on the ImageNet dataset.
pretrained_net.output
Dense(512 -> 1000, linear)
We then build a new neural network to use as the target model. It is defined in the same way as the
pre-trained source model, but the final number of outputs is equal to the number of categories in
the target dataset. In the code below, the model parameters in the member variable features of
the target model instance finetune_net are initialized to model parameters of the corresponding
layer of the source model. Because the model parameters in features are obtained by pre-training
on the ImageNet dataset, it is good enough. Therefore, we generally only need to use small learn-
ing rates to “fine-tune” these parameters. In contrast, model parameters in the member variable
output are randomly initialized and generally require a larger learning rate to learn from scratch.
Assume the learning rate in the Trainer instance is η and use a learning rate of 10η to update the
model parameters in the member variable output.
finetune_net = gluon.model_zoo.vision.resnet18_v2(classes=2)
finetune_net.features = pretrained_net.features
finetune_net.output.initialize(init.Xavier())
# The model parameters in output will be updated using a learning rate ten
# times greater
finetune_net.output.collect_params().setattr('lr_mult', 10)

Fine Tuning the Model
We first define a training function train_fine_tuning that uses fine tuning so it can be called
multiple times.
def train_fine_tuning(net, learning_rate, batch_size=128, num_epochs=5):

train_iter = gluon.data.DataLoader(
train_imgs.transform_first(train_augs), batch_size, shuffle=True)
test_iter = gluon.data.DataLoader(
test_imgs.transform_first(test_augs), batch_size)
net.collect_params().reset_ctx(devices)
net.hybridize()
trainer = gluon.Trainer(net.collect_params(), 'sgd', {
'learning_rate': learning_rate, 'wd': 0.001})
d2l.train_ch13(net, train_iter, test_iter, loss, trainer, num_epochs,
devices)
We set the learning rate in the Trainer instance to a smaller value, such as 0.01, in order to
fine-tune the model parameters obtained in pretraining. Based on the previous settings, we will
train the output layer parameters of the target model from scratch using a learning rate ten times
greater.
train_fine_tuning(finetune_net, 0.01)

For comparison, we define an identical model, but initialize all of its model parameters to random
values. Since the entire model needs to be trained from scratch, we can use a larger learning rate.
scratch_net = gluon.model_zoo.vision.resnet18_v2(classes=2)
scratch_net.initialize(init=init.Xavier())
train_fine_tuning(scratch_net, 0.1)

As you can see, the fine-tuned model tends to achieve higher precision in the same epoch because
the initial values of the parameters are better.
Summary
• Transfer learning migrates the knowledge learned from the source dataset to the target
dataset. Fine tuning is a common technique for transfer learning.
• The target model replicates all model designs and their parameters on the source model,
except the output layer, and fine-tunes these parameters based on the target dataset. In
contrast, the output layer of the target model needs to be trained from scratch.
• Generally, fine tuning parameters use a smaller learning rate, while training the output layer
from scratch can use a larger learning rate.
Exercises
1. Keep increasing the learning rate of finetune_net. How does the precision of the model
change?
2. Further tune the hyperparameters of finetune_net and scratch_net in the comparative ex-
periment. Do they still have different precisions?
3. Set the parameters in finetune_net.features to the parameters of the source model and do
not update them during training. What will happen? You can use the following code.
finetune_net.features.collect_params().setattr('grad_req', 'null')
4. In fact, there is also a “hotdog” class in the ImageNet dataset. Its corresponding weight pa-
rameter at the output layer can be obtained by using the following code. How can we use
this parameter?

weight = pretrained_net.output.weight
hotdog_w = np.split(weight.data(), 1000, axis=0)[713]
hotdog_w.shape
(1, 512)
Discussions181
13.3 Object Detection and Bounding Boxes
In the previous section, we introduced many models for image classification. In image classifica-
tion tasks, we assume that there is only one main target in the image and we only focus on how to
identify the target category. However, in many situations, there are multiple targets in the image
that we are interested in. We not only want to classify them, but also want to obtain their specific
positions in the image. In computer vision, we refer to such tasks as object detection (or object
recognition).
Object detection is widely used in many fields. For example, in self-driving technology, we need
to plan routes by identifying the locations of vehicles, pedestrians, roads, and obstacles in the
captured video image. Robots often perform this type of task to detect targets of interest. Systems
in the security field need to detect abnormal targets, such as intruders or bombs.
In the next few sections, we will introduce multiple deep learning models used for object detec-
tion. Before that, we should discuss the concept of target location. First, import the packages and
modules required for the experiment.
%matplotlib inline
from mxnet import image, npx
npx.set_np()
Next, we will load the sample images that will be used in this section. We can see there is a dog
on the left side of the image and a cat on the right. They are the two main targets in this image.
d2l.set_figsize()
img = image.imread('../img/catdog.jpg').asnumpy()
d2l.plt.imshow(img);
181

13.3.1 Bounding Box
In object detection, we usually use a bounding box to describe the target location. The bounding
box is a rectangular box that can be determined by the x and y axis coordinates in the upper-left
corner and the x and y axis coordinates in the lower-right corner of the rectangle. We will define
the bounding boxes of the dog and the cat in the image based on the coordinate information in the
above image. The origin of the coordinates in the above image is the upper left corner of the image,
and to the right and down are the positive directions of the x axis and the y axis, respectively.
# bbox is the abbreviation for bounding box

dog_bbox, cat_bbox = [60, 45, 378, 516], [400, 112, 655, 493]
We can draw the bounding box in the image to check if it is accurate. Before drawing the box, we
will define a helper function bbox_to_rect. It represents the bounding box in the bounding box
format of matplotlib.
#@save
def bbox_to_rect(bbox, color):
"""Convert bounding box to matplotlib format."""
# Convert the bounding box (top-left x, top-left y, bottom-right x,
# bottom-right y) format to matplotlib format: ((upper-left x,
# upper-left y), width, height)
return d2l.plt.Rectangle(
xy=(bbox[0], bbox[1]), width=bbox[2]-bbox[0], height=bbox[3]-bbox[1],
fill=False, edgecolor=color, linewidth=2)
After loading the bounding box on the image, we can see that the main outline of the target is
basically inside the box.
fig = d2l.plt.imshow(img)
fig.axes.add_patch(bbox_to_rect(dog_bbox, 'blue'))
fig.axes.add_patch(bbox_to_rect(cat_bbox, 'red'));
13.3. Object Detection and Bounding Boxes 553

Summary
• In object detection, we not only need to identify all the objects of interest in the image, but
also their positions. The positions are generally represented by a rectangular bounding box.
Exercises
1. Find some images and try to label a bounding box that contains the target. Compare the
difference between the time it takes to label the bounding box and label the category.
Discussions182
13.4 Anchor Boxes
Object detection algorithms usually sample a large number of regions in the input image, deter-
mine whether these regions contain objects of interest, and adjust the edges of the regions so as
to predict the ground-truth bounding box of the target more accurately. Different models may
use different region sampling methods. Here, we introduce one such method: it generates mul-
tiple bounding boxes with different sizes and aspect ratios while centering on each pixel. These
bounding boxes are called anchor boxes. We will practice object detection based on anchor boxes
in the following sections.
First, import the packages or modules required for this section. Here, we have modified the print-
ing accuracy of NumPy. Because printing tensors actually calls the print function of NumPy, the
floating-point numbers in tensors printed in this section are more concise.
%matplotlib inline
from mxnet import gluon, image, np, npx
np.set_printoptions(2)
npx.set_np()
182

13.4.1 Generating Multiple Anchor Boxes
Assume that the input image has a height of h and width of w. We generate anchor boxes with
different shapes centered on each pixel of the image. Assume the size is s ∈ (0, 1], the aspect ratio
√ √
is r > 0, and the width and height of the anchor box are ws r and hs/ r, respectively. When the
center position is given, an anchor box with known width and height is determined.
Below we set a set of sizes s1 , . . . , sn and a set of aspect ratios r1 , . . . , rm . If we use a combination
of all sizes and aspect ratios with each pixel as the center, the input image will have a total of
whnm anchor boxes. Although these anchor boxes may cover all ground-truth bounding boxes,
the computational complexity is often excessive. Therefore, we are usually only interested in a
combination containing s1 or r1 sizes and aspect ratios, that is:
(s1 , r1 ), (s1 , r2 ), . . . , (s1 , rm ), (s2 , r1 ), (s3 , r1 ), . . . , (sn , r1 ). (13.4.1)
That is, the number of anchor boxes centered on the same pixel is n + m − 1. For the entire input
image, we will generate a total of wh(n + m − 1) anchor boxes.
The above method of generating anchor boxes has been implemented in the multibox_prior func-
tion. We specify the input, a set of sizes, and a set of aspect ratios, and this function will return
all the anchor boxes entered.
img = image.imread('../img/catdog.jpg').asnumpy()
h, w = img.shape[0:2]
print(h, w)
X = np.random.uniform(size=(1, 3, h, w)) # Construct input data
Y = npx.multibox_prior(X, sizes=[0.75, 0.5, 0.25], ratios=[1, 2, 0.5])
Y.shape
561 728
(1, 2042040, 4)
We can see that the shape of the returned anchor box variable y is (batch size, number of an-
chor boxes, 4). After changing the shape of the anchor box variable y to (image height, image
width, number of anchor boxes centered on the same pixel, 4), we can obtain all the anchor boxes
centered on a specified pixel position. In the following example, we access the first anchor box
centered on (250, 250). It has four elements: the x, y axis coordinates in the upper-left corner and
the x, y axis coordinates in the lower-right corner of the anchor box. The coordinate values of the
x and y axis are divided by the width and height of the image, respectively, so the value range is
between 0 and 1.
boxes = Y.reshape(h, w, 5, 4)
boxes[250, 250, 0, :]
array([0.06, 0.07, 0.63, 0.82])
In order to describe all anchor boxes centered on one pixel in the image, we first define the
show_bboxes function to draw multiple bounding boxes on the image.
13.4. Anchor Boxes 555

#@save
def show_bboxes(axes, bboxes, labels=None, colors=None):
"""Show bounding boxes."""
def _make_list(obj, default_values=None):
if obj is None:
obj = default_values
elif not isinstance(obj, (list, tuple)):
obj = [obj]
return obj
labels = _make_list(labels)
colors = _make_list(colors, ['b', 'g', 'r', 'm', 'c'])
for i, bbox in enumerate(bboxes):
color = colors[i % len(colors)]
rect = d2l.bbox_to_rect(bbox.asnumpy(), color)
axes.add_patch(rect)
if labels and len(labels) > i:
text_color = 'k' if color == 'w' else 'w'
axes.text(rect.xy[0], rect.xy[1], labels[i],
va='center', ha='center', fontsize=9, color=text_color,
bbox=dict(facecolor=color, lw=0))
As we just saw, the coordinate values of the x and y axis in the variable boxes have been divided
by the width and height of the image, respectively. When drawing images, we need to restore
the original coordinate values of the anchor boxes and therefore define the variable bbox_scale.
Now, we can draw all the anchor boxes centered on (250, 250) in the image. As you can see, the
blue anchor box with a size of 0.75 and an aspect ratio of 1 covers the dog in the image well.
d2l.set_figsize()
bbox_scale = np.array((w, h, w, h))
show_bboxes(fig.axes, boxes[250, 250, :, :] * bbox_scale,
['s=0.75, r=1', 's=0.5, r=1', 's=0.25, r=1', 's=0.75, r=2',
's=0.75, r=0.5'])

13.4.2 Intersection over Union
We just mentioned that the anchor box covers the dog in the image well. If the ground-truth
bounding box of the target is known, how can “well” here be quantified? An intuitive method
is to measure the similarity between anchor boxes and the ground-truth bounding box. We know
that the Jaccard index can measure the similarity between two sets. Given sets A and B, their
Jaccard index is the size of their intersection divided by the size of their union:
|A ∩ B|
J(A, B) = . (13.4.2)
|A ∪ B|
In fact, we can consider the pixel area of a bounding box as a collection of pixels. In this way,
we can measure the similarity of the two bounding boxes by the Jaccard index of their pixel sets.
When we measure the similarity of two bounding boxes, we usually refer the Jaccard index as
intersection over union (IoU), which is the ratio of the intersecting area to the union area of the
two bounding boxes, as shown in Fig. 13.4.1. The value range of IoU is between 0 and 1: 0 means
that there are no overlapping pixels between the two bounding boxes, while 1 indicates that the
two bounding boxes are equal.
Fig. 13.4.1: IoU is the ratio of the intersecting area to the union area of two bounding boxes.
For the remainder of this section, we will use IoU to measure the similarity between anchor boxes
and ground-truth bounding boxes, and between different anchor boxes.
13.4.3 Labeling Training Set Anchor Boxes
In the training set, we consider each anchor box as a training example. In order to train the object
detection model, we need to mark two types of labels for each anchor box: first, the category of the
target contained in the anchor box (category) and, second, the offset of the ground-truth bounding
box relative to the anchor box (offset). In object detection, we first generate multiple anchor boxes,
predict the categories and offsets for each anchor box, adjust the anchor box position according
to the predicted offset to obtain the bounding boxes to be used for prediction, and finally filter out
the prediction bounding boxes that need to be output.
We know that, in the object detection training set, each image is labelled with the location of the
ground-truth bounding box and the category of the target contained. After the anchor boxes are
generated, we primarily label anchor boxes based on the location and category information of
the ground-truth bounding boxes similar to the anchor boxes. So how do we assign ground-truth
bounding boxes to anchor boxes similar to them?
Assume that the anchor boxes in the image are A1 , A2 , . . . , Ana and the ground-truth bounding
boxes are B1 , B2 , . . . , Bnb and na ≥ nb . Define matrix X ∈ Rna ×nb , where element xij in the ith row

and j th column is the IoU of the anchor box Ai to the ground-truth bounding box Bj . First, we find
the largest element in the matrix X and record the row index and column index of the element
as i1 , j1 . We assign the ground-truth bounding box Bj1 to the anchor box Ai1 . Obviously, anchor
box Ai1 and ground-truth bounding box Bj1 have the highest similarity among all the “anchor
box–ground-truth bounding box” pairings. Next, discard all elements in the i1 th row and the j1 th
column in the matrix X. Find the largest remaining element in the matrix X and record the row
index and column index of the element as i2 , j2 . We assign ground-truth bounding box Bj2 to
anchor box Ai2 and then discard all elements in the i2 th row and the j2 th column in the matrix X.
At this point, elements in two rows and two columns in the matrix X have been discarded.
We proceed until all elements in the nb column in the matrix X are discarded. At this time, we have
assigned a ground-truth bounding box to each of the nb anchor boxes. Next, we only traverse the
remaining na − nb anchor boxes. Given anchor box Ai , find the bounding box Bj with the largest
IoU with Ai according to the ith row of the matrix X, and only assign ground-truth bounding box
Bj to anchor box Ai when the IoU is greater than the predetermined threshold.
As shown in Fig. 13.4.2 (left), assuming that the maximum value in the matrix X is x23 , we will
assign ground-truth bounding box B3 to anchor box A2 . Then, we discard all the elements in row
2 and column 3 of the matrix, find the largest element x71 of the remaining shaded area, and assign
ground-truth bounding box B1 to anchor box A7 . Then, as shown in :numref:fig_anchor_label
(middle), discard all the elements in row 7 and column 1 of the matrix, find the largest element x54
of the remaining shaded area, and assign ground-truth bounding box B4 to anchor box A5 . Finally,
as shown in :numref:fig_anchor_label (right), discard all the elements in row 5 and column 4 of
the matrix, find the largest element x92 of the remaining shaded area, and assign ground-truth
bounding box B2 to anchor box A9 . After that, we only need to traverse the remaining anchor
boxes of A1 , A3 , A4 , A6 , A8 and determine whether to assign ground-truth bounding boxes to the
remaining anchor boxes according to the threshold.
Fig. 13.4.2: Assign ground-truth bounding boxes to anchor boxes.
Now we can label the categories and offsets of the anchor boxes. If an anchor box A is assigned
ground-truth bounding box B, the category of the anchor box A is set to the category of B. And
the offset of the anchor box A is set according to the relative position of the central coordinates of

B and A and the relative sizes of the two boxes. Because the positions and sizes of various boxes
in the dataset may vary, these relative positions and relative sizes usually require some special
transformations to make the offset distribution more uniform and easier to fit. Assume the center
coordinates of anchor box A and its assigned ground-truth bounding box B are (xa , ya ), (xb , yb ),
the widths of A and B are wa , wb , and their heights are ha , hb , respectively. In this case, a common
technique is to label the offset of A as
( x −x yb −ya
)
wa − µx ha − µy log wa − µw log ha − µh
b a wb hb
, , , , (13.4.3)
σx σy σw σh
The default values of the constant are µx = µy = µw = µh = 0, σx = σy = 0.1, andσw = σh = 0.2.

If an anchor box is not assigned a ground-truth bounding box, we only need to set the category of
the anchor box to background. Anchor boxes whose category is background are often referred to
as negative anchor boxes, and the rest are referred to as positive anchor boxes.
Below we demonstrate a detailed example. We define ground-truth bounding boxes for the cat and
dog in the read image, where the first element is category (0 for dog, 1 for cat) and the remaining
four elements are the x, y axis coordinates at top-left corner and x, y axis coordinates at lower-
right corner (the value range is between 0 and 1). Here, we construct five anchor boxes to be
labeled by the coordinates of the upper-left corner and the lower-right corner, which are recorded
as A0 , . . . , A4 , respectively (the index in the program starts from 0). First, draw the positions of
these anchor boxes and the ground-truth bounding boxes in the image.
ground_truth = np.array([[0, 0.1, 0.08, 0.52, 0.92],

[1, 0.55, 0.2, 0.9, 0.88]])
anchors = np.array([[0, 0.1, 0.2, 0.3], [0.15, 0.2, 0.4, 0.4],
[0.63, 0.05, 0.88, 0.98], [0.66, 0.45, 0.8, 0.8],
[0.57, 0.3, 0.92, 0.9]])
show_bboxes(fig.axes, ground_truth[:, 1:] * bbox_scale, ['dog', 'cat'], 'k')
show_bboxes(fig.axes, anchors * bbox_scale, ['0', '1', '2', '3', '4']);
We can label categories and offsets for anchor boxes by using the multibox_target function. This
function sets the background category to 0 and increments the integer index of the target category
from zero by 1 (1 for dog and 2 for cat). We add example dimensions to the anchor boxes and
ground-truth bounding boxes and construct random predicted results with a shape of (batch size,
number of categories including background, number of anchor boxes) by using the expand_dims
function.

labels = npx.multibox_target(np.expand_dims(anchors, axis=0),
np.expand_dims(ground_truth, axis=0),
np.zeros((1, 3, 5)))
There are three items in the returned result, all of which are in the tensor format. The third item
is represented by the category labeled for the anchor box.
labels[2]
array([[0., 1., 2., 0., 2.]])
We analyze these labelled categories based on positions of anchor boxes and ground-truth bound-
ing boxes in the image. First, in all “anchor box–ground-truth bounding box” pairs, the IoU of
anchor box A4 to the ground-truth bounding box of the cat is the largest, so the category of an-
chor box A4 is labeled as cat. Without considering anchor box A4 or the ground-truth bounding
box of the cat, in the remaining “anchor box–ground-truth bounding box” pairs, the pair with the
largest IoU is anchor box A1 and the ground-truth bounding box of the dog, so the category of
anchor box A1 is labeled as dog. Next, traverse the remaining three unlabeled anchor boxes. The
category of the ground-truth bounding box with the largest IoU with anchor box A0 is dog, but
the IoU is smaller than the threshold (the default is 0.5), so the category is labeled as background;
the category of the ground-truth bounding box with the largest IoU with anchor box A2 is cat and
the IoU is greater than the threshold, so the category is labeled as cat; the category of the ground-
truth bounding box with the largest IoU with anchor box A3 is cat, but the IoU is smaller than the
threshold, so the category is labeled as background.
The second item of the return value is a mask variable, with the shape of (batch size, four times
the number of anchor boxes). The elements in the mask variable correspond one-to-one with the
four offset values of each anchor box. Because we do not care about background detection, offsets
of the negative class should not affect the target function. By multiplying by element, the 0 in the
mask variable can filter out negative class offsets before calculating target function.
labels[1]
array([[0., 0., 0., 0., 1., 1., 1., 1., 1., 1., 1., 1., 0., 0., 0., 0.,
1., 1., 1., 1.]])
The first item returned is the four offset values labeled for each anchor box, with the offsets of
negative class anchor boxes labeled as 0.
labels[0]
array([[ 0.00e+00, 0.00e+00, 0.00e+00, 0.00e+00, 1.40e+00, 1.00e+01,

2.59e+00, 7.18e+00, -1.20e+00, 2.69e-01, 1.68e+00, -1.57e+00,
0.00e+00, 0.00e+00, 0.00e+00, 0.00e+00, -5.71e-01, -1.00e+00,
-8.94e-07, 6.26e-01]])

13.4.4 Bounding Boxes for Prediction
During model prediction phase, we first generate multiple anchor boxes for the image and then
predict categories and offsets for these anchor boxes one by one. Then, we obtain prediction
bounding boxes based on anchor boxes and their predicted offsets. When there are many anchor
boxes, many similar prediction bounding boxes may be output for the same target. To simplify
the results, we can remove similar prediction bounding boxes. A commonly used method is called
non-maximum suppression (NMS).
Let us take a look at how NMS works. For a prediction bounding box B, the model calculates
the predicted probability for each category. Assume the largest predicted probability is p, the
category corresponding to this probability is the predicted category of B. We also refer to p as
the confidence level of prediction bounding box B. On the same image, we sort the prediction
bounding boxes with predicted categories other than background by confidence level from high
to low, and obtain the list L. Select the prediction bounding box B1 with highest confidence level
from L as a baseline and remove all non-benchmark prediction bounding boxes with an IoU with
B1 greater than a certain threshold from L. The threshold here is a preset hyperparameter. At this
point, L retains the prediction bounding box with the highest confidence level and removes other
prediction bounding boxes similar to it. Next, select the prediction bounding box B2 with the
second highest confidence level from L as a baseline, and remove all non-benchmark prediction
bounding boxes with an IoU with B2 greater than a certain threshold from L. Repeat this process
until all prediction bounding boxes in L have been used as a baseline. At this time, the IoU of any
pair of prediction bounding boxes in L is less than the threshold. Finally, output all prediction
bounding boxes in the list L.
Next, we will look at a detailed example. First, construct four anchor boxes. For the sake of sim-
plicity, we assume that predicted offsets are all 0. This means that the prediction bounding boxes
are anchor boxes. Finally, we construct a predicted probability for each category.
anchors = np.array([[0.1, 0.08, 0.52, 0.92], [0.08, 0.2, 0.56, 0.95],

[0.15, 0.3, 0.62, 0.91], [0.55, 0.2, 0.9, 0.88]])
offset_preds = np.array([0] * anchors.size)
cls_probs = np.array([[0] * 4, # Predicted probability for background
[0.9, 0.8, 0.7, 0.1], # Predicted probability for dog
[0.1, 0.2, 0.3, 0.9]]) # Predicted probability for cat
Print prediction bounding boxes and their confidence levels on the image.
show_bboxes(fig.axes, anchors * bbox_scale,
['dog=0.9', 'dog=0.8', 'dog=0.7', 'cat=0.9'])

We use the multibox_detection function to perform NMS and set the threshold to 0.5. This adds an
example dimension to the tensor input. We can see that the shape of the returned result is (batch
size, number of anchor boxes, 6). The 6 elements of each row represent the output information
for the same prediction bounding box. The first element is the predicted category index, which
starts from 0 (0 is dog, 1 is cat). The value -1 indicates background or removal in NMS. The second
element is the confidence level of prediction bounding box. The remaining four elements are the
x, y axis coordinates of the upper-left corner and the x, y axis coordinates of the lower-right corner
of the prediction bounding box (the value range is between 0 and 1).
output = npx.multibox_detection(
np.expand_dims(cls_probs, axis=0),
np.expand_dims(offset_preds, axis=0),
np.expand_dims(anchors, axis=0),
nms_threshold=0.5)
output
array([[[ 0. , 0.9 , 0.1 , 0.08, 0.52, 0.92],

[ 1. , 0.9 , 0.55, 0.2 , 0.9 , 0.88],
[-1. , 0.8 , 0.08, 0.2 , 0.56, 0.95],
[-1. , 0.7 , 0.15, 0.3 , 0.62, 0.91]]])
We remove the prediction bounding boxes of category -1 and visualize the results retained by NMS.
for i in output[0].asnumpy():
if i[0] == -1:
continue
label = ('dog=', 'cat=')[int(i[0])] + str(i[1])
show_bboxes(fig.axes, [np.array(i[2:]) * bbox_scale], label)

In practice, we can remove prediction bounding boxes with lower confidence levels before per-
forming NMS, thereby reducing the amount of computation for NMS. We can also filter the output
of NMS, for example, by only retaining results with higher confidence levels as the final output.
Summary
• We generate multiple anchor boxes with different sizes and aspect ratios, centered on each
pixel.
• IoU, also called Jaccard index, measures the similarity of two bounding boxes. It is the ratio
of the intersecting area to the union area of two bounding boxes.
• In the training set, we mark two types of labels for each anchor box: one is the category of the
target contained in the anchor box and the other is the offset of the ground-truth bounding
box relative to the anchor box.
• When predicting, we can use non-maximum suppression (NMS) to remove similar predic-
tion bounding boxes, thereby simplifying the results.
Exercises
1. Change the sizes and ratios values in the multibox_prior function and observe the changes
to the generated anchor boxes.
2. Construct two bounding boxes with an IoU of 0.5, and observe their coincidence.
3. Verify the output of offset labels[0] by marking the anchor box offsets as defined in this
section (the constant is the default value).
4. Modify the variable anchors in the “Labeling Training Set Anchor Boxes” and “Output Bound-
ing Boxes for Prediction” sections. How do the results change?
Discussions183
183

13.5 Multiscale Object Detection
In Section 13.4, we generated multiple anchor boxes centered on each pixel of the input image.
These anchor boxes are used to sample different regions of the input image. However, if anchor
boxes are generated centered on each pixel of the image, soon there will be too many anchor boxes
for us to compute. For example, we assume that the input image has a height and a width of 561
and 728 pixels respectively. If five different shapes of anchor boxes are generated centered on
each pixel, over two million anchor boxes (561 × 728 × 5) need to be predicted and labeled on the
image.
It is not difficult to reduce the number of anchor boxes. An easy way is to apply uniform sampling
on a small portion of pixels from the input image and generate anchor boxes centered on the sam-
pled pixels. In addition, we can generate anchor boxes of varied numbers and sizes on multiple
scales. Notice that smaller objects are more likely to be positioned on the image than larger ones.
Here, we will use a simple example: Objects with shapes of 1 × 1, 1 × 2, and 2 × 2 may have 4, 2,
and 1 possible position(s) on an image with the shape 2×2. Therefore, when using smaller anchor
boxes to detect smaller objects, we can sample more regions; when using larger anchor boxes to
detect larger objects, we can sample fewer regions.
To demonstrate how to generate anchor boxes on multiple scales, let us read an image first. It has
a height and width of 561 × 728 pixels.
%matplotlib inline
from mxnet import image, np, npx
npx.set_np()
img = image.imread('../img/catdog.jpg')
h, w
(561, 728)
In Section 6.2, the 2D array output of the convolutional neural network (CNN) is called a feature
map. We can determine the midpoints of anchor boxes uniformly sampled on any image by defin-
ing the shape of the feature map.
The function display_anchors is defined below. We are going to generate anchor boxes anchors
centered on each unit (pixel) on the feature map fmap. Since the coordinates of axes x and y in
anchor boxes anchors have been divided by the width and height of the feature map fmap, values
between 0 and 1 can be used to represent relative positions of anchor boxes in the feature map.
Since the midpoints of anchor boxes anchors overlap with all the units on feature map fmap, the
relative spatial positions of the midpoints of the anchors on any image must have a uniform dis-
tribution. Specifically, when the width and height of the feature map are set to fmap_w and fmap_h
respectively, the function will conduct uniform sampling for fmap_h rows and fmap_w columns of
pixels and use them as midpoints to generate anchor boxes with size s (we assume that the length
of list s is 1) and different aspect ratios (ratios).
def display_anchors(fmap_w, fmap_h, s):

d2l.set_figsize()
# The values from the first two dimensions will not affect the output

fmap = np.zeros((1, 10, fmap_w, fmap_h))
anchors = npx.multibox_prior(fmap, sizes=s, ratios=[1, 2, 0.5])
bbox_scale = np.array((w, h, w, h))
d2l.show_bboxes(d2l.plt.imshow(img.asnumpy()).axes,
anchors[0] * bbox_scale)
We will first focus on the detection of small objects. In order to make it easier to distinguish upon
display, the anchor boxes with different midpoints here do not overlap. We assume that the size
of the anchor boxes is 0.15 and the height and width of the feature map are 4. We can see that the
midpoints of anchor boxes from the 4 rows and 4 columns on the image are uniformly distributed.
display_anchors(fmap_w=4, fmap_h=4, s=[0.15])
We are going to reduce the height and width of the feature map by half and use a larger anchor
box to detect larger objects. When the size is set to 0.4, overlaps will occur between regions of
some anchor boxes.
Finally, we are going to reduce the height and width of the feature map by half and increase the
anchor box size to 0.8. Now the midpoint of the anchor box is the center of the image.
13.5. Multiscale Object Detection 565

Since we have generated anchor boxes of different sizes on multiple scales, we will use them to
detect objects of various sizes at different scales. Now we are going to introduce a method based
on convolutional neural networks (CNNs).
At a certain scale, suppose we generate h × w sets of anchor boxes with different midpoints based
on ci feature maps with the shape h × w and the number of anchor boxes in each set is a. For
example, for the first scale of the experiment, we generate 16 sets of anchor boxes with different
midpoints based on 10 (number of channels) feature maps with a shape of 4 × 4, and each set
contains 3 anchor boxes. Next, each anchor box is labeled with a category and offset based on
the classification and position of the ground-truth bounding box. At the current scale, the object
detection model needs to predict the category and offset of h×w sets of anchor boxes with different
midpoints based on the input image.
We assume that the ci feature maps are the intermediate output of the CNN based on the input
image. Since each feature map has h × w different spatial positions, the same position will have ci
units. According to the definition of receptive field in the Section 6.2, the ci units of the feature map
at the same spatial position have the same receptive field on the input image. Thus, they represent
the information of the input image in this same receptive field. Therefore, we can transform the ci
units of the feature map at the same spatial position into the categories and offsets of the a anchor
boxes generated using that position as a midpoint. It is not hard to see that, in essence, we use the
information of the input image in a certain receptive field to predict the category and offset of the
anchor boxes close to the field on the input image.
When the feature maps of different layers have receptive fields of different sizes on the input im-
age, they are used to detect objects of different sizes. For example, we can design a network to
have a wider receptive field for each unit in the feature map that is closer to the output layer, to
detect objects with larger sizes in the input image.
We will implement a multiscale object detection model in the following section.

Summary
• We can generate anchor boxes with different numbers and sizes on multiple scales to detect
objects of different sizes on multiple scales.
• The shape of the feature map can be used to determine the midpoint of the anchor boxes
that uniformly sample any image.
• We use the information for the input image from a certain receptive field to predict the cat-
egory and offset of the anchor boxes close to that field on the image.
Exercises
1. Given an input image, assume 1 × ci × h × w to be the shape of the feature map while ci , h, w
are the number, height, and width of the feature map. What methods can you think of to
convert this variable into the anchor boxʼs category and offset? What is the shape of the
output?
Discussions184
13.6 The Object Detection Dataset
There are no small datasets, like MNIST or Fashion-MNIST, in the object detection field. In order
to quickly test models, we are going to assemble a small dataset. First, we generate 1000 banana
images of different angles and sizes using free bananas from our office. Then, we collect a series
of background images and place a banana image at a random position on each image. We use the
im2rec tool185 provided by MXNet to convert the images to binary RecordIO format[1]. This format
can reduce the storage overhead of the dataset on the disk and improve the reading efficiency. If
you want to learn more about how to read images, refer to the documentation for the GluonCV
Toolkit186 .
13.6.1 Downloading the Dataset
The banana detection dataset in RecordIO format can be downloaded directly from the Internet.
%matplotlib inline
import os
npx.set_np()
#@save
d2l.DATA_HUB['bananas'] = (d2l.DATA_URL + 'bananas.zip',
'aadfd1c4c5d7178616799dd1801c9a234ccdaf19')
184
185
https://github.com/apache/incubator-mxnet/blob/master/tools/im2rec.py
186
https://gluon-cv.mxnet.io/
13.6. The Object Detection Dataset 567

We are going to read the object detection dataset by creating the instance ImageDetIter. The “Det”
in the name refers to Detection. We will read the training dataset in random order. Since the for-
mat of the dataset is RecordIO, we need the image index file 'train.idx' to read random mini-
batches. In addition, for each image of the training set, we will use random cropping and require
the cropped image to cover at least 95% of each object. Since the cropping is random, this re-
quirement is not always satisfied. We preset the maximum number of random cropping attempts
to 200. If none of them meets the requirement, the image will not be cropped. To ensure the cer-
tainty of the output, we will not randomly crop the images in the test dataset. We also do not need
to read the test dataset in random order.
#@save
def load_data_bananas(batch_size, edge_size=256):
"""Load the bananas dataset."""
data_dir = d2l.download_extract('bananas')
train_iter = image.ImageDetIter(
path_imgrec=os.path.join(data_dir, 'train.rec'),
path_imgidx=os.path.join(data_dir, 'train.idx'),
batch_size=batch_size,
data_shape=(3, edge_size, edge_size), # The shape of the output image
shuffle=True, # Read the dataset in random order
rand_crop=1, # The probability of random cropping is 1
min_object_covered=0.95, max_attempts=200)
val_iter = image.ImageDetIter(
path_imgrec=os.path.join(data_dir, 'val.rec'), batch_size=batch_size,
data_shape=(3, edge_size, edge_size), shuffle=False)
return train_iter, val_iter
Below, we read a minibatch and print the shape of the image and label. The shape of the image
is the same as in the previous experiment (batch size, number of channels, height, width). The
shape of the label is (batch size, m, 5), where m is equal to the maximum number of bounding
boxes contained in a single image in the dataset. Although computation for the minibatch is very
efficient, it requires each image to contain the same number of bounding boxes so that they can be
placed in the same batch. Since each image may have a different number of bounding boxes, we
can add illegal bounding boxes to images that have less than m bounding boxes until each image
contains m bounding boxes. Thus, we can read a minibatch of images each time. The label of each
bounding box in the image is represented by an array of length 5. The first element in the array is
the category of the object contained in the bounding box. When the value is -1, the bounding box
is an illegal bounding box for filling purpose. The remaining four elements of the array represent
the x, y axis coordinates of the upper-left corner of the bounding box and the x, y axis coordinates
of the lower-right corner of the bounding box (the value range is between 0 and 1). The banana
dataset here has only one bounding box per image, so m = 1.
batch_size, edge_size = 32, 256

train_iter, _ = load_data_bananas(batch_size, edge_size)
batch = train_iter.next()
batch.data[0].shape, batch.label[0].shape
Downloading ../data/bananas.zip from http://d2l-data.s3-accelerate.amazonaws.com/bananas.zip.

,→..

((32, 3, 256, 256), (32, 1, 5))
13.6.3 Demonstration
We have ten images with bounding boxes on them. We can see that the angle, size, and position of
banana are different in each image. Of course, this is a simple artificial dataset. In actual practice,
the data are usually much more complicated.
imgs = (batch.data[0][0:10].transpose(0, 2, 3, 1)) / 255

axes = d2l.show_images(imgs, 2, 5, scale=2)
for ax, label in zip(axes, batch.label[0][0:10]):
d2l.show_bboxes(ax, [label[0][1:5] * edge_size], colors=['w'])
Summary
• The banana detection dataset we synthesized can be used to test object detection models.
• The data reading for object detection is similar to that for image classification. However,
after we introduce bounding boxes, the label shape and image augmentation (e.g., random
cropping) are changed.
Exercises
1. Referring to the MXNet documentation, what are the parameters for the constructors of the
image.ImageDetIter and image.CreateDetAugmenter classes? What is their significance?
Discussions187
187
13.6. The Object Detection Dataset 569

13.7 Single Shot Multibox Detection (SSD)
In the previous few sections, we have introduced bounding boxes, anchor boxes, multiscale ob-
ject detection, and datasets. Now, we will use this background knowledge to construct an object
detection model: single shot multibox detection (SSD) (Liu et al., 2016). This quick and easy model
is already widely used. Some of the design concepts and implementation details of this model are
also applicable to other object detection models.
13.7.1 Model
Fig. 13.7.1 shows the design of an SSD model. The modelʼs main components are a base network
block and several multiscale feature blocks connected in a series. Here, the base network block is
used to extract features of original images, and it generally takes the form of a deep convolutional
neural network. The paper on SSDs chooses to place a truncated VGG before the classification
layer (Liu et al., 2016), but this is now commonly replaced by ResNet. We can design the base
network so that it outputs larger heights and widths. In this way, more anchor boxes are gen-
erated based on this feature map, allowing us to detect smaller objects. Next, each multiscale
feature block reduces the height and width of the feature map provided by the previous layer (for
example, it may reduce the sizes by half). The blocks then use each element in the feature map to
expand the receptive field on the input image. In this way, the closer a multiscale feature block
is to the top of Fig. 13.7.1 the smaller its output feature map, and the fewer the anchor boxes that
are generated based on the feature map. In addition, the closer a feature block is to the top, the
larger the receptive field of each element in the feature map and the better suited it is to detect
larger objects. As the SSD generates different numbers of anchor boxes of different sizes based
on the base network block and each multiscale feature block and then predicts the categories and
offsets (i.e., predicted bounding boxes) of the anchor boxes in order to detect objects of different
sizes, SSD is a multiscale object detection model.
Fig. 13.7.1: The SSD is composed of a base network block and several multiscale feature blocks
connected in a series.
Next, we will describe the implementation of the modules in Fig. 13.7.1. First, we need to discuss

the implementation of category prediction and bounding box prediction.
Category Prediction Layer
Set the number of object categories to q. In this case, the number of anchor box categories is
q + 1, with 0 indicating an anchor box that only contains background. For a certain scale, set
the height and width of the feature map to h and w, respectively. If we use each element as the
center to generate a anchor boxes, we need to classify a total of hwa anchor boxes. If we use a
fully connected layer (FCN) for the output, this will likely result in an excessive number of model
parameters. Recall how we used convolutional layer channels to output category predictions in
Section 7.3. SSD uses the same method to reduce the model complexity.
Specifically, the category prediction layer uses a convolutional layer that maintains the input
height and width. Thus, the output and input have a one-to-one correspondence to the spatial
coordinates along the width and height of the feature map. Assuming that the output and input
have the same spatial coordinates (x, y), the channel for the coordinates (x, y) on the output fea-
ture map contains the category predictions for all anchor boxes generated using the input feature
map coordinates (x, y) as the center. Therefore, there are a(q + 1) output channels, with the out-
put channels indexed as i(q + 1) + j (0 ≤ j ≤ q) representing the predictions of the category index
j for the anchor box index i.
Now, we will define a category prediction layer of this type. After we specify the parameters a and
q, it uses a 3 × 3 convolutional layer with a padding of 1. The heights and widths of the input and
output of this convolutional layer remain unchanged.
%matplotlib inline
npx.set_np()
def cls_predictor(num_anchors, num_classes):

return nn.Conv2D(num_anchors * (num_classes + 1), kernel_size=3,
padding=1)
Bounding Box Prediction Layer
The design of the bounding box prediction layer is similar to that of the category prediction layer.
The only difference is that, here, we need to predict 4 offsets for each anchor box, rather than q +1
categories.
def bbox_predictor(num_anchors):
return nn.Conv2D(num_anchors * 4, kernel_size=3, padding=1)
13.7. Single Shot Multibox Detection (SSD) 571

Concatenating Predictions for Multiple Scales
As we mentioned, SSD uses feature maps based on multiple scales to generate anchor boxes and
predict their categories and offsets. Because the shapes and number of anchor boxes centered on
the same element differ for the feature maps of different scales, the prediction outputs at different
scales may have different shapes.
In the following example, we use the same batch of data to construct feature maps of two different
scales, Y1 and Y2. Here, Y2 has half the height and half the width of Y1. Using category prediction
as an example, we assume that each element in the Y1 and Y2 feature maps generates five (Y1) or
three (Y2) anchor boxes. When there are 10 object categories, the number of category prediction
output channels is either 5 × (10 + 1) = 55 or 3 × (10 + 1) = 33. The format of the prediction
output is (batch size, number of channels, height, width). As you can see, except for the batch
size, the sizes of the other dimensions are different. Therefore, we must transform them into a
consistent format and concatenate the predictions of the multiple scales to facilitate subsequent
computation.
def forward(x, block):

block.initialize()
return block(x)
Y1 = forward(np.zeros((2, 8, 20, 20)), cls_predictor(5, 10))

Y2 = forward(np.zeros((2, 16, 10, 10)), cls_predictor(3, 10))
(Y1.shape, Y2.shape)
((2, 55, 20, 20), (2, 33, 10, 10))
The channel dimension contains the predictions for all anchor boxes with the same center. We
first move the channel dimension to the final dimension. Because the batch size is the same for
all scales, we can convert the prediction results to binary format (batch size, height × width ×
number of channels) to facilitate subsequent concatenation on the 1st dimension.
def flatten_pred(pred):
return npx.batch_flatten(pred.transpose(0, 2, 3, 1))
def concat_preds(preds):
return np.concatenate([flatten_pred(p) for p in preds], axis=1)
Thus, regardless of the different shapes of Y1 and Y2, we can still concatenate the prediction results
for the two different scales of the same batch.
concat_preds([Y1, Y2]).shape
(2, 25300)

Height and Width Downsample Block
For multiscale object detection, we define the following down_sample_blk block, which reduces
the height and width by 50%. This block consists of two 3 × 3 convolutional layers with a padding
of 1 and a 2 × 2 maximum pooling layer with a stride of 2 connected in a series. As we know, 3 × 3
convolutional layers with a padding of 1 do not change the shape of feature maps. However, the
subsequent pooling layer directly reduces the size of the feature map by half. Because 1 × 2 +
(3 − 1) + (3 − 1) = 6, each element in the output feature map has a receptive field on the input
feature map of the shape 6 × 6. As you can see, the height and width downsample block enlarges
the receptive field of each element in the output feature map.
def down_sample_blk(num_channels):
for _ in range(2):
blk.add(nn.Conv2D(num_channels, kernel_size=3, padding=1),
nn.BatchNorm(in_channels=num_channels),
nn.Activation('relu'))
blk.add(nn.MaxPool2D(2))
return blk
By testing forward computation in the height and width downsample block, we can see that it
changes the number of input channels and halves the height and width.
forward(np.zeros((2, 3, 20, 20)), down_sample_blk(10)).shape
(2, 10, 10, 10)
Base Network Block
The base network block is used to extract features from original images. To simplify the compu-
tation, we will construct a small base network. This network consists of three height and width
downsample blocks connected in a series, so it doubles the number of channels at each step.
When we input an original image with the shape 256 × 256, the base network block outputs a
feature map with the shape 32 × 32.
def base_net():
for num_filters in [16, 32, 64]:
blk.add(down_sample_blk(num_filters))
return blk
forward(np.zeros((2, 3, 256, 256)), base_net()).shape
(2, 64, 32, 32)

The Complete Model
The SSD model contains a total of five modules. Each module outputs a feature map used to gen-
erate anchor boxes and predict the categories and offsets of these anchor boxes. The first module
is the base network block, modules two to four are height and width downsample blocks, and the
fifth module is a global maximum pooling layer that reduces the height and width to 1. Therefore,
modules two to five are all multiscale feature blocks shown in Fig. 13.7.1.
def get_blk(i):
if i == 0:
blk = base_net()
elif i == 4:
blk = nn.GlobalMaxPool2D()
else:
blk = down_sample_blk(128)
return blk
Now, we will define the forward computation process for each module. In contrast to the
previously-described convolutional neural networks, this module not only returns feature map
Y output by convolutional computation, but also the anchor boxes of the current scale generated
from Y and their predicted categories and offsets.
def blk_forward(X, blk, size, ratio, cls_predictor, bbox_predictor):

Y = blk(X)
anchors = npx.multibox_prior(Y, sizes=size, ratios=ratio)
cls_preds = cls_predictor(Y)
bbox_preds = bbox_predictor(Y)
return (Y, anchors, cls_preds, bbox_preds)
As we mentioned, the closer a multiscale feature block is to the top in Fig. 13.7.1, the larger the
objects it detects and the larger the anchor boxes it must generate. Here, we first divide the interval
from 0.2 to 1.05 into five equal parts to determine√ the sizes of smaller√anchor boxes at different
scales: 0.2, 0.37, 0.54, etc. Then, according to 0.2 × 0.37 = 0.272, 0.37 × 0.54 = 0.447, and
similar formulas, we determine the sizes of larger anchor boxes at the different scales.
sizes = [[0.2, 0.272], [0.37, 0.447], [0.54, 0.619], [0.71, 0.79],

[0.88, 0.961]]
ratios = [[1, 2, 0.5]] * 5
num_anchors = len(sizes[0]) + len(ratios[0]) - 1
Now, we can define the complete model, TinySSD.
class TinySSD(nn.Block):
def __init__(self, num_classes, **kwargs):
super(TinySSD, self).__init__(**kwargs)
self.num_classes = num_classes
for i in range(5):
# The assignment statement is self.blk_i = get_blk(i)
setattr(self, f'blk_{i}', get_blk(i))
setattr(self, f'cls_{i}', cls_predictor(num_anchors, num_classes))
setattr(self, f'bbox_{i}', bbox_predictor(num_anchors))


anchors, cls_preds, bbox_preds = [None] * 5, [None] * 5, [None] * 5
for i in range(5):
# getattr(self, 'blk_%d' % i) accesses self.blk_i
X, anchors[i], cls_preds[i], bbox_preds[i] = blk_forward(
X, getattr(self, f'blk_{i}'), sizes[i], ratios[i],
getattr(self, f'cls_{i}'), getattr(self, f'bbox_{i}'))
# In the reshape function, 0 indicates that the batch size remains
# unchanged
anchors = np.concatenate(anchors, axis=1)
cls_preds = concat_preds(cls_preds)
cls_preds = cls_preds.reshape(
cls_preds.shape[0], -1, self.num_classes + 1)
bbox_preds = concat_preds(bbox_preds)
return anchors, cls_preds, bbox_preds
We now create an SSD model instance and use it to perform forward computation on image mini-
batch X, which has a height and width of 256 pixels. As we verified previously, the first module
outputs a feature map with the shape 32 × 32. Because modules two to four are height and width
downsample blocks, module five is a global pooling layer, and each element in the feature map is
used as the center for 4 anchor boxes, a total of (322 + 162 + 82 + 42 + 1) × 4 = 5444 anchor boxes
are generated for each image at the five scales.
net = TinySSD(num_classes=1)
net.initialize()
X = np.zeros((32, 3, 256, 256))
anchors, cls_preds, bbox_preds = net(X)
print('output anchors:', anchors.shape)

print('output class preds:', cls_preds.shape)
print('output bbox preds:', bbox_preds.shape)
output anchors: (1, 5444, 4)

output class preds: (32, 5444, 2)
output bbox preds: (32, 21776)
13.7.2 Training
Now, we will explain, step by step, how to train the SSD model for object detection.
Data Reading and Initialization
We read the banana detection dataset we created in the previous section.
batch_size = 32
train_iter, _ = d2l.load_data_bananas(batch_size)
There is 1 category in the banana detection dataset. After defining the module, we need to initial-
ize the model parameters and define the optimization algorithm.

device, net = d2l.try_gpu(), TinySSD(num_classes=1)
net.initialize(init=init.Xavier(), ctx=device)
{'learning_rate': 0.2, 'wd': 5e-4})
Defining Loss and Evaluation Functions
Object detection is subject to two types of losses. The first is anchor box category loss. For this, we
can simply reuse the cross-entropy loss function we used in image classification. The second loss
is positive anchor box offset loss. Offset prediction is a normalization problem. However, here,
we do not use the squared loss introduced previously. Rather, we use the L1 norm loss, which
is the absolute value of the difference between the predicted value and the ground-truth value.
The mask variable bbox_masks removes negative anchor boxes and padding anchor boxes from
the loss calculation. Finally, we add the anchor box category and offset losses to find the final loss
function for the model.
cls_loss = gluon.loss.SoftmaxCrossEntropyLoss()
bbox_loss = gluon.loss.L1Loss()
def calc_loss(cls_preds, cls_labels, bbox_preds, bbox_labels, bbox_masks):

cls = cls_loss(cls_preds, cls_labels)
bbox = bbox_loss(bbox_preds * bbox_masks, bbox_labels * bbox_masks)
return cls + bbox
We can use the accuracy rate to evaluate the classification results. As we use the L1 norm loss, we
will use the average absolute error to evaluate the bounding box prediction results.
def cls_eval(cls_preds, cls_labels):

# Because the category prediction results are placed in the final
# dimension, argmax must specify this dimension
return float((cls_preds.argmax(axis=-1).astype(
cls_labels.dtype) == cls_labels).sum())
def bbox_eval(bbox_preds, bbox_labels, bbox_masks):

return float((np.abs((bbox_labels - bbox_preds) * bbox_masks)).sum())
Training the Model
During model training, we must generate multiscale anchor boxes (anchors) in the modelʼs for-
ward computation process and predict the category (cls_preds) and offset (bbox_preds) for each
anchor box. Afterwards, we label the category (cls_labels) and offset (bbox_labels) of each gen-
erated anchor box based on the label information Y. Finally, we calculate the loss function using
the predicted and labeled category and offset values. To simplify the code, we do not evaluate the
training dataset here.
num_epochs, timer = 20, d2l.Timer()

legend=['class error', 'bbox mae'])

# accuracy_sum, mae_sum, num_examples, num_labels
train_iter.reset() # Read data from the start.
for batch in train_iter:
timer.start()
X = batch.data[0].as_in_ctx(device)
Y = batch.label[0].as_in_ctx(device)
# Generate multiscale anchor boxes and predict the category and
# offset of each
anchors, cls_preds, bbox_preds = net(X)
# Label the category and offset of each anchor box
bbox_labels, bbox_masks, cls_labels = npx.multibox_target(
anchors, Y, cls_preds.transpose(0, 2, 1))
# Calculate the loss function using the predicted and labeled
# category and offset values
l = calc_loss(cls_preds, cls_labels, bbox_preds, bbox_labels,
bbox_masks)
l.backward()
metric.add(cls_eval(cls_preds, cls_labels), cls_labels.size,
bbox_eval(bbox_preds, bbox_labels, bbox_masks),
bbox_labels.size)
cls_err, bbox_mae = 1-metric[0]/metric[1], metric[2]/metric[3]
animator.add(epoch+1, (cls_err, bbox_mae))
print(f'class err {cls_err:.2e}, bbox mae {bbox_mae:.2e}')
print(f'{train_iter.num_image/timer.stop():.1f} examples/sec on '
f'{str(device)}')
class err 5.07e-03, bbox mae 5.51e-03


13.7.3 Prediction
In the prediction stage, we want to detect all objects of interest in the image. Below, we read the
test image and transform its size. Then, we convert it to the four-dimensional format required by
the convolutional layer.
img = image.imread('../img/banana.jpg')
feature = image.imresize(img, 256, 256).astype('float32')
X = np.expand_dims(feature.transpose(2, 0, 1), axis=0)
Using the MultiBoxDetection function, we predict the bounding boxes based on the anchor boxes
and their predicted offsets. Then, we use non-maximum suppression to remove similar bounding
boxes.
def predict(X):
anchors, cls_preds, bbox_preds = net(X.as_in_ctx(device))
cls_probs = npx.softmax(cls_preds).transpose(0, 2, 1)
output = npx.multibox_detection(cls_probs, bbox_preds, anchors)
idx = [i for i, row in enumerate(output[0]) if row[0] != -1]
return output[0, idx]
output = predict(X)
Finally, we take all the bounding boxes with a confidence level of at least 0.3 and display them as
the final output.
def display(img, output, threshold):

fig = d2l.plt.imshow(img.asnumpy())
for row in output:
score = float(row[1])
if score < threshold:
continue
bbox = [row[2:6] * np.array((w, h, w, h), ctx=row.ctx)]
d2l.show_bboxes(fig.axes, bbox, '%.2f' % score, 'w')
display(img, output, threshold=0.9)

Summary
• SSD is a multiscale object detection model. This model generates different numbers of an-
chor boxes of different sizes based on the base network block and each multiscale feature
block and predicts the categories and offsets of the anchor boxes to detect objects of different
sizes.
• During SSD model training, the loss function is calculated using the predicted and labeled
category and offset values.
Exercises
1. Due to space limitations, we have ignored some of the implementation details of the SSD
model in this experiment. Can you further improve the model in the following areas?
Loss Function
A. For the predicted offsets, replace L1 norm loss with L1 regularization loss. This loss function
uses a square function around zero for greater smoothness. This is the regularized area controlled
by the hyperparameter σ:
{
(σx)2 /2, if |x| < 1/σ 2
f (x) = (13.7.1)
|x| − 0.5/σ 2 , otherwise
When σ is large, this loss is similar to the L1 norm loss. When the value is small, the loss function
is smoother.

sigmas = [10, 1, 0.5]
lines = ['-', '--', '-.']
x = np.arange(-2, 2, 0.1)
d2l.set_figsize()
for l, s in zip(lines, sigmas):

y = npx.smooth_l1(x, scalar=s)
d2l.plt.plot(x.asnumpy(), y.asnumpy(), l, label='sigma=%.1f' % s)
d2l.plt.legend();
In the experiment, we used cross-entropy loss for category prediction. Now, assume that the pre-
diction probability of the actual category j is pj and the cross-entropy loss is − log pj . We can also
use the focal loss (Lin et al., 2017). Given the positive hyperparameters γ and α, this loss is defined
as:
−α(1 − pj )γ log pj . (13.7.2)
As you can see, by increasing γ, we can effectively reduce the loss when the probability of predict-
ing the correct category is high.
def focal_loss(gamma, x):

return -(1 - x) ** gamma * np.log(x)
x = np.arange(0.01, 1, 0.01)
for l, gamma in zip(lines, [0, 1, 5]):
y = d2l.plt.plot(x.asnumpy(), focal_loss(gamma, x).asnumpy(), l,
label='gamma=%.1f' % gamma)
d2l.plt.legend();

B. When an object is relatively large compared to the image, the model normally adopts a larger
input image size.
C. This generally produces a large number of negative anchor boxes when labeling anchor box
categories. We can sample the negative anchor boxes to better balance the data categories. To do
this, we can set the MultiBoxTarget functionʼs negative_mining_ratio parameter.
D. Assign hyperparameters with different weights to the anchor box category loss and positive
anchor box offset loss in the loss function.
E. Refer to the SSD paper. What methods can be used to evaluate the precision of object detection
models (Liu et al., 2016)?
Discussions188
13.8 Region-based CNNs (R-CNNs)
Region-based convolutional neural networks or regions with CNN features (R-CNNs) are a pio-
neering approach that applies deep models to object detection (Girshick et al., 2014). In this sec-
tion, we will discuss R-CNNs and a series of improvements made to them: Fast R-CNN (Girshick,
2015), Faster R-CNN (Ren et al., 2015), and Mask R-CNN (He et al., 2017a). Due to space limitations,
we will confine our discussion to the designs of these models.
13.8.1 R-CNNs
R-CNN models first select several proposed regions from an image (for example, anchor boxes are
one type of selection method) and then label their categories and bounding boxes (e.g., offsets).
Then, they use a CNN to perform forward computation to extract features from each proposed
area. Afterwards, we use the features of each proposed region to predict their categories and
bounding boxes. Fig. 13.8.1 shows an R-CNN model.
188
13.8. Region-based CNNs (R-CNNs) 581

Fig. 13.8.1: R-CNN model.
Specifically, R-CNNs are composed of four main parts:

1. Selective search is performed on the input image to select multiple high-quality proposed
regions (Uijlings et al., 2013). These proposed regions are generally selected on multiple
scales and have different shapes and sizes. The category and ground-truth bounding box of
each proposed region is labeled.
2. A pre-trained CNN is selected and placed, in truncated form, before the output layer. It
transforms each proposed region into the input dimensions required by the network and
uses forward computation to output the features extracted from the proposed regions.
3. The features and labeled category of each proposed region are combined as an example to
train multiple support vector machines for object classification. Here, each support vector
machine is used to determine whether an example belongs to a certain category.
4. The features and labeled bounding box of each proposed region are combined as an example
to train a linear regression model for ground-truth bounding box prediction.
Although R-CNN models use pre-trained CNNs to effectively extract image features, the main
downside is the slow speed. As you can imagine, we can select thousands of proposed regions
from a single image, requiring thousands of forward computations from the CNN to perform ob-
ject detection. This massive computing load means that R-CNNs are not widely used in actual
applications.
13.8.2 Fast R-CNN
The main performance bottleneck of an R-CNN model is the need to independently extract fea-
tures for each proposed region. As these regions have a high degree of overlap, independent
feature extraction results in a high volume of repetitive computations. Fast R-CNN improves on
the R-CNN by only performing CNN forward computation on the image as a whole.

Fig. 13.8.2: Fast R-CNN model.
Fig. 13.8.2 shows a Fast R-CNN model. It is primary computation steps are described below:
1. Compared to an R-CNN model, a Fast R-CNN model uses the entire image as the CNN input
for feature extraction, rather than each proposed region. Moreover, this network is gener-
ally trained to update the model parameters. As the input is an entire image, the CNN output
shape is 1 × c × h1 × w1 .
2. Assuming selective search generates n proposed regions, their different shapes indicate re-
gions of interests (RoIs) of different shapes on the CNN output. Features of the same shapes
must be extracted from these RoIs (here we assume that the height is h2 and the width is
w2 ). Fast R-CNN introduces RoI pooling, which uses the CNN output and RoIs as input to
output a concatenation of the features extracted from each proposed region with the shape
n × c × h2 × w2 .
3. A fully connected layer is used to transform the output shape to n × d, where d is determined
by the model design.
4. During category prediction, the shape of the fully connected layer output is again trans-
formed to n × q and we use softmax regression (q is the number of categories). During
bounding box prediction, the shape of the fully connected layer output is again transformed
to n×4. This means that we predict the category and bounding box for each proposed region.
The RoI pooling layer in Fast R-CNN is somewhat different from the pooling layers we have dis-
cussed before. In a normal pooling layer, we set the pooling window, padding, and stride to con-
trol the output shape. In an RoI pooling layer, we can directly specify the output shape of each
region, such as specifying the height and width of each region as h2 , w2 . Assuming that the height
and width of the RoI window are h and w, this window is divided into a grid of sub-windows with
the shape h2 × w2 . The size of each sub-window is about (h/h2 ) × (w/w2 ). The sub-window height
and width must always be integers and the largest element is used as the output for a given sub-
window. This allows the RoI pooling layer to extract features of the same shape from RoIs of
different shapes.
In Fig. 13.8.3, we select an 3 × 3 region as an RoI of the 4 × 4 input. For this RoI, we use a 2 × 2 RoI
pooling layer to obtain a single 2 × 2 output. When we divide the region into four sub-windows,
they respectively contain the elements 0, 1, 4, and 5 (5 is the largest); 2 and 6 (6 is the largest); 8

and 9 (9 is the largest); and 10.
Fig. 13.8.3: 2 × 2 RoI pooling layer.
We use the ROIPooling function to demonstrate the RoI pooling layer computation. Assume that
the CNN extracts the feature X with both a height and width of 4 and only a single channel.
npx.set_np()
X = np.arange(16).reshape(1, 1, 4, 4)
X
array([[[[ 0., 1., 2., 3.],

[ 4., 5., 6., 7.],
[ 8., 9., 10., 11.],
[12., 13., 14., 15.]]]])
Assume that the height and width of the image are both 40 pixels and that selective search gener-
ates two proposed regions on the image. Each region is expressed as five elements: the regionʼs
object category and the x, y coordinates of its upper-left and bottom-right corners.
rois = np.array([[0, 0, 0, 20, 20], [0, 0, 10, 30, 30]])
Because the height and width of X are 1/10 of the height and width of the image, the coordinates
of the two proposed regions are multiplied by 0.1 according to the spatial_scale, and then the
RoIs are labeled on X as X[:, :, 0:3, 0:3] and X[:, :, 1:4, 0:4], respectively. Finally, we
divide the two RoIs into a sub-window grid and extract features with a height and width of 2.
npx.roi_pooling(X, rois, pooled_size=(2, 2), spatial_scale=0.1)
array([[[[ 5., 6.],

[ 9., 10.]]],
[[[ 9., 11.],

[13., 15.]]]])

13.8.3 Faster R-CNN
In order to obtain precise object detection results, Fast R-CNN generally requires that many pro-
posed regions be generated in selective search. Faster R-CNN replaces selective search with a re-
gion proposal network. This reduces the number of proposed regions generated, while ensuring
precise object detection.
Fig. 13.8.4: Faster R-CNN model.
Fig. 13.8.4 shows a Faster R-CNN model. Compared to Fast R-CNN, Faster R-CNN only changes the
method for generating proposed regions from selective search to region proposal network. The
other parts of the model remain unchanged. The detailed region proposal network computation
process is described below:
1. We use a 3 × 3 convolutional layer with a padding of 1 to transform the CNN output and set
the number of output channels to c. This way, each element in the feature map the CNN
extracts from the image is a new feature with a length of c.
2. We use each element in the feature map as a center to generate multiple anchor boxes of
different sizes and aspect ratios and then label them.
3. We use the features of the elements of length c at the center on the anchor boxes to predict
the binary category (object or background) and bounding box for their respective anchor
boxes.
4. Then, we use non-maximum suppression to remove similar bounding box results that corre-
spond to category predictions of “object”. Finally, we output the predicted bounding boxes
as the proposed regions required by the RoI pooling layer.
It is worth noting that, as a part of the Faster R-CNN model, the region proposal network is trained
together with the rest of the model. In addition, the Faster R-CNN object functions include the
category and bounding box predictions in object detection, as well as the binary category and
bounding box predictions for the anchor boxes in the region proposal network. Finally, the region
proposal network can learn how to generate high-quality proposed regions, which reduces the
number of proposed regions while maintaining the precision of object detection.

13.8.4 Mask R-CNN
If training data is labeled with the pixel-level positions of each object in an image, a Mask R-CNN
model can effectively use these detailed labels to further improve the precision of object detection.
Fig. 13.8.5: Mask R-CNN model.
As shown in Fig. 13.8.5, Mask R-CNN is a modification to the Faster R-CNN model. Mask R-CNN
models replace the RoI pooling layer with an RoI alignment layer. This allows the use of bilinear
interpolation to retain spatial information on feature maps, making Mask R-CNN better suited
for pixel-level predictions. The RoI alignment layer outputs feature maps of the same shape for
all RoIs. This not only predicts the categories and bounding boxes of RoIs, but allows us to use
an additional fully convolutional network to predict the pixel-level positions of objects. We will
describe how to use fully convolutional networks to predict pixel-level semantics in images later
in this chapter.
Summary
• An R-CNN model selects several proposed regions and uses a CNN to perform forward com-
putation and extract the features from each proposed region. It then uses these features to
predict the categories and bounding boxes of proposed regions.
• Fast R-CNN improves on the R-CNN by only performing CNN forward computation on the
image as a whole. It introduces an RoI pooling layer to extract features of the same shape
from RoIs of different shapes.
• Faster R-CNN replaces the selective search used in Fast R-CNN with a region proposal net-
work. This reduces the number of proposed regions generated, while ensuring precise ob-
ject detection.
• Mask R-CNN uses the same basic structure as Faster R-CNN, but adds a fully convolution
layer to help locate objects at the pixel level and further improve the precision of object
detection.

Exercises
1. Study the implementation of each model in the GluonCV toolkit189 related to this section.
Discussions190
13.9 Semantic Segmentation and the Dataset
In our discussion of object detection issues in the previous sections, we only used rectangular
bounding boxes to label and predict objects in images. In this section, we will look at seman-
tic segmentation, which attempts to segment images into regions with different semantic cate-
gories. These semantic regions label and predict objects at the pixel level. Fig. 13.9.1 shows a
semantically-segmented image, with areas labeled “dog”, “cat”, and “background”. As you can
see, compared to object detection, semantic segmentation labels areas with pixel-level borders,
for significantly greater precision.
Fig. 13.9.1: Semantically-segmented image, with areas labeled “dog”, “cat”, and “background”.
13.9.1 Image Segmentation and Instance Segmentation
In the computer vision field, there are two important methods related to semantic segmentation:
image segmentation and instance segmentation. Here, we will distinguish these concepts from
semantic segmentation as follows:
• Image segmentation divides an image into several constituent regions. This method gen-
erally uses the correlations between pixels in an image. During training, labels are not
needed for image pixels. However, during prediction, this method cannot ensure that the
segmented regions have the semantics we want. If we input the image in 9.10, image seg-
mentation might divide the dog into two regions, one covering the dogʼs mouth and eyes
where black is the prominent color and the other covering the rest of the dog where yellow
is the prominent color.
• Instance segmentation is also called simultaneous detection and segmentation. This
method attempts to identify the pixel-level regions of each object instance in an image. In
contrast to semantic segmentation, instance segmentation not only distinguishes semantics,
but also different object instances. If an image contains two dogs, instance segmentation
will distinguish which pixels belong to which dog.
189
https://github.com/dmlc/gluon-cv/
190
13.9. Semantic Segmentation and the Dataset 587

13.9.2 The Pascal VOC2012 Semantic Segmentation Dataset
In the semantic segmentation field, one important dataset is Pascal VOC2012191 . To better under-
stand this dataset, we must first import the package or module needed for the experiment.
%matplotlib inline
import os
npx.set_np()
The original site might be unstable, so we download the data from a mirror site. The archive is
about 2 GB, so it will take some time to download. After you decompress the archive, the dataset
is located in the ../data/VOCdevkit/VOC2012 path.
#@save
d2l.DATA_HUB['voc2012'] = (d2l.DATA_URL + 'VOCtrainval_11-May-2012.tar',
'4e443f8a2eca6b1dac8a6c57641b67dd40621a49')
voc_dir = d2l.download_extract('voc2012', 'VOCdevkit/VOC2012')
Go to ../data/VOCdevkit/VOC2012 to see the different parts of the dataset. The ImageSets/

Segmentation path contains text files that specify the training and testing examples. The JPEGIm-
ages and SegmentationClass paths contain the example input images and labels, respectively.
These labels are also in image format, with the same dimensions as the input images to which
they correspond. In the labels, pixels with the same color belong to the same semantic category.
The read_voc_images function defined below reads all input images and labels to the memory.
#@save
def read_voc_images(voc_dir, is_train=True):
"""Read all VOC feature and label images."""
txt_fname = os.path.join(voc_dir, 'ImageSets', 'Segmentation',
'train.txt' if is_train else 'val.txt')
with open(txt_fname, 'r') as f:
images = f.read().split()
features, labels = [], []
for i, fname in enumerate(images):
features.append(image.imread(os.path.join(
voc_dir, 'JPEGImages', f'{fname}.jpg')))
labels.append(image.imread(os.path.join(
voc_dir, 'SegmentationClass', f'{fname}.png')))
return features, labels
train_features, train_labels = read_voc_images(voc_dir, True)
We draw the first five input images and their labels. In the label images, white represents borders
and black represents the background. Other colors correspond to different categories.
n = 5
imgs = train_features[0:n] + train_labels[0:n]
d2l.show_images(imgs, 2, n);
191
http://host.robots.ox.ac.uk/pascal/VOC/voc2012/

Next, we list each RGB color value in the labels and the categories they label.
#@save
VOC_COLORMAP = [[0, 0, 0], [128, 0, 0], [0, 128, 0], [128, 128, 0],
[0, 0, 128], [128, 0, 128], [0, 128, 128], [128, 128, 128],
[64, 0, 0], [192, 0, 0], [64, 128, 0], [192, 128, 0],
[64, 0, 128], [192, 0, 128], [64, 128, 128], [192, 128, 128],
[0, 64, 0], [128, 64, 0], [0, 192, 0], [128, 192, 0],
[0, 64, 128]]
#@save
VOC_CLASSES = ['background', 'aeroplane', 'bicycle', 'bird', 'boat',
'bottle', 'bus', 'car', 'cat', 'chair', 'cow',
'diningtable', 'dog', 'horse', 'motorbike', 'person',
'potted plant', 'sheep', 'sofa', 'train', 'tv/monitor']
After defining the two constants above, we can easily find the category index for each pixel in the
labels.
#@save
def build_colormap2label():
"""Build an RGB color to label mapping for segmentation."""
colormap2label = np.zeros(256 ** 3)
for i, colormap in enumerate(VOC_COLORMAP):
colormap2label[(colormap[0]*256 + colormap[1])*256 + colormap[2]] = i
return colormap2label
#@save
def voc_label_indices(colormap, colormap2label):
"""Map an RGB color to a label."""
colormap = colormap.astype(np.int32)
idx = ((colormap[:, :, 0] * 256 + colormap[:, :, 1]) * 256
+ colormap[:, :, 2])
return colormap2label[idx]
For example, in the first example image, the category index for the front part of the airplane is 1
and the index for the background is 0.
y = voc_label_indices(train_labels[0], build_colormap2label())
y[105:115, 130:140], VOC_CLASSES[1]

(array([[0., 0., 0., 0., 0., 0., 0., 0., 0., 1.],
[0., 0., 0., 0., 0., 0., 0., 1., 1., 1.],
[0., 0., 0., 0., 0., 0., 1., 1., 1., 1.],
[0., 0., 0., 0., 0., 1., 1., 1., 1., 1.],
[0., 0., 0., 0., 0., 1., 1., 1., 1., 1.],
[0., 0., 0., 0., 1., 1., 1., 1., 1., 1.],
[0., 0., 0., 0., 0., 1., 1., 1., 1., 1.],
[0., 0., 0., 0., 0., 1., 1., 1., 1., 1.],
[0., 0., 0., 0., 0., 0., 1., 1., 1., 1.],
[0., 0., 0., 0., 0., 0., 0., 0., 1., 1.]]),
'aeroplane')
Data Preprocessing
In the preceding chapters, we scaled images to make them fit the input shape of the model. In
semantic segmentation, this method would require us to re-map the predicted pixel categories
back to the original-size input image. It would be very difficult to do this precisely, especially in
segmented regions with different semantics. To avoid this problem, we crop the images to set
dimensions and do not scale them. Specifically, we use the random cropping method used in
image augmentation to crop the same region from input images and their labels.
#@save
def voc_rand_crop(feature, label, height, width):
"""Randomly crop for both feature and label images."""
feature, rect = image.random_crop(feature, (width, height))
label = image.fixed_crop(label, *rect)
return feature, label
imgs = []
for _ in range(n):
imgs += voc_rand_crop(train_features[0], train_labels[0], 200, 300)
d2l.show_images(imgs[::2] + imgs[1::2], 2, n);

Dataset Classes for Custom Semantic Segmentation
We use the inherited Dataset class provided by Gluon to customize the semantic segmentation
dataset class VOCSegDataset. By implementing the __getitem__ function, we can arbitrarily access
the input image with the index idx and the category indexes for each of its pixels from the dataset.
As some images in the dataset may be smaller than the output dimensions specified for random
cropping, we must remove these example by using a custom filter function. In addition, we
define the normalize_image function to normalize each of the three RGB channels of the input
images.
#@save
class VOCSegDataset(gluon.data.Dataset):
"""A customized dataset to load VOC dataset."""
def __init__(self, is_train, crop_size, voc_dir):

self.rgb_mean = np.array([0.485, 0.456, 0.406])
self.rgb_std = np.array([0.229, 0.224, 0.225])
self.crop_size = crop_size
features, labels = read_voc_images(voc_dir, is_train=is_train)
self.features = [self.normalize_image(feature)
for feature in self.filter(features)]
self.labels = self.filter(labels)
self.colormap2label = build_colormap2label()
print('read ' + str(len(self.features)) + ' examples')
def normalize_image(self, img):

return (img.astype('float32') / 255 - self.rgb_mean) / self.rgb_std
def filter(self, imgs):

return [img for img in imgs if (
img.shape[0] >= self.crop_size[0] and
img.shape[1] >= self.crop_size[1])]

feature, label = voc_rand_crop(self.features[idx], self.labels[idx],
*self.crop_size)
return (feature.transpose(2, 0, 1),
voc_label_indices(label, self.colormap2label))
def __len__(self):
return len(self.features)
Reading the Dataset
Using the custom VOCSegDataset class, we create the training set and testing set instances. We
assume the random cropping operation output images in the shape 320 × 480. Below, we can see
the number of examples retained in the training and testing sets.
crop_size = (320, 480)

voc_train = VOCSegDataset(True, crop_size, voc_dir)
voc_test = VOCSegDataset(False, crop_size, voc_dir)

read 1114 examples
read 1078 examples
We set the batch size to 64 and define the iterators for the training and testing sets. Print the shape
of the first minibatch. In contrast to image classification and object recognition, labels here are
three-dimensional arrays.
batch_size = 64
train_iter = gluon.data.DataLoader(voc_train, batch_size, shuffle=True,
last_batch='discard',
print(X.shape)
print(Y.shape)
break
(64, 3, 320, 480)

(64, 320, 480)
Putting All Things Together
Finally, we define a function load_data_voc that downloads and loads this dataset, and then re-
turns the data iterators.
#@save
def load_data_voc(batch_size, crop_size):
"""Download and load the VOC2012 semantic dataset."""
voc_dir = d2l.download_extract('voc2012', os.path.join(
'VOCdevkit', 'VOC2012'))
num_workers = d2l.get_dataloader_workers()
VOCSegDataset(True, crop_size, voc_dir), batch_size,
shuffle=True, last_batch='discard', num_workers=num_workers)
VOCSegDataset(False, crop_size, voc_dir), batch_size,
last_batch='discard', num_workers=num_workers)
return train_iter, test_iter
Summary
• Semantic segmentation looks at how images can be segmented into regions with different
semantic categories.
• In the semantic segmentation field, one important dataset is Pascal VOC2012.
• Because the input images and labels in semantic segmentation have a one-to-one correspon-
dence at the pixel level, we randomly crop them to a fixed size, rather than scaling them.

Exercises
1. Recall the content we covered in Section 13.1. Which of the image augmentation methods
used in image classification would be hard to use in semantic segmentation?
Discussions192
13.10 Transposed Convolution
The layers we introduced so far for convolutional neural networks, including convolutional lay-
ers (Section 6.2) and pooling layers (Section 6.5), often reduce the input width and height, or
keep them unchanged. Applications such as semantic segmentation (Section 13.9) and generative
adversarial networks (Section 17.2), however, require to predict values for each pixel and there-
fore needs to increase input width and height. Transposed convolution, also named fractionally-
strided convolution (Dumoulin & Visin, 2016) or deconvolution (Long et al., 2015), serves this pur-
pose.
from mxnet import np, npx, init

npx.set_np()
13.10.1 Basic 2D Transposed Convolution
Let us consider a basic case that both input and output channels are 1, with 0 padding and 1 stride.
Fig. 13.10.1 illustrates how transposed convolution with a 2 × 2 kernel is computed on the 2 × 2
input matrix.
Fig. 13.10.1: Transposed convolution layer with a 2 × 2 kernel.
We can implement this operation by giving matrix kernel K and matrix input X.
def trans_conv(X, K):

h, w = K.shape
Y = np.zeros((X.shape[0] + h - 1, X.shape[1] + w - 1))
for i in range(X.shape[0]):
for j in range(X.shape[1]):
Y[i: i + h, j: j + w] += X[i, j] * K
return Y
192
13.10. Transposed Convolution 593

Remember the convolution computes results by Y[i, j] = (X[i: i + h, j: j + w] * K).
sum() (refer to corr2d in Section 6.2), which summarizes input values through the kernel. While
the transposed convolution broadcasts input values through the kernel, which results in a larger
output shape.
Verify the results in Fig. 13.10.1.
X = np.array([[0, 1], [2, 3]])

K = np.array([[0, 1], [2, 3]])
trans_conv(X, K)
array([[ 0., 0., 1.],

[ 0., 4., 6.],
[ 4., 12., 9.]])
Or we can use nn.Conv2DTranspose to obtain the same results. As nn.Conv2D, both input and kernel
should be 4-D tensors.
X, K = X.reshape(1, 1, 2, 2), K.reshape(1, 1, 2, 2)

tconv = nn.Conv2DTranspose(1, kernel_size=2)
tconv.initialize(init.Constant(K))
tconv(X)
array([[[[ 0., 0., 1.],

[ 0., 4., 6.],
[ 4., 12., 9.]]]])
13.10.2 Padding, Strides, and Channels
We apply padding elements to the input in convolution, while they are applied to the output in
transposed convolution. A 1 × 1 padding means we first compute the output as normal, then
remove the first/last rows and columns.
tconv = nn.Conv2DTranspose(1, kernel_size=2, padding=1)

tconv(X)
array([[[[4.]]]])
Similarly, strides are applied to outputs as well.
tconv = nn.Conv2DTranspose(1, kernel_size=2, strides=2)

tconv(X)
array([[[[0., 0., 0., 1.],

[0., 0., 2., 3.],
[0., 2., 0., 3.],
[4., 6., 6., 9.]]]])

The multi-channel extension of the transposed convolution is the same as the convolution. When
the input has multiple channels, denoted by ci , the transposed convolution assigns a kh ×kw kernel
matrix to each input channel. If the output has a channel size co , then we have a ci ×kh ×kw kernel
for each output channel.
As a result, if we feed X into a convolutional layer f to compute Y = f (X) and create a transposed
convolution layer g with the same hyperparameters as f except for the output channel set to be
the channel size of X, then g(Y ) should has the same shape as X. Let us verify this statement.

conv = nn.Conv2D(20, kernel_size=5, padding=2, strides=3)
tconv = nn.Conv2DTranspose(10, kernel_size=5, padding=2, strides=3)
conv.initialize()
tconv.initialize()
tconv(conv(X)).shape == X.shape
True
13.10.3 Analogy to Matrix Transposition
The transposed convolution takes its name from the matrix transposition. In fact, convolution
operations can also be achieved by matrix multiplication. In the example below, we define a 3×
input X with a 2 × 2 kernel K, and then use corr2d to compute the convolution output.
K = np.array([[0, 1], [2, 3]])
Y = d2l.corr2d(X, K)
Y
array([[19., 25.],
[37., 43.]])
Next, we rewrite convolution kernel K as a matrix W . Its shape will be (4, 9), where the ith row
present applying the kernel to the input to generate the ith output element.
def kernel2matrix(K):
k, W = np.zeros(5), np.zeros((4, 9))
k[:2], k[3:5] = K[0, :], K[1, :]
W[0, :5], W[1, 1:6], W[2, 3:8], W[3, 4:] = k, k, k, k
return W
W = kernel2matrix(K)
W
array([[0., 1., 0., 2., 3., 0., 0., 0., 0.],

[0., 0., 1., 0., 2., 3., 0., 0., 0.],
[0., 0., 0., 0., 1., 0., 2., 3., 0.],
[0., 0., 0., 0., 0., 1., 0., 2., 3.]])
Then the convolution operator can be implemented by matrix multiplication with proper reshap-
ing.
13.10. Transposed Convolution 595

Y == np.dot(W, X.reshape(-1)).reshape(2, 2)
array([[ True, True],

[ True, True]])
We can implement transposed convolution as a matrix multiplication as well by reusing ker-

nel2matrix. To reuse the generated W , we construct a 2 × 2 input, so the corresponding weight
matrix will have a shape (9, 4), which is W ⊤ . Let us verify the results.
X = np.array([[0, 1], [2, 3]])

Y = trans_conv(X, K)
Y == np.dot(W.T, X.reshape(-1)).reshape(3, 3)
array([[ True, True, True],

[ True, True, True],
[ True, True, True]])
Summary
• Compared to convolutions that reduce inputs through kernels, transposed convolutions

broadcast inputs.
• If a convolution layer reduces the input width and height by nw and hh time, respectively.
Then a transposed convolution layer with the same kernel sizes, padding and strides will
increase the input width and height by nw and nh , respectively.
• We can implement convolution operations by the matrix multiplication, the corresponding
transposed convolutions can be done by transposed matrix multiplication.
Exercises
1. Is it efficient to use matrix multiplication to implement convolution operations? Why?

Discussions193
13.11 Fully Convolutional Networks (FCN)
We previously discussed semantic segmentation using each pixel in an image for category predic-
tion. A fully convolutional network (FCN) (Long et al., 2015) uses a convolutional neural network
to transform image pixels to pixel categories. Unlike the convolutional neural networks previously
introduced, an FCN transforms the height and width of the intermediate layer feature map back
to the size of input image through the transposed convolution layer, so that the predictions have a
one-to-one correspondence with input image in spatial dimension (height and width). Given a po-
sition on the spatial dimension, the output of the channel dimension will be a category prediction
of the pixel corresponding to the location.
193

We will first import the package or module needed for the experiment and then explain the trans-
posed convolution layer.
%matplotlib inline
from mxnet import gluon, image, init, np, npx
npx.set_np()
13.11.1 Constructing a Model
Here, we demonstrate the most basic design of a fully convolutional network model. As shown in
Fig. 13.11.1, the fully convolutional network first uses the convolutional neural network to extract
image features, then transforms the number of channels into the number of categories through
the 1 × 1 convolution layer, and finally transforms the height and width of the feature map to
the size of the input image by using the transposed convolution layer Section 13.10. The model
output has the same height and width as the input image and has a one-to-one correspondence
in spatial positions. The final output channel contains the category prediction of the pixel of the
corresponding spatial position.
Fig. 13.11.1: Fully convolutional network.
Below, we use a ResNet-18 model pre-trained on the ImageNet dataset to extract image features
and record the network instance as pretrained_net. As you can see, the last two layers of the
model member variable features are the global maximum pooling layer GlobalAvgPool2D and
example flattening layer Flatten. The output module contains the fully connected layer used for
output. These layers are not required for a fully convolutional network.
13.11. Fully Convolutional Networks (FCN) 597

pretrained_net = gluon.model_zoo.vision.resnet18_v2(pretrained=True)
pretrained_net.features[-4:], pretrained_net.output
(HybridSequential(
(0): BatchNorm(axis=1, eps=1e-05, momentum=0.9, fix_gamma=False, use_global_stats=False,␣
,→in_channels=512)
(1): Activation(relu)
(2): GlobalAvgPool2D(size=(1, 1), stride=(1, 1), padding=(0, 0), ceil_mode=True, global_
,→pool=True, pool_type=avg, layout=NCHW)
(3): Flatten
),
Dense(512 -> 1000, linear))
Next, we create the fully convolutional network instance net. It duplicates all the neural layers
except the last two layers of the instance member variable features of pretrained_net and the
model parameters obtained after pre-training.
for layer in pretrained_net.features[:-2]:
net.add(layer)
Given an input of a height and width of 320 and 480 respectively, the forward computation of net
will reduce the height and width of the input to 1/32 of the original, i.e., 10 and 15.

net(X).shape
(1, 512, 10, 15)
Next, we transform the number of output channels to the number of categories of Pascal VOC2012
(21) through the 1 × 1 convolution layer. Finally, we need to magnify the height and width of the
feature map by a factor of 32 to change them back to the height and width of the input image.
Recall the calculation method for the convolution layer output shape described in Section 6.3.
Because (320 − 64 + 16 × 2 + 32)/32 = 10 and (480 − 64 + 16 × 2 + 32)/32 = 15, we construct a
transposed convolution layer with a stride of 32 and set the height and width of the convolution
kernel to 64 and the padding to 16. It is not difficult to see that, if the stride is s, the padding is
s/2 (assuming s/2 is an integer), and the height and width of the convolution kernel are 2s, the
transposed convolution kernel will magnify both the height and width of the input by a factor of
s.
num_classes = 21
net.add(nn.Conv2D(num_classes, kernel_size=1),
nn.Conv2DTranspose(
num_classes, kernel_size=64, padding=16, strides=32))

13.11.2 Initializing the Transposed Convolution Layer
We already know that the transposed convolution layer can magnify a feature map. In image pro-
cessing, sometimes we need to magnify the image, i.e., upsampling. There are many methods
for upsampling, and one common method is bilinear interpolation. Simply speaking, in order to
get the pixel of the output image at the coordinates (x, y), the coordinates are first mapped to the
coordinates of the input image (x′ , y ′ ). This can be done based on the ratio of the size of three
input to the size of the output. The mapped values x′ and y ′ are usually real numbers. Then,
we find the four pixels closest to the coordinate (x′ , y ′ ) on the input image. Finally, the pixels
of the output image at coordinates (x, y) are calculated based on these four pixels on the input
image and their relative distances to (x′ , y ′ ). Upsampling by bilinear interpolation can be imple-
mented by transposed convolution layer of the convolution kernel constructed using the following
bilinear_kernel function. Due to space limitations, we only give the implementation of the bi-
linear_kernel function and will not discuss the principles of the algorithm.
def bilinear_kernel(in_channels, out_channels, kernel_size):

factor = (kernel_size + 1) // 2
if kernel_size % 2 == 1:
center = factor - 1
else:
center = factor - 0.5
og = (np.arange(kernel_size).reshape(-1, 1),
np.arange(kernel_size).reshape(1, -1))
filt = (1 - np.abs(og[0] - center) / factor) * \
(1 - np.abs(og[1] - center) / factor)
weight = np.zeros((in_channels, out_channels, kernel_size, kernel_size))
weight[range(in_channels), range(out_channels), :, :] = filt
return np.array(weight)
Now, we will experiment with bilinear interpolation upsampling implemented by transposed con-
volution layers. Construct a transposed convolution layer that magnifies height and width of input
by a factor of 2 and initialize its convolution kernel with the bilinear_kernel function.
conv_trans = nn.Conv2DTranspose(3, kernel_size=4, padding=1, strides=2)

conv_trans.initialize(init.Constant(bilinear_kernel(3, 3, 4)))
Read the image X and record the result of upsampling as Y. In order to print the image, we need
to adjust the position of the channel dimension.
img = image.imread('../img/catdog.jpg')
X = np.expand_dims(img.astype('float32').transpose(2, 0, 1), axis=0) / 255
Y = conv_trans(X)
out_img = Y[0].transpose(1, 2, 0)
As you can see, the transposed convolution layer magnifies both the height and width of the image
by a factor of 2. It is worth mentioning that, besides to the difference in coordinate scale, the image
magnified by bilinear interpolation and original image printed in Section 13.3 look the same.
d2l.set_figsize()
print('input image shape:', img.shape)
d2l.plt.imshow(img.asnumpy());
print('output image shape:', out_img.shape)
d2l.plt.imshow(out_img.asnumpy());

input image shape: (561, 728, 3)
output image shape: (1122, 1456, 3)
In a fully convolutional network, we initialize the transposed convolution layer for upsampled
bilinear interpolation. For a 1 × 1 convolution layer, we use Xavier for randomly initialization.
W = bilinear_kernel(num_classes, num_classes, 64)

net[-1].initialize(init.Constant(W))
net[-2].initialize(init=init.Xavier())
We read the dataset using the method described in the previous section. Here, we specify shape
of the randomly cropped output image as 320 × 480, so both the height and width are divisible by
32.
batch_size, crop_size = 32, (320, 480)

train_iter, test_iter = d2l.load_data_voc(batch_size, crop_size)
Downloading ../data/VOCtrainval_11-May-2012.tar from http://d2l-data.s3-accelerate.amazonaws.

,→com/VOCtrainval_11-May-2012.tar...
read 1114 examples
read 1078 examples
13.11.4 Training
Now we can start training the model. The loss function and accuracy calculation here are not
substantially different from those used in image classification. Because we use the channel of the
transposed convolution layer to predict pixel categories, the axis=1 (channel dimension) option
is specified in SoftmaxCrossEntropyLoss. In addition, the model calculates the accuracy based on
whether the prediction category of each pixel is correct.

num_epochs, lr, wd, devices = 5, 0.1, 1e-3, d2l.try_all_gpus()
loss = gluon.loss.SoftmaxCrossEntropyLoss(axis=1)
net.collect_params().reset_ctx(devices)
{'learning_rate': lr, 'wd': wd})
d2l.train_ch13(net, train_iter, test_iter, loss, trainer, num_epochs, devices)

13.11.5 Prediction
During predicting, we need to standardize the input image in each channel and transform them
into the four-dimensional input format required by the convolutional neural network.
def predict(img):
X = test_iter._dataset.normalize_image(img)
X = np.expand_dims(X.transpose(2, 0, 1), axis=0)
pred = net(X.as_in_ctx(devices[0])).argmax(axis=1)
return pred.reshape(pred.shape[1], pred.shape[2])
To visualize the predicted categories for each pixel, we map the predicted categories back to their
labeled colors in the dataset.
def label2image(pred):
colormap = np.array(d2l.VOC_COLORMAP, ctx=devices[0], dtype='uint8')
X = pred.astype('int32')
return colormap[X, :]
The size and shape of the images in the test dataset vary. Because the model uses a transposed
convolution layer with a stride of 32, when the height or width of the input image is not divisible
by 32, the height or width of the transposed convolution layer output deviates from the size of the
input image. In order to solve this problem, we can crop multiple rectangular areas in the image
with heights and widths as integer multiples of 32, and then perform forward computation on the
pixels in these areas. When combined, these areas must completely cover the input image. When
a pixel is covered by multiple areas, the average of the transposed convolution layer output in the

forward computation of the different areas can be used as an input for the softmax operation to
predict the category.
For the sake of simplicity, we only read a few large test images and crop an area with a shape of
320 × 480 from the top-left corner of the image. Only this area is used for prediction. For the input
image, we print the cropped area first, then print the predicted result, and finally print the labeled
category.
voc_dir = d2l.download_extract('voc2012', 'VOCdevkit/VOC2012')

test_images, test_labels = d2l.read_voc_images(voc_dir, False)
n, imgs = 4, []
for i in range(n):
crop_rect = (0, 0, 480, 320)
X = image.fixed_crop(test_images[i], *crop_rect)
pred = label2image(predict(X))
imgs += [X, pred, image.fixed_crop(test_labels[i], *crop_rect)]
d2l.show_images(imgs[::3] + imgs[1::3] + imgs[2::3], 3, n, scale=2);
Summary
• The fully convolutional network first uses the convolutional neural network to extract image
features, then transforms the number of channels into the number of categories through the
1 × 1 convolution layer, and finally transforms the height and width of the feature map to
the size of the input image by using the transposed convolution layer to output the category
of each pixel.
• In a fully convolutional network, we initialize the transposed convolution layer for upsam-
pled bilinear interpolation.

Exercises
1. If we use Xavier to randomly initialize the transposed convolution layer, what will happen
to the result?
2. Can you further improve the accuracy of the model by tuning the hyperparameters?
3. Predict the categories of all pixels in the test image.
4. The outputs of some intermediate layers of the convolutional neural network are also used
in the paper on fully convolutional networks[1]. Try to implement this idea.
Discussions194
13.12 Neural Style Transfer
If you use social sharing apps or happen to be an amateur photographer, you are familiar with
filters. Filters can alter the color styles of photos to make the background sharper or peopleʼs
faces whiter. However, a filter generally can only change one aspect of a photo. To create the
ideal photo, you often need to try many different filter combinations. This process is as complex
as tuning the hyperparameters of a model.
In this section, we will discuss how we can use convolution neural networks (CNNs) to automat-
ically apply the style of one image to another image, an operation known as style transfer (Gatys
et al., 2016). Here, we need two input images, one content image and one style image. We use
a neural network to alter the content image so that its style mirrors that of the style image. In
Fig. 13.12.1, the content image is a landscape photo the author took in Mount Rainier National
Part near Seattle. The style image is an oil painting of oak trees in autumn. The output composite
image retains the overall shapes of the objects in the content image, but applies the oil painting
brushwork of the style image and makes the overall color more vivid.
Fig. 13.12.1: Content and style input images and composite image produced by style transfer.
194
13.12. Neural Style Transfer 603

13.12.1 Technique
The CNN-based style transfer model is shown in Fig. 13.12.2. First, we initialize the composite
image. For example, we can initialize it as the content image. This composite image is the only
variable that needs to be updated in the style transfer process, i.e., the model parameter to be
updated in style transfer. Then, we select a pre-trained CNN to extract image features. These
model parameters do not need to be updated during training. The deep CNN uses multiple neu-
ral layers that successively extract image features. We can select the output of certain layers to
use as content features or style features. If we use the structure in Fig. 13.12.2, the pre-trained
neural network contains three convolutional layers. The second layer outputs the image content
features, while the outputs of the first and third layers are used as style features. Next, we use for-
ward propagation (in the direction of the solid lines) to compute the style transfer loss function
and backward propagation (in the direction of the dotted lines) to update the model parameter,
constantly updating the composite image. The loss functions used in style transfer generally have
three parts: 1. Content loss is used to make the composite image approximate the content im-
age as regards content features. 2. Style loss is used to make the composite image approximate
the style image in terms of style features. 3. Total variation loss helps reduce the noise in the
composite image. Finally, after we finish training the model, we output the style transfer model
parameters to obtain the final composite image.
Fig. 13.12.2: CNN-based style transfer process. Solid lines show the direction of forward propaga-
tion and dotted lines show backward propagation.
Next, we will perform an experiment to help us better understand the technical details of style
transfer.

13.12.2 Reading the Content and Style Images
First, we read the content and style images. By printing out the image coordinate axes, we can see
that they have different dimensions.
%matplotlib inline
npx.set_np()
d2l.set_figsize()
content_img = image.imread('../img/rainier.jpg')
d2l.plt.imshow(content_img.asnumpy());
style_img = image.imread('../img/autumn_oak.jpg')
d2l.plt.imshow(style_img.asnumpy());

13.12.3 Preprocessing and Postprocessing
Below, we define the functions for image preprocessing and postprocessing. The preprocess func-
tion normalizes each of the three RGB channels of the input images and transforms the results to
a format that can be input to the CNN. The postprocess function restores the pixel values in the
output image to their original values before normalization. Because the image printing function
requires that each pixel has a floating point value from 0 to 1, we use the clip function to replace
values smaller than 0 or greater than 1 with 0 or 1, respectively.
rgb_mean = np.array([0.485, 0.456, 0.406])

rgb_std = np.array([0.229, 0.224, 0.225])
def preprocess(img, image_shape):

img = image.imresize(img, *image_shape)
img = (img.astype('float32') / 255 - rgb_mean) / rgb_std
return np.expand_dims(img.transpose(2, 0, 1), axis=0)
def postprocess(img):
img = img[0].as_in_ctx(rgb_std.ctx)
return (img.transpose(1, 2, 0) * rgb_std + rgb_mean).clip(0, 1)
13.12.4 Extracting Features
We use the VGG-19 model pre-trained on the ImageNet dataset to extract image features[1].
pretrained_net = gluon.model_zoo.vision.vgg19(pretrained=True)
To extract image content and style features, we can select the outputs of certain layers in the VGG
network. In general, the closer an output is to the input layer, the easier it is to extract image
detail information. The farther away an output is, the easier it is to extract global information. To
prevent the composite image from retaining too many details from the content image, we select
a VGG network layer near the output layer to output the image content features. This layer is
called the content layer. We also select the outputs of different layers from the VGG network for
matching local and global styles. These are called the style layers. As we mentioned in Section 7.2,
VGG networks have five convolutional blocks. In this experiment, we select the last convolutional
layer of the fourth convolutional block as the content layer and the first layer of each block as style
layers. We can obtain the indexes for these layers by printing the pretrained_net instance.
style_layers, content_layers = [0, 5, 10, 19, 28], [25]
During feature extraction, we only need to use all the VGG layers from the input layer to the content
or style layer nearest the output layer. Below, we build a new network, net, which only retains the
layers in the VGG network we need to use. We then use net to extract features.
for i in range(max(content_layers + style_layers) + 1):
net.add(pretrained_net.features[i])
Given input X, if we simply call the forward computation net(X), we can only obtain the output of
the last layer. Because we also need the outputs of the intermediate layers, we need to perform
layer-by-layer computation and retain the content and style layer outputs.

def extract_features(X, content_layers, style_layers):
contents = []
styles = []
for i in range(len(net)):
X = net[i](X)
if i in style_layers:
styles.append(X)
if i in content_layers:
contents.append(X)
return contents, styles
Next, we define two functions: The get_contents function obtains the content features extracted
from the content image, while the get_styles function obtains the style features extracted from
the style image. Because we do not need to change the parameters of the pre-trained VGG model
during training, we can extract the content features from the content image and style features
from the style image before the start of training. As the composite image is the model parameter
that must be updated during style transfer, we can only call the extract_features function during
training to extract the content and style features of the composite image.
def get_contents(image_shape, device):

content_X = preprocess(content_img, image_shape).copyto(device)
contents_Y, _ = extract_features(content_X, content_layers, style_layers)
return content_X, contents_Y
def get_styles(image_shape, device):

style_X = preprocess(style_img, image_shape).copyto(device)
_, styles_Y = extract_features(style_X, content_layers, style_layers)
return style_X, styles_Y
Next, we will look at the loss function used for style transfer. The loss function includes the content
loss, style loss, and total variation loss.
Content Loss
Similar to the loss function used in linear regression, content loss uses a square error function
to measure the difference in content features between the composite image and content image.
The two inputs of the square error function are both content layer outputs obtained from the ex-
tract_features function.
def content_loss(Y_hat, Y):

return np.square(Y_hat - Y).mean()

Style Loss
Style loss, similar to content loss, uses a square error function to measure the difference in style
between the composite image and style image. To express the styles output by the style layers,
we first use the extract_features function to compute the style layer output. Assuming that the
output has 1 example, c channels, and a height and width of h and w, we can transform the output
into the matrix X, which has c rows and h·w columns. You can think of matrix X as the combination
of the c vectors x1 , . . . , xc , which have a length of hw. Here, the vector xi represents the style
feature of channel i. In the Gram matrix of these vectors XX⊤ ∈ Rc×c , element xij in row i column
j is the inner product of vectors xi and xj . It represents the correlation of the style features of
channels i and j. We use this type of Gram matrix to represent the style output by the style layers.
You must note that, when the h·w value is large, this often leads to large values in the Gram matrix.
In addition, the height and width of the Gram matrix are both the number of channels c. To ensure
that the style loss is not affected by the size of these values, we define the gram function below to
divide the Gram matrix by the number of its elements, i.e., c · h · w.
def gram(X):
num_channels, n = X.shape[1], X.size // X.shape[1]
X = X.reshape(num_channels, n)
return np.dot(X, X.T) / (num_channels * n)
Naturally, the two Gram matrix inputs of the square error function for style loss are taken from
the composite image and style image style layer outputs. Here, we assume that the Gram matrix
of the style image, gram_Y, has been computed in advance.
def style_loss(Y_hat, gram_Y):

return np.square(gram(Y_hat) - gram_Y).mean()
Total Variance Loss
Sometimes, the composite images we learn have a lot of high-frequency noise, particularly bright
or dark pixels. One common noise reduction method is total variation denoising. We assume that
xi,j represents the pixel value at the coordinate (i, j), so the total variance loss is:
∑
|xi,j − xi+1,j | + |xi,j − xi,j+1 | . (13.12.1)
i,j
We try to make the values of neighboring pixels as similar as possible.
def tv_loss(Y_hat):
return 0.5 * (np.abs(Y_hat[:, :, 1:, :] - Y_hat[:, :, :-1, :]).mean() +
np.abs(Y_hat[:, :, :, 1:] - Y_hat[:, :, :, :-1]).mean())

The Loss Function
The loss function for style transfer is the weighted sum of the content loss, style loss, and total
variance loss. By adjusting these weight hyperparameters, we can balance the retained content,
transferred style, and noise reduction in the composite image according to their relative impor-
tance.
content_weight, style_weight, tv_weight = 1, 1e3, 10
def compute_loss(X, contents_Y_hat, styles_Y_hat, contents_Y, styles_Y_gram):

# Calculate the content, style, and total variance losses respectively
contents_l = [content_loss(Y_hat, Y) * content_weight for Y_hat, Y in zip(
contents_Y_hat, contents_Y)]
styles_l = [style_loss(Y_hat, Y) * style_weight for Y_hat, Y in zip(
styles_Y_hat, styles_Y_gram)]
tv_l = tv_loss(X) * tv_weight
# Add up all the losses
l = sum(styles_l + contents_l + [tv_l])
return contents_l, styles_l, tv_l, l
13.12.6 Creating and Initializing the Composite Image
In style transfer, the composite image is the only variable that needs to be updated. Therefore, we
can define a simple model, GeneratedImage, and treat the composite image as a model parameter.
In the model, forward computation only returns the model parameter.
class GeneratedImage(nn.Block):
def __init__(self, img_shape, **kwargs):
super(GeneratedImage, self).__init__(**kwargs)
self.weight = self.params.get('weight', shape=img_shape)
def forward(self):
return self.weight.data()
Next, we define the get_inits function. This function creates a composite image model instance
and initializes it to the image X. The Gram matrix for the various style layers of the style image,
styles_Y_gram, is computed prior to training.
def get_inits(X, device, lr, styles_Y):

gen_img = GeneratedImage(X.shape)
gen_img.initialize(init.Constant(X), ctx=device, force_reinit=True)
trainer = gluon.Trainer(gen_img.collect_params(), 'adam',
styles_Y_gram = [gram(Y) for Y in styles_Y]
return gen_img(), styles_Y_gram, trainer

13.12.7 Training
During model training, we constantly extract the content and style features of the composite im-
age and calculate the loss function. Recall our discussion of how synchronization functions force
the front end to wait for computation results in Section 12.2. Because we only call the asscalar
synchronization function every 50 epochs, the process may occupy a great deal of memory. There-
fore, we call the waitall synchronization function during every epoch.
def train(X, contents_Y, styles_Y, device, lr, num_epochs, lr_decay_epoch):

X, styles_Y_gram, trainer = get_inits(X, device, lr, styles_Y)
xlim=[1, num_epochs],
legend=['content', 'style', 'TV'],
ncols=2, figsize=(7, 2.5))
for epoch in range(1, num_epochs+1):
contents_Y_hat, styles_Y_hat = extract_features(
X, content_layers, style_layers)
contents_l, styles_l, tv_l, l = compute_loss(
X, contents_Y_hat, styles_Y_hat, contents_Y, styles_Y_gram)
l.backward()
trainer.step(1)
npx.waitall()
if epoch % lr_decay_epoch == 0:
trainer.set_learning_rate(trainer.learning_rate * 0.1)
if epoch % 10 == 0:
animator.axes[1].imshow(postprocess(X).asnumpy())
animator.add(epoch, [float(sum(contents_l)),
float(sum(styles_l)),
float(tv_l)])
return X
Next, we start to train the model. First, we set the height and width of the content and style images
to 150 by 225 pixels. We use the content image to initialize the composite image.
device, image_shape = d2l.try_gpu(), (225, 150)

net.collect_params().reset_ctx(device)
content_X, contents_Y = get_contents(image_shape, device)
_, styles_Y = get_styles(image_shape, device)
output = train(content_X, contents_Y, styles_Y, device, 0.01, 500, 200)

As you can see, the composite image retains the scenery and objects of the content image, while
introducing the color of the style image. Because the image is relatively small, the details are a bit
fuzzy.
To obtain a clearer composite image, we train the model using a larger image size: 900 × 600. We
increase the height and width of the image used before by a factor of four and initialize a larger
composite image.
image_shape = (900, 600)

_, content_Y = get_contents(image_shape, device)
_, style_Y = get_styles(image_shape, device)
X = preprocess(postprocess(output) * 255, image_shape)
output = train(X, content_Y, style_Y, device, 0.01, 300, 100)
d2l.plt.imsave('../img/neural-style.jpg', postprocess(output).asnumpy())
As you can see, each epoch takes more time due to the larger image size. As shown in Fig. 13.12.3,
the composite image produced retains more detail due to its larger size. The composite image not
only has large blocks of color like the style image, but these blocks even have the subtle texture of
brush strokes.

Fig. 13.12.3: 900 × 600 composite image.
Summary
• The loss functions used in style transfer generally have three parts: 1. Content loss is used
to make the composite image approximate the content image as regards content features. 2.
Style loss is used to make the composite image approximate the style image in terms of style
features. 3. Total variation loss helps reduce the noise in the composite image.
• We can use a pre-trained CNN to extract image features and minimize the loss function to
continuously update the composite image.
• We use a Gram matrix to represent the style output by the style layers.
Exercises
1. How does the output change when you select different content and style layers?
2. Adjust the weight hyperparameters in the loss function. Does the output retain more content
or have less noise?
3. Use different content and style images. Can you create more interesting composite images?
Discussions195
195

13.13 Image Classification (CIFAR-10) on Kaggle
So far, we have been using Gluonʼs data package to directly obtain image datasets in the tensor for-
mat. In practice, however, image datasets often exist in the format of image files. In this section,
we will start with the original image files and organize, read, and convert the files to the tensor
format step by step.
We performed an experiment on the CIFAR-10 dataset in Section 13.1. This is an important data set
in the computer vision field. Now, we will apply the knowledge we learned in the previous sections
in order to participate in the Kaggle competition, which addresses CIFAR-10 image classification
problems. The competitionʼs web address is
https://www.kaggle.com/c/cifar-10
Fig. 13.13.1 shows the information on the competitionʼs webpage. In order to submit the results,
please register an account on the Kaggle website first.
Fig. 13.13.1: CIFAR-10 image classification competition webpage information. The dataset for the
competition can be accessed by clicking the “Data” tab.
First, import the packages or modules required for the competition.
import collections
import math
from mxnet import autograd, gluon, init, npx
import os
import pandas as pd
import shutil
import time
npx.set_np()
13.13. Image Classification (CIFAR-10) on Kaggle 613

13.13.1 Obtaining and Organizing the Dataset
The competition data is divided into a training set and testing set. The training set contains 50, 000
images. The testing set contains 300, 000 images, of which 10, 000 images are used for scoring,
while the other 290, 000 non-scoring images are included to prevent the manual labeling of the
testing set and the submission of labeling results. The image formats in both datasets are PNG,
with heights and widths of 32 pixels and three color channels (RGB). The images cover 10 cate-
gories: planes, cars, birds, cats, deer, dogs, frogs, horses, boats, and trucks. The upper-left corner
of Figure 9.16 shows some images of planes, cars, and birds in the dataset.
Downloading the Dataset
After logging in to Kaggle, we can click on the “Data” tab on the CIFAR-10 image classification
competition webpage shown in Fig. 13.13.1 and download the dataset by clicking the “Download
All” button. After unzipping the downloaded file in ../data, and unzipping train.7z and test.7z
inside it, you will find the entire dataset in the following paths:
• ../data/cifar-10/train/[1-50000].png
• ../data/cifar-10/test/[1-300000].png
• ../data/cifar-10/trainLabels.csv
• ../data/cifar-10/sampleSubmission.csv
Here folders train and test contain the training and testing images respectively, trainLabels.csv
has labels for the training images, and sample_submission.csv is a sample of submission.
To make it easier to get started, we provide a small-scale sample of the dataset: it contains the first
1000 training images and 5 random testing images. To use the full dataset of the Kaggle competi-
tion, you need to set the following demo variable to False.
#@save
d2l.DATA_HUB['cifar10_tiny'] = (d2l.DATA_URL + 'kaggle_cifar10_tiny.zip',
'2068874e4b9a9f0fb07ebe0ad2b29754449ccacd')
# If you use the full dataset downloaded for the Kaggle competition, set
# `demo` to False
demo = True
if demo:
data_dir = d2l.download_extract('cifar10_tiny')
else:
data_dir = '../data/cifar-10/'
Downloading ../data/kaggle_cifar10_tiny.zip from http://d2l-data.s3-accelerate.amazonaws.com/

,→kaggle_cifar10_tiny.zip...

Organizing the Dataset
We need to organize datasets to facilitate model training and testing. Let us first read the labels
from the csv file. The following function returns a dictionary that maps the filename without
extension to its label.
#@save
def read_csv_labels(fname):
"""Read fname to return a name to label dictionary."""
with open(fname, 'r') as f:
# Skip the file header line (column name)
lines = f.readlines()[1:]
tokens = [l.rstrip().split(',') for l in lines]
return dict(((name, label) for name, label in tokens))
labels = read_csv_labels(os.path.join(data_dir, 'trainLabels.csv'))

print('# training examples:', len(labels))
print('# classes:', len(set(labels.values())))
# training examples: 1000

# classes: 10
Next, we define the reorg_train_valid function to segment the validation set from the original
training set. The argument valid_ratio in this function is the ratio of the number of examples
in the validation set to the number of examples in the original training set. In particular, let n
be the number of images of the class with the least examples, and r be the ratio, then we will
use max(⌊nr⌋, 1) images for each class as the validation set. Let us use valid_ratio=0.1 as an
example. Since the original training set has 50, 000 images, there will be 45, 000 images used for
training and stored in the path “train_valid_test/train” when tuning hyperparameters, while
the other 5, 000 images will be stored as validation set in the path “train_valid_test/valid”. After
organizing the data, images of the same class will be placed under the same folder so that we can
read them later.
#@save
def copyfile(filename, target_dir):
"""Copy a file into a target directory."""
d2l.mkdir_if_not_exist(target_dir)
shutil.copy(filename, target_dir)
#@save
def reorg_train_valid(data_dir, labels, valid_ratio):
# The number of examples of the class with the least examples in the
# training dataset
n = collections.Counter(labels.values()).most_common()[-1][1]
# The number of examples per class for the validation set
n_valid_per_label = max(1, math.floor(n * valid_ratio))
label_count = {}
for train_file in os.listdir(os.path.join(data_dir, 'train')):
label = labels[train_file.split('.')[0]]
fname = os.path.join(data_dir, 'train', train_file)
# Copy to train_valid_test/train_valid with a subfolder per class
copyfile(fname, os.path.join(data_dir, 'train_valid_test',
'train_valid', label))

if label not in label_count or label_count[label] < n_valid_per_label:
# Copy to train_valid_test/valid
'valid', label))
label_count[label] = label_count.get(label, 0) + 1
else:
# Copy to train_valid_test/train
'train', label))
return n_valid_per_label
The reorg_test function below is used to organize the testing set to facilitate the reading during
prediction.
#@save
def reorg_test(data_dir):
for test_file in os.listdir(os.path.join(data_dir, 'test')):
copyfile(os.path.join(data_dir, 'test', test_file),
os.path.join(data_dir, 'train_valid_test', 'test',
'unknown'))
Finally, we use a function to call the previously defined read_csv_labels, reorg_train_valid, and
reorg_test functions.
def reorg_cifar10_data(data_dir, valid_ratio):

labels = read_csv_labels(os.path.join(data_dir, 'trainLabels.csv'))
reorg_train_valid(data_dir, labels, valid_ratio)
reorg_test(data_dir)
We only set the batch size to 4 for the demo dataset. During actual training and testing, the com-
plete dataset of the Kaggle competition should be used and batch_size should be set to a larger
integer, such as 128. We use 10% of the training examples as the validation set for tuning hyper-
parameters.
batch_size = 4 if demo else 128

valid_ratio = 0.1
reorg_cifar10_data(data_dir, valid_ratio)
13.13.2 Image Augmentation
To cope with overfitting, we use image augmentation. For example, by adding transforms.
RandomFlipLeftRight(), the images can be flipped at random. We can also perform normalization
for the three RGB channels of color images using transforms.Normalize(). Below, we list some
of these operations that you can choose to use or modify depending on requirements.
transform_train = gluon.data.vision.transforms.Compose([
# Magnify the image to a square of 40 pixels in both height and width
# Randomly crop a square image of 40 pixels in both height and width to
# produce a small square of 0.64 to 1 times the area of the original

# image, and then shrink it to a square of 32 pixels in both height and
# width
gluon.data.vision.transforms.RandomResizedCrop(32, scale=(0.64, 1.0),
ratio=(1.0, 1.0)),
# Normalize each channel of the image
gluon.data.vision.transforms.Normalize([0.4914, 0.4822, 0.4465],
[0.2023, 0.1994, 0.2010])])
In order to ensure the certainty of the output during testing, we only perform normalization on
the image.
transform_test = gluon.data.vision.transforms.Compose([
[0.2023, 0.1994, 0.2010])])
Next, we can create the ImageFolderDataset instance to read the organized dataset containing the
original image files, where each example includes the image and label.
train_ds, valid_ds, train_valid_ds, test_ds = [

gluon.data.vision.ImageFolderDataset(
os.path.join(data_dir, 'train_valid_test', folder))
for folder in ['train', 'valid', 'train_valid', 'test']]
We specify the defined image augmentation operation in DataLoader. During training, we only use
the validation set to evaluate the model, so we need to ensure the certainty of the output. During
prediction, we will train the model on the combined training set and validation set to make full
use of all labelled data.
train_iter, train_valid_iter = [gluon.data.DataLoader(

dataset.transform_first(transform_train), batch_size, shuffle=True,
last_batch='discard') for dataset in (train_ds, train_valid_ds)]
valid_iter = gluon.data.DataLoader(
valid_ds.transform_first(transform_test), batch_size, shuffle=False,
last_batch='discard')
test_ds.transform_first(transform_test), batch_size, shuffle=False,
last_batch='keep')

Here, we build the residual blocks based on the HybridBlock class, which is slightly different than
the implementation described in Section 7.6. This is done to improve execution efficiency.
class Residual(nn.HybridBlock):
def __init__(self, num_channels, use_1x1conv=False, strides=1, **kwargs):
super(Residual, self).__init__(**kwargs)
self.conv1 = nn.Conv2D(num_channels, kernel_size=3, padding=1,
strides=strides)
self.conv2 = nn.Conv2D(num_channels, kernel_size=3, padding=1)
if use_1x1conv:
self.conv3 = nn.Conv2D(num_channels, kernel_size=1,
strides=strides)
else:
self.conv3 = None
def hybrid_forward(self, F, X):

Y = F.npx.relu(self.bn1(self.conv1(X)))
Y = self.bn2(self.conv2(Y))
if self.conv3:
X = self.conv3(X)
return F.npx.relu(Y + X)
Next, we define the ResNet-18 model.
def resnet18(num_classes):
nn.BatchNorm(), nn.Activation('relu'))

blk = nn.HybridSequential()
blk.add(Residual(num_channels, use_1x1conv=True, strides=2))
else:
blk.add(Residual(num_channels))
return blk
net.add(nn.GlobalAvgPool2D(), nn.Dense(num_classes))
return net
The CIFAR-10 image classification challenge uses 10 categories. We will perform Xavier random
initialization on the model before training begins.
def get_net(devices):
num_classes = 10
net = resnet18(num_classes)

net.initialize(ctx=devices, init=init.Xavier())
return net
13.13.5 Defining the Training Functions
We will select the model and tune hyperparameters according to the modelʼs performance on the
validation set. Next, we define the model training function train. We record the training time of
each epoch, which helps us compare the time costs of different models.
def train(net, train_iter, valid_iter, num_epochs, lr, wd, devices, lr_period,

lr_decay):
{'learning_rate': lr, 'momentum': 0.9, 'wd': wd})
legend=['train loss', 'train acc', 'valid acc'])
if epoch > 0 and epoch % lr_period == 0:
trainer.set_learning_rate(trainer.learning_rate * lr_decay)
timer.start()
l, acc = d2l.train_batch_ch13(
net, features, labels.astype('float32'), loss, trainer,
devices, d2l.split_batch)
metric.add(l, acc, labels.shape[0])
timer.stop()
if (i + 1) % (num_batches // 5) == 0:
(metric[0] / metric[2], metric[1] / metric[2],
None))
if valid_iter is not None:
valid_acc = d2l.evaluate_accuracy_gpus(net, valid_iter, d2l.split_batch)
animator.add(epoch + 1, (None, None, valid_acc))
print(f'loss {metric[0] / metric[2]:.3f}, '
f'train acc {metric[1] / metric[2]:.3f}, '
f'valid acc {valid_acc:.3f}')
else:
f'train acc {metric[1] / metric[2]:.3f}')

13.13.6 Training and Validating the Model
Now, we can train and validate the model. The following hyperparameters can be tuned. For
example, we can increase the number of epochs. Because lr_period and lr_decay are set to 80
and 0.1 respectively, the learning rate of the optimization algorithm will be multiplied by 0.1 after
every 80 epochs. For simplicity, we only train one epoch here.
devices, num_epochs, lr, wd = d2l.try_all_gpus(), 5, 0.1, 5e-4

lr_period, lr_decay, net = 50, 0.1, get_net(devices)
net.hybridize()
train(net, train_iter, valid_iter, num_epochs, lr, wd, devices, lr_period,
lr_decay)
loss 2.342, train acc 0.113, valid acc 0.113

13.13.7 Classifying the Testing Set and Submitting Results on Kaggle
After obtaining a satisfactory model design and hyperparameters, we use all training datasets (in-
cluding validation sets) to retrain the model and classify the testing set.
net, preds = get_net(devices), []

net.hybridize()
train(net, train_valid_iter, None, num_epochs, lr, wd, devices, lr_period,
lr_decay)
for X, _ in test_iter:
y_hat = net(X.as_in_ctx(devices[0]))
preds.extend(y_hat.argmax(axis=1).astype(int).asnumpy())
sorted_ids = list(range(1, len(test_ds) + 1))
sorted_ids.sort(key=lambda x: str(x))
df = pd.DataFrame({'id': sorted_ids, 'label': preds})
df['label'] = df['label'].apply(lambda x: train_valid_ds.synsets[x])
df.to_csv('submission.csv', index=False)

loss 2.316, train acc 0.128
After executing the above code, we will get a “submission.csv” file. The format of this file is con-
sistent with the Kaggle competition requirements. The method for submitting results is similar
to method in Section 4.10.
Summary
• We can create an ImageFolderDataset instance to read the dataset containing the original
image files.
• We can use convolutional neural networks, image augmentation, and hybrid programming
to take part in an image classification competition.
Exercises
1. Use the complete CIFAR-10 dataset for the Kaggle competition. Change the batch_size and
number of epochs num_epochs to 128 and 100, respectively. See what accuracy and ranking
you can achieve in this competition.
2. What accuracy can you achieve when not using image augmentation?
3. Scan the QR code to access the relevant discussions and exchange ideas about the meth-
ods used and the results obtained with the community. Can you come up with any better
techniques?
Discussions196
196

13.14 Dog Breed Identification (ImageNet Dogs) on Kaggle
In this section, we will tackle the dog breed identification challenge in the Kaggle Competition.
The competitionʼs web address is
https://www.kaggle.com/c/dog-breed-identification
In this competition, we attempt to identify 120 different breeds of dogs. The dataset used in this
competition is actually a subset of the famous ImageNet dataset. Different from the images in the
CIFAR-10 dataset used in the previous section, the images in the ImageNet dataset are higher and
wider and their dimensions are inconsistent.
Fig. 13.14.1 shows the information on the competitionʼs webpage. In order to submit the results,
please register an account on the Kaggle website first.
Fig. 13.14.1: Dog breed identification competition website. The dataset for the competition can
be accessed by clicking the “Data” tab.
First, import the packages or modules required for the competition.
import collections
import math
from mxnet import autograd, gluon, init, npx
import os
import time
npx.set_np()

13.14.1 Obtaining and Organizing the Dataset
The competition data is divided into a training set and testing set. The training set contains 10, 222
images and the testing set contains 10, 357 images. The images in both sets are in JPEG format.
These images contain three RGB channels (color) and they have different heights and widths.
There are 120 breeds of dogs in the training set, including Labradors, Poodles, Dachshunds,
Samoyeds, Huskies, Chihuahuas, and Yorkshire Terriers.
Downloading the Dataset
After logging in to Kaggle, we can click on the “Data” tab on the dog breed identification com-
petition webpage shown in Fig. 13.14.1 and download the dataset by clicking the “Download All”
button. After unzipping the downloaded file in ../data, you will find the entire dataset in the
following paths:
• ../data/dog-breed-identification/labels.csv
• ../data/dog-breed-identification/sample_submission.csv
• ../data/dog-breed-identification/train
• ../data/dog-breed-identification/test
You may have noticed that the above structure is quite similar to that of the CIFAR-10 competition
in Section 13.13, where folders train/ and test/ contain training and testing dog images respec-
tively, and labels.csv has the labels for the training images.
Similarly, to make it easier to get started, we provide a small-scale sample of the dataset men-
tioned above, “train_valid_test_tiny.zip”. If you are going to use the full dataset for the Kaggle
competition, you will also need to change the demo variable below to False.
#@save
d2l.DATA_HUB['dog_tiny'] = (d2l.DATA_URL + 'kaggle_dog_tiny.zip',
'0cb91d09b814ecdc07b50f31f8dcad3e81d6a86d')
# If you use the full dataset downloaded for the Kaggle competition, change
# the variable below to False
demo = True
if demo:
data_dir = d2l.download_extract('dog_tiny')
else:
data_dir = os.path.join('..', 'data', 'dog-breed-identification')
Downloading ../data/kaggle_dog_tiny.zip from http://d2l-data.s3-accelerate.amazonaws.com/

,→kaggle_dog_tiny.zip...
13.14. Dog Breed Identification (ImageNet Dogs) on Kaggle 623

Organizing the Dataset
We can organize the dataset similarly to what we did in Section 13.13, namely separating a valida-
tion set from the training set, and moving images into subfolders grouped by labels.
The reorg_dog_data function below is used to read the training data labels, segment the validation
set, and organize the training set.
def reorg_dog_data(data_dir, valid_ratio):

labels = d2l.read_csv_labels(os.path.join(data_dir, 'labels.csv'))
d2l.reorg_train_valid(data_dir, labels, valid_ratio)
d2l.reorg_test(data_dir)
batch_size = 4 if demo else 128

valid_ratio = 0.1
reorg_dog_data(data_dir, valid_ratio)
13.14.2 Image Augmentation
The size of the images in this section are larger than the images in the previous section. Here are
some more image augmentation operations that might be useful.
transform_train = gluon.data.vision.transforms.Compose([
# Randomly crop the image to obtain an image with an area of 0.08 to 1 of
# the original area and height to width ratio between 3/4 and 4/3. Then,
# scale the image to create a new image with a height and width of 224
# pixels each
gluon.data.vision.transforms.RandomResizedCrop(224, scale=(0.08, 1.0),
ratio=(3.0/4.0, 4.0/3.0)),
# Randomly change the brightness, contrast, and saturation
gluon.data.vision.transforms.RandomColorJitter(brightness=0.4,
contrast=0.4,
saturation=0.4),
# Add random noise
gluon.data.vision.transforms.RandomLighting(0.1),
# Standardize each channel of the image
[0.229, 0.224, 0.225])])
During testing, we only use definite image preprocessing operations.
transform_test = gluon.data.vision.transforms.Compose([
# Crop a square of 224 by 224 from the center of the image
gluon.data.vision.transforms.CenterCrop(224),
[0.229, 0.224, 0.225])])

As in the previous section, we can create an ImageFolderDataset instance to read the dataset con-
taining the original image files.
train_ds, valid_ds, train_valid_ds, test_ds = [

gluon.data.vision.ImageFolderDataset(
os.path.join(data_dir, 'train_valid_test', folder))
for folder in ('train', 'valid', 'train_valid', 'test')]
Here, we create DataLoader instances, just like in Section 13.13.
train_iter, train_valid_iter = [gluon.data.DataLoader(

dataset.transform_first(transform_train), batch_size, shuffle=True,
last_batch='discard') for dataset in (train_ds, train_valid_ds)]
valid_iter = gluon.data.DataLoader(
valid_ds.transform_first(transform_test), batch_size, shuffle=False,
last_batch='discard')
test_ds.transform_first(transform_test), batch_size, shuffle=False,
last_batch='keep')
The dataset for this competition is a subset of the ImageNet data set. Therefore, we can use the
approach discussed in Section 13.2 to select a model pre-trained on the entire ImageNet dataset
and use it to extract image features to be input in the custom small-scale output network. Gluon
provides a wide range of pre-trained models. Here, we will use the pre-trained ResNet-34 model.
Because the competition dataset is a subset of the pre-training dataset, we simply reuse the in-
put of the pre-trained modelʼs output layer, i.e., the extracted features. Then, we can replace the
original output layer with a small custom output network that can be trained, such as two fully
connected layers in a series. Different from the experiment in Section 13.2, here, we do not re-
train the pre-trained model used for feature extraction. This reduces the training time and the
memory required to store model parameter gradients.
You must note that, during image augmentation, we use the mean values and standard deviations
of the three RGB channels for the entire ImageNet dataset for normalization. This is consistent
with the normalization of the pre-trained model.
def get_net(devices):
finetune_net = gluon.model_zoo.vision.resnet34_v2(pretrained=True)
# Define a new output network
finetune_net.output_new = nn.HybridSequential(prefix='')
finetune_net.output_new.add(nn.Dense(256, activation='relu'))
# There are 120 output categories
finetune_net.output_new.add(nn.Dense(120))
# Initialize the output network
finetune_net.output_new.initialize(init.Xavier(), ctx=devices)
# Distribute the model parameters to the CPUs or GPUs used for computation
finetune_net.collect_params().reset_ctx(devices)
return finetune_net

When calculating the loss, we first use the member variable features to obtain the input of the
pre-trained modelʼs output layer, i.e., the extracted feature. Then, we use this feature as the input
for our small custom output network and compute the output.
def evaluate_loss(data_iter, net, devices):

l_sum, n = 0.0, 0
for features, labels in data_iter:
X_shards, y_shards = d2l.split_batch(features, labels, devices)
output_features = [net.features(X_shard) for X_shard in X_shards]
outputs = [net.output_new(feature) for feature in output_features]
ls = [loss(output, y_shard).sum() for output, y_shard
in zip(outputs, y_shards)]
l_sum += sum([float(l.sum()) for l in ls])
n += labels.size
return l_sum / n
13.14.5 Defining the Training Functions
We will select the model and tune hyperparameters according to the modelʼs performance on the
validation set. The model training function train only trains the small custom output network.
def train(net, train_iter, valid_iter, num_epochs, lr, wd, devices, lr_period,

lr_decay):
# Only train the small custom output network
trainer = gluon.Trainer(net.output_new.collect_params(), 'sgd',
{'learning_rate': lr, 'momentum': 0.9, 'wd': wd})
legend=['train loss', 'valid loss'])
if epoch > 0 and epoch % lr_period == 0:
trainer.set_learning_rate(trainer.learning_rate * lr_decay)
timer.start()
X_shards, y_shards = d2l.split_batch(features, labels, devices)
output_features = [net.features(X_shard) for X_shard in X_shards]
outputs = [net.output_new(feature) for feature in output_features]
ls = [loss(output, y_shard).sum() for output, y_shard
in zip(outputs, y_shards)]
for l in ls:
l.backward()
metric.add(sum([float(l.sum()) for l in ls]), labels.shape[0])
timer.stop()
if (i + 1) % (num_batches // 5) == 0:
(metric[0] / metric[1], None))
valid_loss = evaluate_loss(valid_iter, net, devices)
animator.add(epoch + 1, (None, valid_loss))

print(f'train loss {metric[0] / metric[1]:.3f}, '
f'valid loss {valid_loss:.3f}')
else:
print(f'train loss {metric[0] / metric[1]:.3f}')
13.14.6 Training and Validating the Model
Now, we can train and validate the model. The following hyperparameters can be tuned. For
example, we can increase the number of epochs. Because lr_period and lr_decay are set to 10
and 0.1 respectively, the learning rate of the optimization algorithm will be multiplied by 0.1 after
every 10 epochs.
devices, num_epochs, lr, wd = d2l.try_all_gpus(), 5, 0.01, 1e-4

lr_period, lr_decay, net = 10, 0.1, get_net(devices)
net.hybridize()
train(net, train_iter, valid_iter, num_epochs, lr, wd, devices, lr_period,
lr_decay)
train loss 2.338, valid loss 2.340

13.14.7 Classifying the Testing Set and Submitting Results on Kaggle
After obtaining a satisfactory model design and hyperparameters, we use all training datasets (in-
cluding validation sets) to retrain the model and then classify the testing set. Note that predictions
are made by the output network we just trained.
net = get_net(devices)
net.hybridize()
train(net, train_valid_iter, None, num_epochs, lr, wd, devices, lr_period,

lr_decay)
preds = []
for data, label in test_iter:
output_features = net.features(data.as_in_ctx(devices[0]))
output = npx.softmax(net.output_new(output_features))
preds.extend(output.asnumpy())
ids = sorted(os.listdir(
os.path.join(data_dir, 'train_valid_test', 'test', 'unknown')))
with open('submission.csv', 'w') as f:
f.write('id,' + ','.join(train_valid_ds.synsets) + '\n')
for i, output in zip(ids, preds):
f.write(i.split('.')[0] + ',' + ','.join(
[str(num) for num in output]) + '\n')
train loss 2.446

After executing the above code, we will generate a “submission.csv” file. The format of this file
is consistent with the Kaggle competition requirements. The method for submitting results is
similar to method in Section 4.10.

Summary
• We can use a model pre-trained on the ImageNet dataset to extract features and only train a
small custom output network. This will allow us to classify a subset of the ImageNet dataset
with lower computing and storage overhead.
Exercises
1. When using the entire Kaggle dataset, what kind of results do you get when you increase the
batch_size (batch size) and num_epochs (number of epochs)?
2. Do you get better results if you use a deeper pre-trained model?
3. Scan the QR code to access the relevant discussions and exchange ideas about the meth-
ods used and the results obtained with the community. Can you come up with any better
techniques?
Discussions197
197

14 | Natural Language Processing: Pre-
training
Humans need to communicate. Out of this basic need of the human condition, a vast amount of
written text has been generated on an everyday basis. Given rich text in social media, chat apps,
emails, product reviews, news articles, research papers, and books, it becomes vital to enable
computers to understand them to offer assistance or make decisions based on human languages.
Natural language processing studies interactions between computers and humans using natural
languages. In practice, it is very common to use natural language processing techniques to pro-
cess and analyze text (human natural language) data, such as language models in Section 8.3 and
machine translation models in Section 9.5.
To understand text, we can begin with its representation, such as treating each word or subword
as an individual text token. As we will see in this chapter, the representation of each token can
be pretrained on a large corpus, using word2vec, GloVe, or subword embedding models. After
pretraining, representation of each token can be a vector, however, it remains the same no matter
what the context is. For instance, the vector representation of “bank” is the same in both “go
to the bank to deposit some money” and “go to the bank to sit down”. Thus, many more recent
pretraining models adapt representation of the same token to different contexts. Among them is
BERT, a much deeper model based on the Transformer encoder. In this chapter, we will focus on
how to pretrain such representations for text, as highlighted in Fig. 14.1.
Fig. 14.1: Pretrained text representations can be fed to various deep learning architectures for
different downstream natural language processing applications. This chapter focuses on the up-
stream text representation pretraining.
631
As shown in Fig. 14.1, the pretrained text representations can be fed to a variety of deep learning
architectures for different downstream natural language processing applications. We will cover
them in Chapter 15.
14.1 Word Embedding (word2vec)
A natural language is a complex system that we use to express meanings. In this system, words
are the basic unit of linguistic meaning. As its name implies, a word vector is a vector used to
represent a word. It can also be thought of as the feature vector of a word. The technique of
mapping words to vectors of real numbers is also known as word embedding. Over the last few
years, word embedding has gradually become basic knowledge in natural language processing.
14.1.1 Why Not Use One-hot Vectors?
We used one-hot vectors to represent words (characters are words) in Section 8.5 . Recall that
when we assume the number of different words in a dictionary (the dictionary size) is N , each
word can correspond one-to-one with consecutive integers from 0 to N − 1. These integers that
correspond to words are called the indices of the words. We assume that the index of a word is i.
In order to get the one-hot vector representation of the word, we create a vector of all 0s with a
length of N and set element i to 1. In this way, each word is represented as a vector of length N
that can be used directly by the neural network.
Although one-hot word vectors are easy to construct, they are usually not a good choice. One of the
major reasons is that the one-hot word vectors cannot accurately express the similarity between
different words, such as the cosine similarity that we commonly use. For the vectors x, y ∈ Rd ,
their cosine similarities are the cosines of the angles between them:
x⊤ y
∈ [−1, 1]. (14.1.1)
∥x∥∥y∥
Since the cosine similarity between the one-hot vectors of any two different words is 0, it is difficult
to use the one-hot vector to accurately represent the similarity between multiple different words.
Word2vec198 is a tool that we came up with to solve the problem above. It represents each word
with a fixed-length vector and uses these vectors to better indicate the similarity and analogy rela-
tionships between different words. The Word2vec tool contains two models: skip-gram (Mikolov
et al., 2013b) and continuous bag of words (CBOW) (Mikolov et al., 2013a). Next, we will take a
look at the two models and their training methods.
14.1.2 The Skip-Gram Model
The skip-gram model assumes that a word can be used to generate the words that surround it in a
text sequence. For example, we assume that the text sequence is “the”, “man”, “loves”, “his”, and
“son”. We use “loves” as the central target word and set the context window size to 2. As shown
in Fig. 14.1.1, given the central target word “loves”, the skip-gram model is concerned with the
conditional probability for generating the context words, “the”, “man”, “his” and “son”, that are
within a distance of no more than 2 words, which is
P ("the", "man", "his", "son" | "loves"). (14.1.2)

198
https://code.google.com/archive/p/word2vec/
632 Chapter 14. Natural Language Processing: Pretraining

We assume that, given the central target word, the context words are generated independently of
each other. In this case, the formula above can be rewritten as
P ("the" | "loves") · P ("man" | "loves") · P ("his" | "loves") · P ("son" | "loves"). (14.1.3)
Fig. 14.1.1: The skip-gram model cares about the conditional probability of generating context
words for a given central target word.
In the skip-gram model, each word is represented as two d-dimension vectors, which are used to
compute the conditional probability. We assume that the word is indexed as i in the dictionary, its
vector is represented as vi ∈ Rd when it is the central target word, and ui ∈ Rd when it is a context
word. Let the central target word wc and context word wo be indexed as c and o respectively in the
dictionary. The conditional probability of generating the context word for the given central target
word can be obtained by performing a softmax operation on the vector inner product:
exp(u⊤
o vc )
P (wo | wc ) = ∑ ⊤
, (14.1.4)
i∈V exp(ui vc )
where vocabulary index set V = {0, 1, . . . , |V|−1}. Assume that a text sequence of length T is given,
where the word at timestep t is denoted as w(t) . Assume that context words are independently
generated given center words. When context window size is m, the likelihood function of the
skip-gram model is the joint probability of generating all the context words given any center word
∏
T ∏
P (w(t+j) | w(t) ), (14.1.5)
t=1 −m≤j≤m, j̸=0
Here, any timestep that is less than 1 or greater than T can be ignored.
Skip-Gram Model Training
The skip-gram model parameters are the central target word vector and context word vector for
each individual word. In the training process, we are going to learn the model parameters by
maximizing the likelihood function, which is also known as maximum likelihood estimation. This
is equivalent to minimizing the following loss function:
∑
T ∑
− log P (w(t+j) | w(t) ). (14.1.6)
t=1 −m≤j≤m, j̸=0
If we use the SGD, in each iteration we are going to pick a shorter subsequence through random
sampling to compute the loss for that subsequence, and then compute the gradient to update the
14.1. Word Embedding (word2vec) 633

model parameters. The key of gradient computation is to compute the gradient of the logarithmic
conditional probability for the central word vector and the context word vector. By definition, we
first have
( )
∑
log P (wo | wc ) = u⊤
o vc − log exp(u⊤
i vc ) . (14.1.7)
i∈V
Through differentiation, we can get the gradient vc from the formula above.
∑ ⊤
∂log P (wo | wc ) j∈V exp(uj vc )uj
= uo − ∑ ⊤
∂vc i∈V exp(ui vc )
( )
∑ exp(u⊤j v c )
= uo − ∑ ⊤v )
uj (14.1.8)
j∈V i∈V exp(u i c
∑
= uo − P (wj | wc )uj .
j∈V
Its computation obtains the conditional probability for all the words in the dictionary given the
central target word wc . We then use the same method to obtain the gradients for other word vec-
tors.
After the training, for any word in the dictionary with index i, we are going to get its two word
vector sets vi and ui . In applications of natural language processing, the central target word vector
in the skip-gram model is generally used as the representation vector of a word.
14.1.3 The Continuous Bag of Words (CBOW) Model
The continuous bag of words (CBOW) model is similar to the skip-gram model. The biggest dif-
ference is that the CBOW model assumes that the central target word is generated based on the
context words before and after it in the text sequence. With the same text sequence “the”, “man”,
“loves”, “his” and “son”, in which “loves” is the central target word, given a context window size
of 2, the CBOW model is concerned with the conditional probability of generating the target word
“loves” based on the context words “the”, “man”, “his” and “son”(as shown in Fig. 14.1.2), such as
P ("loves" | "the", "man", "his", "son"). (14.1.9)
Fig. 14.1.2: The CBOW model cares about the conditional probability of generating the central
target word from given context words.
Since there are multiple context words in the CBOW model, we will average their word vectors
and then use the same method as the skip-gram model to compute the conditional probability.

We assume that vi ∈ Rd and ui ∈ Rd are the context word vector and central target word vector
of the word with index i in the dictionary (notice that the symbols are opposite to the ones in the
skip-gram model). Let central target word wc be indexed as c, and context words wo1 , . . . , wo2m be
indexed as o1 , . . . , o2m in the dictionary. Thus, the conditional probability of generating a central
target word from the given context word is
( 1 ⊤ )
exp 2m uc (vo1 + . . . , +vo2m )
P (wc | wo1 , . . . , wo2m ) = ∑ ( 1 ⊤ ). (14.1.10)
i∈V exp 2m ui (vo1 + . . . , +vo2m )
For brevity, denote Wo = {wo1 , . . . , wo2m }, and v̄o = (vo1 + . . . , +vo2m ) /(2m). The equation above
can be simplified as
( )
exp u⊤c v̄o
P (wc | Wo ) = ∑ ( ⊤ ). (14.1.11)
i∈V exp ui v̄o
Given a text sequence of length T , we assume that the word at timestep t is w(t) , and the context
window size is m. The likelihood function of the CBOW model is the probability of generating any
central target word from the context words.
∏
T
P (w(t) | w(t−m) , . . . , w(t−1) , w(t+1) , . . . , w(t+m) ). (14.1.12)
t=1
CBOW Model Training
CBOW model training is quite similar to skip-gram model training. The maximum likelihood es-
timation of the CBOW model is equivalent to minimizing the loss function.
∑
T
− log P (w(t) | w(t−m) , . . . , w(t−1) , w(t+1) , . . . , w(t+m) ). (14.1.13)
t=1
Notice that
( )
∑ ( )
log P (wc | Wo ) = u⊤
c v̄o − log exp u⊤
i v̄o . (14.1.14)
i∈V
Through differentiation, we can compute the logarithm of the conditional probability of the gra-
dient of any context word vector voi (i = 1, . . . , 2m) in the formula above.
   
∂ log P (wc | Wo ) 1  ∑ exp(u⊤ v̄ o )u j 1 ∑
= uc − ∑
j = uc − P (wj | Wo )uj  .
∂voi 2m exp(u ⊤ v̄ ) 2m
i∈V
j∈V i o j∈V
(14.1.15)
We then use the same method to obtain the gradients for other word vectors. Unlike the skip-gram
model, we usually use the context word vector as the representation vector for a word in the CBOW
model.
14.1. Word Embedding (word2vec) 635

Summary
• A word vector is a vector used to represent a word. The technique of mapping words to
vectors of real numbers is also known as word embedding.
• Word2vec includes both the continuous bag of words (CBOW) and skip-gram models. The
skip-gram model assumes that context words are generated based on the central target word.
The CBOW model assumes that the central target word is generated based on the context
words.
Exercises
1. What is the computational complexity of each gradient? If the dictionary contains a large
volume of words, what problems will this cause?
2. There are some fixed phrases in the English language which consist of multiple words, such
as “new york”. How can you train their word vectors? Hint: See section 4 in the Word2vec
paper[2].
3. Use the skip-gram model as an example to think about the design of a word2vec model. What
is the relationship between the inner product of two word vectors and the cosine similarity
in the skip-gram model? For a pair of words with close semantical meaning, why it is likely
for their word vector cosine similarity to be high?
Discussions199
14.2 Approximate Training
Recall content of the last section. The core feature of the skip-gram model is the use of softmax
operations to compute the conditional probability of generating context word wo based on the
given central target word wc .
exp(u⊤
o vc )
P (wo | wc ) = ∑ ⊤
. (14.2.1)
i∈V exp(ui vc )
The logarithmic loss corresponding to the conditional probability is given as

( )
∑
⊤ ⊤
− log P (wo | wc ) = −uo vc + log exp(ui vc ) . (14.2.2)
i∈V
Because the softmax operation has considered that the context word could be any word in the
dictionary V, the loss mentioned above actually includes the sum of the number of items in the
dictionary size. From the last section, we know that for both the skip-gram model and CBOW
model, because they both get the conditional probability using a softmax operation, the gradient
computation for each step contains the sum of the number of items in the dictionary size. For
larger dictionaries with hundreds of thousands or even millions of words, the overhead for com-
puting each gradient may be too high. In order to reduce such computational complexity, we will
introduce two approximate training methods in this section: negative sampling and hierarchical
softmax. Since there is no major difference between the skip-gram model and the CBOW model,
we will only use the skip-gram model as an example to introduce these two training methods in
this section.
199

14.2.1 Negative Sampling
Negative sampling modifies the original objective function. Given a context window for the central
target word wc , we will treat it as an event for context word wo to appear in the context window
and compute the probability of this event from
P (D = 1 | wc , wo ) = σ(u⊤
o vc ), (14.2.3)
Here, the σ function has the same definition as the sigmoid activation function:
1
σ(x) = . (14.2.4)
1 + exp(−x)
We will first consider training the word vector by maximizing the joint probability of all events in
the text sequence. Given a text sequence of length T , we assume that the word at timestep t is w(t)
and the context window size is m. Now we consider maximizing the joint probability
∏
T ∏
P (D = 1 | w(t) , w(t+j) ). (14.2.5)
t=1 −m≤j≤m, j̸=0
However, the events included in the model only consider positive examples. In this case, only
when all the word vectors are equal and their values approach infinity can the joint probabil-
ity above be maximized to 1. Obviously, such word vectors are meaningless. Negative sampling
makes the objective function more meaningful by sampling with an addition of negative exam-
ples. Assume that event P occurs when context word wo appears in the context window of central
target word wc , and we sample K words that do not appear in the context window according to
the distribution P (w) to act as noise words. We assume the event for noise word wk (k = 1, . . . , K)
to not appear in the context window of central target word wc is Nk . Suppose that events P and
N1 , . . . , NK for both positive and negative examples are independent of each other. By considering
negative sampling, we can rewrite the joint probability above, which only considers the positive
examples, as
∏
T ∏
P (w(t+j) | w(t) ), (14.2.6)
t=1 −m≤j≤m, j̸=0
Here, the conditional probability is approximated to be

∏
K
P (w (t+j)
| w ) = P (D = 1 | w , w
(t) (t) (t+j)
) P (D = 0 | w(t) , wk ). (14.2.7)
k=1, wk ∼P (w)
Let the text sequence index of word w(t) at timestep t be it and hk for noise word wk in the dictio-
nary. The logarithmic loss for the conditional probability above is
∑
K
− log P (w (t+j)
| w ) = − log P (D = 1 | w , w
(t) (t) (t+j)
)− log P (D = 0 | w(t) , wk )
k=1, wk ∼P (w)
( ) ∑
K ( ( ))
= − log σ u⊤
it+j t −
v i log 1 − σ u⊤ v
hk ti
k=1, wk ∼P (w)
( ) ∑
K ( )
= − log σ u⊤
it+j it −
v log σ −u⊤ v
hk it .
k=1, wk ∼P (w)
(14.2.8)
Here, the gradient computation in each step of the training is no longer related to the dictionary
size, but linearly related to K. When K takes a smaller constant, the negative sampling has a lower
computational overhead for each step.
14.2. Approximate Training 637

14.2.2 Hierarchical Softmax
Hierarchical softmax is another type of approximate training method. It uses a binary tree for
data structure as illustrated in Fig. 14.2.1, with the leaf nodes of the tree representing every word
in the dictionary V.
Fig. 14.2.1: Hierarchical Softmax. Each leaf node of the tree represents a word in the dictionary.
We assume that L(w) is the number of nodes on the path (including the root and leaf nodes) from
the root node of the binary tree to the leaf node of word w. Let n(w, j) be the j th node on this path,
with the context word vector un(w,j) . We use Figure 10.3 as an example, so L(w3 ) = 4. Hierarchical
softmax will approximate the conditional probability in the skip-gram model as
∏
L(wo )−1 ( )
P (wo | wc ) = σ [[n(wo , j + 1) = leftChild(n(wo , j))]] · u⊤
n(wo ,j) v c , (14.2.9)
j=1
Here the σ function has the same definition as the sigmoid activation function, and leftChild(n)
is the left child node of node n. If x is true, [[x]] = 1; otherwise [[x]] = −1. Now, we will compute
the conditional probability of generating word w3 based on the given word wc in Figure 10.3. We
need to find the inner product of word vector vc (for word wc ) and each non-leaf node vector on
the path from the root node to w3 . Because, in the binary tree, the path from the root node to leaf
node w3 needs to be traversed left, right, and left again (the path with the bold line in Figure 10.3),
we get
P (w3 | wc ) = σ(u⊤ ⊤ ⊤
n(w3 ,1) vc ) · σ(−un(w3 ,2) vc ) · σ(un(w3 ,3) vc ). (14.2.10)
Because σ(x) + σ(−x) = 1, the condition that the sum of the conditional probability of any word
generated based on the given central target word wc in dictionary V be 1 will also suffice:
∑
P (w | wc ) = 1. (14.2.11)
w∈V
In addition, because the order of magnitude for L(wo )−1 is O(log2 |V|), when the size of dictionary
V is large, the computational overhead for each step in the hierarchical softmax training is greatly
reduced compared to situations where we do not use approximate training.

Summary
• Negative sampling constructs the loss function by considering independent events that con-
tain both positive and negative examples. The gradient computational overhead for each
step in the training process is linearly related to the number of noise words we sample.
• Hierarchical softmax uses a binary tree and constructs the loss function based on the path
from the root node to the leaf node. The gradient computational overhead for each step in
the training process is related to the logarithm of the dictionary size.
Exercises
1. Before reading the next section, think about how we should sample noise words in negative
sampling.
2. What makes the last formula in this section hold?
3. How can we apply negative sampling and hierarchical softmax in the skip-gram model?
Discussions200
14.3 The Dataset for Pretraining Word Embedding
In this section, we will introduce how to preprocess a dataset with negative sampling Section 14.2
and load into minibatches for word2vec training. The dataset we use is Penn Tree Bank (PTB)201 ,
which is a small but commonly-used corpus. It takes samples from Wall Street Journal articles
and includes training sets, validation sets, and test sets.
First, import the packages and modules required for the experiment.

import math
from mxnet import gluon, np
import os
import random
14.3.1 Reading and Preprocessing the Dataset
This dataset has already been preprocessed. Each line of the dataset acts as a sentence. All the
words in a sentence are separated by spaces. In the word embedding task, each word is a token.
#@save
d2l.DATA_HUB['ptb'] = (d2l.DATA_URL + 'ptb.zip',
'319d85e578af0cdc590547f26231e4e31cdf1e42')
#@save
def read_ptb():
data_dir = d2l.download_extract('ptb')
200
201
https://catalog.ldc.upenn.edu/LDC99T42
14.3. The Dataset for Pretraining Word Embedding 639

with open(os.path.join(data_dir, 'ptb.train.txt')) as f:
raw_text = f.read()
return [line.split() for line in raw_text.split('\n')]
sentences = read_ptb()
f'# sentences: {len(sentences)}'
Downloading ../data/ptb.zip from http://d2l-data.s3-accelerate.amazonaws.com/ptb.zip...
'# sentences: 42069'
Next we build a vocabulary with words appeared not greater than 10 times mapped into a “<unk>”
token. Note that the preprocessed PTB data also contains “<unk>” tokens presenting rare words.
vocab = d2l.Vocab(sentences, min_freq=10)

f'vocab size: {len(vocab)}'
'vocab size: 6719'
14.3.2 Subsampling
In text data, there are generally some words that appear at high frequencies, such “the”, “a”, and
“in” in English. Generally speaking, in a context window, it is better to train the word embedding
model when a word (such as “chip”) and a lower-frequency word (such as “microprocessor”) ap-
pear at the same time, rather than when a word appears with a higher-frequency word (such as
“the”). Therefore, when training the word embedding model, we can perform subsampling[2] on
the words. Specifically, each indexed word wi in the dataset will drop out at a certain probability.
The dropout probability is given as:
( √ )
t
P (wi ) = max 1 − ,0 , (14.3.1)
f (wi )
Here, f (wi ) is the ratio of the instances of word wi to the total number of words in the dataset,
and the constant t is a hyperparameter (set to 10−4 in this experiment). As we can see, it is only
possible to drop out the word wi in subsampling when f (wi ) > t. The higher the wordʼs frequency,
the higher its dropout probability.
#@save
def subsampling(sentences, vocab):
# Map low frequency words into <unk>
sentences = [[vocab.idx_to_token[vocab[tk]] for tk in line]
for line in sentences]
# Count the frequency for each word
counter = d2l.count_corpus(sentences)
num_tokens = sum(counter.values())
# Return True if to keep this token during subsampling

def keep(token):

return(random.uniform(0, 1) <
math.sqrt(1e-4 / counter[token] * num_tokens))
# Now do the subsampling

return [[tk for tk in line if keep(tk)] for line in sentences]
subsampled = subsampling(sentences, vocab)
Compare the sequence lengths before and after sampling, we can see subsampling significantly
reduced the sequence length.
d2l.set_figsize()
d2l.plt.hist([[len(line) for line in sentences],
[len(line) for line in subsampled]])
d2l.plt.xlabel('# tokens per sentence')
d2l.plt.ylabel('count')
d2l.plt.legend(['origin', 'subsampled']);
For individual tokens, the sampling rate of the high-frequency word “the” is less than 1/20.
def compare_counts(token):
return (f'# of "{token}": '
f'before={sum([line.count(token) for line in sentences])}, '
f'after={sum([line.count(token) for line in subsampled])}')
compare_counts('the')
'# of "the": before=50770, after=2083'
But the low-frequency word “join” is completely preserved.
compare_counts('join')
'# of "join": before=45, after=45'
Last, we map each token into an index to construct the corpus.

corpus = [vocab[line] for line in subsampled]
corpus[0:3]
[[0], [392, 2132, 406], [5464, 3080, 1595, 95]]
14.3.3 Loading the Dataset
Next we read the corpus with token indicies into data batches for training.
Extracting Central Target Words and Context Words
We use words with a distance from the central target word not exceeding the context window size
as the context words of the given center target word. The following definition function extracts all
the central target words and their context words. It uniformly and randomly samples an integer to
be used as the context window size between integer 1 and the max_window_size (maximum context
window).
#@save
def get_centers_and_contexts(corpus, max_window_size):
centers, contexts = [], []
for line in corpus:
# Each sentence needs at least 2 words to form a "central target word
# - context word" pair
if len(line) < 2:
continue
centers += line
for i in range(len(line)): # Context window centered at i
window_size = random.randint(1, max_window_size)
indices = list(range(max(0, i - window_size),
min(len(line), i + 1 + window_size)))
# Exclude the central target word from the context words
indices.remove(i)
contexts.append([line[idx] for idx in indices])
return centers, contexts
Next, we create an artificial dataset containing two sentences of 7 and 3 words, respectively. As-
sume the maximum context window is 2 and print all the central target words and their context
words.
tiny_dataset = [list(range(7)), list(range(7, 10))]

print('dataset', tiny_dataset)
for center, context in zip(*get_centers_and_contexts(tiny_dataset, 2)):
print('center', center, 'has contexts', context)
dataset [[0, 1, 2, 3, 4, 5, 6], [7, 8, 9]]

center 0 has contexts [1, 2]
center 1 has contexts [0, 2, 3]
center 3 has contexts [1, 2, 4, 5]

center 5 has contexts [3, 4, 6]
center 7 has contexts [8]
We set the maximum context window size to 5. The following extracts all the central target words
and their context words in the dataset.
all_centers, all_contexts = get_centers_and_contexts(corpus, 5)

f'# center-context pairs: {len(all_centers)}'
'# center-context pairs: 353005'
Negative Sampling
We use negative sampling for approximate training. For a central and context word pair, we ran-
domly sample K noise words (K = 5 in the experiment). According to the suggestion in the
Word2vec paper, the noise word sampling probability P (w) is the ratio of the word frequency of
w to the total word frequency raised to the power of 0.75 [2].
We first define a class to draw a candidate according to the sampling weights. It caches a 10000
size random number bank instead of calling random.choices every time.
#@save
class RandomGenerator:
"""Draw a random int in [0, n] according to n sampling weights."""
def __init__(self, sampling_weights):
self.population = list(range(len(sampling_weights)))
self.sampling_weights = sampling_weights
self.candidates = []
self.i = 0
def draw(self):
if self.i == len(self.candidates):
self.candidates = random.choices(
self.population, self.sampling_weights, k=10000)
self.i = 0
self.i += 1
return self.candidates[self.i-1]
generator = RandomGenerator([2, 3, 4])

[generator.draw() for _ in range(10)]
[2, 1, 1, 0, 0, 2, 0, 0, 2, 1]
#@save
def get_negatives(all_contexts, corpus, K):

counter = d2l.count_corpus(corpus)
sampling_weights = [counter[i]**0.75 for i in range(len(counter))]
all_negatives, generator = [], RandomGenerator(sampling_weights)
for contexts in all_contexts:
negatives = []
while len(negatives) < len(contexts) * K:
neg = generator.draw()
# Noise words cannot be context words
if neg not in contexts:
negatives.append(neg)
all_negatives.append(negatives)
return all_negatives
all_negatives = get_negatives(all_contexts, corpus, 5)
Reading into Batches
We extract all central target words all_centers, and the context words all_contexts and noise
words all_negatives of each central target word from the dataset. We will read them in random
minibatches.
In a minibatch of data, the ith example includes a central word and its corresponding ni con-
text words and mi noise words. Since the context window size of each example may be differ-
ent, the sum of context words and noise words, ni + mi , will be different. When constructing a
minibatch, we concatenate the context words and noise words of each example, and add 0s for
padding until the length of the concatenations are the same, that is, the length of all concate-
nations is maxi ni + mi (max_len). In order to avoid the effect of padding on the loss function
calculation, we construct the mask variable masks, each element of which corresponds to an el-
ement in the concatenation of context and noise words, contexts_negatives. When an element
in the variable contexts_negatives is a padding, the element in the mask variable masks at the
same position will be 0. Otherwise, it takes the value 1. In order to distinguish between positive
and negative examples, we also need to distinguish the context words from the noise words in the
contexts_negatives variable. Based on the construction of the mask variable, we only need to
create a label variable labels with the same shape as the contexts_negatives variable and set the
elements corresponding to context words (positive examples) to 1, and the rest to 0.
Next, we will implement the minibatch reading function batchify. Its minibatch input data is a
list whose length is the batch size, each element of which contains central target words center,
context words context, and noise words negative. The minibatch data returned by this function
conforms to the format we need, for example, it includes the mask variable.
#@save
def batchify(data):
max_len = max(len(c) + len(n) for _, c, n in data)
centers, contexts_negatives, masks, labels = [], [], [], []
for center, context, negative in data:
cur_len = len(context) + len(negative)
centers += [center]
contexts_negatives += [context + negative + [0] * (max_len - cur_len)]
masks += [[1] * cur_len + [0] * (max_len - cur_len)]
labels += [[1] * len(context) + [0] * (max_len - len(context))]

return (np.array(centers).reshape(-1, 1), np.array(contexts_negatives),
np.array(masks), np.array(labels))
Construct two simple examples:
x_1 = (1, [2, 2], [3, 3, 3, 3])

x_2 = (1, [2, 2, 2], [3, 3])
batch = batchify((x_1, x_2))
names = ['centers', 'contexts_negatives', 'masks', 'labels']

for name, data in zip(names, batch):
print(name, '=', data)
centers = [[1.]
[1.]]
contexts_negatives = [[2. 2. 3. 3. 3. 3.]
[2. 2. 2. 3. 3. 0.]]
masks = [[1. 1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1. 0.]]
labels = [[1. 1. 0. 0. 0. 0.]
[1. 1. 1. 0. 0. 0.]]
We use the batchify function just defined to specify the minibatch reading method in the Dat-
aLoader instance.
Last, we define the load_data_ptb function that read the PTB dataset and return the data iterator.
#@save
def load_data_ptb(batch_size, max_window_size, num_noise_words):
sentences = read_ptb()
vocab = d2l.Vocab(sentences, min_freq=10)
subsampled = subsampling(sentences, vocab)
corpus = [vocab[line] for line in subsampled]
all_centers, all_contexts = get_centers_and_contexts(
corpus, max_window_size)
all_negatives = get_negatives(all_contexts, corpus, num_noise_words)
dataset = gluon.data.ArrayDataset(
all_centers, all_contexts, all_negatives)
data_iter = gluon.data.DataLoader(dataset, batch_size, shuffle=True,
batchify_fn=batchify,
num_workers=num_workers)
return data_iter, vocab
Let us print the first minibatch of the data iterator.
data_iter, vocab = load_data_ptb(512, 5, 5)

for batch in data_iter:
for name, data in zip(names, batch):

print(name, 'shape:', data.shape)
break
centers shape: (512, 1)

contexts_negatives shape: (512, 60)
masks shape: (512, 60)
labels shape: (512, 60)
Summary
• Subsampling attempts to minimize the impact of high-frequency words on the training of a

word embedding model.
• We can pad examples of different lengths to create minibatches with examples of all the
same length and use mask variables to distinguish between padding and non-padding ele-
ments, so that only non-padding elements participate in the calculation of the loss function.
Exercises
1. We use the batchify function to specify the minibatch reading method in the DataLoader
instance and print the shape of each variable in the first batch read. How should these shapes
be calculated?
Discussions202
14.4 Pretraining word2vec
In this section, we will train a skip-gram model defined in Section 14.1.

First, import the packages and modules required for the experiment, and load the PTB dataset.

npx.set_np()
batch_size, max_window_size, num_noise_words = 512, 5, 5

data_iter, vocab = d2l.load_data_ptb(batch_size, max_window_size,
num_noise_words)
202

14.4.1 The Skip-Gram Model
We will implement the skip-gram model by using embedding layers and minibatch multiplication.
These methods are also often used to implement other natural language processing applications.
Embedding Layer
The layer in which the obtained word is embedded is called the embedding layer, which can be
obtained by creating an nn.Embedding instance in Gluon. The weight of the embedding layer is a
matrix whose number of rows is the dictionary size (input_dim) and whose number of columns
is the dimension of each word vector (output_dim). We set the dictionary size to 20 and the word
vector dimension to 4.
embed = nn.Embedding(input_dim=20, output_dim=4)

embed.initialize()
embed.weight
Parameter embedding0_weight (shape=(20, 4), dtype=float32)
The input of the embedding layer is the index of the word. When we enter the index i of a word,
the embedding layer returns the ith row of the weight matrix as its word vector. Below we enter
an index of shape (2, 3) into the embedding layer. Because the dimension of the word vector is 4,
we obtain a word vector of shape (2, 3, 4).
x = np.array([[1, 2, 3], [4, 5, 6]])

embed(x)
array([[[ 0.01438687, 0.05011239, 0.00628365, 0.04861524],

[-0.01068833, 0.01729892, 0.02042518, -0.01618656],
[-0.00873779, -0.02834515, 0.05484822, -0.06206018]],
[[ 0.06491279, -0.03182812, -0.01631819, -0.00312688],

[ 0.0408415 , 0.04370362, 0.00404529, -0.0028032 ],
[ 0.00952624, -0.01501013, 0.05958354, 0.04705103]]])
Minibatch Multiplication
We can multiply the matrices in two minibatches one by one, by the minibatch multiplication
operation batch_dot. Suppose the first batch contains n matrices X1 , . . . , Xn with a shape of a × b,
and the second batch contains n matrices Y1 , . . . , Yn with a shape of b × c. The output of matrix
multiplication on these two batches are n matrices X1 Y1 , . . . , Xn Yn with a shape of a×c. Therefore,
given two tensors of shape (n, a, b) and (n, b, c), the shape of the minibatch multiplication output
is (n, a, c).
X = np.ones((2, 1, 4))
Y = np.ones((2, 4, 6))
npx.batch_dot(X, Y).shape
14.4. Pretraining word2vec 647

(2, 1, 6)
Skip-gram Model Forward Calculation
In forward calculation, the input of the skip-gram model contains the central target word index
center and the concatenated context and noise word index contexts_and_negatives. In which,
the center variable has the shape (batch size, 1), while the contexts_and_negatives variable has
the shape (batch size, max_len). These two variables are first transformed from word indexes to
word vectors by the word embedding layer, and then the output of shape (batch size, 1, max_len)
is obtained by minibatch multiplication. Each element in the output is the inner product of the
central target word vector and the context word vector or noise word vector.
def skip_gram(center, contexts_and_negatives, embed_v, embed_u):

v = embed_v(center)
u = embed_u(contexts_and_negatives)
pred = npx.batch_dot(v, u.swapaxes(1, 2))
return pred
Verify that the output shape should be (batch size, 1, max_len).
skip_gram(np.ones((2, 1)), np.ones((2, 4)), embed, embed).shape
(2, 1, 4)
14.4.2 Training
Before training the word embedding model, we need to define the loss function of the model.
Binary Cross Entropy Loss Function
According to the definition of the loss function in negative sampling, we can directly use Gluonʼs
binary cross-entropy loss function SigmoidBinaryCrossEntropyLoss.
loss = gluon.loss.SigmoidBinaryCrossEntropyLoss()
It is worth mentioning that we can use the mask variable to specify the partial predicted value
and label that participate in loss function calculation in the minibatch: when the mask is 1, the
predicted value and label of the corresponding position will participate in the calculation of the
loss function; When the mask is 0, the predicted value and label of the corresponding position do
not participate in the calculation of the loss function. As we mentioned earlier, mask variables
can be used to avoid the effect of padding on loss function calculations.
Given two identical examples, different masks lead to different loss values.
pred = np.array([[.5]*4]*2)
label = np.array([[1, 0, 1, 0]]*2)
mask = np.array([[1, 1, 1, 1], [1, 1, 0, 0]])
loss(pred, label, mask)

array([0.724077 , 0.3620385])
We can normalize the loss in each example due to various lengths in each example.
loss(pred, label, mask) / mask.sum(axis=1) * mask.shape[1]
array([0.724077, 0.724077])
We construct the embedding layers of the central and context words, respectively, and set the
hyperparameter word vector dimension embed_size to 100.
embed_size = 100
net.add(nn.Embedding(input_dim=len(vocab), output_dim=embed_size),
nn.Embedding(input_dim=len(vocab), output_dim=embed_size))
Training
The training function is defined below. Because of the existence of padding, the calculation of the
loss function is slightly different compared to the previous training functions.
def train(net, data_iter, lr, num_epochs, device=d2l.try_gpu()):

net.initialize(ctx=device, force_reinit=True)
xlim=[0, num_epochs])
timer = d2l.Timer()
metric = d2l.Accumulator(2) # Sum of losses, no. of tokens
for i, batch in enumerate(data_iter):
center, context_negative, mask, label = [
data.as_in_ctx(device) for data in batch]
pred = skip_gram(center, context_negative, net[0], net[1])
l = (loss(pred.reshape(label.shape), label, mask)
/ mask.sum(axis=1) * mask.shape[1])
l.backward()
metric.add(l.sum(), l.size)
if (i+1) % 50 == 0:
animator.add(epoch+(i+1)/len(data_iter),
(metric[0]/metric[1],))
f'{metric[1] / timer.stop():.1f} tokens/sec on {str(device)}')
Now, we can train a skip-gram model using negative sampling.
14.4. Pretraining word2vec 649

train(net, data_iter, lr, num_epochs)
14.4.3 Applying the Word Embedding Model
After training the word embedding model, we can represent similarity in meaning between words
based on the cosine similarity of two word vectors. As we can see, when using the trained word
embedding model, the words closest in meaning to the word “chip” are mostly related to chips.
def get_similar_tokens(query_token, k, embed):

W = embed.weight.data()
x = W[vocab[query_token]]
# Compute the cosine similarity. Add 1e-9 for numerical stability
cos = np.dot(W, x) / np.sqrt(np.sum(W * W, axis=1) * np.sum(x * x) + 1e-9)
topk = npx.topk(cos, k=k+1, ret_typ='indices').asnumpy().astype('int32')
for i in topk[1:]: # Remove the input words
print(f'cosine sim={float(cos[i]):.3f}: {vocab.idx_to_token[i]}')
get_similar_tokens('chip', 3, net[0])
cosine sim=0.572: intel

cosine sim=0.453: microprocessor
cosine sim=0.442: bugs

Summary
• We can pretrain a skip-gram model through negative sampling.
Exercises
1. Set sparse_grad=True when creating an instance of nn.Embedding. Does it accelerate train-

ing? Look up MXNet documentation to learn the meaning of this argument.
2. Try to find synonyms for other words.
3. Tune the hyperparameters and observe and analyze the experimental results.
4. When the dataset is large, we usually sample the context words and the noise words for the
central target word in the current minibatch only when updating the model parameters. In
other words, the same central target word may have different context words or noise words
in different epochs. What are the benefits of this sort of training? Try to implement this
training method.
Discussions203
14.5 Word Embedding with Global Vectors (GloVe)
First, we should review the skip-gram model in word2vec. The conditional probability P (wj | wi )
expressed in the skip-gram model using the softmax operation will be recorded as qij , that is:
exp(u⊤
j vi )
qij = ∑ , (14.5.1)
k∈V exp(u⊤
k vi )
where vi and ui are the vector representations of word wi of index i as the center word and context
word respectively, and V = {0, 1, . . . , |V| − 1} is the vocabulary index set.
For word wi , it may appear in the dataset for multiple times. We collect all the context words
every time when wi is a center word and keep duplicates, denoted as multiset Ci . The number
of an element in a multiset is called the multiplicity of the element. For instance, suppose that
word wi appears twice in the dataset: the context windows when these two wi become center
words in the text sequence contain context word indices 2, 1, 5, 2 and 2, 3, 2, 1. Then, multiset
Ci = {1, 1, 2, 2, 2, 2, 3, 5}, where multiplicity of element 1 is 2, multiplicity of element 2 is 4, and
multiplicities of elements 3 and 5 are both 1. Denote multiplicity of element j in multiset Ci as xij :
it is the number of word wj in all the context windows for center word wi in the entire dataset. As
a result, the loss function of the skip-gram model can be expressed in a different way:
∑∑
− xij log qij . (14.5.2)
i∈V j∈V
We add up the number of all the context words for the central target word wi to get xi , and record
the conditional probability xij /xi for generating context word wj based on central target word wi
as pij . We can rewrite the loss function of the skip-gram model as
∑ ∑
− xi pij log qij . (14.5.3)
i∈V j∈V
203
14.5. Word Embedding with Global Vectors (GloVe) 651

∑
In the formula above, j∈V pij log qij computes the conditional probability distribution pij for
context word generation based on the central target word wi and the cross-entropy of conditional
probability distribution qij predicted by the model. The loss function is weighted using the sum
of the number of context words with the central target word wi . If we minimize the loss function
from the formula above, we will be able to allow the predicted conditional probability distribution
to approach as close as possible to the true conditional probability distribution.
However, although the most common type of loss function, the cross-entropy loss function is
sometimes not a good choice. On the one hand, as we mentioned in Section 14.2 the cost of letting
the model prediction qij become the legal probability distribution has the sum of all items in the
entire dictionary in its denominator. This can easily lead to excessive computational overhead.
On the other hand, there are often a lot of uncommon words in the dictionary, and they appear
rarely in the dataset. In the cross-entropy loss function, the final prediction of the conditional
probability distribution on a large number of uncommon words is likely to be inaccurate.
14.5.1 The GloVe Model
To address this, GloVe (Pennington et al., 2014), a word embedding model that came after
word2vec, adopts square loss and makes three changes to the skip-gram model based on this loss.
1. Here, we use the non-probability distribution variables p′ij = xij and qij
′ = exp(u⊤ v ) and
j i
( )2 ( )2
take their logs. Therefore, we get the square loss log p′ij − log qij
′ = u⊤
j vi − log xij .
2. We add two scalar model parameters for each word wi : the bias terms bi (for central target
words) and ci ( for context words).
3. Replace the weight of each loss with the function h(xij ). The weight function h(x) is a mono-
tone increasing function with the range [0, 1].
Therefore, the goal of GloVe is to minimize the loss function.
∑∑ ( )2
h(xij ) u⊤
j vi + bi + cj − log xij . (14.5.4)
i∈V j∈V
Here, we have a suggestion for the choice of weight function h(x): when x < c (e.g c = 100), make
h(x) = (x/c)α (e.g α = 0.75), otherwise make h(x) = 1. Because h(0) = 0, the squared loss term
for xij = 0 can be simply ignored. When we use minibatch SGD for training, we conduct random
sampling to get a non-zero minibatch xij from each timestep and compute the gradient to update
the model parameters. These non-zero xij are computed in advance based on the entire dataset
and they contain global statistics for the dataset. Therefore, the name GloVe is taken from “Global
Vectors”.
Notice that if word wi appears in the context window of word wj , then word wj will also appear in
the context window of word wi . Therefore, xij = xji . Unlike word2vec, GloVe fits the symmetric
log xij in lieu of the asymmetric conditional probability pij . Therefore, the central target word
vector and context word vector of any word are equivalent in GloVe. However, the two sets of word
vectors that are learned by the same word may be different in the end due to different initialization
values. After learning all the word vectors, GloVe will use the sum of the central target word vector
and the context word vector as the final word vector for the word.

14.5.2 Understanding GloVe from Conditional Probability Ratios
We can also try to understand GloVe word embedding from another perspective. We will continue
the use of symbols from earlier in this section, P (wj | wi ) represents the conditional probability
of generating context word wj with central target word wi in the dataset, and it will be recorded as
pij . From a real example from a large corpus, here we have the following two sets of conditional
probabilities with “ice” and “steam” as the central target words and the ratio between them:
wk = “solid” “gas” “water” “fashion”

p1 = P (wk | “ice” ) 0.00019 0.000066 0.003 0.000017
p2 = P (wk | “steam” ) 0.000022 0.00078 0.0022 0.000018
p1 /p2 8.9 0.085 1.36 0.96
We will be able to observe phenomena such as:

• For a word wk that is related to “ice” but not to “steam”, such as wk =“solid”, we would expect
a larger conditional probability ratio, like the value 8.9 in the last row of the table above.
• For a word wk that is related to “steam” but not to “ice”, such as wk =“gas”, we would expect
a smaller conditional probability ratio, like the value 0.085 in the last row of the table above.
• For a word wk that is related to both “ice” and “steam”, such as wk =“water”, we would expect
a conditional probability ratio close to 1, like the value 1.36 in the last row of the table above.
• For a word wk that is related to neither “ice” or “steam”, such as wk =“fashion”, we would
expect a conditional probability ratio close to 1, like the value 0.96 in the last row of the table
above.
We can see that the conditional probability ratio can represent the relationship between different
words more intuitively. We can construct a word vector function to fit the conditional probability
ratio more effectively. As we know, to obtain any ratio of this type requires three words wi , wj ,
and wk . The conditional probability ratio with wi as the central target word is pij /pik . We can find
a function that uses word vectors to fit this conditional probability ratio.
pij
f (uj , uk , vi ) ≈ . (14.5.5)
pik
The possible design of function f here will not be unique. We only need to consider a more reason-
able possibility. Notice that(the conditional ) probability ratio is a scalar, we can limit f to be a scalar
⊤
function: f (uj , uk , vi ) = f (uj − uk ) vi . After exchanging index j with k, we will be able to see
that function f satisfies the condition f (x)f (−x) = 1, so one possibility could be f (x) = exp(x).
Thus:
( )
exp u⊤
j v i p
f (uj , uk , vi ) = ( ⊤ ) ≈ ij . (14.5.6)
exp uk vi pik
( )
One possibility that satisfies the right side of the approximation sign is exp u⊤ j i ≈ αpij , where
v
α is a constant. Considering that pij = xij /xi , after taking the logarithm we get u⊤ j vi ≈ log α +
log xij − log xi . We use additional bias terms to fit − log α + log xi , such as the central target word
bias term bi and context word bias term cj :
u⊤
j vi + bi + cj ≈ log(xij ). (14.5.7)
By taking the square error and weighting the left and right sides of the formula above, we can get
the loss function of GloVe.
14.5. Word Embedding with Global Vectors (GloVe) 653

Summary
• In some cases, the cross-entropy loss function may have a disadvantage. GloVe uses squared
loss and the word vector to fit global statistics computed in advance based on the entire
dataset.
• The central target word vector and context word vector of any word are equivalent in GloVe.
Exercises
1. If a word appears in the context window of another word, how can we use the distance be-
tween them in the text sequence to redesign the method for computing the conditional prob-
ability pij ? Hint: See section 4.2 from the paper GloVe (Pennington et al., 2014).
2. For any word, will its central target word bias term and context word bias term be equivalent
to each other in GloVe? Why?
Discussions204
14.6 Subword Embedding
English words usually have internal structures and formation methods. For example, we can de-
duce the relationship between “dog”, “dogs”, and “dogcatcher” by their spelling. All these words
have the same root, “dog”, but they use different suffixes to change the meaning of the word.
Moreover, this association can be extended to other words. For example, the relationship be-
tween “dog” and “dogs” is just like the relationship between “cat” and “cats”. The relationship
between “boy” and “boyfriend” is just like the relationship between “girl” and “girlfriend”. This
characteristic is not unique to English. In French and Spanish, a lot of verbs can have more than
40 different forms depending on the context. In Finnish, a noun may have more than 15 forms. In
fact, morphology, which is an important branch of linguistics, studies the internal structure and
formation of words.
14.6.1 fastText
In word2vec, we did not directly use morphology information. In both the skip-gram model and
continuous bag-of-words model, we use different vectors to represent words with different forms.
For example, “dog” and “dogs” are represented by two different vectors, while the relationship
between these two vectors is not directly represented in the model. In view of this, fastText (Bo-
janowski et al., 2017) proposes the method of subword embedding, thereby attempting to intro-
duce morphological information in the skip-gram model in word2vec.
In fastText, each central word is represented as a collection of subwords. Below we use the word
“where” as an example to understand how subwords are formed. First, we add the special charac-
ters “<” and “>” at the beginning and end of the word to distinguish the subwords used as prefixes
and suffixes. Then, we treat the word as a sequence of characters to extract the n-grams. For
example, when n = 3, we can get all subwords with a length of 3:
"<wh", "whe", "her", "ere", "re>", (14.6.1)

204

and the special subword "<where>".
In fastText, for a word w, we record the union of all its subwords with length of 3 to 6 and special
subwords as Gw . Thus, the dictionary is the union of the collection of subwords of all words.
Assume the vector of the subword g in the dictionary is zg . Then, the central word vector uw for
the word w in the skip-gram model can be expressed as
∑
uw = zg . (14.6.2)
g∈Gw
The rest of the fastText process is consistent with the skip-gram model, so it is not repeated here.
As we can see, compared with the skip-gram model, the dictionary in fastText is larger, resulting
in more model parameters. Also, the vector of one word requires the summation of all subword
vectors, which results in higher computation complexity. However, we can obtain better vectors
for more uncommon complex words, even words not existing in the dictionary, by looking at other
words with similar structures.
14.6.2 Byte Pair Encoding
In fastText, all the extracted subwords have to be of the specified lengths, such as 3 to 6, thus the
vocabulary size cannot be predefined. To allow for variable-length subwords in a fixed-size vocab-
ulary, we can apply a compression algorithm called byte pair encoding (BPE) to extract subwords
(Sennrich et al., 2015).
Byte pair encoding performs a statistical analysis of the training dataset to discover common sym-
bols within a word, such as consecutive characters of arbitrary length. Starting from symbols of
length 1, byte pair encoding iteratively merges the most frequent pair of consecutive symbols to
produce new longer symbols. Note that for efficiency, pairs crossing word boundaries are not con-
sidered. In the end, we can use such symbols as subwords to segment words. Byte pair encoding
and its variants has been used for input representations in popular natural language processing
pretraining models such as GPT-2 (Radford et al., 2019) and RoBERTa (Liu et al., 2019). In the
following, we will illustrate how byte pair encoding works.
First, we initialize the vocabulary of symbols as all the English lowercase characters, a special
end-of-word symbol '_', and a special unknown symbol '[UNK]'.
import collections
symbols = ['a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'm',
'n', 'o', 'p', 'q', 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z',
'_', '[UNK]']
Since we do not consider symbol pairs that cross boundaries of words, we only need a dictionary
raw_token_freqs that maps words to their frequencies (number of occurrences) in a dataset. Note
that the special symbol '_' is appended to each word so that we can easily recover a word sequence
(e.g., “a taller man”) from a sequence of output symbols ( e.g., “a_ tall er_ man”). Since we start
the merging process from a vocabulary of only single characters and special symbols, space is
inserted between every pair of consecutive characters within each word (keys of the dictionary
token_freqs). In other words, space is the delimiter between symbols within a word.
raw_token_freqs = {'fast_': 4, 'faster_': 3, 'tall_': 5, 'taller_': 4}

token_freqs = {}
14.6. Subword Embedding 655

for token, freq in raw_token_freqs.items():
token_freqs[' '.join(list(token))] = raw_token_freqs[token]
token_freqs
{'f a s t _': 4, 'f a s t e r _': 3, 't a l l _': 5, 't a l l e r _': 4}
We define the following get_max_freq_pair function that returns the most frequent pair of con-
secutive symbols within a word, where words come from keys of the input dictionary token_freqs.
def get_max_freq_pair(token_freqs):
pairs = collections.defaultdict(int)
for token, freq in token_freqs.items():
symbols = token.split()
for i in range(len(symbols) - 1):
# Key of `pairs` is a tuple of two consecutive symbols
pairs[symbols[i], symbols[i + 1]] += freq
return max(pairs, key=pairs.get) # Key of `pairs` with the max value
As a greedy approach based on frequency of consecutive symbols, byte pair encoding will use
the following merge_symbols function to merge the most frequent pair of consecutive symbols to
produce new symbols.
def merge_symbols(max_freq_pair, token_freqs, symbols):

symbols.append(''.join(max_freq_pair))
new_token_freqs = dict()
for token, freq in token_freqs.items():
new_token = token.replace(' '.join(max_freq_pair),
''.join(max_freq_pair))
new_token_freqs[new_token] = token_freqs[token]
return new_token_freqs
Now we iteratively perform the byte pair encoding algorithm over the keys of the dictionary to-
ken_freqs. In the first iteration, the most frequent pair of consecutive symbols are 't' and 'a',
thus byte pair encoding merges them to produce a new symbol 'ta'. In the second iteration, byte
pair encoding continues to merge 'ta' and 'l' to result in another new symbol 'tal'.
num_merges = 10
for i in range(num_merges):
max_freq_pair = get_max_freq_pair(token_freqs)
token_freqs = merge_symbols(max_freq_pair, token_freqs, symbols)
print(f'merge #{i + 1}:', max_freq_pair)
merge #1: ('t', 'a')

merge #2: ('ta', 'l')
merge #3: ('tal', 'l')
merge #4: ('f', 'a')
merge #5: ('fa', 's')
merge #6: ('fas', 't')
merge #7: ('e', 'r')
merge #8: ('er', '_')
merge #9: ('tall', '_')
merge #10: ('fast', '_')

After 10 iterations of byte pair encoding, we can see that list symbols now contains 10 more sym-
bols that are iteratively merged from other symbols.
print(symbols)
['a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', 'r', 's
,→', 't', 'u', 'v', 'w', 'x', 'y', 'z', '_', '[UNK]', 'ta', 'tal', 'tall', 'fa', 'fas', 'fast
,→', 'er', 'er_', 'tall_', 'fast_']
For the same dataset specified in the keys of the dictionary raw_token_freqs, each word in the
dataset is now segmented by subwords “fast_”, “fast”, “er_”, “tall_”, and “tall” as a result of the byte
pair encoding algorithm. For instance, words “faster_” and “taller_” are segmented as “fast er_”
and “tall er_”, respectively.
print(list(token_freqs.keys()))
['fast_', 'fast er_', 'tall_', 'tall er_']
Note that the result of byte pair encoding depends on the dataset being used. We can also use the
subwords learned from one dataset to segment words of another dataset. As a greedy approach,
the following segment_BPE function tries to break words into the longest possible subwords from
the input argument symbols.
def segment_BPE(tokens, symbols):

outputs = []
for token in tokens:
start, end = 0, len(token)
cur_output = []
# Segment token with the longest possible subwords from symbols
while start < len(token) and start < end:
if token[start: end] in symbols:
cur_output.append(token[start: end])
start = end
end = len(token)
else:
end -= 1
if start < len(token):
cur_output.append('[UNK]')
outputs.append(' '.join(cur_output))
return outputs
In the following, we use the subwords in list symbols, which is learned from the aforementioned
dataset, to segment tokens that represent another dataset.
tokens = ['tallest_', 'fatter_']

print(segment_BPE(tokens, symbols))
['tall e s t _', 'fa t t er_']
14.6. Subword Embedding 657

Summary
• FastText proposes a subword embedding method. Based on the skip-gram model in

word2vec, it represents the central word vector as the sum of the subword vectors of the
word.
• Subword embedding utilizes the principles of morphology, which usually improves the qual-
ity of representations of uncommon words.
• Byte pair encoding performs a statistical analysis of the training dataset to discover common
symbols within a word. As a greedy approach, byte pair encoding iteratively merges the most
frequent pair of consecutive symbols.
Exercises
1. When there are too many subwords (for example, 6 words in English result in about 3 × 108
combinations), what problems arise? Can you think of any methods to solve them? Hint:
Refer to the end of section 3.2 of the fastText paper[1].
2. How can you design a subword embedding model based on the continuous bag-of-words
model?
3. To get a vocabulary of size m, how many merging operations are needed when the initial
symbol vocabulary size is n?
4. How can we extend the idea of byte pair encoding to extract phrases?
Discussions205
14.7 Finding Synonyms and Analogies
In Section 14.4 we trained a word2vec word embedding model on a small-scale dataset and
searched for synonyms using the cosine similarity of word vectors. In practice, word vectors pre-
trained on a large-scale corpus can often be applied to downstream natural language processing
tasks. This section will demonstrate how to use these pretrained word vectors to find synonyms
and analogies. We will continue to apply pretrained word vectors in subsequent sections.

import os
npx.set_np()
205

14.7.1 Using Pretrained Word Vectors
Below lists pretrained GloVe embeddings of dimensions 50, 100, and 300, which can be down-
loaded from the GloVe website206 . The pretrained fastText embeddings are available in multiple
languages. Here we consider one English version (300-dimensional “wiki.en”) that can be down-
loaded from the fastText website207 .
#@save
d2l.DATA_HUB['glove.6b.50d'] = (d2l.DATA_URL + 'glove.6B.50d.zip',
'0b8703943ccdb6eb788e6f091b8946e82231bc4d')
#@save
'cd43bfb07e44e6f27cbcc7bc9ae3d80284fdaf5a')
#@save
'b5116e234e9eb9076672cfeabf5469f3eec904fa')
#@save
d2l.DATA_HUB['wiki.en'] = (d2l.DATA_URL + 'wiki.en.zip',
'c1816da3821ae9f43899be655002f6c723e91b88')
We define the following TokenEmbedding class to load the above pretrained Glove and fastText em-
beddings.
#@save
class TokenEmbedding:
"""Token Embedding."""
def __init__(self, embedding_name):
self.idx_to_token, self.idx_to_vec = self._load_embedding(
embedding_name)
self.unknown_idx = 0
self.token_to_idx = {token: idx for idx, token in
enumerate(self.idx_to_token)}
def _load_embedding(self, embedding_name):

idx_to_token, idx_to_vec = ['<unk>'], []
data_dir = d2l.download_extract(embedding_name)
# GloVe website: https://nlp.stanford.edu/projects/glove/
# fastText website: https://fasttext.cc/
with open(os.path.join(data_dir, 'vec.txt'), 'r') as f:
for line in f:
elems = line.rstrip().split(' ')
token, elems = elems[0], [float(elem) for elem in elems[1:]]
# Skip header information, such as the top row in fastText
if len(elems) > 1:
idx_to_token.append(token)
idx_to_vec.append(elems)
idx_to_vec = [[0] * len(idx_to_vec[0])] + idx_to_vec
return idx_to_token, np.array(idx_to_vec)

206
https://nlp.stanford.edu/projects/glove/
207
https://fasttext.cc/
14.7. Finding Synonyms and Analogies 659

def __getitem__(self, tokens):
indices = [self.token_to_idx.get(token, self.unknown_idx)
for token in tokens]
vecs = self.idx_to_vec[np.array(indices)]
return vecs
def __len__(self):
return len(self.idx_to_token)
Next, we use 50-dimensional GloVe embeddings pretrained on a subset of the Wikipedia. The
corresponding word embedding is automatically downloaded the first time we create a pretrained
word embedding instance.
glove_6b50d = TokenEmbedding('glove.6b.50d')
Downloading ../data/glove.6B.50d.zip from http://d2l-data.s3-accelerate.amazonaws.com/glove.

,→6B.50d.zip...
Output the dictionary size. The dictionary contains 400, 000 words and a special unknown token.
len(glove_6b50d)
400001
We can use a word to get its index in the dictionary, or we can get the word from its index.
glove_6b50d.token_to_idx['beautiful'], glove_6b50d.idx_to_token[3367]
(3367, 'beautiful')
14.7.2 Applying Pretrained Word Vectors
Below, we demonstrate the application of pretrained word vectors, using GloVe as an example.
Finding Synonyms
Here, we re-implement the algorithm used to search for synonyms by cosine similarity introduced
in Section 14.1
In order to reuse the logic for seeking the k nearest neighbors when seeking analogies, we encap-
sulate this part of the logic separately in the knn (k-nearest neighbors) function.
def knn(W, x, k):

# The added 1e-9 is for numerical stability
cos = np.dot(W, x.reshape(-1,)) / (
np.sqrt(np.sum(W * W, axis=1) + 1e-9) * np.sqrt((x * x).sum()))
topk = npx.topk(cos, k=k, ret_typ='indices')
return topk, [cos[int(i)] for i in topk]

Then, we search for synonyms by pre-training the word vector instance embed.
def get_similar_tokens(query_token, k, embed):

topk, cos = knn(embed.idx_to_vec, embed[[query_token]], k + 1)
for i, c in zip(topk[1:], cos[1:]): # Remove input words
print(f'cosine sim={float(c):.3f}: {embed.idx_to_token[int(i)]}')
The dictionary of pretrained word vector instance glove_6b50d already created contains 400,000
words and a special unknown token. Excluding input words and unknown words, we search for
the three words that are the most similar in meaning to “chip”.
get_similar_tokens('chip', 3, glove_6b50d)
cosine sim=0.856: chips

cosine sim=0.749: intel
cosine sim=0.749: electronics
Next, we search for the synonyms of “baby” and “beautiful”.
get_similar_tokens('baby', 3, glove_6b50d)
cosine sim=0.839: babies

cosine sim=0.800: boy
cosine sim=0.792: girl
get_similar_tokens('beautiful', 3, glove_6b50d)
cosine sim=0.921: lovely

cosine sim=0.893: gorgeous
cosine sim=0.830: wonderful
Finding Analogies
In addition to seeking synonyms, we can also use the pretrained word vector to seek the analo-
gies between words. For example, “man”:“woman”::“son”:“daughter” is an example of analogy,
“man” is to “woman” as “son” is to “daughter”. The problem of seeking analogies can be defined
as follows: for four words in the analogical relationship a : b :: c : d, given the first three words,
a, b and c, we want to find d. Assume the word vector for the word w is vec(w). To solve the
analogy problem, we need to find the word vector that is most similar to the result vector of
vec(c) + vec(b) − vec(a).
def get_analogy(token_a, token_b, token_c, embed):

vecs = embed[[token_a, token_b, token_c]]
x = vecs[1] - vecs[0] + vecs[2]
topk, cos = knn(embed.idx_to_vec, x, 1)
return embed.idx_to_token[int(topk[0])] # Remove unknown words
Verify the “male-female” analogy.
14.7. Finding Synonyms and Analogies 661

get_analogy('man', 'woman', 'son', glove_6b50d)
'daughter'
“Capital-country” analogy: “beijing” is to “china” as “tokyo” is to what? The answer should be

“japan”.
get_analogy('beijing', 'china', 'tokyo', glove_6b50d)
'japan'
“Adjective-superlative adjective” analogy: “bad” is to “worst” as “big” is to what? The answer

should be “biggest”.
get_analogy('bad', 'worst', 'big', glove_6b50d)
'biggest'
“Present tense verb-past tense verb” analogy: “do” is to “did” as “go” is to what? The answer should
be “went”.
get_analogy('do', 'did', 'go', glove_6b50d)
'went'
Summary
• Word vectors pre-trained on a large-scale corpus can often be applied to downstream natural
language processing tasks.
• We can use pre-trained word vectors to seek synonyms and analogies.
Exercises
1. Test the fastText results using TokenEmbedding('wiki.en').

2. If the dictionary is extremely large, how can we accelerate finding synonyms and analogies?
Discussions208
208

14.8 Bidirectional Encoder Representations from Transformers (BERT)
We have introduced several word embedding models for natural language understanding. Af-
ter pretraining, the output can be thought of as a matrix where each row is a vector that repre-
sents a word of a predefined vocabulary. In fact, these word embedding models are all context-
independent. Let us begin by illustrating this property.
14.8.1 From Context-Independent to Context-Sensitive
Recall the experiments in Section 14.4 and Section 14.7. For instance, word2vec and GloVe both
assign the same pretrained vector to the same word regardless of the context of the word (if any).
Formally, a context-independent representation of any token x is a function f (x) that only takes
x as its input. Given the abundance of polysemy and complex semantics in natural languages,
context-independent representations have obvious limitations. For instance, the word “crane” in
contexts “a crane is flying” and “a crane driver came” has completely different meanings; thus,
the same word may be assigned different representations depending on contexts.
This motivates the development of context-sensitive word representations, where representations
of words depend on their contexts. Hence, a context-sensitive representation of token x is a func-
tion f (x, c(x)) depending on both x and its context c(x). Popular context-sensitive representations
include TagLM (language-model-augmented sequence tagger) (Peters et al., 2017b), CoVe (Con-
text Vectors) (McCann et al., 2017), and ELMo (Embeddings from Language Models) (Peters et al.,
2018).
For example, by taking the entire sequence as the input, ELMo is a function that assigns a rep-
resentation to each word from the input sequence. Specifically, ELMo combines all the inter-
mediate layer representations from pretrained bidirectional LSTM as the output representation.
Then the ELMo representation will be added to a downstream taskʼs existing supervised model as
additional features, such as by concatenating ELMo representation and the original representa-
tion (e.g., GloVe) of tokens in the existing model. On one hand, all the weights in the pretrained
bidirectional LSTM model are frozen after ELMo representations are added. On the other hand,
the existing supervised model is specifically customized for a given task. Leveraging different
best models for different tasks at that time, adding ELMo improved the state of the art across six
natural language processing tasks: sentiment analysis, natural language inference, semantic role
labeling, coreference resolution, named entity recognition, and question answering.
14.8.2 From Task-Specific to Task-Agnostic
Although ELMo has significantly improved solutions to a diverse set of natural language process-
ing tasks, each solution still hinges on a task-specific architecture. However, it is practically non-
trivial to craft a specific architecture for every natural language processing task. The GPT (Gen-
erative Pre-Training) model represents an effort in designing a general task-agnostic model for
context-sensitive representations (Radford et al., 2018). Built on a Transformer decoder, GPT pre-
trains a language model that will be used to represent text sequences. When applying GPT to a
downstream task, the output of the language model will be fed into an added linear output layer to
predict the label of the task. In sharp contrast to ELMo that freezes parameters of the pretrained
model, GPT fine-tunes all the parameters in the pretrained Transformer decoder during super-
vised learning of the downstream task. GPT was evaluated on twelve tasks of natural language
inference, question answering, sentence similarity, and classification, and improved the state of
the art in nine of them with minimal changes to the model architecture.
14.8. Bidirectional Encoder Representations from Transformers (BERT) 663

However, due to the autoregressive nature of language models, GPT only looks forward (left-to-
right). In contexts “i went to the bank to deposit cash” and “i went to the bank to sit down”, as
“bank” is sensitive to the context to its left, GPT will return the same representation for “bank”,
though it has different meanings.
14.8.3 BERT: Combining the Best of Both Worlds
As we have seen, ELMo encodes context bidirectionally but uses task-specific architectures; while
GPT is task-agnostic but encodes context left-to-right. Combining the best of both worlds, BERT
(Bidirectional Encoder Representations from Transformers) encodes context bidirectionally and
requires minimal architecture changes for a wide range of natural language processing tasks (De-
vlin et al., 2018). Using a pretrained Transformer encoder, BERT is able to represent any token
based on its bidirectional context. During supervised learning of downstream tasks, BERT is sim-
ilar to GPT in two aspects. First, BERT representations will be fed into an added output layer,
with minimal changes to the model architecture depending on nature of tasks, such as predict-
ing for every token vs. predicting for the entire sequence. Second, all the parameters of the pre-
trained Transformer encoder are fine-tuned, while the additional output layer will be trained from
scratch. Fig. 14.8.1 depicts the differences among ELMo, GPT, and BERT.
Fig. 14.8.1: A comparison of ELMo, GPT, and BERT.
BERT further improved the state of the art on eleven natural language processing tasks under
broad categories of i) single text classification (e.g., sentiment analysis), ii) text pair classifica-
tion (e.g., natural language inference), iii) question answering, iv) text tagging (e.g., named entity
recognition). All proposed in 2018, from context-sensitive ELMo to task-agnostic GPT and BERT,
conceptually simple yet empirically powerful pretraining of deep representations for natural lan-
guages have revolutionized solutions to various natural language processing tasks.
In the rest of this chapter, we will dive into the pretraining of BERT. When natural language pro-
cessing applications are explained in Chapter 15, we will illustrate fine-tuning of BERT for down-
stream applications.

npx.set_np()
14.8.4 Input Representation
In natural language processing, some tasks (e.g., sentiment analysis) take single text as the input,
while in some other tasks (e.g., natural language inference), the input is a pair of text sequences.
The BERT input sequence unambiguously represents both single text and text pairs. In the former,
the BERT input sequence is the concatenation of the special classification token “<cls>”, tokens of
a text sequence, and the special separation token “<sep>”. In the latter, the BERT input sequence
is the concatenation of “<cls>”, tokens of the first text sequence, “<sep>”, tokens of the second text
sequence, and “<sep>”. We will consistently distinguish the terminology “BERT input sequence”
from other types of “sequences”. For instance, one BERT input sequence may include either one
text sequence or two text sequences.
To distinguish text pairs, the learned segment embeddings eA and eB are added to the token em-
beddings of the first sequence and the second sequence, respectively. For single text inputs, only
eA is used.
The following get_tokens_and_segments takes either one sentence or two sentences as the input,
then returns tokens of the BERT input sequence and their corresponding segment IDs.
#@save
def get_tokens_and_segments(tokens_a, tokens_b=None):
tokens = ['<cls>'] + tokens_a + ['<sep>']
# 0 and 1 are marking segment A and B, respectively
segments = [0] * (len(tokens_a) + 2)
if tokens_b is not None:
tokens += tokens_b + ['<sep>']
segments += [1] * (len(tokens_b) + 1)
return tokens, segments
BERT chooses the Transformer encoder as its bidirectional architecture. Common in the Trans-
former encoder, positional embeddings are added at every position of the BERT input sequence.
However, different from the original Transformer encoder, BERT uses learnable positional em-
beddings. To sum up, Fig. 14.8.2 shows that the embeddings of the BERT input sequence are the
sum of the token embeddings, segment embeddings, and positional embeddings.

Fig. 14.8.2: The embeddings of the BERT input sequence are the sum of the token embeddings,
segment embeddings, and positional embeddings.
The following BERTEncoder class is similar to the TransformerEncoder class as implemented in Sec-
tion 10.3. Different from TransformerEncoder, BERTEncoder uses segment embeddings and learn-
able positional embeddings.
#@save
class BERTEncoder(nn.Block):
def __init__(self, vocab_size, num_hiddens, ffn_num_hiddens, num_heads,
num_layers, dropout, max_len=1000, **kwargs):
super(BERTEncoder, self).__init__(**kwargs)
self.token_embedding = nn.Embedding(vocab_size, num_hiddens)
self.segment_embedding = nn.Embedding(2, num_hiddens)
for _ in range(num_layers):
self.blks.add(d2l.EncoderBlock(
num_hiddens, ffn_num_hiddens, num_heads, dropout, True))
# In BERT, positional embeddings are learnable, thus we create a
# parameter of positional embeddings that are long enough
self.pos_embedding = self.params.get('pos_embedding',
shape=(1, max_len, num_hiddens))
def forward(self, tokens, segments, valid_lens):

# Shape of `X` remains unchanged in the following code snippet:
# (batch size, max sequence length, `num_hiddens`)
X = self.token_embedding(tokens) + self.segment_embedding(segments)
X = X + self.pos_embedding.data(ctx=X.ctx)[:, :X.shape[1], :]
X = blk(X, valid_lens)
return X
Suppose that the vocabulary size is 10,000. To demonstrate forward inference of BERTEncoder, let
us create an instance of it and initialize its parameters.
vocab_size, num_hiddens, ffn_num_hiddens, num_heads = 10000, 768, 1024, 4

num_layers, dropout = 2, 0.2
encoder = BERTEncoder(vocab_size, num_hiddens, ffn_num_hiddens, num_heads,
num_layers, dropout)
We define tokens to be 2 BERT input sequences of length 8, where each token is an index of the vo-

cabulary. The forward inference of BERTEncoder with the input tokens returns the encoded result
where each token is represented by a vector whose length is predefined by the hyperparameter
num_hiddens. This hyperparameter is usually referred to as the hidden size (number of hidden
units) of the Transformer encoder.
tokens = np.random.randint(0, vocab_size, (2, 8))

segments = np.array([[0, 0, 0, 0, 1, 1, 1, 1], [0, 0, 0, 1, 1, 1, 1, 1]])
encoded_X = encoder(tokens, segments, None)
encoded_X.shape
(2, 8, 768)
14.8.5 Pretraining Tasks
The forward inference of BERTEncoder gives the BERT representation of each token of the input
text and the inserted special tokens “<cls>” and “<seq>”. Next, we will use these representations
to compute the loss function for pretraining BERT. The pretraining is composed of the following
two tasks: masked language modeling and next sentence prediction.
Masked Language Modeling
As illustrated in Section 8.3, a language model predicts a token using the context on its left. To
encode context bidirectionally for representing each token, BERT randomly masks tokens and
uses tokens from the bidirectional context to predict the masked tokens. This task is referred to
as a masked language model.
In this pretraining task, 15% of tokens will be selected at random as the masked tokens for predic-
tion. To predict a masked token without cheating by using the label, one straightforward approach
is to always replace it with a special “<mask>” token in the BERT input sequence. However, the
artificial special token “<mask>” will never appear in fine-tuning. To avoid such a mismatch be-
tween pretraining and fine-tuning, if a token is masked for prediction (e.g., “great” is selected to
be masked and predicted in “this movie is great”), in the input it will be replaced with:
• a special “<mask>” token for 80% of the time (e.g., “this movie is great” becomes “this movie
is <mask>”);
• a random token for 10% of the time (e.g., “this movie is great” becomes “this movie is drink”);
• the unchanged label token for 10% of the time (e.g., “this movie is great” becomes “this
movie is great”).
Note that for 10% of 15% time a random token is inserted. This occasional noise encourages BERT
to be less biased towards the masked token (especially when the label token remains unchanged)
in its bidirectional context encoding.
We implement the following MaskLM class to predict masked tokens in the masked language model
task of BERT pretraining. The prediction uses a one-hidden-layer MLP (self.mlp). In forward
inference, it takes two inputs: the encoded result of BERTEncoder and the token positions for pre-
diction. The output is the prediction results at these positions.

#@save
class MaskLM(nn.Block):
def __init__(self, vocab_size, num_hiddens, **kwargs):
super(MaskLM, self).__init__(**kwargs)
self.mlp = nn.Sequential()
self.mlp.add(
nn.Dense(num_hiddens, flatten=False, activation='relu'))
self.mlp.add(nn.LayerNorm())
self.mlp.add(nn.Dense(vocab_size, flatten=False))
def forward(self, X, pred_positions):

num_pred_positions = pred_positions.shape[1]
pred_positions = pred_positions.reshape(-1)
batch_size = X.shape[0]
batch_idx = np.arange(0, batch_size)
# Suppose that `batch_size` = 2, `num_pred_positions` = 3, then
# `batch_idx` is `np.array([0, 0, 0, 1, 1, 1])`
batch_idx = np.repeat(batch_idx, num_pred_positions)
masked_X = X[batch_idx, pred_positions]
masked_X = masked_X.reshape((batch_size, num_pred_positions, -1))
mlm_Y_hat = self.mlp(masked_X)
return mlm_Y_hat
To demonstrate the forward inference of MaskLM, we create its instance mlm and initialize it. Recall
that encoded_X from the forward inference of BERTEncoder represents 2 BERT input sequences.
We define mlm_positions as the 3 indices to predict in either BERT input sequence of encoded_X.
The forward inference of mlm returns prediction results mlm_Y_hat at all the masked positions
mlm_positions of encoded_X. For each prediction, the size of the result is equal to the vocabulary
size.
mlm = MaskLM(vocab_size, num_hiddens)

mlm.initialize()
mlm_positions = np.array([[1, 5, 2], [6, 1, 5]])
mlm_Y_hat = mlm(encoded_X, mlm_positions)
mlm_Y_hat.shape
(2, 3, 10000)
With the ground truth labels mlm_Y of the predicted tokens mlm_Y_hat under masks, we can calcu-
late the cross entropy loss of the masked language model task in BERT pretraining.
mlm_Y = np.array([[7, 8, 9], [10, 20, 30]])

mlm_l = loss(mlm_Y_hat.reshape((-1, vocab_size)), mlm_Y.reshape(-1))
mlm_l.shape
(6,)

Next Sentence Prediction
Although masked language modeling is able to encode bidirectional context for representing
words, it does not explicitly model the logical relationship between text pairs. To help under-
stand the relationship between two text sequences, BERT considers a binary classification task,
next sentence prediction, in its pretraining. When generating sentence pairs for pretraining, for
half of the time they are indeed consecutive sentences with the label “True”; while for the other
half of the time the second sentence is randomly sampled from the corpus with the label “False”.
The following NextSentencePred class uses a one-hidden-layer MLP to predict whether the second
sentence is the next sentence of the first in the BERT input sequence. Due to self-attention in the
Transformer encoder, the BERT representation of the special token “<cls>” encodes both the two
sentences from the input. Hence, the output layer (self.output) of the MLP classifier takes X as
the input, where X is the output of the MLP hidden layer whose input is the encoded “<cls>” token.
#@save
class NextSentencePred(nn.Block):
super(NextSentencePred, self).__init__(**kwargs)

# `X` shape: (batch size, `num_hiddens`)
return self.output(X)
We can see that the forward inference of an NextSentencePred instance returns binary predictions
for each BERT input sequence.
nsp = NextSentencePred()
nsp.initialize()
nsp_Y_hat = nsp(encoded_X)
nsp_Y_hat.shape
(2, 2)
The cross-entropy loss of the 2 binary classifications can also be computed.
nsp_y = np.array([0, 1])

nsp_l = loss(nsp_Y_hat, nsp_y)
nsp_l.shape
(2,)
It is noteworthy that all the labels in both the aforementioned pretraining tasks can be trivially
obtained from the pretraining corpus without manual labeling effort. The original BERT has been
pretrained on the concatenation of BookCorpus (Zhu et al., 2015) and English Wikipedia. These
two text corpora are huge: they have 800 million words and 2.5 billion words, respectively.

When pretraining BERT, the final loss function is a linear combination of both the loss functions
for masked language modeling and next sentence prediction. Now we can define the BERTModel
class by instantiating the three classes BERTEncoder, MaskLM, and NextSentencePred. The forward
inference returns the encoded BERT representations encoded_X, predictions of masked language
modeling mlm_Y_hat, and next sentence predictions nsp_Y_hat.
#@save
class BERTModel(nn.Block):
def __init__(self, vocab_size, num_hiddens, ffn_num_hiddens, num_heads,
num_layers, dropout, max_len=1000):
super(BERTModel, self).__init__()
self.encoder = BERTEncoder(vocab_size, num_hiddens, ffn_num_hiddens,
num_heads, num_layers, dropout, max_len)
self.hidden = nn.Dense(num_hiddens, activation='tanh')
self.mlm = MaskLM(vocab_size, num_hiddens)
self.nsp = NextSentencePred()
def forward(self, tokens, segments, valid_lens=None, pred_positions=None):

encoded_X = self.encoder(tokens, segments, valid_lens)
if pred_positions is not None:
mlm_Y_hat = self.mlm(encoded_X, pred_positions)
else:
mlm_Y_hat = None
# The hidden layer of the MLP classifier for next sentence prediction.
# 0 is the index of the '<cls>' token
nsp_Y_hat = self.nsp(self.hidden(encoded_X[:, 0, :]))
return encoded_X, mlm_Y_hat, nsp_Y_hat
Summary
• Word embedding models such as word2vec and GloVe are context-independent. They assign
the same pretrained vector to the same word regardless of the context of the word (if any).
It is hard for them to handle well polysemy or complex semantics in natural languages.
• For context-sensitive word representations such as ELMo and GPT, representations of words
depend on their contexts.
• ELMo encodes context bidirectionally but uses task-specific architectures (however, it is
practically non-trivial to craft a specific architecture for every natural language processing
task); while GPT is task-agnostic but encodes context left-to-right.
• BERT combines the best of both worlds: it encodes context bidirectionally and requires min-
imal architecture changes for a wide range of natural language processing tasks.
• The embeddings of the BERT input sequence are the sum of the token embeddings, segment
embeddings, and positional embeddings.
• Pretraining BERT is composed of two tasks: masked language modeling and next sentence
prediction. The former is able to encode bidirectional context for representing words, while
the later explicitly models the logical relationship between text pairs.

Exercises
1. Why does BERT succeed?

2. All other things being equal, will a masked language model require more or fewer pretrain-
ing steps to converge than a left-to-right language model? Why?
3. In the original implementation of BERT, the position-wise feed-forward network in BERTEn-
coder (via d2l.EncoderBlock) and the fully-connected layer in MaskLM both use the Gaussian
error linear unit (GELU) (Hendrycks & Gimpel, 2016) as the activation function. Research
into the difference between GELU and ReLU.
Discussions209
14.9 The Dataset for Pretraining BERT
To pretrain the BERT model as implemented in Section 14.8, we need to generate the dataset in the
ideal format to facilitate the two pretraining tasks: masked language modeling and next sentence
prediction. On one hand, the original BERT model is pretrained on the concatenation of two huge
corpora BookCorpus and English Wikipedia (see Section 14.8.5), making it hard to run for most
readers of this book. On the other hand, the off-the-shelf pretrained BERT model may not fit for
applications from specific domains like medicine. Thus, it is getting popular to pretrain BERT on a
customized dataset. To facilitate the demonstration of BERT pretraining, we use a smaller corpus
WikiText-2 (Merity et al., 2016).
Comparing with the PTB dataset used for pretraining word2vec in Section 14.3, WikiText-2 i) re-
tains the original punctuation, making it suitable for next sentence prediction; ii) retrains the
original case and numbers; iii) is over twice larger.
import collections
import mxnet as mx
import os
import random
import time
import zipfile
npx.set_np()
In the WikiText-2 dataset, each line represents a paragraph where space is inserted between any
punctuation and its preceding token. Paragraphs with at least two sentences are retained. To split
sentences, we only use the period as the delimiter for simplicity. We leave discussions of more
complex sentence splitting techniques in the exercises at the end of this section.
#@save
d2l.DATA_HUB['wikitext-2'] = (
'https://s3.amazonaws.com/research.metamind.io/wikitext/'
'wikitext-2-v1.zip', '3c914d17d80b1459be871a5039ac23e752a53cbe')
#@save
209
14.9. The Dataset for Pretraining BERT 671

def _read_wiki(data_dir):
file_name = os.path.join(data_dir, 'wiki.train.tokens')
with open(file_name, 'r') as f:
lines = f.readlines()
# Uppercase letters are converted to lowercase ones
paragraphs = [line.strip().lower().split(' . ')
for line in lines if len(line.split(' . ')) >= 2]
random.shuffle(paragraphs)
return paragraphs
14.9.1 Defining Helper Functions for Pretraining Tasks
In the following, we begin by implementing helper functions for the two BERT pretraining tasks:
next sentence prediction and masked language modeling. These helper functions will be invoked
later when transforming the raw text corpus into the dataset of the ideal format to pretrain BERT.
Generating the Next Sentence Prediction Task
According to descriptions of .. _subsec_nsp:, the _get_next_sentence function generates a train-

ing example for the binary classification task.
#@save
def _get_next_sentence(sentence, next_sentence, paragraphs):
if random.random() < 0.5:
is_next = True
else:
# `paragraphs` is a list of lists of lists
next_sentence = random.choice(random.choice(paragraphs))
is_next = False
return sentence, next_sentence, is_next
The following function generates training examples for next sentence prediction from the input
paragraph by invoking the _get_next_sentence function. Here paragraph is a list of sentences,
where each sentence is a list of tokens. The argument max_len specifies the maximum length of a
BERT input sequence during pretraining.
#@save
def _get_nsp_data_from_paragraph(paragraph, paragraphs, vocab, max_len):
nsp_data_from_paragraph = []
for i in range(len(paragraph) - 1):
tokens_a, tokens_b, is_next = _get_next_sentence(
paragraph[i], paragraph[i + 1], paragraphs)
# Consider 1 '<cls>' token and 2 '<sep>' tokens
if len(tokens_a) + len(tokens_b) + 3 > max_len:
continue
tokens, segments = d2l.get_tokens_and_segments(tokens_a, tokens_b)
nsp_data_from_paragraph.append((tokens, segments, is_next))
return nsp_data_from_paragraph

Generating the Masked Language Modeling Task
In order to generate training examples for the masked language modeling task from a BERT input
sequence, we define the following _replace_mlm_tokens function. In its inputs, tokens is a list of
tokens representing a BERT input sequence, candidate_pred_positions is a list of token indices
of the BERT input sequence excluding those of special tokens (special tokens are not predicted
in the masked language modeling task), and num_mlm_preds indicates the number of predictions
(recall 15% random tokens to predict). Following the definition of the masked language modeling
task in Section 14.8.5, at each prediction position, the input may be replaced by a special “<mask>”
token or a random token, or remain unchanged. In the end, the function returns the input tokens
after possible replacement, the token indices where predictions take place and labels for these
predictions.
#@save
def _replace_mlm_tokens(tokens, candidate_pred_positions, num_mlm_preds,
vocab):
# Make a new copy of tokens for the input of a masked language model,
# where the input may contain replaced '<mask>' or random tokens
mlm_input_tokens = [token for token in tokens]
pred_positions_and_labels = []
# Shuffle for getting 15% random tokens for prediction in the masked
# language modeling task
random.shuffle(candidate_pred_positions)
for mlm_pred_position in candidate_pred_positions:
if len(pred_positions_and_labels) >= num_mlm_preds:
break
masked_token = None
# 80% of the time: replace the word with the '<mask>' token
masked_token = '<mask>'
else:
# 10% of the time: keep the word unchanged
masked_token = tokens[mlm_pred_position]
# 10% of the time: replace the word with a random word
else:
masked_token = random.randint(0, len(vocab) - 1)
mlm_input_tokens[mlm_pred_position] = masked_token
pred_positions_and_labels.append(
(mlm_pred_position, tokens[mlm_pred_position]))
return mlm_input_tokens, pred_positions_and_labels
By invoking the aforementioned _replace_mlm_tokens function, the following function takes a

BERT input sequence (tokens) as an input and returns indices of the input tokens (after possible
token replacement as described in Section 14.8.5), the token indices where predictions take place,
and label indices for these predictions.
#@save
def _get_mlm_data_from_tokens(tokens, vocab):
candidate_pred_positions = []
# `tokens` is a list of strings
for i, token in enumerate(tokens):
# Special tokens are not predicted in the masked language modeling
# task

if token in ['<cls>', '<sep>']:
continue
candidate_pred_positions.append(i)
# 15% of random tokens are predicted in the masked language modeling task
num_mlm_preds = max(1, round(len(tokens) * 0.15))
mlm_input_tokens, pred_positions_and_labels = _replace_mlm_tokens(
tokens, candidate_pred_positions, num_mlm_preds, vocab)
pred_positions_and_labels = sorted(pred_positions_and_labels,
key=lambda x: x[0])
pred_positions = [v[0] for v in pred_positions_and_labels]
mlm_pred_labels = [v[1] for v in pred_positions_and_labels]
return vocab[mlm_input_tokens], pred_positions, vocab[mlm_pred_labels]
14.9.2 Transforming Text into the Pretraining Dataset
Now we are almost ready to customize a Dataset class for pretraining BERT. Before that, we
still need to define a helper function _pad_bert_inputs to append the special “<mask>” to-
kens to the inputs. Its argument examples contain the outputs from the helper functions
_get_nsp_data_from_paragraph and _get_mlm_data_from_tokens for the two pretraining tasks.
#@save
def _pad_bert_inputs(examples, max_len, vocab):
max_num_mlm_preds = round(max_len * 0.15)
all_token_ids, all_segments, valid_lens, = [], [], []
all_pred_positions, all_mlm_weights, all_mlm_labels = [], [], []
nsp_labels = []
for (token_ids, pred_positions, mlm_pred_label_ids, segments,
is_next) in examples:
all_token_ids.append(np.array(token_ids + [vocab['<pad>']] * (
max_len - len(token_ids)), dtype='int32'))
all_segments.append(np.array(segments + [0] * (
max_len - len(segments)), dtype='int32'))
# `valid_lens` excludes count of '<pad>' tokens
valid_lens.append(np.array(len(token_ids), dtype='float32'))
all_pred_positions.append(np.array(pred_positions + [0] * (
max_num_mlm_preds - len(pred_positions)), dtype='int32'))
# Predictions of padded tokens will be filtered out in the loss via
# multiplication of 0 weights
all_mlm_weights.append(
np.array([1.0] * len(mlm_pred_label_ids) + [0.0] * (
max_num_mlm_preds - len(pred_positions)), dtype='float32'))
all_mlm_labels.append(np.array(mlm_pred_label_ids + [0] * (
max_num_mlm_preds - len(mlm_pred_label_ids)), dtype='int32'))
nsp_labels.append(np.array(is_next))
return (all_token_ids, all_segments, valid_lens, all_pred_positions,
all_mlm_weights, all_mlm_labels, nsp_labels)
Putting the helper functions for generating training examples of the two pretraining tasks, and the
helper function for padding inputs together, we customize the following _WikiTextDataset class
as the WikiText-2 dataset for pretraining BERT. By implementing the __getitem__function, we
can arbitrarily access the pretraining (masked language modeling and next sentence prediction)
examples generated from a pair of sentences from the WikiText-2 corpus.

The original BERT model uses WordPiece embeddings whose vocabulary size is 30,000 (Wu et al.,
2016). The tokenization method of WordPiece is a slight modification of the original byte pair
encoding algorithm in Section 14.6.2. For simplicity, we use the d2l.tokenize function for tok-
enization. Infrequent tokens that appear less than five times are filtered out.
#@save
class _WikiTextDataset(gluon.data.Dataset):
def __init__(self, paragraphs, max_len):
# Input `paragraphs[i]` is a list of sentence strings representing a
# paragraph; while output `paragraphs[i]` is a list of sentences
# representing a paragraph, where each sentence is a list of tokens
paragraphs = [d2l.tokenize(
paragraph, token='word') for paragraph in paragraphs]
sentences = [sentence for paragraph in paragraphs
for sentence in paragraph]
self.vocab = d2l.Vocab(sentences, min_freq=5, reserved_tokens=[
'<pad>', '<mask>', '<cls>', '<sep>'])
# Get data for the next sentence prediction task
examples = []
for paragraph in paragraphs:
examples.extend(_get_nsp_data_from_paragraph(
paragraph, paragraphs, self.vocab, max_len))
# Get data for the masked language model task
examples = [(_get_mlm_data_from_tokens(tokens, self.vocab)
+ (segments, is_next))
for tokens, segments, is_next in examples]
# Pad inputs
(self.all_token_ids, self.all_segments, self.valid_lens,
self.all_pred_positions, self.all_mlm_weights,
self.all_mlm_labels, self.nsp_labels) = _pad_bert_inputs(
examples, max_len, self.vocab)

return (self.all_token_ids[idx], self.all_segments[idx],
self.valid_lens[idx], self.all_pred_positions[idx],
self.all_mlm_weights[idx], self.all_mlm_labels[idx],
self.nsp_labels[idx])
def __len__(self):
return len(self.all_token_ids)
By using the _read_wiki function and the _WikiTextDataset class, we define the following
load_data_wiki to download and WikiText-2 dataset and generate pretraining examples from it.
#@save
def load_data_wiki(batch_size, max_len):
data_dir = d2l.download_extract('wikitext-2', 'wikitext-2')
paragraphs = _read_wiki(data_dir)
train_set = _WikiTextDataset(paragraphs, max_len)
train_iter = gluon.data.DataLoader(train_set, batch_size, shuffle=True,
return train_iter, train_set.vocab
Setting the batch size to 512 and the maximum length of a BERT input sequence to be 64, we
print out the shapes of a minibatch of BERT pretraining examples. Note that in each BERT input

sequence, 10 (64 × 0.15) positions are predicted for the masked language modeling task.
batch_size, max_len = 512, 64

train_iter, vocab = load_data_wiki(batch_size, max_len)
for (tokens_X, segments_X, valid_lens_x, pred_positions_X, mlm_weights_X,

mlm_Y, nsp_y) in train_iter:
print(tokens_X.shape, segments_X.shape, valid_lens_x.shape,
pred_positions_X.shape, mlm_weights_X.shape, mlm_Y.shape,
nsp_y.shape)
break
Downloading ../data/wikitext-2-v1.zip from https://s3.amazonaws.com/research.metamind.io/

,→wikitext/wikitext-2-v1.zip...
(512, 64) (512, 64) (512,) (512, 10) (512, 10) (512, 10) (512,)
In the end, let us take a look at the vocabulary size. Even after filtering out infrequent tokens, it is
still over twice larger than that of the PTB dataset.
len(vocab)
20256
Summary
• Comparing with the PTB dataset, the WikiText-2 dateset retains the original punctuation,
case and numbers, and is over twice larger.
• We can arbitrarily access the pretraining (masked language modeling and next sentence
prediction) examples generated from a pair of sentences from the WikiText-2 corpus.
Exercises
1. For simplicity, the period is used as the only delimiter for splitting sentences. Try other sen-
tence splitting techniques, such as the spaCy and NLTK. Take NLTK as an example. You need
to install NLTK first: pip install nltk. In the code, first import nltk. Then, download the
Punkt sentence tokenizer: nltk.download('punkt'). To split sentences such as sentences
= 'This is great ! Why not ?', invoking nltk.tokenize.sent_tokenize(sentences) will
return a list of two sentence strings: ['This is great !', 'Why not ?'].
2. What is the vocabulary size if we do not filter out any infrequent token?
Discussions210
210

14.10 Pretraining BERT
With the BERT model implemented in Section 14.8 and the pretraining examples generated from
the WikiText-2 dataset in Section 14.9, we will pretrain BERT on the WikiText-2 dataset in this
section.

npx.set_np()
To start, we load the WikiText-2 dataset as minibatches of pretraining examples for masked lan-
guage modeling and next sentence prediction. The batch size is 512 and the maximum length of
a BERT input sequence is 64. Note that in the original BERT model, the maximum length is 512.
batch_size, max_len = 512, 64

train_iter, vocab = d2l.load_data_wiki(batch_size, max_len)
14.10.1 Pretraining BERT
The original BERT has two versions of different model sizes (Devlin et al., 2018). The base model
(BERTBASE ) uses 12 layers (Transformer encoder blocks) with 768 hidden units (hidden size) and
12 self-attention heads. The large model (BERTLARGE ) uses 24 layers with 1024 hidden units and
16 self-attention heads. Notably, the former has 110 million parameters while the latter has 340
million parameters. For demonstration with ease, we define a small BERT, using 2 layers, 128
hidden units, and 2 self-attention heads.
net = d2l.BERTModel(len(vocab), num_hiddens=128, ffn_num_hiddens=256,

num_heads=2, num_layers=2, dropout=0.2)
net.initialize(init.Xavier(), ctx=devices)
loss = gluon.loss.SoftmaxCELoss()
Before defining the training loop, we define a helper function _get_batch_loss_bert. Given the
shard of training examples, this function computes the loss for both the masked language model-
ing and next sentence prediction tasks. Note that the final loss of BERT pretraining is just the sum
of both the masked language modeling loss and the next sentence prediction loss.
#@save
def _get_batch_loss_bert(net, loss, vocab_size, tokens_X_shards,
segments_X_shards, valid_lens_x_shards,
pred_positions_X_shards, mlm_weights_X_shards,
mlm_Y_shards, nsp_y_shards):
mlm_ls, nsp_ls, ls = [], [], []
for (tokens_X_shard, segments_X_shard, valid_lens_x_shard,
pred_positions_X_shard, mlm_weights_X_shard, mlm_Y_shard,
nsp_y_shard) in zip(
tokens_X_shards, segments_X_shards, valid_lens_x_shards,
pred_positions_X_shards, mlm_weights_X_shards, mlm_Y_shards,
nsp_y_shards):
# Forward pass
14.10. Pretraining BERT 677

_, mlm_Y_hat, nsp_Y_hat = net(
tokens_X_shard, segments_X_shard, valid_lens_x_shard.reshape(-1),
pred_positions_X_shard)
# Compute masked language model loss
mlm_l = loss(
mlm_Y_hat.reshape((-1, vocab_size)), mlm_Y_shard.reshape(-1),
mlm_weights_X_shard.reshape((-1, 1)))
mlm_l = mlm_l.sum() / (mlm_weights_X_shard.sum() + 1e-8)
# Compute next sentence prediction loss
nsp_l = loss(nsp_Y_hat, nsp_y_shard)
nsp_l = nsp_l.mean()
mlm_ls.append(mlm_l)
nsp_ls.append(nsp_l)
ls.append(mlm_l + nsp_l)
npx.waitall()
return mlm_ls, nsp_ls, ls
Invoking the two aforementioned helper functions, the following train_bert function defines the
procedure to pretrain BERT (net) on the WikiText-2 (train_iter) dataset. Training BERT can take
very long. Instead of specifying the number of epochs for training as in the train_ch13 function
(see Section 13.1), the input num_steps of the following function specifies the number of iteration
steps for training.
#@save
def train_bert(train_iter, net, loss, vocab_size, devices, log_interval,
num_steps):
{'learning_rate': 1e-3})
step, timer = 0, d2l.Timer()
animator = d2l.Animator(xlabel='step', ylabel='loss',
xlim=[1, num_steps], legend=['mlm', 'nsp'])
# Sum of masked language modeling losses, sum of next sentence prediction
# losses, no. of sentence pairs, count
num_steps_reached = False
while step < num_steps and not num_steps_reached:
for batch in train_iter:
(tokens_X_shards, segments_X_shards, valid_lens_x_shards,
pred_positions_X_shards, mlm_weights_X_shards,
mlm_Y_shards, nsp_y_shards) = [gluon.utils.split_and_load(
elem, devices, even_split=False) for elem in batch]
timer.start()
mlm_ls, nsp_ls, ls = _get_batch_loss_bert(
net, loss, vocab_size, tokens_X_shards, segments_X_shards,
valid_lens_x_shards, pred_positions_X_shards,
mlm_weights_X_shards, mlm_Y_shards, nsp_y_shards)
for l in ls:
l.backward()
trainer.step(1)
mlm_l_mean = sum([float(l) for l in mlm_ls]) / len(mlm_ls)
nsp_l_mean = sum([float(l) for l in nsp_ls]) / len(nsp_ls)
metric.add(mlm_l_mean, nsp_l_mean, batch[0].shape[0], 1)
timer.stop()

if (step + 1) % log_interval == 0:
animator.add(step + 1,
(metric[0] / metric[3], metric[1] / metric[3]))
step += 1
if step == num_steps:
num_steps_reached = True
break
print(f'MLM loss {metric[0] / metric[3]:.3f}, '

f'NSP loss {metric[1] / metric[3]:.3f}')
print(f'{metric[2] / timer.sum():.1f} sentence pairs/sec on '
f'{str(devices)}')
We can plot both the masked language modeling loss and the next sentence prediction loss during
BERT pretraining.
train_bert(train_iter, net, loss, len(vocab), devices, 1, 50)
MLM loss 7.906, NSP loss 0.732

8803.2 sentence pairs/sec on [gpu(0), gpu(1)]
14.10.2 Representing Text with BERT
After pretraining BERT, we can use it to represent single text, text pairs, or any token in them. The
following function returns the BERT (net) representations for all tokens in tokens_a and tokens_b.
def get_bert_encoding(net, tokens_a, tokens_b=None):

tokens, segments = d2l.get_tokens_and_segments(tokens_a, tokens_b)
token_ids = np.expand_dims(np.array(vocab[tokens], ctx=devices[0]),
axis=0)
segments = np.expand_dims(np.array(segments, ctx=devices[0]), axis=0)
valid_len = np.expand_dims(np.array(len(tokens), ctx=devices[0]), axis=0)
encoded_X, _, _ = net(token_ids, segments, valid_len)
return encoded_X

Consider the sentence “a crane is flying”. Recall the input representation of BERT as discussed in
Section 14.8.4. After inserting special tokens “<cls>” (used for classification) and “<sep>” (used for
separation), the BERT input sequence has a length of six. Since zero is the index of the “<cls>” to-
ken, encoded_text[:, 0, :] is the BERT representation of the entire input sentence. To evaluate
the polysemy token “crane”, we also print out the first three elements of the BERT representation
of the token.
tokens_a = ['a', 'crane', 'is', 'flying']

encoded_text = get_bert_encoding(net, tokens_a)
# Tokens: '<cls>', 'a', 'crane', 'is', 'flying', '<sep>'
encoded_text_cls = encoded_text[:, 0, :]
encoded_text_crane = encoded_text[:, 2, :]
encoded_text.shape, encoded_text_cls.shape, encoded_text_crane[0][:3]
((1, 6, 128),
(1, 128),
array([ 0.8918758, 1.0964735, -1.0611488], ctx=gpu(0)))
Now consider a sentence pair “a crane driver came” and “he just left”. Similarly, encoded_pair[:,
0, :] is the encoded result of the entire sentence pair from the pretrained BERT. Note that the
first three elements of the polysemy token “crane” are different from those when the context is
different. This supports that BERT representations are context-sensitive.
tokens_a, tokens_b = ['a', 'crane', 'driver', 'came'], ['he', 'just', 'left']

encoded_pair = get_bert_encoding(net, tokens_a, tokens_b)
# Tokens: '<cls>', 'a', 'crane', 'driver', 'came', '<sep>', 'he', 'just',
# 'left', '<sep>'
encoded_pair_cls = encoded_pair[:, 0, :]
encoded_pair_crane = encoded_pair[:, 2, :]
encoded_pair.shape, encoded_pair_cls.shape, encoded_pair_crane[0][:3]
((1, 10, 128),

(1, 128),
array([ 0.91807 , 1.1650459, -1.0900582], ctx=gpu(0)))
In Chapter 15, we will fine-tune a pretrained BERT model for downstream natural language pro-
cessing applications.
Summary
• The original BERT has two versions, where the base model has 110 million parameters and
the large model has 340 million parameters.
• After pretraining BERT, we can use it to represent single text, text pairs, or any token in them.
• In the experiment, the same token has different BERT representation when their contexts
are different. This supports that BERT representations are context-sensitive.

Exercises
1. In the experiment, we can see that the masked language modeling loss is significantly higher
than the next sentence prediction loss. Why?
2. Set the maximum length of a BERT input sequence to be 512 (same as the original BERT
model). Use the configurations of the original BERT model such as BERTLARGE . Do you en-
counter any error when running this section? Why?
Discussions211
211

15 | Natural Language Processing: Ap-
plications
We have seen how to represent text tokens and train their representations in Chapter 14. Such
pretrained text representations can be fed to various models for different downstream natural
language processing tasks.
This book does not intend to cover natural language processing applications in a comprehensive
manner. Our focus is on how to apply (deep) representation learning of languages to addressing natu-
ral language processing problems. Nonetheless, we have already discussed several natural language
processing applications without pretraining in earlier chapters, just for explaining deep learning
architectures. For instance, in Chapter 8, we have relied on RNNs to design language models to
generate novella-like text. In Chapter 9 and Chapter 10, we have also designed models based on
RNNs and attention mechanisms for machine translation. Given pretrained text representations,
in this chapter, we will consider two more downstream natural language processing tasks: sen-
timent analysis and natural language inference. These are popular and representative natural
language processing applications: the former analyzes single text and the latter analyzes relation-
ships of text pairs.
Fig. 15.1: Pretrained text representations can be fed to various deep learning architectures for
different downstream natural language processing applications. This chapter focuses on how to
design models for different downstream natural language processing applications.
As depicted in Fig. 15.1, this chapter focuses on describing the basic ideas of designing natural
language processing models using different types of deep learning architectures, such as MLPs,
CNNs, RNNs, and attention. Though it is possible to combine any pretrained text representations
683
with any architecture for either downstream natural language processing task in Fig. 15.1, we
select a few representative combinations. Specifically, we will explore popular architectures based
on RNNs and CNNs for sentiment analysis. For natural language inference, we choose attention
and MLPs to demonstrate how to analyze text pairs. In the end, we introduce how to fine-tune
a pretrained BERT model for a wide range of natural language processing applications, such as
on a sequence level (single text classification and text pair classification) and a token level (text
tagging and question answering). As a concrete empirical case, we will fine-tune BERT for natural
language processing.
As we have introduced in Section 14.8, BERT requires minimal architecture changes for a wide
range of natural language processing applications. However, this benefit comes at the cost of fine-
tuning a huge number of BERT parameters for the downstream applications. When space or time
is limited, those crafted models based on MLPs, CNNs, RNNs, and attention are more feasible. In
the following, we start by the sentiment analysis application and illustrate the model design based
on RNNs and CNNs, respectively.
15.1 Sentiment Analysis and the Dataset
Text classification is a common task in natural language processing, which transforms a sequence
of text of indefinite length into a category of text. It is similar to the image classification, the most
frequently used application in this book, e.g., Section 18.9. The only difference is that, rather than
an image, text classificationʼs example is a text sentence.
This section will focus on loading data for one of the sub-questions in this field: using text senti-
ment classification to analyze the emotions of the textʼs author. This problem is also called senti-
ment analysis and has a wide range of applications. For example, we can analyze user reviews of
products to obtain user satisfaction statistics, or analyze user sentiments about market conditions
and use it to predict future trends.

import os
npx.set_np()
15.1.1 The Sentiment Analysis Dataset
We use Stanfordʼs Large Movie Review Dataset212 as the dataset for sentiment analysis. This
dataset is divided into two datasets for training and testing purposes, each containing 25,000 movie
reviews downloaded from IMDb. In each dataset, the number of comments labeled as “positive”
and “negative” is equal.
212
https://ai.stanford.edu/~amaas/data/sentiment/
684 Chapter 15. Natural Language Processing: Applications

Reading the Dataset
We first download this dataset to the “../data” path and extract it to “../data/aclImdb”.
#@save
d2l.DATA_HUB['aclImdb'] = (
'http://ai.stanford.edu/~amaas/data/sentiment/aclImdb_v1.tar.gz',
'01ada507287d82875905620988597833ad4e0903')
data_dir = d2l.download_extract('aclImdb', 'aclImdb')
Downloading ../data/aclImdb_v1.tar.gz from http://ai.stanford.edu/~amaas/data/sentiment/

,→aclImdb_v1.tar.gz...
Next, read the training and test datasets. Each example is a review and its corresponding label: 1
indicates “positive” and 0 indicates “negative”.
#@save
def read_imdb(data_dir, is_train):
data, labels = [], []
for label in ('pos', 'neg'):
folder_name = os.path.join(data_dir, 'train' if is_train else 'test',
label)
for file in os.listdir(folder_name):
with open(os.path.join(folder_name, file), 'rb') as f:
review = f.read().decode('utf-8').replace('\n', '')
data.append(review)
labels.append(1 if label == 'pos' else 0)
return data, labels
train_data = read_imdb(data_dir, is_train=True)

print('# trainings:', len(train_data[0]))
for x, y in zip(train_data[0][:3], train_data[1][:3]):
print('label:', y, 'review:', x[0:60])
# trainings: 25000
label: 1 review: Normally the best way to annoy me in a film is to include so
label: 1 review: The Bible teaches us that the love of money is the root of a
label: 1 review: Being someone who lists Night of the Living Dead at number t
Tokenization and Vocabulary
We use a word as a token, and then create a dictionary based on the training dataset.
train_tokens = d2l.tokenize(train_data[0], token='word')

vocab = d2l.Vocab(train_tokens, min_freq=5, reserved_tokens=['<pad>'])
d2l.set_figsize()
d2l.plt.hist([len(line) for line in train_tokens], bins=range(0, 1000, 50));
15.1. Sentiment Analysis and the Dataset 685

Padding to the Same Length
Because the reviews have different lengths, so they cannot be directly combined into minibatches.
Here we fix the length of each comment to 500 by truncating or adding “<unk>” indices.
num_steps = 500 # sequence length

train_features = np.array([d2l.truncate_pad(
vocab[line], num_steps, vocab['<pad>']) for line in train_tokens])
train_features.shape
(25000, 500)
Creating the Data Iterator
Now, we will create a data iterator. Each iteration will return a minibatch of data.
train_iter = d2l.load_array((train_features, train_data[1]), 64)
print('X', X.shape, 'y', y.shape)
break
'# batches:', len(train_iter)
X (64, 500) y (64,)
('# batches:', 391)

Last, we will save a function load_data_imdb into d2l, which returns the vocabulary and data iter-
ators.
#@save
def load_data_imdb(batch_size, num_steps=500):
data_dir = d2l.download_extract('aclImdb', 'aclImdb')
train_data = read_imdb(data_dir, True)
test_data = read_imdb(data_dir, False)
train_tokens = d2l.tokenize(train_data[0], token='word')
test_tokens = d2l.tokenize(test_data[0], token='word')
vocab = d2l.Vocab(train_tokens, min_freq=5)
train_features = np.array([d2l.truncate_pad(
vocab[line], num_steps, vocab.unk) for line in train_tokens])
test_features = np.array([d2l.truncate_pad(
vocab[line], num_steps, vocab.unk) for line in test_tokens])
train_iter = d2l.load_array((train_features, train_data[1]), batch_size)
test_iter = d2l.load_array((test_features, test_data[1]), batch_size,
is_train=False)
return train_iter, test_iter, vocab
Summary
• Text classification can classify a text sequence into a category.

• To classify a text sentiment, we load an IMDb dataset and tokenize its words. Then we pad
the text sequence for short reviews and create a data iterator.
Exercises
1. Discover a different natural language dataset (such as Amazon reviews213 ) and build a similar
data_loader function as load_data_imdb.
Discussions214
15.2 Sentiment Analysis: Using Recurrent Neural Networks
Similar to search synonyms and analogies, text classification is also a downstream application of
word embedding. In this section, we will apply pre-trained word vectors (GloVe) and bidirectional
recurrent neural networks with multiple hidden layers (Maas et al., 2011), as shown in Fig. 15.2.1.
We will use the model to determine whether a text sequence of indefinite length contains positive
or negative emotion.
213
https://snap.stanford.edu/data/web-Amazon.html
214
15.2. Sentiment Analysis: Using Recurrent Neural Networks 687

Fig. 15.2.1: This section feeds pretrained GloVe to an RNN-based architecture for sentiment anal-
ysis.

npx.set_np()
batch_size = 64
train_iter, test_iter, vocab = d2l.load_data_imdb(batch_size)
15.2.1 Using a Recurrent Neural Network Model
In this model, each word first obtains a feature vector from the embedding layer. Then, we further
encode the feature sequence using a bidirectional recurrent neural network to obtain sequence
information. Finally, we transform the encoded sequence information to output through the fully
connected layer. Specifically, we can concatenate hidden states of bidirectional long-short term
memory in the initial timestep and final timestep and pass it to the output layer classification
as encoded feature sequence information. In the BiRNN class implemented below, the Embedding
instance is the embedding layer, the LSTM instance is the hidden layer for sequence encoding, and
the Dense instance is the output layer for generated classification results.
class BiRNN(nn.Block):
def __init__(self, vocab_size, embed_size, num_hiddens,
num_layers, **kwargs):
super(BiRNN, self).__init__(**kwargs)
# Set `bidirectional` to True to get a bidirectional recurrent neural
# network
self.encoder = rnn.LSTM(num_hiddens, num_layers=num_layers,
bidirectional=True, input_size=embed_size)
self.decoder = nn.Dense(2)
def forward(self, inputs):

# The shape of ìnputs` is (batch size, no. of words). Because LSTM

# needs to use sequence as the first dimension, the input is
# transformed and the word feature is then extracted. The output shape
# is (no. of words, batch size, word vector dimension).
embeddings = self.embedding(inputs.T)
# Since the input (embeddings) is the only argument passed into
# rnn.LSTM, it only returns the hidden states of the last hidden layer
# at different timestep (outputs). The shape of òutputs` is
# (no. of words, batch size, 2 * no. of hidden units).
outputs = self.encoder(embeddings)
# Concatenate the hidden states of the initial timestep and final
# timestep to use as the input of the fully connected layer. Its
# shape is (batch size, 4 * no. of hidden units)
encoding = np.concatenate((outputs[0], outputs[-1]), axis=1)
outs = self.decoder(encoding)
return outs
Create a bidirectional recurrent neural network with two hidden layers.
embed_size, num_hiddens, num_layers, devices = 100, 100, 2, d2l.try_all_gpus()

net = BiRNN(len(vocab), embed_size, num_hiddens, num_layers)
Loading Pre-trained Word Vectors
Because the training dataset for sentiment classification is not very large, in order to deal with
overfitting, we will directly use word vectors pre-trained on a larger corpus as the feature vectors
of all words. Here, we load a 100-dimensional GloVe word vector for each word in the dictionary
vocab.
glove_embedding = d2l.TokenEmbedding('glove.6b.100d')
Query the word vectors that in our vocabulary.
embeds = glove_embedding[vocab.idx_to_token]
embeds.shape
(49339, 100)
Then, we will use these word vectors as feature vectors for each word in the reviews. Note that
the dimensions of the pre-trained word vectors need to be consistent with the embedding layer
output size embed_size in the created model. In addition, we no longer update these word vectors
during training.
net.embedding.weight.set_data(embeds)
net.embedding.collect_params().setattr('grad_req', 'null')
15.2. Sentiment Analysis: Using Recurrent Neural Networks 689

Training and Evaluating the Model
Now, we can start training.

trainer = gluon.Trainer(net.collect_params(), 'adam', {'learning_rate': lr})

Finally, define the prediction function.
#@save
def predict_sentiment(net, vocab, sentence):
sentence = np.array(vocab[sentence.split()], ctx=d2l.try_gpu())
label = np.argmax(net(sentence.reshape(1, -1)), axis=1)
return 'positive' if label == 1 else 'negative'
Then, use the trained model to classify the sentiments of two simple sentences.
predict_sentiment(net, vocab, 'this movie is so great')
'positive'
predict_sentiment(net, vocab, 'this movie is so bad')
'negative'

Summary
• Text classification transforms a sequence of text of indefinite length into a category of text.
This is a downstream application of word embedding.
• We can apply pre-trained word vectors and recurrent neural networks to classify the emo-
tions in a text.
Exercises
1. Increase the number of epochs. What accuracy rate can you achieve on the training and
testing datasets? What about trying to re-tune other hyperparameters?
2. Will using larger pre-trained word vectors, such as 300-dimensional GloVe word vectors, im-
prove classification accuracy?
3. Can we improve the classification accuracy by using the spaCy word tokenization tool? You
need to install spaCy: pip install spacy and install the English package: python -m spacy
download en. In the code, first import spacy: import spacy. Then, load the spacy English
package: spacy_en = spacy.load('en'). Finally, define the function def tokenizer(text):
return [tok.text for tok in spacy_en.tokenizer(text)] and replace the original tok-
enizer function. It should be noted that GloVeʼs word vector uses “-” to connect each word
when storing noun phrases. For example, the phrase “new york” is represented as “new-
york” in GloVe. After using spaCy tokenization, “new york” may be stored as “new york”.
Discussions215
15.3 Sentiment Analysis: Using Convolutional Neural Networks
In Chapter 6, we explored how to process two-dimensional image data with two-dimensional con-
volutional neural networks. In the previous language models and text classification tasks, we
treated text data as a time series with only one dimension, and naturally, we used recurrent neu-
ral networks to process such data. In fact, we can also treat text as a one-dimensional image, so
that we can use one-dimensional convolutional neural networks to capture associations between
adjacent words. As described in .. _fig_nlp-map-sa-cnn: This section describes a groundbreaking
approach to applying convolutional neural networks to sentiment analysis: textCNN (Kim, 2014).
215
15.3. Sentiment Analysis: Using Convolutional Neural Networks 691

Fig. 15.3.1: This section feeds pretrained GloVe to a CNN-based architecture for sentiment analy-
sis.
First, import the packages and modules required for the experiment.

npx.set_np()
batch_size = 64
train_iter, test_iter, vocab = d2l.load_data_imdb(batch_size)
15.3.1 One-Dimensional Convolutional Layer
Before introducing the model, let us explain how a one-dimensional convolutional layer works.
Like a two-dimensional convolutional layer, a one-dimensional convolutional layer uses a one-
dimensional cross-correlation operation. In the one-dimensional cross-correlation operation, the
convolution window starts from the leftmost side of the input array and slides on the input array
from left to right successively. When the convolution window slides to a certain position, the
input subarray in the window and kernel array are multiplied and summed by element to get the
element at the corresponding location in the output array. As shown in Fig. 15.3.2, the input is a
one-dimensional array with a width of 7 and the width of the kernel array is 2. As we can see, the
output width is 7 − 2 + 1 = 6 and the first element is obtained by performing multiplication by
element on the leftmost input subarray with a width of 2 and kernel array and then summing the
results.
Fig. 15.3.2: One-dimensional cross-correlation operation. The shaded parts are the first output
element as well as the input and kernel array elements used in its calculation: 0 × 1 + 1 × 2 = 2.

Next, we implement one-dimensional cross-correlation in the corr1d function. It accepts the in-
put array X and kernel array K and outputs the array Y.
def corr1d(X, K):

w = K.shape[0]
Y = np.zeros((X.shape[0] - w + 1))
Y[i] = (X[i: i + w] * K).sum()
return Y
Now, we will reproduce the results of the one-dimensional cross-correlation operation in Fig.
15.3.2.
X, K = np.array([0, 1, 2, 3, 4, 5, 6]), np.array([1, 2])

corr1d(X, K)
array([ 2., 5., 8., 11., 14., 17.])
The one-dimensional cross-correlation operation for multiple input channels is also similar to
the two-dimensional cross-correlation operation for multiple input channels. On each channel,
it performs the one-dimensional cross-correlation operation on the kernel and its corresponding
input and adds the results of the channels to get the output. Fig. 15.3.3 shows a one-dimensional
cross-correlation operation with three input channels.
Fig. 15.3.3: One-dimensional cross-correlation operation with three input channels. The shaded
parts are the first output element as well as the input and kernel array elements used in its calcu-
lation: 0 × 1 + 1 × 2 + 1 × 3 + 2 × 4 + 2 × (−1) + 3 × (−3) = 2.
Now, we reproduce the results of the one-dimensional cross-correlation operation with multi-
input channel in Fig. 15.3.3.
def corr1d_multi_in(X, K):

# First, we traverse along the 0th dimension (channel dimension) of `X`
# and `K`. Then, we add them together by using * to turn the result list
# into a positional argument of the àdd_n` function
return sum(corr1d(x, k) for x, k in zip(X, K))
X = np.array([[0, 1, 2, 3, 4, 5, 6],
[1, 2, 3, 4, 5, 6, 7],
[2, 3, 4, 5, 6, 7, 8]])
K = np.array([[1, 2], [3, 4], [-1, -3]])
corr1d_multi_in(X, K)

array([ 2., 8., 14., 20., 26., 32.])
The definition of a two-dimensional cross-correlation operation tells us that a one-dimensional

cross-correlation operation with multiple input channels can be regarded as a two-dimensional
cross-correlation operation with a single input channel. As shown in Fig. 15.3.4, we can also
present the one-dimensional cross-correlation operation with multiple input channels in Fig.
15.3.3 as the equivalent two-dimensional cross-correlation operation with a single input channel.
Here, the height of the kernel is equal to the height of the input.
Fig. 15.3.4: Two-dimensional cross-correlation operation with a single input channel. The high-
lighted parts are the first output element and the input and kernel array elements used in its cal-
culation: 2 × (−1) + 3 × (−3) + 1 × 3 + 2 × 4 + 0 × 1 + 1 × 2 = 2.
Both the outputs in Fig. 15.3.2 and Fig. 15.3.3 have only one channel. We discussed how to specify
multiple output channels in a two-dimensional convolutional layer in Section 6.4. Similarly, we
can also specify multiple output channels in the one-dimensional convolutional layer to extend
the model parameters in the convolutional layer.
15.3.2 Max-Over-Time Pooling Layer
Similarly, we have a one-dimensional pooling layer. The max-over-time pooling layer used in
TextCNN actually corresponds to a one-dimensional global maximum pooling layer. Assum-
ing that the input contains multiple channels, and each channel consists of values on different
timesteps, the output of each channel will be the largest value of all timesteps in the channel.
Therefore, the input of the max-over-time pooling layer can have different timesteps on each chan-
nel.
To improve computing performance, we often combine timing examples of different lengths into
a minibatch and make the lengths of each timing example in the batch consistent by appending
special characters (such as 0) to the end of shorter examples. Naturally, the added special char-
acters have no intrinsic meaning. Because the main purpose of the max-over-time pooling layer
is to capture the most important features of timing, it usually allows the model to be unaffected
by the manually added characters.

15.3.3 The TextCNN Model
TextCNN mainly uses a one-dimensional convolutional layer and max-over-time pooling layer.
Suppose the input text sequence consists of n words, and each word is represented by a d-
dimension word vector. Then the input example has a width of n, a height of 1, and d input
channels. The calculation of textCNN can be mainly divided into the following steps:
1. Define multiple one-dimensional convolution kernels and use them to perform convolution
calculations on the inputs. Convolution kernels with different widths may capture the cor-
relation of different numbers of adjacent words.
2. Perform max-over-time pooling on all output channels, and then concatenate the pooling
output values of these channels in a vector.
3. The concatenated vector is transformed into the output for each category through the fully
connected layer. A dropout layer can be used in this step to deal with overfitting.
Fig. 15.3.5: TextCNN design.
Fig. 15.3.5 gives an example to illustrate the textCNN. The input here is a sentence with 11 words,
with each word represented by a 6-dimensional word vector. Therefore, the input sequence has a
width of 11 and 6 input channels. We assume there are two one-dimensional convolution kernels
with widths of 2 and 4, and 4 and 5 output channels, respectively. Therefore, after one-dimensional
convolution calculation, the width of the four output channels is 11 − 2 + 1 = 10, while the width
of the other five channels is 11 − 4 + 1 = 8. Even though the width of each channel is different, we
can still perform max-over-time pooling for each channel and concatenate the pooling outputs of
the 9 channels into a 9-dimensional vector. Finally, we use a fully connected layer to transform

the 9-dimensional vector into a 2-dimensional output: positive sentiment and negative sentiment
predictions.
Next, we will implement a textCNN model. Compared with the previous section, in addition to
replacing the recurrent neural network with a one-dimensional convolutional layer, here we use
two embedding layers, one with a fixed weight and another that participates in training.
class TextCNN(nn.Block):
def __init__(self, vocab_size, embed_size, kernel_sizes, num_channels,
**kwargs):
super(TextCNN, self).__init__(**kwargs)
# The embedding layer does not participate in training
self.constant_embedding = nn.Embedding(vocab_size, embed_size)
self.dropout = nn.Dropout(0.5)
self.decoder = nn.Dense(2)
# The max-over-time pooling layer has no weight, so it can share an
# instance
self.pool = nn.GlobalMaxPool1D()
# Create multiple one-dimensional convolutional layers
self.convs = nn.Sequential()
for c, k in zip(num_channels, kernel_sizes):
self.convs.add(nn.Conv1D(c, k, activation='relu'))

# Concatenate the output of two embedding layers with shape of
# (batch size, no. of words, word vector dimension) by word vector
embeddings = np.concatenate((
self.embedding(inputs), self.constant_embedding(inputs)), axis=2)
# According to the input format required by Conv1D, the word vector
# dimension, that is, the channel dimension of the one-dimensional
# convolutional layer, is transformed into the previous dimension
embeddings = embeddings.transpose(0, 2, 1)
# For each one-dimensional convolutional layer, after max-over-time
# pooling, an ndarray with the shape of (batch size, channel size, 1)
# can be obtained. Use the flatten function to remove the last
# dimension and then concatenate on the channel dimension
encoding = np.concatenate([
np.squeeze(self.pool(conv(embeddings)), axis=-1)
for conv in self.convs], axis=1)
# After applying the dropout method, use a fully connected layer to
# obtain the output
outputs = self.decoder(self.dropout(encoding))
return outputs
Create a TextCNN instance. It has 3 convolutional layers with kernel widths of 3, 4, and 5, all with
100 output channels.
embed_size, kernel_sizes, nums_channels = 100, [3, 4, 5], [100, 100, 100]

net = TextCNN(len(vocab), embed_size, kernel_sizes, nums_channels)

Load Pre-trained Word Vectors
As in the previous section, load pre-trained 100-dimensional GloVe word vectors and initialize the
embedding layers embedding and constant_embedding. Here, the former participates in training
while the latter has a fixed weight.
net.constant_embedding.weight.set_data(embeds)
net.constant_embedding.collect_params().setattr('grad_req', 'null')
Train and Evaluate the Model
Now we can train the model.


Below, we use the trained model to classify sentiments of two simple sentences.
d2l.predict_sentiment(net, vocab, 'this movie is so great')
'negative'
d2l.predict_sentiment(net, vocab, 'this movie is so bad')

'negative'
Summary
• We can use one-dimensional convolution to process and analyze timing data.

• A one-dimensional cross-correlation operation with multiple input channels can be re-
garded as a two-dimensional cross-correlation operation with a single input channel.
• The input of the max-over-time pooling layer can have different numbers of timesteps on
each channel.
• TextCNN mainly uses a one-dimensional convolutional layer and max-over-time pooling
layer.
Exercises
1. Tune the hyperparameters and compare the two sentiment analysis methods, using recur-
rent neural networks and using convolutional neural networks, as regards accuracy and op-
erational efficiency.
2. Can you further improve the accuracy of the model on the test set by using the three methods
introduced in the previous section: tuning hyperparameters, using larger pre-trained word
vectors, and using the spaCy word tokenization tool?
3. What other natural language processing tasks can you use textCNN for?
Discussions216
15.4 Natural Language Inference and the Dataset
In Section 15.1, we discussed the problem of sentiment analysis. This task aims to classify a sin-
gle text sequence into predefined categories, such as a set of sentiment polarities. However, when
there is a need to decide whether one sentence can be inferred form another, or eliminate redun-
dancy by identifying sentences that are semantically equivalent, knowing how to classify one text
sequence is insufficient. Instead, we need to be able to reason over pairs of text sequences.
15.4.1 Natural Language Inference
Natural language inference studies whether a hypothesis can be inferred from a premise, where both
are a text sequence. In other words, natural language inference determines the logical relation-
ship between a pair of text sequences. Such relationships usually fall into three types:
• Entailment: the hypothesis can be inferred from the premise.
• Contradiction: the negation of the hypothesis can be inferred from the premise.
• Neutral: all the other cases.
216

Natural language inference is also known as the recognizing textual entailment task. For example,
the following pair will be labeled as entailment because “showing affection” in the hypothesis can
be inferred from “hugging one another” in the premise.
Premise: Two women are hugging each other.
Hypothesis: Two women are showing affection.
The following is an example of contradiction as “running the coding example” indicates “not sleep-
ing” rather than “sleeping”.
Premise: A man is running the coding example from Dive into Deep Learning.
Hypothesis: The man is sleeping.
The third example shows a neutrality relationship because neither “famous” nor “not famous” can
be inferred from the fact that “are performing for us”.
Premise: The musicians are performing for us.
Hypothesis: The musicians are famous.
Natural language inference has been a central topic for understanding natural language. It en-
joys wide applications ranging from information retrieval to open-domain question answering.
To study this problem, we will begin by investigating a popular natural language inference bench-
mark dataset.
15.4.2 The Stanford Natural Language Inference (SNLI) Dataset
Stanford Natural Language Inference (SNLI) Corpus is a collection of over 500, 000 labeled English
sentence pairs (Bowman et al., 2015). We download and store the extracted SNLI dataset in the
path ../data/snli_1.0.
import collections
import os
import re
import zipfile
npx.set_np()
#@save
d2l.DATA_HUB['SNLI'] = (
'https://nlp.stanford.edu/projects/snli/snli_1.0.zip',
'9fcde07509c7e87ec61c640c1b2753d9041758e4')
data_dir = d2l.download_extract('SNLI')
Downloading ../data/snli_1.0.zip from https://nlp.stanford.edu/projects/snli/snli_1.0.zip...
15.4. Natural Language Inference and the Dataset 699

Reading the Dataset
The original SNLI dataset contains much richer information than what we really need in our ex-
periments. Thus, we define a function read_snli to only extract part of the dataset, then return
lists of premises, hypotheses, and their labels.
#@save
def read_snli(data_dir, is_train):
"""Read the SNLI dataset into premises, hypotheses, and labels."""
def extract_text(s):
# Remove information that will not be used by us
s = re.sub('\$', '', s)
s = re.sub('\$', '', s)
# Substitute two or more consecutive whitespace with space
s = re.sub('\\s{2,}', ' ', s)
return s.strip()
label_set = {'entailment': 0, 'contradiction': 1, 'neutral': 2}
file_name = os.path.join(data_dir, 'snli_1.0_train.txt'
if is_train else 'snli_1.0_test.txt')
with open(file_name, 'r') as f:
rows = [row.split('\t') for row in f.readlines()[1:]]
premises = [extract_text(row[1]) for row in rows if row[0] in label_set]
hypotheses = [extract_text(row[2]) for row in rows if row[0] in label_set]
labels = [label_set[row[0]] for row in rows if row[0] in label_set]
return premises, hypotheses, labels
Now let us print the first 3 pairs of premise and hypothesis, as well as their labels (“0”, “1”, and “2”
correspond to “entailment”, “contradiction”, and “neutral”, respectively ).
train_data = read_snli(data_dir, is_train=True)

for x0, x1, y in zip(train_data[0][:3], train_data[1][:3], train_data[2][:3]):
print('premise:', x0)
print('hypothesis:', x1)
print('label:', y)
premise: A person on a horse jumps over a broken down airplane .

hypothesis: A person is training his horse for a competition .
label: 2
hypothesis: A person is at a diner , ordering an omelette .
label: 1
hypothesis: A person is outdoors , on a horse .
label: 0
The training set has about 550, 000 pairs, and the testing set has about 10, 000 pairs. The following
shows that the three labels “entailment”, “contradiction”, and “neutral” are balanced in both the
training set and the testing set.
test_data = read_snli(data_dir, is_train=False)

for data in [train_data, test_data]:
print([[row for row in data[2]].count(i) for i in range(3)])

[183416, 183187, 182764]
[3368, 3237, 3219]
Defining a Class for Loading the Dataset
Below we define a class for loading the SNLI dataset by inheriting from the Dataset class in Gluon.
The argument num_steps in the class constructor specifies the length of a text sequence so that
each minibatch of sequences will have the same shape. In other words, tokens after the first
num_steps ones in longer sequence are trimmed, while special tokens “<pad>” will be appended
to shorter sequences until their length becomes num_steps. By implementing the __getitem__
function, we can arbitrarily access the premise, hypothesis, and label with the index idx.
#@save
class SNLIDataset(gluon.data.Dataset):
"""A customized dataset to load the SNLI dataset."""
def __init__(self, dataset, num_steps, vocab=None):
self.num_steps = num_steps
all_premise_tokens = d2l.tokenize(dataset[0])
all_hypothesis_tokens = d2l.tokenize(dataset[1])
if vocab is None:
self.vocab = d2l.Vocab(all_premise_tokens + all_hypothesis_tokens,
min_freq=5, reserved_tokens=['<pad>'])
else:
self.vocab = vocab
self.premises = self._pad(all_premise_tokens)
self.hypotheses = self._pad(all_hypothesis_tokens)
self.labels = np.array(dataset[2])
print('read ' + str(len(self.premises)) + ' examples')
def _pad(self, lines):

return np.array([d2l.truncate_pad(
self.vocab[line], self.num_steps, self.vocab['<pad>'])
for line in lines])

return (self.premises[idx], self.hypotheses[idx]), self.labels[idx]
def __len__(self):
return len(self.premises)
Now we can invoke the read_snli function and the SNLIDataset class to download the SNLI dataset
and return DataLoader instances for both training and testing sets, together with the vocabulary
of the training set. It is noteworthy that we must use the vocabulary constructed from the training
set as that of the testing set. As a result, any new token from the testing set will be unknown to
the model trained on the training set.
#@save
def load_data_snli(batch_size, num_steps=50):
15.4. Natural Language Inference and the Dataset 701

"""Download the SNLI dataset and return data iterators and vocabulary."""
train_data = read_snli(data_dir, True)
test_data = read_snli(data_dir, False)
train_set = SNLIDataset(train_data, num_steps)
test_set = SNLIDataset(test_data, num_steps, train_set.vocab)
test_iter = gluon.data.DataLoader(test_set, batch_size, shuffle=False,
return train_iter, test_iter, train_set.vocab
Here we set the batch size to 128 and sequence length to 50, and invoke the load_data_snli func-
tion to get the data iterators and vocabulary. Then we print the vocabulary size.
train_iter, test_iter, vocab = load_data_snli(128, 50)

len(vocab)
read 549367 examples

read 9824 examples
18678
Now we print the shape of the first minibatch. Contrary to sentiment analysis, we have 2 inputs
X[0] and X[1] representing pairs of premises and hypotheses.
print(X[0].shape)
print(X[1].shape)
print(Y.shape)
break
(128, 50)
(128, 50)
(128,)
Summary
• Natural language inference studies whether a hypothesis can be inferred from a premise,
where both are a text sequence.
• In natural language inference, relationships between premises and hypotheses include en-
tailment, contradiction, and neutral.
• Stanford Natural Language Inference (SNLI) Corpus is a popular benchmark dataset of nat-
ural language inference.

Exercises
1. Machine translation has long been evaluated based on superficial n-gram matching between
an output translation and a ground-truth translation. Can you design a measure for evaluat-
ing machine translation results by using natural language inference?
2. How can we change hyperparameters to reduce the vocabulary size?
Discussions217
15.5 Natural Language Inference: Using Attention
We introduced the natural language inference task and the SNLI dataset in Section 15.4. In view of
many models that are based on complex and deep architectures, Parikh et al. proposed to address
natural language inference with attention mechanisms and called it a “decomposable attention
model” (Parikh et al., 2016). This results in a model without recurrent or convolutional layers,
achieving the best result at the time on the SNLI dataset with much fewer parameters. In this
section, we will describe and implement this attention-based method (with MLPs) for natural lan-
guage inference, as depicted in Fig. 15.5.1.
Fig. 15.5.1: This section feeds pretrained GloVe to an architecture based on attention and MLPs
for natural language inference.
15.5.1 The Model
Simpler than preserving the order of words in premises and hypotheses, we can just align words
in one text sequence to every word in the other, and vice versa, then compare and aggregate such
information to predict the logical relationships between premises and hypotheses. Similar to
alignment of words between source and target sentences in machine translation, the alignment of
words between premises and hypotheses can be neatly accomplished by attention mechanisms.
217
15.5. Natural Language Inference: Using Attention 703

Fig. 15.5.2: Natural language inference using attention mechanisms.
Fig. 15.5.2 depicts the natural language inference method using attention mechanisms. At a high
level, it consists of three jointly trained steps: attending, comparing, and aggregating. We will
illustrate them step by step in the following.

import mxnet as mx
npx.set_np()
Attending
The first step is to align words in one text sequence to each word in the other sequence. Suppose
that the premise is “i do need sleep” and the hypothesis is “i am tired”. Due to semantical simi-
larity, we may wish to align “i” in the hypothesis with “i” in the premise, and align “tired” in the
hypothesis with “sleep” in the premise. Likewise, we may wish to align “i” in the premise with
“i” in the hypothesis, and align “need” and “sleep” in the premise with “tired” in the hypothesis.
Note that such alignment is soft using weighted average, where ideally large weights are associated
with the words to be aligned. For ease of demonstration, Fig. 15.5.2 shows such alignment in a
hard way.
Now we describe the soft alignment using attention mechanisms in more detail. Denote by A =
(a1 , . . . , am ) and B = (b1 , . . . , bn ) the premise and hypothesis, whose number of words are m and
n, respectively, where ai , bj ∈ Rd (i = 1, . . . , m, j = 1, . . . , n) is a d-dimensional word embedding
vector. For soft alignment, we compute the attention weights eij ∈ R as
eij = f (ai )⊤ f (bj ), (15.5.1)
where the function f is a multilayer perceptron defined in the following mlp function. The output
dimension of f is specified by the num_hiddens argument of mlp.

def mlp(num_hiddens, flatten):
net.add(nn.Dropout(0.2))
net.add(nn.Dense(num_hiddens, activation='relu', flatten=flatten))
net.add(nn.Dropout(0.2))
net.add(nn.Dense(num_hiddens, activation='relu', flatten=flatten))
return net
It should be highlighted that, in (15.5.1) f takes inputs ai and bj separately rather than takes a pair
of them together as the input. This decomposition trick leads to only m + n applications (linear
complexity) of f rather than mn applications (quadratic complexity).
Normalizing the attention weights in (15.5.1), we compute the weighted average of all the word
embeddings in the hypothesis to obtain representation of the hypothesis that is softly aligned with
the word indexed by i in the premise:
∑
n
exp(eij )
βi = ∑n bj . (15.5.2)
j=1 k=1 exp(eik )
Likewise, we compute soft alignment of premise words for each word indexed by j in the hypoth-
esis:
∑
m
exp(eij )
αj = ∑m ai . (15.5.3)
i=1 k=1 exp(ekj )
Below we define the Attend class to compute the soft alignment of hypotheses (beta) with input
premises A and soft alignment of premises (alpha) with input hypotheses B.
class Attend(nn.Block):
def __init__(self, num_hiddens, **kwargs):
super(Attend, self).__init__(**kwargs)
self.f = mlp(num_hiddens=num_hiddens, flatten=False)
def forward(self, A, B):

# Shape of À`/`B`: (bàtch_size`, no. of words in sequence A/B,
# èmbed_size`)
# Shape of `f_A`/`f_B`: (`batch_size`, no. of words in sequence A/B,
# `num_hiddens`)
f_A = self.f(A)
f_B = self.f(B)
# Shape of è`: (`batch_size`, no. of words in sequence A,
# no. of words in sequence B)
e = npx.batch_dot(f_A, f_B, transpose_b=True)
# Shape of `beta`: (`batch_size`, no. of words in sequence A,
# èmbed_size`), where sequence B is softly aligned with each word
# (axis 1 of `beta`) in sequence A
beta = npx.batch_dot(npx.softmax(e), B)
# Shape of àlpha`: (`batch_size`, no. of words in sequence B,
# èmbed_size`), where sequence A is softly aligned with each word
# (axis 1 of àlpha`) in sequence B
alpha = npx.batch_dot(npx.softmax(e.transpose(0, 2, 1)), A)
return beta, alpha

Comparing
In the next step, we compare a word in one sequence with the other sequence that is softly aligned
with that word. Note that in soft alignment, all the words from one sequence, though with prob-
ably different attention weights, will be compared with a word in the other sequence. For easy of
demonstration, Fig. 15.5.2 pairs words with aligned words in a hard way. For example, suppose
that the attending step determines that “need” and “sleep” in the premise are both aligned with
“tired” in the hypothesis, the pair “tired–need sleep” will be compared.
In the comparing step, we feed the concatenation (operator [·, ·]) of words from one sequence and
aligned words from the other sequence into a function g (a multilayer perceptron):
vA,i = g([ai , β i ]), i = 1, . . . , m

(15.5.4)
vB,j = g([bj , αj ]), j = 1, . . . , n.
In (15.5.4), vA,i is the comparison between word i in the premise and all the hypothesis words that
are softly aligned with word i; while vB,j is the comparison between word j in the hypothesis and
all the premise words that are softly aligned with word j. The following Compare class defines such
as comparing step.
class Compare(nn.Block):
def __init__(self, num_hiddens, **kwargs):
super(Compare, self).__init__(**kwargs)
self.g = mlp(num_hiddens=num_hiddens, flatten=False)
def forward(self, A, B, beta, alpha):

V_A = self.g(np.concatenate([A, beta], axis=2))
V_B = self.g(np.concatenate([B, alpha], axis=2))
return V_A, V_B
Aggregating
With two sets of comparison vectors vA,i (i = 1, . . . , m) and vB,j (j = 1, . . . , n) on hand, in the last
step we will aggregate such information to infer the logical relationship. We begin by summing
up both sets:
∑
m ∑
n
vA = vA,i , vB = vB,j . (15.5.5)
i=1 j=1
Next we feed the concatenation of both summarization results into function h (a multilayer per-
ceptron) to obtain the classification result of the logical relationship:
ŷ = h([vA , vB ]). (15.5.6)
The aggregation step is defined in the following Aggregate class.
class Aggregate(nn.Block):
def __init__(self, num_hiddens, num_outputs, **kwargs):
super(Aggregate, self).__init__(**kwargs)
self.h = mlp(num_hiddens=num_hiddens, flatten=True)
self.h.add(nn.Dense(num_outputs))

def forward(self, V_A, V_B):

# Sum up both sets of comparison vectors
V_A = V_A.sum(axis=1)
V_B = V_B.sum(axis=1)
# Feed the concatenation of both summarization results into an MLP
Y_hat = self.h(np.concatenate([V_A, V_B], axis=1))
return Y_hat
By putting the attending, comparing, and aggregating steps together, we define the decomposable
attention model to jointly train these three steps.
class DecomposableAttention(nn.Block):
def __init__(self, vocab, embed_size, num_hiddens, **kwargs):
super(DecomposableAttention, self).__init__(**kwargs)
self.embedding = nn.Embedding(len(vocab), embed_size)
self.attend = Attend(num_hiddens)
self.compare = Compare(num_hiddens)
# There are 3 possible outputs: entailment, contradiction, and neutral
self.aggregate = Aggregate(num_hiddens, 3)

premises, hypotheses = X
A = self.embedding(premises)
B = self.embedding(hypotheses)
beta, alpha = self.attend(A, B)
V_A, V_B = self.compare(A, B, beta, alpha)
Y_hat = self.aggregate(V_A, V_B)
return Y_hat
15.5.2 Training and Evaluating the Model
Now we will train and evaluate the defined decomposable attention model on the SNLI dataset.
We begin by reading the dataset.
Reading the dataset
We download and read the SNLI dataset using the function defined in Section 15.4. The batch size
and sequence length are set to 256 and 50, respectively.

train_iter, test_iter, vocab = d2l.load_data_snli(batch_size, num_steps)

read 9824 examples

Creating the Model
We use the pretrained 100-dimensional GloVe embedding to represent the input tokens. Thus, we
predefine the dimension of vectors ai and bj in :eqref:eq_nli_e as 100. The output dimension of
functions f in (15.5.1) and g in :eqref:eq_nli_v_ab is set to 200. Then we create a model instance,
initialize its parameters, and load the GloVe embedding to initialize vectors of input tokens.
embed_size, num_hiddens, devices = 100, 200, d2l.try_all_gpus()

net = DecomposableAttention(vocab, embed_size, num_hiddens)
Downloading ../data/glove.6B.100d.zip from http://d2l-data.s3-accelerate.amazonaws.com/glove.

,→6B.100d.zip...
Training and Evaluating the Model
In contrast to the split_batch function in Section 12.5 that takes single inputs such as text se-
quences (or images), we define a split_batch_multi_inputs function to take multiple inputs such
as premises and hypotheses in minibatches.
#@save
def split_batch_multi_inputs(X, y, devices):
"""Split multi-input `X` and `y` into multiple devices."""
X = list(zip(*[gluon.utils.split_and_load(
feature, devices, even_split=False) for feature in X]))
return (X, gluon.utils.split_and_load(y, devices, even_split=False))
Now we can train and evaluate the model on the SNLI dataset.

d2l.train_ch13(net, train_iter, test_iter, loss, trainer, num_epochs, devices,
split_batch_multi_inputs)


Using the Model
Finally, define the prediction function to output the logical relationship between a pair of premise
and hypothesis.
#@save
def predict_snli(net, vocab, premise, hypothesis):
premise = np.array(vocab[premise], ctx=d2l.try_gpu())
hypothesis = np.array(vocab[hypothesis], ctx=d2l.try_gpu())
label = np.argmax(net([premise.reshape((1, -1)),
hypothesis.reshape((1, -1))]), axis=1)
return 'entailment' if label == 0 else 'contradiction' if label == 1 \
else 'neutral'
We can use the trained model to obtain the natural language inference result for a sample pair of
sentences.
predict_snli(net, vocab, ['he', 'is', 'good', '.'], ['he', 'is', 'bad', '.'])
'contradiction'
Summary
• The decomposable attention model consists of three steps for predicting the logical relation-
ships between premises and hypotheses: attending, comparing, and aggregating.
• With attention mechanisms, we can align words in one text sequence to every word in the
other, and vice versa. Such alignment is soft using weighted average, where ideally large
weights are associated with the words to be aligned.
• The decomposition trick leads to a more desirable linear complexity than quadratic com-
plexity when computing attention weights.
• We can use pretrained word embedding as the input representation for downstream natural
language processing task such as natural language inference.

Exercises
1. Train the model with other combinations of hyperparameters. Can you get better accuracy
on the test set?
2. What are major drawbacks of the decomposable attention model for natural language infer-
ence?
3. Suppose that we want to get the level of semantical similarity (e.g., a continuous value be-
tween 0 and 1) for any pair of sentences. How shall we collect and label the dataset? Can you
design a model with attention mechanisms?
Discussions218
15.6 Fine-Tuning BERT for Sequence-Level and Token-Level Applications
In the previous sections of this chapter, we have designed different models for natural language
processing applications, such as based on RNNs, CNNs, attention, and MLPs. These models are
helpful when there is space or time constraint, however, crafting a specific model for every natu-
ral language processing task is practically infeasible. In Section 14.8, we introduced a pretraining
model, BERT, that requires minimal architecture changes for a wide range of natural language
processing tasks. One one hand, at the time of its proposal, BERT improved the state of the art
on various natural language processing tasks. On the other hand, as noted in Section 14.10, the
two versions of the original BERT model come with 110 million and 340 million parameters. Thus,
when there are sufficient computational resources, we may consider fine-tuning BERT for down-
stream natural language processing applications.
In the following, we generalize a subset of natural language processing applications as sequence-
level and token-level. On the sequence level, we introduce how to transform the BERT represen-
tation of the text input to the output label in single text classification and text pair classification
or regression. On the token level, we will briefly introduce new applications such as text tagging
and question answering and shed light on how BERT can represent their inputs and get trans-
formed into output labels. During fine-tuning, the “minimal architecture changes” required by
BERT across different applications are the extra fully-connected layers. During supervised learn-
ing of a downstream application, parameters of the extra layers are learned from scratch while all
the parameters in the pretrained BERT model are fine-tuned.
15.6.1 Single Text Classification
Single text classification takes a single text sequence as the input and outputs its classification re-
sult. Besides sentiment analysis that we have studied in this chapter, the Corpus of Linguistic
Acceptability (CoLA) is also a dataset for single text classification, judging whether a given sen-
tence is grammatically acceptable or not (Warstadt et al., 2019). For instance, “I should study.” is
acceptable but “I should studying.” is not.
218

Fig. 15.6.1: Fine-tuning BERT for single text classification applications, such as sentiment analysis
and testing linguistic acceptability. Suppose that the input single text has six tokens.
Section 14.8 describes the input representation of BERT. The BERT input sequence unambiguously
represents both single text and text pairs, where the special classification token “<cls>” is used for
sequence classification and the special classification token “<sep>” marks the end of single text
or separates a pair of text. As shown in Fig. 15.6.1, in single text classification applications, the
BERT representation of the special classification token “<cls>” encodes the information of the
entire input text sequence. As the representation of the input single text, it will be fed into a small
MLP consisting of fully-connected (dense) layers to output the distribution of all the discrete label
values.
15.6.2 Text Pair Classification or Regression
We have also examined natural language inference in this chapter. It belongs to text pair classifi-
cation, a type of application classifying a pair of text.
Taking a pair of text as the input but outputting a continuous value, semantic textual similarity
is a popular text pair regression task. This task measures semantic similarity of sentences. For
instance, in the Semantic Textual Similarity Benchmark dataset, the similarity score of a pair of
sentences is an ordinal scale ranging from 0 (no meaning overlap) to 5 (meaning equivalence) (Cer
et al., 2017). The goal is to predict these scores. Examples from the Semantic Textual Similarity
Benchmark dataset include (sentence 1, sentence 2, similarity score):
• “A plane is taking off.”, “An air plane is taking off.”, 5.000;
• “A woman is eating something.”, “A woman is eating meat.”, 3.000;
• “A woman is dancing.”, “A man is talking.”, 0.000.
15.6. Fine-Tuning BERT for Sequence-Level and Token-Level Applications 711

Fig. 15.6.2: Fine-tuning BERT for text pair classification or regression applications, such as natural
language inference and semantic textual similarity. Suppose that the input text pair has two and
three tokens.
Comparing with single text classification in Fig. 15.6.1, fine-tuning BERT for text pair classifica-
tion in Fig. 15.6.2 is different in the input representation. For text pair regression tasks such as
semantic textual similarity, trivial changes can be applied such as outputting a continuous label
value and using the mean squared loss: they are common for regression.
15.6.3 Text Tagging
Now let us consider token-level tasks, such as text tagging, where each token is assigned a label.
Among text tagging tasks, part-of-speech tagging assigns each word a part-of-speech tag (e.g., ad-
jective and determiner) according to the role of the word in the sentence. For example, according
to the Penn Treebank II tag set, the sentence “John Smith ʼs car is new” should be tagged as “NNP
(noun, proper singular) NNP POS (possessive ending) NN (noun, singular or mass) VB (verb, base
form) JJ (adjective)”.

Fig. 15.6.3: Fine-tuning BERT for text tagging applications, such as part-of-speech tagging. Sup-
pose that the input single text has six tokens.
Fine-tuning BERT for text tagging applications is illustrated in Fig. 15.6.3. Comparing with Fig.
15.6.1, the only distinction lies in that in text tagging, the BERT representation of every token of the
input text is fed into the same extra fully-connected layers to output the label of the token, such
as a part-of-speech tag.
15.6.4 Question Answering
As another token-level application, question answering reflects capabilities of reading comprehen-

sion. For example, the Stanford Question Answering Dataset (SQuAD v1.1) consists of reading
passages and questions, where the answer to every question is just a segment of text (text span)
from the passage that the question is about (Rajpurkar et al., 2016). To explain, consider a passage
“Some experts report that a maskʼs efficacy is inconclusive. However, mask makers insist that
their products, such as N95 respirator masks, can guard against the virus.” and a question “Who
say that N95 respirator masks can guard against the virus?”. The answer should be the text span
“mask makers” in the passage. Thus, the goal in SQuAD v1.1 is to predict the start and end of the
text span in the passage given a pair of question and passage.
15.6. Fine-Tuning BERT for Sequence-Level and Token-Level Applications 713

Fig. 15.6.4: Fine-tuning BERT for question answering. Suppose that the input text pair has two
and three tokens.
To fine-tune BERT for question answering, the question and passage are packed as the first and
second text sequence, respectively, in the input of BERT. To predict the position of the start of
the text span, the same additional fully-connected layer will transform the BERT representation
of any token from the passage of position i into a scalar score si . Such scores of all the passage
tokens are further transformed by the softmax operation into a probability distribution, so that
each token position i in the passage is assigned a probability pi of being the start of the text span.
Predicting the end of the text span is the same as above, except that parameters in its additional
fully-connected layer are independent from those for predicting the start. When predicting the
end, any passage token of position i is transformed by the same fully-connected layer into a scalar
score ei . :numref:fig_bert-qa depicts fine-tuning BERT for question answering.
For question answering, the supervised learningʼs training objective is as straightforward as maxi-
mizing the log-likelihoods of the ground-truth start and end positions. When predicting the span,
we can compute the score si + ej for a valid span from position i to position j (i ≤ j), and output
the span with the highest score.
Summary
• BERT requires minimal architecture changes (extra fully-connected layers) for sequence-
level and token-level natural language processing applications, such as single text classifi-
cation (e.g., sentiment analysis and testing linguistic acceptability), text pair classification
or regression (e.g., natural language inference and semantic textual similarity), text tagging
(e.g., part-of-speech tagging), and question answering.
• During supervised learning of a downstream application, parameters of the extra layers are
learned from scratch while all the parameters in the pretrained BERT model are fine-tuned.

Exercises
1. Let us design a search engine algorithm for news articles. When the system receives an
query (e.g., “oil industry during the coronavirus outbreak”), it should return a ranked list
of news articles that are most relevant to the query. Suppose that we have a huge pool of
news articles and a large number of queries. To simplify the problem, suppose that the
most relevant article has been labeled for each query. How can we apply negative sampling
(see Section 14.2.1) and BERT in the algorithm design?
2. How can we leverage BERT in training language models?
3. Can we leverage BERT in machine translation?
Discussions219
15.7 Natural Language Inference: Fine-Tuning BERT
In earlier sections of this chapter, we have designed an attention-based architecture (in Section
15.5) for the natural language inference task on the SNLI dataset (as described in Section 15.4).
Now we revisit this task by fine-tuning BERT. As discussed in Section 15.6, natural language infer-
ence is a sequence-level text pair classification problem, and fine-tuning BERT only requires an
additional MLP-based architecture, as illustrated in Fig. 15.7.1.
Fig. 15.7.1: This section feeds pretrained BERT to an MLP-based architecture for natural language
inference.
In this section, we will download a pretrained small version of BERT, then fine-tune it for natural
language inference on the SNLI dataset.

import json
import multiprocessing
219
15.7. Natural Language Inference: Fine-Tuning BERT 715

import os
npx.set_np()
15.7.1 Loading Pretrained BERT
We have explained how to pretrain BERT on the WikiText-2 dataset in Section 14.9 and Section
14.10 (note that the original BERT model is pretrained on much bigger corpora). As discussed in
Section 14.10, the original BERT model has hundreds of millions of parameters. In the following,
we provide two versions of pretrained BERT: “bert.base” is about as big as the original BERT base
model that requires a lot of computational resources to fine-tune, while “bert.small” is a small
version to facilitate demonstration.
d2l.DATA_HUB['bert.base'] = (d2l.DATA_URL + 'bert.base.zip',

'7b3820b35da691042e5d34c0971ac3edbd80d3f4')
d2l.DATA_HUB['bert.small'] = (d2l.DATA_URL + 'bert.small.zip',
'a4e718a47137ccd1809c9107ab4f5edd317bae2c')
Either pretrained BERT model contains a “vocab.json” file that defines the vocabulary set
and a “pretrained.params” file of the pretrained parameters. We implement the following
load_pretrained_model function to load pretrained BERT parameters.
def load_pretrained_model(pretrained_model, num_hiddens, ffn_num_hiddens,

num_heads, num_layers, dropout, max_len, devices):
data_dir = d2l.download_extract(pretrained_model)
# Define an empty vocabulary to load the predefined vocabulary
vocab = d2l.Vocab([])
vocab.idx_to_token = json.load(open(os.path.join(data_dir, 'vocab.json')))
vocab.token_to_idx = {token: idx for idx, token in enumerate(
vocab.idx_to_token)}
bert = d2l.BERTModel(len(vocab), num_hiddens, ffn_num_hiddens, num_heads,
num_layers, dropout, max_len)
# Load pretrained BERT parameters
bert.load_parameters(os.path.join(data_dir, 'pretrained.params'),
ctx=devices)
return bert, vocab
To facilitate demonstration on most of machines, we will load and fine-tune the small version
(“bert.small”) of the pretrained BERT in this section. In the exercise, we will show how to fine-
tune the much larger “bert.base” to significantly improve the testing accuracy.
bert, vocab = load_pretrained_model(
'bert.small', num_hiddens=256, ffn_num_hiddens=512, num_heads=4,
num_layers=2, dropout=0.1, max_len=512, devices=devices)
Downloading ../data/bert.small.zip from http://d2l-data.s3-accelerate.amazonaws.com/bert.

,→small.zip...

15.7.2 The Dataset for Fine-Tuning BERT
For the downstream task natural language inference on the SNLI dataset, we define a customized
dataset class SNLIBERTDataset. In each example, the premise and hypothesis form a pair of text
sequence and is packed into one BERT input sequence as depicted in Fig. 15.6.2. Recall Section
14.8.4 that segment IDs are used to distinguish the premise and the hypothesis in a BERT input se-
quence. With the predefined maximum length of a BERT input sequence (max_len), the last token
of the longer of the input text pair keeps getting removed until max_len is met. To accelerate gen-
eration of the SNLI dataset for fine-tuning BERT, we use 4 worker processes to generate training
or testing examples in parallel.
class SNLIBERTDataset(gluon.data.Dataset):
def __init__(self, dataset, max_len, vocab=None):
all_premise_hypothesis_tokens = [[
p_tokens, h_tokens] for p_tokens, h_tokens in zip(
*[d2l.tokenize([s.lower() for s in sentences])
for sentences in dataset[:2]])]
self.labels = np.array(dataset[2])
self.vocab = vocab
self.max_len = max_len
(self.all_token_ids, self.all_segments,
self.valid_lens) = self._preprocess(all_premise_hypothesis_tokens)
print('read ' + str(len(self.all_token_ids)) + ' examples')
def _preprocess(self, all_premise_hypothesis_tokens):

pool = multiprocessing.Pool(4) # Use 4 worker processes
out = pool.map(self._mp_worker, all_premise_hypothesis_tokens)
all_token_ids = [
token_ids for token_ids, segments, valid_len in out]
all_segments = [segments for token_ids, segments, valid_len in out]
valid_lens = [valid_len for token_ids, segments, valid_len in out]
return (np.array(all_token_ids, dtype='int32'),
np.array(all_segments, dtype='int32'),
np.array(valid_lens))
def _mp_worker(self, premise_hypothesis_tokens):

p_tokens, h_tokens = premise_hypothesis_tokens
self._truncate_pair_of_tokens(p_tokens, h_tokens)
tokens, segments = d2l.get_tokens_and_segments(p_tokens, h_tokens)
token_ids = self.vocab[tokens] + [self.vocab['<pad>']] \
* (self.max_len - len(tokens))
segments = segments + [0] * (self.max_len - len(segments))
valid_len = len(tokens)
return token_ids, segments, valid_len
def _truncate_pair_of_tokens(self, p_tokens, h_tokens):

# Reserve slots for '<CLS>', '<SEP>', and '<SEP>' tokens for the BERT
# input
while len(p_tokens) + len(h_tokens) > self.max_len - 3:
if len(p_tokens) > len(h_tokens):
p_tokens.pop()
else:
h_tokens.pop()

return (self.all_token_ids[idx], self.all_segments[idx],
self.valid_lens[idx]), self.labels[idx]
def __len__(self):
return len(self.all_token_ids)
After downloading the SNLI dataset, we generate training and testing examples by instantiating
the SNLIBERTDataset class. Such examples will be read in minibatches during training and testing
of natural language inference.
# Reduce `batch_size` if there is an out of memory error. In the original BERT

# model, `max_len` = 512
batch_size, max_len, num_workers = 512, 128, d2l.get_dataloader_workers()
train_set = SNLIBERTDataset(d2l.read_snli(data_dir, True), max_len, vocab)
test_set = SNLIBERTDataset(d2l.read_snli(data_dir, False), max_len, vocab)
test_iter = gluon.data.DataLoader(test_set, batch_size,

read 9824 examples
15.7.3 Fine-Tuning BERT
As Fig. 15.6.2 indicates, fine-tuning BERT for natural language inference requires only an extra
MLP consisting of two fully-connected layers (see self.hidden and self.output in the following
BERTClassifier class). This MLP transforms the BERT representation of the special “<cls>” to-
ken, which encodes the information of both the premise and the hypothesis, into three outputs of
natural language inference: entailment, contradiction, and neutral.
class BERTClassifier(nn.Block):
def __init__(self, bert):
super(BERTClassifier, self).__init__()
self.encoder = bert.encoder
self.hidden = bert.hidden

tokens_X, segments_X, valid_lens_x = inputs
encoded_X = self.encoder(tokens_X, segments_X, valid_lens_x)
return self.output(self.hidden(encoded_X[:, 0, :]))
In the following, the pretrained BERT model bert is fed into the BERTClassifier instance net for
the downstream application. In common implementations of BERT fine-tuning, only the param-
eters of the output layer of the additional MLP (net.output) will be learned from scratch. All the
parameters of the pretrained BERT encoder (net.encoder) and the hidden layer of the additional
MLP (net.hidden) will be fine-tuned.

net = BERTClassifier(bert)
net.output.initialize(ctx=devices)
Recall that in Section 14.8 both the MaskLM class and the NextSentencePred class have parameters
in their employed MLPs. These parameters are part of those in the pretrained BERT model bert,
and thus part of parameters in net. However, such parameters are only for computing the masked
language modeling loss and the next sentence prediction loss during pretraining. These two loss
functions are irrelevant to fine-tuning downstream applications, thus the parameters of the em-
ployed MLPs in MaskLM and NextSentencePred are not updated (staled) when BERT is fine-tuned.
To allow parameters with stale gradients, the flag ignore_stale_grad=True is set in the step func-
tion of d2l.train_batch_ch13. We use this function to train and evaluate the model net using the
training set (train_iter) and the testing set (test_iter) of SNLI. Due to the limited computational
resources, the training and testing accuracy can be further improved: we leave its discussions in
the exercises.
lr, num_epochs = 1e-4, 5

d2l.train_ch13(net, train_iter, test_iter, loss, trainer, num_epochs, devices,
d2l.split_batch_multi_inputs)

Summary
• We can fine-tune the pretrained BERT model for downstream applications, such as natural
language inference on the SNLI dataset.
• During fine-tuning, the BERT model becomes part of the model for the downstream appli-
cation. Parameters that are only related to pretraining loss will not be updated during fine-
tuning.

Exercises
1. Fine-tune a much larger pretrained BERT model that is about as big as the original BERT base
model if your computational resource allows. Set arguments in the load_pretrained_model
function as: replacing ʻbert.smallʼ with ʻbert.baseʼ, increasing values of num_hiddens=256,
ffn_num_hiddens=512, num_heads=4, num_layers=2 to 768, 3072, 12, 12, respectively. By in-
creasing fine-tuning epochs (and possibly tuning other hyperparameters), can you get a test-
ing accuracy higher than 0.86?
2. How to truncate a pair of sequences according to their ratio of length? Compare this pair
truncation method and the one used in the SNLIBERTDataset class. What are their pros and
cons?
Discussions220
220

16 | Recommender Systems
Shuai Zhang (Amazon), Aston Zhang (Amazon), and Yi Tay (Google)

Recommender systems are widely employed in industry and are ubiquitous in our daily lives.
These systems are utilized in a number of areas such as online shopping sites (e.g., amazon.com),
music/movie services site (e.g., Netflix and Spotify), mobile application stores (e.g., IOS app store
and google play), online advertising, just to name a few.
The major goal of recommender systems is to help users discover relevant items such as movies
to watch, text to read or products to buy, so as to create a delightful user experience. Moreover,
recommender systems are among the most powerful machine learning systems that online retail-
ers implement in order to drive incremental revenue. Recommender systems are replacements
of search engines by reducing the efforts in proactive searches and surprising users with offers
they never searched for. Many companies managed to position themselves ahead of their com-
petitors with the help of more effective recommender systems. As such, recommender systems
are central to not only our everyday lives but also highly indispensable in some industries.
In this chapter, we will cover the fundamentals and advancements of recommender systems,
along with exploring some common fundamental techniques for building recommender systems
with different data sources available and their implementations. Specifically, you will learn how
to predict the rating a user might give to a prospective item, how to generate a recommendation
list of items and how to predict the click-through rate from abundant features. These tasks are
commonplace in real-world applications. By studying this chapter, you will get hands-on experi-
ence pertaining to solving real world recommendation problems with not only classical methods
but the more advanced deep learning based models as well.
16.1 Overview of Recommender Systems
In the last decade, the Internet has evolved into a platform for large-scale online services, which
profoundly changed the way we communicate, read news, buy products, and watch movies. In
the meanwhile, the unprecedented number of items (we use the term item to refer to movies,
news, books, and products.) offered online requires a system that can help us discover items that
we preferred. Recommender systems are therefore powerful information filtering tools that can
facilitate personalized services and provide tailored experience to individual users. In short, rec-
ommender systems play a pivotal role in utilizing the wealth of data available to make choices
manageable. Nowadays, recommender systems are at the core of a number of online services
providers such as Amazon, Netflix, and YouTube. Recall the example of Deep learning books rec-
ommended by Amazon in Fig. 1.3.3. The benefits of employing recommender systems are two-
folds: On the one hand, it can largely reduce usersʼ effort in finding items and alleviate the issue of
information overload. On the other hand, it can add business value to online service providers and
is an important source of revenue. This chapter will introduce the fundamental concepts, classic
721
models and recent advances with deep learning in the field of recommender systems, together
with implemented examples.
Fig. 16.1.1: Illustration of the Recommendation Process
16.1.1 Collaborative Filtering
We start the journey with the important concept in recommender systems—collaborative filtering
(CF), which was first coined by the Tapestry system (Goldberg et al., 1992), referring to “people
collaborate to help one another perform the filtering process in order to handle the large amounts
of email and messages posted to newsgroups”. This term has been enriched with more senses. In
a broad sense, it is the process of filtering for information or patterns using techniques involving
collaboration among multiple users, agents, and data sources. CF has many forms and numerous
CF methods proposed since its advent.
Overall, CF techniques can be categorized into: memory-based CF, model-based CF, and their
hybrid (Su & Khoshgoftaar, 2009). Representative memory-based CF techniques are nearest
neighbor-based CF such as user-based CF and item-based CF (Sarwar et al., 2001). Latent factor
models such as matrix factorization are examples of model-based CF. Memory-based CF has lim-
itations in dealing with sparse and large-scale data since it computes the similarity values based
on common items. Model-based methods become more popular with its better capability in deal-
ing with sparsity and scalability. Many model-based CF approaches can be extended with neu-
ral networks, leading to more flexible and scalable models with the computation acceleration
in deep learning (Zhang et al., 2019). In general, CF only uses the user-item interaction data to
make predictions and recommendations. Besides CF, content-based and context-based recom-
mender systems are also useful in incorporating the content descriptions of items/users and con-
textual signals such as timestamps and locations. Obviously, we may need to adjust the model
types/structures when different input data is available.
722 Chapter 16. Recommender Systems

16.1.2 Explicit Feedback and Implicit Feedback
To learn the preference of users, the system shall collect feedback from them. The feedback can be
either explicit or implicit (Hu et al., 2008). For example, IMDB221 collects star ratings ranging from
one to ten stars for movies. YouTube provides the thumbs-up and thumbs-down buttons for users
to show their preferences. It is apparent that gathering explicit feedback requires users to indicate
their interests proactively. Nonetheless, explicit feedback is not always readily available as many
users may be reluctant to rate products. Relatively speaking, implicit feedback is often readily
available since it is mainly concerned with modeling implicit behavior such user clicks. As such,
many recommender systems are centered on implicit feedback which indirectly reflects userʼs
opinion through observing user behavior. There are diverse forms of implicit feedback including
purchase history, browsing history, watches and even mouse movements. For example, a user that
purchased many books by the same author probably likes that author. Note that implicit feedback
is inherently noisy. We can only guess their preferences and true motives. A user watched a movie
does not necessarily indicate a positive view of that movie.
16.1.3 Recommendation Tasks
A number of recommendation tasks have been investigated in the past decades. Based on the
domain of applications, there are movies recommendation, news recommendations, point-of-
interest recommendation (Ye et al., 2011) and so forth. It is also possible to differentiate the tasks
based on the types of feedback and input data, for example, the rating prediction task aims to
predict the explicit ratings. Top-n recommendation (item ranking) ranks all items for each user
personally based on the implicit feedback. If time-stamp information is also included, we can
build sequence-aware recommendation (Quadrana et al., 2018). Another popular task is called
click-through rate prediction, which is also based on implicit feedback, but various categorical
features can be utilized. Recommending for new users and recommending new items to existing
users are called cold-start recommendation (Schein et al., 2002).
Summary
• Recommender systems are important for individual users and industries. Collaborative fil-
tering is a key concept in recommendation.
• There are two types of feedbacks: implicit feedback and explicit feedback. A number of
recommendation tasks have been explored during the last decade.
Exercises
1. Can you explain how recommender systems influence your daily life?
2. What interesting recommendation tasks do you think can be investigated?
Discussions222
221
https://www.imdb.com/
222
16.1. Overview of Recommender Systems 723

16.2 The MovieLens Dataset
There are a number of datasets that are available for recommendation research. Amongst
them, the MovieLens223 dataset is probably one of the more popular ones. MovieLens is a non-
commercial web-based movie recommender system. It is created in 1997 and run by GroupLens,
a research lab at the University of Minnesota, in order to gather movie rating data for research
purposes. MovieLens data has been critical for several research studies including personalized
recommendation and social psychology.
16.2.1 Getting the Data
The MovieLens dataset is hosted by the GroupLens224 website. Several versions are available. We
will use the MovieLens 100K dataset (Herlocker et al., 1999). This dataset is comprised of 100, 000
ratings, ranging from 1 to 5 stars, from 943 users on 1682 movies. It has been cleaned up so that
each user has rated at least 20 movies. Some simple demographic information such as age, gender,
genres for the users and items are also available. We can download the ml-100k.zip225 and extract
the u.data file, which contains all the 100, 000 ratings in the csv format. There are many other
files in the folder, a detailed description for each file can be found in the README226 file of the
dataset.
To begin with, let us import the packages required to run this sectionʼs experiments.

import os
import pandas as pd
Then, we download the MovieLens 100k dataset and load the interactions as DataFrame.
#@save
d2l.DATA_HUB['ml-100k'] = (
'http://files.grouplens.org/datasets/movielens/ml-100k.zip',
'cd4dcac4241c8a4ad7badc7ca635da8a69dddb83')
#@save
def read_data_ml100k():
data_dir = d2l.download_extract('ml-100k')
names = ['user_id', 'item_id', 'rating', 'timestamp']
data = pd.read_csv(os.path.join(data_dir, 'u.data'), '\t', names=names,
engine='python')
num_users = data.user_id.unique().shape[0]
num_items = data.item_id.unique().shape[0]
return data, num_users, num_items
223
https://movielens.org/
224
https://grouplens.org/datasets/movielens/
225
http://files.grouplens.org/datasets/movielens/ml-100k.zip
226
http://files.grouplens.org/datasets/movielens/ml-100k-README.txt

16.2.2 Statistics of the Dataset
Let us load up the data and inspect the first five records manually. It is an effective way to learn
the data structure and verify that they have been loaded properly.
data, num_users, num_items = read_data_ml100k()

sparsity = 1 - len(data) / (num_users * num_items)
print(f'number of users: {num_users}, number of items: {num_items}')
print(f'matrix sparsity: {sparsity:f}')
print(data.head(5))
number of users: 943, number of items: 1682

matrix sparsity: 0.936953
user_id item_id rating timestamp
0 196 242 3 881250949
1 186 302 3 891717742
2 22 377 1 878887116
3 244 51 2 880606923
4 166 346 1 886397596
We can see that each line consists of four columns, including “user id” 1-943, “item id” 1-1682,
“rating” 1-5 and “timestamp”. We can construct an interaction matrix of size n × m, where n and
m are the number of users and the number of items respectively. This dataset only records the
existing ratings, so we can also call it rating matrix and we will use interaction matrix and rating
matrix interchangeably in case that the values of this matrix represent exact ratings. Most of the
values in the rating matrix are unknown as users have not rated the majority of movies. We also
show the sparsity of this dataset. The sparsity is defined as 1 - number of nonzero entries / (
number of users * number of items). Clearly, the interaction matrix is extremely sparse (i.e.,
sparsity = 93.695%). Real world datasets may suffer from a greater extent of sparsity and has been
a long-standing challenge in building recommender systems. A viable solution is to use additional
side information such as user/item features to alleviate the sparsity.
We then plot the distribution of the count of different ratings. As expected, it appears to be a
normal distribution, with most ratings centered at 3-4.
d2l.plt.hist(data['rating'], bins=5, ec='black')

d2l.plt.xlabel('Rating')
d2l.plt.ylabel('Count')
d2l.plt.title('Distribution of Ratings in MovieLens 100K')
d2l.plt.show()
16.2. The MovieLens Dataset 725

16.2.3 Splitting the dataset
We split the dataset into training and test sets. The following function provides two split modes
including random and seq-aware. In the random mode, the function splits the 100k interactions
randomly without considering timestamp and uses the 90% of the data as training samples and the
rest 10% as test samples by default. In the seq-aware mode, we leave out the item that a user rated
most recently for test, and usersʼ historical interactions as training set. User historical interactions
are sorted from oldest to newest based on timestamp. This mode will be used in the sequence-
aware recommendation section.
#@save
def split_data_ml100k(data, num_users, num_items,
split_mode='random', test_ratio=0.1):
"""Split the dataset in random mode or seq-aware mode."""
if split_mode == 'seq-aware':
train_items, test_items, train_list = {}, {}, []
for line in data.itertuples():
u, i, rating, time = line[1], line[2], line[3], line[4]
train_items.setdefault(u, []).append((u, i, rating, time))
if u not in test_items or test_items[u][-1] < time:
test_items[u] = (i, rating, time)
for u in range(1, num_users + 1):
train_list.extend(sorted(train_items[u], key=lambda k: k[3]))
test_data = [(key, *value) for key, value in test_items.items()]
train_data = [item for item in train_list if item not in test_data]
train_data = pd.DataFrame(train_data)
test_data = pd.DataFrame(test_data)
else:
mask = [True if x == 1 else False for x in np.random.uniform(
0, 1, (len(data))) < 1 - test_ratio]

neg_mask = [not x for x in mask]
train_data, test_data = data[mask], data[neg_mask]
return train_data, test_data
Note that it is good practice to use a validation set in practice, apart from only a test set. However,
we omit that for the sake of brevity. In this case, our test set can be regarded as our held-out
validation set.
16.2.4 Loading the data
After dataset splitting, we will convert the training set and test set into lists and dictionaries/matrix
for the sake of convenience. The following function reads the dataframe line by line and enu-
merates the index of users/items start from zero. The function then returns lists of users, items,
ratings and a dictionary/matrix that records the interactions. We can specify the type of feedback
to either explicit or implicit.
#@save
def load_data_ml100k(data, num_users, num_items, feedback='explicit'):
users, items, scores = [], [], []
inter = np.zeros((num_items, num_users)) if feedback == 'explicit' else {}
for line in data.itertuples():
user_index, item_index = int(line[1] - 1), int(line[2] - 1)
score = int(line[3]) if feedback == 'explicit' else 1
users.append(user_index)
items.append(item_index)
scores.append(score)
if feedback == 'implicit':
inter.setdefault(user_index, []).append(item_index)
else:
inter[item_index, user_index] = score
return users, items, scores, inter
Afterwards, we put the above steps together and it will be used in the next section. The results are
wrapped with Dataset and DataLoader. Note that the last_batch of DataLoader for training data
is set to the rollover mode (The remaining samples are rolled over to the next epoch.) and orders
are shuffled.
#@save
def split_and_load_ml100k(split_mode='seq-aware', feedback='explicit',
test_ratio=0.1, batch_size=256):
data, num_users, num_items = read_data_ml100k()
train_data, test_data = split_data_ml100k(
data, num_users, num_items, split_mode, test_ratio)
train_u, train_i, train_r, _ = load_data_ml100k(
train_data, num_users, num_items, feedback)
test_u, test_i, test_r, _ = load_data_ml100k(
test_data, num_users, num_items, feedback)
train_set = gluon.data.ArrayDataset(
np.array(train_u), np.array(train_i), np.array(train_r))
test_set = gluon.data.ArrayDataset(
np.array(test_u), np.array(test_i), np.array(test_r))
16.2. The MovieLens Dataset 727

train_set, shuffle=True, last_batch='rollover',
batch_size=batch_size)
test_set, batch_size=batch_size)
return num_users, num_items, train_iter, test_iter
Summary
• MovieLens datasets are widely used for recommendation research. It is public available and
free to use.
• We define functions to download and preprocess the MovieLens 100k dataset for further use
in later sections.
Exercises
• What other similar recommendation datasets can you find?

• Go through the https://movielens.org/ site for more information about MovieLens.
Discussions227
16.3 Matrix Factorization
Matrix Factorization (Koren et al., 2009) is a well-established algorithm in the recommender sys-
tems literature. The first version of matrix factorization model is proposed by Simon Funk in a
famous blog post228 in which he described the idea of factorizing the interaction matrix. It then
became widely known due to the Netflix contest which was held in 2006. At that time, Netflix, a
media-streaming and video-rental company, announced a contest to improve its recommender
system performance. The best team that can improve on the Netflix baseline, i.e., Cinematch),
by 10 percent would win a one million USD prize. As such, this contest attracted a lot of atten-
tion to the field of recommender system research. Subsequently, the grand prize was won by the
BellKorʼs Pragmatic Chaos team, a combined team of BellKor, Pragmatic Theory, and BigChaos
(you do not need to worry about these algorithms now). Although the final score was the result
of an ensemble solution (i.e., a combination of many algorithms), the matrix factorization algo-
rithm played a critical role in the final blend. The technical report the Netflix Grand Prize solution
(Toscher et al., 2009) provides a detailed introduction to the adopted model. In this section, we
will dive into the details of the matrix factorization model and its implementation.
227
228
https://sifter.org/~simon/journal/20061211.html

16.3.1 The Matrix Factorization Model
Matrix factorization is a class of collaborative filtering models. Specifically, the model factorizes
the user-item interaction matrix (e.g., rating matrix) into the product of two lower-rank matrices,
capturing the low-rank structure of the user-item interactions.
Let R ∈ Rm×n denote the interaction matrix with m users and n items, and the values of R
represent explicit ratings. The user-item interaction will be factorized into a user latent matrix
P ∈ Rm×k and an item latent matrix Q ∈ Rn×k , where k ≪ m, n, is the latent factor size. Let pu
denote the uth row of P and qi denote the ith row of Q. For a given item i, the elements of qi mea-
sure the extent to which the item possesses those characteristics such as the genres and languages
of a movie. For a given user u, the elements of pu measure the extent of interest the user has in
itemsʼ corresponding characteristics. These latent factors might measure obvious dimensions as
mentioned in those examples or are completely uninterpretable. The predicted ratings can be
estimated by
R̂ = PQ⊤ (16.3.1)
where R̂ ∈ Rm×n is the predicted rating matrix which has the same shape as R. One major problem
of this prediction rule is that users/items biases can not be modeled. For example, some users
tend to give higher ratings or some items always get lower ratings due to poorer quality. These
biases are commonplace in real-world applications. To capture these biases, user specific and
item specific bias terms are introduced. Specifically, the predicted rating user u gives to item i is
calculated by
R̂ui = pu q⊤
i + bu + bi (16.3.2)
Then, we train the matrix factorization model by minimizing the mean squared error between
predicted rating scores and real rating scores. The objective function is defined as follows:
∑
argmin ∥Rui − R̂ui ∥2 + λ(∥P∥2F + ∥Q∥2F + b2u + b2i ) (16.3.3)
P,Q,b (u,i)∈K
where λ denotes the regularization rate. The regularizing term λ(∥P∥2F + ∥Q∥2F + b2u + b2i ) is used
to avoid over-fitting by penalizing the magnitude of the parameters. The (u, i) pairs for which Rui
is known are stored in the set K = {(u, i) | Rui is known}. The model parameters can be learned
with an optimization algorithm, such as Stochastic Gradient Descent and Adam.
An intuitive illustration of the matrix factorization model is shown below:
Fig. 16.3.1: Illustration of matrix factorization model
16.3. Matrix Factorization 729

In the rest of this section, we will explain the implementation of matrix factorization and train the
model on the MovieLens dataset.

import mxnet as mx
npx.set_np()
16.3.2 Model Implementation
First, we implement the matrix factorization model described above. The user and item latent
factors can be created with the nn.Embedding. The input_dim is the number of items/users and
the (output_dim) is the dimension of the latent factors (k). We can also use nn.Embedding to create
the user/item biases by setting the output_dim to one. In the forward function, user and item ids
are used to look up the embeddings.
class MF(nn.Block):
def __init__(self, num_factors, num_users, num_items, **kwargs):
super(MF, self).__init__(**kwargs)
self.P = nn.Embedding(input_dim=num_users, output_dim=num_factors)
self.Q = nn.Embedding(input_dim=num_items, output_dim=num_factors)
self.user_bias = nn.Embedding(num_users, 1)
self.item_bias = nn.Embedding(num_items, 1)
def forward(self, user_id, item_id):

P_u = self.P(user_id)
Q_i = self.Q(item_id)
b_u = self.user_bias(user_id)
b_i = self.item_bias(item_id)
outputs = (P_u * Q_i).sum(axis=1) + np.squeeze(b_u) + np.squeeze(b_i)
return outputs.flatten()
16.3.3 Evaluation Measures
We then implement the RMSE (root-mean-square error) measure, which is commonly used to
measure the differences between rating scores predicted by the model and the actually observed
ratings (ground truth) (Gunawardana & Shani, 2015). RMSE is defined as:
v
u
u 1 ∑
RMSE = t (Rui − R̂ui )2 (16.3.4)
|T |
(u,i)∈T
where T is the set consisting of pairs of users and items that you want to evaluate on. |T | is the
size of this set. We can use the RMSE function provided by mx.metric.
def evaluator(net, test_iter, devices):

rmse = mx.metric.RMSE() # Get the RMSE
rmse_list = []
for idx, (users, items, ratings) in enumerate(test_iter):
u = gluon.utils.split_and_load(users, devices, even_split=False)

i = gluon.utils.split_and_load(items, devices, even_split=False)
r_ui = gluon.utils.split_and_load(ratings, devices, even_split=False)
r_hat = [net(u, i) for u, i in zip(u, i)]
rmse.update(labels=r_ui, preds=r_hat)
rmse_list.append(rmse.get()[1])
return float(np.mean(np.array(rmse_list)))
In the training function, we adopt the L2 loss with weight decay. The weight decay mechanism
has the same effect as the L2 regularization.
#@save
def train_recsys_rating(net, train_iter, test_iter, loss, trainer, num_epochs,
devices=d2l.try_all_gpus(), evaluator=None,
**kwargs):
timer = d2l.Timer()
legend=['train loss', 'test RMSE'])
metric, l = d2l.Accumulator(3), 0.
for i, values in enumerate(train_iter):
timer.start()
input_data = []
values = values if isinstance(values, list) else [values]
for v in values:
input_data.append(gluon.utils.split_and_load(v, devices))
train_feat = input_data[0:-1] if len(values) > 1 else input_data
train_label = input_data[-1]
preds = [net(*t) for t in zip(*train_feat)]
ls = [loss(p, s) for p, s in zip(preds, train_label)]
[l.backward() for l in ls]
l += sum([l.asnumpy() for l in ls]).mean() / len(devices)
trainer.step(values[0].shape[0])
metric.add(l, values[0].shape[0], values[0].size)
timer.stop()
if len(kwargs) > 0: # It will be used in section AutoRec
test_rmse = evaluator(net, test_iter, kwargs['inter_mat'],
devices)
else:
test_rmse = evaluator(net, test_iter, devices)
train_l = l / (i + 1)
animator.add(epoch + 1, (train_l, test_rmse))
f'test RMSE {test_rmse:.3f}')
Finally, let us put all things together and train the model. Here, we set the latent factor dimension
to 30.
16.3. Matrix Factorization 731

num_users, num_items, train_iter, test_iter = d2l.split_and_load_ml100k(
test_ratio=0.1, batch_size=512)
net = MF(30, num_users, num_items)
net.initialize(ctx=devices, force_reinit=True, init=mx.init.Normal(0.01))
lr, num_epochs, wd, optimizer = 0.002, 20, 1e-5, 'adam'
trainer = gluon.Trainer(net.collect_params(), optimizer,
{"learning_rate": lr, 'wd': wd})
train_recsys_rating(net, train_iter, test_iter, loss, trainer, num_epochs,
devices, evaluator)
train loss 0.065, test RMSE 1.057

Below, we use the trained model to predict the rating that a user (ID 20) might give to an item (ID
30).
scores = net(np.array([20], dtype='int', ctx=devices[0]),

np.array([30], dtype='int', ctx=devices[0]))
scores
array([3.0418706], ctx=gpu(0))
Summary
• The matrix factorization model is widely used in recommender systems. It can be used to
predict ratings that a user might give to an item.
• We can implement and train matrix factorization for recommender systems.

Exercises
• Vary the size of latent factors. How does the size of latent factors influence the model per-
formance?
• Try different optimizers, learning rates, and weight decay rates.
• Check the predicted rating scores of other users for a specific movie.
Discussions229
16.4 AutoRec: Rating Prediction with Autoencoders
Although the matrix factorization model achieves decent performance on the rating prediction
task, it is essentially a linear model. Thus, such models are not capable of capturing complex
nonlinear and intricate relationships that may be predictive of usersʼ preferences. In this section,
we introduce a nonlinear neural network collaborative filtering model, AutoRec (Sedhain et al.,
2015). It identifies collaborative filtering (CF) with an autoencoder architecture and aims to inte-
grate nonlinear transformations into CF on the basis of explicit feedback. Neural networks have
been proven to be capable of approximating any continuous function, making it suitable to ad-
dress the limitation of matrix factorization and enrich the expressiveness of matrix factorization.
On one hand, AutoRec has the same structure as an autoencoder which consists of an input layer,
a hidden layer, and a reconstruction (output) layer. An autoencoder is a neural network that
learns to copy its input to its output in order to code the inputs into the hidden (and usually
low-dimensional) representations. In AutoRec, instead of explicitly embedding users/items into
low-dimensional space, it uses the column/row of the interaction matrix as the input, then recon-
structs the interaction matrix in the output layer.
On the other hand, AutoRec differs from a traditional autoencoder: rather than learning the hid-
den representations, AutoRec focuses on learning/reconstructing the output layer. It uses a par-
tially observed interaction matrix as the input, aiming to reconstruct a completed rating matrix.
In the meantime, the missing entries of the input are filled in the output layer via reconstruction
for the purpose of recommendation.
There are two variants of AutoRec: user-based and item-based. For brevity, here we only introduce
the item-based AutoRec. User-based AutoRec can be derived accordingly.
16.4.1 Model
Let R∗i denote the ith column of the rating matrix, where unknown ratings are set to zeros by
default. The neural architecture is defined as:
h(R∗i ) = f (W · g(VR∗i + µ) + b) (16.4.1)
where f (·) and g(·) represent activation functions, W and V are weight matrices, µ and b are biases.
Let h(·) denote the whole network of AutoRec. The output h(R∗i ) is the reconstruction of the ith
column of the rating matrix.
229
https://discuss.d2l.ai/t/
16.4. AutoRec: Rating Prediction with Autoencoders 733

The following objective function aims to minimize the reconstruction error:
∑
M
argmin ∥ R∗i − h(R∗i ) ∥2O + λ(∥W∥2F + ∥V∥2F ) (16.4.2)
W,V,µ,b i=1
where ∥ · ∥O means only the contribution of observed ratings are considered, that is, only weights
that are associated with observed inputs are updated during back-propagation.

import mxnet as mx
import sys
npx.set_np()
16.4.2 Implementing the Model
A typical autoencoder consists of an encoder and a decoder. The encoder projects the input to
hidden representations and the decoder maps the hidden layer to the reconstruction layer. We
follow this practice and create the encoder and decoder with dense layers. The activation of en-
coder is set to sigmoid by default and no activation is applied for decoder. Dropout is included
after the encoding transformation to reduce over-fitting. The gradients of unobserved inputs are
masked out to ensure that only observed ratings contribute to the model learning process.
class AutoRec(nn.Block):
def __init__(self, num_hidden, num_users, dropout=0.05):
super(AutoRec, self).__init__()
self.encoder = nn.Dense(num_hidden, activation='sigmoid',
use_bias=True)
self.decoder = nn.Dense(num_users, use_bias=True)
def forward(self, input):

hidden = self.dropout(self.encoder(input))
pred = self.decoder(hidden)
if autograd.is_training(): # Mask the gradient during training
return pred * np.sign(input)
else:
return pred
16.4.3 Reimplementing the Evaluator
Since the input and output have been changed, we need to reimplement the evaluation function,
while we still use RMSE as the accuracy measure.
def evaluator(network, inter_matrix, test_data, devices):

scores = []
for values in inter_matrix:
feat = gluon.utils.split_and_load(values, devices, even_split=False)
scores.extend([network(i).asnumpy() for i in feat])

recons = np.array([item for sublist in scores for item in sublist])
# Calculate the test RMSE
rmse = np.sqrt(np.sum(np.square(test_data - np.sign(test_data) * recons))
/ np.sum(np.sign(test_data)))
return float(rmse)
Now, let us train and evaluate AutoRec on the MovieLens dataset. We can clearly see that the
test RMSE is lower than the matrix factorization model, confirming the effectiveness of neural
networks in the rating prediction task.
# Load the MovieLens 100K dataset
df, num_users, num_items = d2l.read_data_ml100k()
train_data, test_data = d2l.split_data_ml100k(df, num_users, num_items)
_, _, _, train_inter_mat = d2l.load_data_ml100k(train_data, num_users,
num_items)
_, _, _, test_inter_mat = d2l.load_data_ml100k(test_data, num_users,
num_items)
train_iter = gluon.data.DataLoader(train_inter_mat, shuffle=True,
last_batch="rollover", batch_size=256,
test_iter = gluon.data.DataLoader(np.array(train_inter_mat), shuffle=False,
last_batch="keep", batch_size=1024,
# Model initialization, training, and evaluation
net = AutoRec(500, num_users)
d2l.train_recsys_rating(net, train_iter, test_iter, loss, trainer, num_epochs,
devices, evaluator, inter_mat=test_inter_mat)
train loss 0.000, test RMSE 0.899

16.4. AutoRec: Rating Prediction with Autoencoders 735

Summary
• We can frame the matrix factorization algorithm with autoencoders, while integrating non-
linear layers and dropout regularization.
• Experiments on the MovieLens 100K dataset show that AutoRec achieves superior perfor-
mance than matrix factorization.
Exercises
• Vary the hidden dimension of AutoRec to see its impact on the model performance.
• Try to add more hidden layers. Is it helpful to improve the model performance?
• Can you find a better combination of decoder and encoder activation functions?
Discussions230
16.5 Personalized Ranking for Recommender Systems
In the former sections, only explicit feedback was considered and models were trained and tested
on observed ratings. There are two demerits of such methods: First, most feedback is not explicit
but implicit in real-world scenarios, and explicit feedback can be more expensive to collect. Sec-
ond, non-observed user-item pairs which may be predictive for usersʼ interests are totally ignored,
making these methods unsuitable for cases where ratings are not missing at random but because
of usersʼ preferences. Non-observed user-item pairs are a mixture of real negative feedback (users
are not interested in the items) and missing values (the user might interact with the items in the
future). We simply ignore the non-observed pairs in matrix factorization and AutoRec. Clearly,
these models are incapable of distinguishing between observed and non-observed pairs and are
usually not suitable for personalized ranking tasks.
To this end, a class of recommendation models targeting at generating ranked recommendation
lists from implicit feedback have gained popularity. In general, personalized ranking models can
be optimized with pointwise, pairwise or listwise approaches. Pointwise approaches considers
230

a single interaction at a time and train a classifier or a regressor to predict individual prefer-
ences. Matrix factorization and AutoRec are optimized with pointwise objectives. Pairwise ap-
proaches consider a pair of items for each user and aim to approximate the optimal ordering for
that pair. Usually, pairwise approaches are more suitable for the ranking task because predict-
ing relative order is reminiscent to the nature of ranking. Listwise approaches approximate the
ordering of the entire list of items, for example, direct optimizing the ranking measures such
as Normalized Discounted Cumulative Gain (NDCG231 ). However, listwise approaches are more
complex and compute-intensive than pointwise or pairwise approaches. In this section, we will
introduce two pairwise objectives/losses, Bayesian Personalized Ranking loss and Hinge loss, and
their respective implementations.
16.5.1 Bayesian Personalized Ranking Loss and its Implementation
Bayesian personalized ranking (BPR) (Rendle et al., 2009) is a pairwise personalized ranking loss
that is derived from the maximum posterior estimator. It has been widely used in many existing
recommendation models. The training data of BPR consists of both positive and negative pairs
(missing values). It assumes that the user prefers the positive item over all other non-observed
items.
In formal, the training data is constructed by tuples in the form of (u, i, j), which represents that
the user u prefers the item i over the item j. The Bayesian formulation of BPR which aims to
maximize the posterior probability is given below:
p(Θ |>u ) ∝ p(>u | Θ)p(Θ) (16.5.1)
Where Θ represents the parameters of an arbitrary recommendation model, >u represents the de-
sired personalized total ranking of all items for user u. We can formulate the maximum posterior
estimator to derive the generic optimization criterion for the personalized ranking task.
BPR-OPT : = ln p(Θ |>u )

∝ ln p(>u | Θ)p(Θ)
∏
= ln σ(ŷui − ŷuj )p(Θ)
(u,i,j∈D)
∑ (16.5.2)
= ln σ(ŷui − ŷuj ) + ln p(Θ)
(u,i,j∈D)
∑
= ln σ(ŷui − ŷuj ) − λΘ ∥Θ∥2
(u,i,j∈D)
where D := {(u, i, j) | i ∈ Iu+ ∧ j ∈ I\Iu+ } is the training set, with Iu+ denoting the items the user
u liked, I denoting all items, and I\Iu+ indicating all other items excluding items the user liked.
ŷui and ŷuj are the predicted scores of the user u to item i and j, respectively. The prior p(Θ) is a
normal distribution with zero mean and variance-covariance matrix ΣΘ . Here, we let ΣΘ = λΘ I.
231
https://en.wikipedia.org/wiki/Discounted_cumulative_gain
16.5. Personalized Ranking for Recommender Systems 737

We will implement the base class mxnet.gluon.loss.Loss and override the forward method to
construct the Bayesian personalized ranking loss. We begin by importing the Loss class and the
np module.

npx.set_np()
The implementation of BPR loss is as follows.
#@save
class BPRLoss(gluon.loss.Loss):
def __init__(self, weight=None, batch_axis=0, **kwargs):
super(BPRLoss, self).__init__(weight=None, batch_axis=0, **kwargs)
def forward(self, positive, negative):

distances = positive - negative
loss = - np.sum(np.log(npx.sigmoid(distances)), 0, keepdims=True)
return loss
16.5.2 Hinge Loss and its Implementation
The Hinge loss for ranking has different form to the hinge loss232 provided within the gluon library
that is often used in classifiers such as SVMs. The loss used for ranking in recommender systems
has the following form.
∑
max(m − ŷui + ŷuj , 0) (16.5.3)
(u,i,j∈D)
where m is the safety margin size. It aims to push negative items away from positive items. Similar
to BPR, it aims to optimize for relevant distance between positive and negative samples instead of
absolute outputs, making it well suited to recommender systems.
#@save
class HingeLossbRec(gluon.loss.Loss):
232
https://mxnet.incubator.apache.org/api/python/gluon/loss.html#mxnet.gluon.loss.HingeLoss

def __init__(self, weight=None, batch_axis=0, **kwargs):
super(HingeLossbRec, self).__init__(weight=None, batch_axis=0,
**kwargs)
def forward(self, positive, negative, margin=1):

distances = positive - negative
loss = np.sum(np.maximum(- distances + margin, 0))
return loss
These two losses are interchangeable for personalized ranking in recommendation.
Summary
• There are three types of ranking losses available for the personalized ranking task in recom-
mender systems, namely, pointwise, pairwise and listwise methods.
• The two pairwise loses, Bayesian personalized ranking loss and hinge loss, can be used in-
terchangeably.
Exercises
• Are there any variants of BPR and hinge loss available?

• Can you find any recommendation models that use BPR or hinge loss?
Discussions233
16.6 Neural Collaborative Filtering for Personalized Ranking
This section moves beyond explicit feedback, introducing the neural collaborative filtering (NCF)
framework for recommendation with implicit feedback. Implicit feedback is pervasive in recom-
mender systems. Actions such as Clicks, buys, and watches are common implicit feedback which
are easy to collect and indicative of usersʼ preferences. The model we will introduce, titled NeuMF
(He et al., 2017b), short for neural matrix factorization, aims to address the personalized rank-
ing task with implicit feedback. This model leverages the flexibility and non-linearity of neural
networks to replace dot products of matrix factorization, aiming at enhancing the model expres-
siveness. In specific, this model is structured with two subnetworks including generalized matrix
factorization (GMF) and MLP and models the interactions from two pathways instead of simple in-
ner products. The outputs of these two networks are concatenated for the final prediction scores
calculation. Unlike the rating prediction task in AutoRec, this model generates a ranked recom-
mendation list to each user based on the implicit feedback. We will use the personalized ranking
loss introduced in the last section to train this model.
233
16.6. Neural Collaborative Filtering for Personalized Ranking 739

16.6.1 The NeuMF model
As aforementioned, NeuMF fuses two subnetworks. The GMF is a generic neural network version
of matrix factorization where the input is the elementwise product of user and item latent factors.
It consists of two neural layers:
x = p u ⊙ qi
(16.6.1)
ŷui = α(h⊤ x),
where ⊙ denotes the Hadamard product of vectors. P ∈ Rm×k and Q ∈ Rn×k corespond to user
and item latent matrix respectively. pu ∈ Rk is the uth row of P and qi ∈ Rk is the ith row of Q. α
and h denote the activation function and weight of the output layer. ŷui is the prediction score of
the user u might give to the item i.
Another component of this model is MLP. To enrich model flexibility, the MLP subnetwork does
not share user and item embeddings with GMF. It uses the concatenation of user and item embed-
dings as input. With the complicated connections and nonlinear transformations, it is capable of
eastimating the intricate interactions between users and items. More precisely, the MLP subnet-
work is defined as:
z (1) = ϕ1 (Uu , Vi ) = [Uu , Vi ]

ϕ(2) (z (1) ) = α1 (W(2) z (1) + b(2) )
... (16.6.2)
(L) (L−1) L (L) (L−1) (L)
ϕ (z ) = α (W z +b ))
ŷui = α(h⊤ ϕL (z (L) ))
where W∗ , b∗ and α∗ denote the weight matrix, bias vector, and activation function. ϕ∗ denotes
the function of the corresponding layer. z∗ denotes the output of corresponding layer.
To fuse the results of GMF and MLP, instead of simple addition, NeuMF concatenates the second
last layers of two subnetworks to create a feature vector which can be passed to the further lay-
ers. Afterwards, the ouputs are projected with matrix h and a sigmoid activation function. The
prediction layer is formulated as:
ŷui = σ(h⊤ [x, ϕL (z (L) )]). (16.6.3)
The following figure illustrates the model architecture of NeuMF.

Fig. 16.6.1: Illustration of the NeuMF model

import mxnet as mx
import random
import sys
npx.set_np()
The following code implements the NeuMF model. It consists of a generalized matrix factoriza-
tion model and a multi-layered perceptron with different user and item embedding vectors. The
structure of the MLP is controlled with the parameter nums_hiddens. ReLU is used as the default
activation function.
class NeuMF(nn.Block):
def __init__(self, num_factors, num_users, num_items, nums_hiddens,
**kwargs):
super(NeuMF, self).__init__(**kwargs)
self.P = nn.Embedding(num_users, num_factors)
self.Q = nn.Embedding(num_items, num_factors)
self.U = nn.Embedding(num_users, num_factors)
self.V = nn.Embedding(num_items, num_factors)
for num_hiddens in nums_hiddens:
self.mlp.add(nn.Dense(num_hiddens, activation='relu',
use_bias=True))
def forward(self, user_id, item_id):


p_mf = self.P(user_id)
q_mf = self.Q(item_id)
gmf = p_mf * q_mf
p_mlp = self.U(user_id)
q_mlp = self.V(item_id)
mlp = self.mlp(np.concatenate([p_mlp, q_mlp], axis=1))
con_res = np.concatenate([gmf, mlp], axis=1)
return np.sum(con_res, axis=-1)
16.6.3 Customized Dataset with Negative Sampling
For pairwise ranking loss, an important step is negative sampling. For each user, the items that
a user has not interacted with are candidate items (unobserved entries). The following function
takes users identity and candidate items as input, and samples negative items randomly for each
user from the candidate set of that user. During the training stage, the model ensures that the
items that a user likes to be ranked higher than items he dislikes or has not interacted with.
class PRDataset(gluon.data.Dataset):
def __init__(self, users, items, candidates, num_items):
self.users = users
self.items = items
self.cand = candidates
self.all = set([i for i in range(num_items)])
def __len__(self):
return len(self.users)

neg_items = list(self.all - set(self.cand[int(self.users[idx])]))
indices = random.randint(0, len(neg_items) - 1)
return self.users[idx], self.items[idx], neg_items[indices]
16.6.4 Evaluator
In this section, we adopt the splitting by time strategy to construct the training and test sets. Two
evaluation measures including hit rate at given cutting off ℓ (Hit@ℓ) and area under the ROC curve
(AUC) are used to assess the model effectiveness. Hit rate at given position ℓ for each user indicates
that whether the recommended item is included in the top ℓ ranked list. The formal definition is
as follows:
1 ∑
Hit@ℓ = 1(ranku,gu <= ℓ), (16.6.4)
m
u∈U
where 1 denotes an indicator function that is equal to one if the ground truth item is ranked in the
top ℓ list, otherwise it is equal to zero. ranku,gu denotes the ranking of the ground truth item gu of
the user u in the recommendation list (The ideal ranking is 1). m is the number of users. U is the
user set.
The definition of AUC is as follows:
1 ∑ 1 ∑
AUC = 1(ranku,gu < ranku,j ), (16.6.5)
m |I\Su |
u∈U j∈I\Su

where I is the item set. Su is the candidate items of user u. Note that many other evaluation
protocols such as precision, recall and normalized discounted cumulative gain (NDCG) can also
be used.
The following function calculates the hit counts and AUC for each user.
#@save
def hit_and_auc(rankedlist, test_matrix, k):
hits_k = [(idx, val) for idx, val in enumerate(rankedlist[:k])
if val in set(test_matrix)]
hits_all = [(idx, val) for idx, val in enumerate(rankedlist)
if val in set(test_matrix)]
max = len(rankedlist) - 1
auc = 1.0 * (max - hits_all[0][0]) / max if len(hits_all) > 0 else 0
return len(hits_k), auc
Then, the overall Hit rate and AUC are calculated as follows.
#@save
def evaluate_ranking(net, test_input, seq, candidates, num_users, num_items,
devices):
ranked_list, ranked_items, hit_rate, auc = {}, {}, [], []
all_items = set([i for i in range(num_users)])
for u in range(num_users):
neg_items = list(all_items - set(candidates[int(u)]))
user_ids, item_ids, x, scores = [], [], [], []
[item_ids.append(i) for i in neg_items]
[user_ids.append(u) for _ in neg_items]
x.extend([np.array(user_ids)])
if seq is not None:
x.append(seq[user_ids, :])
x.extend([np.array(item_ids)])
test_data_iter = gluon.data.DataLoader(
gluon.data.ArrayDataset(*x), shuffle=False, last_batch="keep",
batch_size=1024)
for index, values in enumerate(test_data_iter):
x = [gluon.utils.split_and_load(v, devices, even_split=False)
for v in values]
scores.extend([list(net(*t).asnumpy()) for t in zip(*x)])
scores = [item for sublist in scores for item in sublist]
item_scores = list(zip(item_ids, scores))
ranked_list[u] = sorted(item_scores, key=lambda t: t[1], reverse=True)
ranked_items[u] = [r[0] for r in ranked_list[u]]
temp = hit_and_auc(ranked_items[u], test_input[u], 50)
hit_rate.append(temp[0])
auc.append(temp[1])
return np.mean(np.array(hit_rate)), np.mean(np.array(auc))

The training function is defined below. We train the model in the pairwise manner.
#@save
def train_ranking(net, train_iter, test_iter, loss, trainer, test_seq_iter,
num_users, num_items, num_epochs, devices, evaluator,
candidates, eval_step=1):
timer, hit_rate, auc = d2l.Timer(), 0, 0
legend=['test hit rate', 'test AUC'])
metric, l = d2l.Accumulator(3), 0.
for i, values in enumerate(train_iter):
input_data = []
for v in values:
input_data.append(gluon.utils.split_and_load(v, devices))
p_pos = [net(*t) for t in zip(*input_data[0:-1])]
p_neg = [net(*t) for t in zip(*input_data[0:-2],
input_data[-1])]
ls = [loss(p, n) for p, n in zip(p_pos, p_neg)]
[l.backward(retain_graph=False) for l in ls]
l += sum([l.asnumpy() for l in ls]).mean()/len(devices)
trainer.step(values[0].shape[0])
metric.add(l, values[0].shape[0], values[0].size)
timer.stop()
with autograd.predict_mode():
if (epoch + 1) % eval_step == 0:
hit_rate, auc = evaluator(net, test_iter, test_seq_iter,
candidates, num_users, num_items,
devices)
animator.add(epoch + 1, (hit_rate, auc))
f'test hit rate {float(hit_rate):.3f}, test AUC {float(auc):.3f}')
Now, we can load the MovieLens 100k dataset and train the model. Since there are only ratings in
the MovieLens dataset, with some losses of accuracy, we binarize these ratings to zeros and ones.
If a user rated an item, we consider the implicit feedback as one, otherwise as zero. The action of
rating an item can be treated as a form of providing implicit feedback. Here, we split the dataset
in the seq-aware mode where usersʼ latest interacted items are left out for test.
batch_size = 1024
train_data, test_data = d2l.split_data_ml100k(df, num_users, num_items,
'seq-aware')
users_train, items_train, ratings_train, candidates = d2l.load_data_ml100k(
train_data, num_users, num_items, feedback="implicit")
users_test, items_test, ratings_test, test_iter = d2l.load_data_ml100k(
test_data, num_users, num_items, feedback="implicit")
PRDataset(users_train, items_train, candidates, num_items ), batch_size,
True, last_batch="rollover", num_workers=d2l.get_dataloader_workers())

We then create and initialize the model. we use a three-layer MLP with constant hidden size 10.
net = NeuMF(10, num_users, num_items, nums_hiddens=[10, 10, 10])
The following code trains the model.

loss = d2l.BPRLoss()
train_ranking(net, train_iter, test_iter, loss, trainer, None, num_users,
num_items, num_epochs, devices, evaluate_ranking, candidates)
train loss 4.399, test hit rate 0.338, test AUC 0.736
Summary
• Adding nonlinearity to matrix factorization model is beneficial for improving the model ca-
pability and effectiveness.
• NeuMF is a combination of matrix factorization and Multilayer perceptron. The multilayer
perceptron takes the concatenation of user and item embeddings as the input.
Exercises
• Vary the size of latent factors. How the size of latent factors impact the model performance?
• Vary the architectures (e.g., number of layers, number of neurons of each layer) of the MLP
to check the its impact on the performance.
• Try different optimizers, learning rate and weight decay rate.
• Try to use hinge loss defined in the last section to optimize this model.

Discussions234
16.7 Sequence-Aware Recommender Systems
In previous sections, we abstract the recommendation task as a matrix completion problem with-
out considering usersʼ short-term behaviors. In this section, we will introduce a recommendation
model that takes the sequentially-ordered user interaction logs into account. It is a sequence-
aware recommender (Quadrana et al., 2018) where the input is an ordered and often timestamped
list of past user actions. A number of recent literatures have demonstrated the usefulness of in-
corporating such information in modeling usersʼ temporal behavioral patterns and discovering
their interest drift.
The model we will introduce, Caser (Tang & Wang, 2018), short for convolutional sequence embed-
ding recommendation model, adopts convolutional neural networks capture the dynamic pattern
influences of usersʼ recent activities. The main component of Caser consists of a horizontal con-
volutional network and a vertical convolutional network, aiming to uncover the union-level and
point-level sequence patterns, respectively. Point-level pattern indicates the impact of single item
in the historical sequence on the target item, while union level pattern implies the influences of
several previous actions on the subsequent target. For example, buying both milk and butter to-
gether leads to higher probability of buying flour than just buying one of them. Moreover, usersʼ
general interests, or long term preferences are also modeled in the last fully-connected layers,
resulting in a more comprehensive modeling of user interests. Details of the model are described
as follows.
16.7.1 Model Architectures
In sequence-aware recommendation system, each user is associated with a sequence of some

items from the item set. Let S u = (S1u , ...S|S
u ) denotes the ordered sequence. The goal of Caser is
u|
to recommend item by considering user general tastes as well as short-term intention. Suppose
we take the previous L items into consideration, an embedding matrix that represents the former
interactions for timestep t can be constructed:
⊤
E(u,t) = [qSt−L
u , ..., qSt−2
u , qS u ] ,
t−1
(16.7.1)
where Q ∈ Rn×k represents item embeddings and qi denotes the ith row. E(u,t) ∈ RL×k can be
used to infer the transient interest of user u at time-step t. We can view the input matrix E(u,t) as
an image which is the input of the subsequent two convolutional components.
The horizontal convolutional layer has d horizontal filters Fj ∈ Rh×k , 1 ≤ j ≤ d, h = {1, ..., L},
and the vertical convolutional layer has d′ vertical filters Gj ∈ RL×1 , 1 ≤ j ≤ d′ . After a series of
convolutional and pool operations, we get the two outputs:
o = HConv(E(u,t) , F)
(16.7.2)
o′ = VConv(E(u,t) , G),
′
where o ∈ Rd is the output of horizontal convolutional network and o′ ∈ Rkd is the output of verti-
cal convolutional network. For simplicity, we omit the details of convolution and pool operations.
234

They are concatenated and fed into a fully-connected neural network layer to get more high-level
representations.
z = ϕ(W[o, o′ ]⊤ + b), (16.7.3)

′
where W ∈ Rk×(d+kd ) is the weight matrix and b ∈ Rk is the bias. The learned vector z ∈ Rk is the
representation of userʼs short-term intent.
At last, the prediction function combines usersʼ short-term and general taste together, which is
defined as:
ŷuit = vi · [z, pu ]⊤ + b′i , (16.7.4)
where V ∈ Rn×2k is another item embedding matrix. b′ ∈ Rn is the item specific bias. P ∈ Rm×k
is the user embedding matrix for usersʼ general tastes. pu ∈ Rk is the uth row of P and vi ∈ R2k is
the ith row of V.
The model can be learned with BPR or Hinge loss. The architecture of Caser is shown below:
Fig. 16.7.1: Illustration of the Caser Model
We first import the required libraries.

import mxnet as mx
import random
import sys
npx.set_np()
16.7. Sequence-Aware Recommender Systems 747

The following code implements the Caser model. It consists of a vertical convolutional layer, a
horizontal convolutional layer, and a full-connected layer.
class Caser(nn.Block):
def __init__(self, num_factors, num_users, num_items, L=5, d=16,
d_prime=4, drop_ratio=0.05, **kwargs):
super(Caser, self).__init__(**kwargs)
self.P = nn.Embedding(num_users, num_factors)
self.Q = nn.Embedding(num_items, num_factors)
self.d_prime, self.d = d_prime, d
# Vertical convolution layer
self.conv_v = nn.Conv2D(d_prime, (L, 1), in_channels=1)
# Horizontal convolution layer
h = [i + 1 for i in range(L)]
self.conv_h, self.max_pool = nn.Sequential(), nn.Sequential()
for i in h:
self.conv_h.add(nn.Conv2D(d, (i, num_factors), in_channels=1))
self.max_pool.add(nn.MaxPool1D(L - i + 1))
# Fully-connected layer
self.fc1_dim_v, self.fc1_dim_h = d_prime * num_factors, d * len(h)
self.fc = nn.Dense(in_units=d_prime * num_factors + d * L,
activation='relu', units=num_factors)
self.Q_prime = nn.Embedding(num_items, num_factors * 2)
self.b = nn.Embedding(num_items, 1)
self.dropout = nn.Dropout(drop_ratio)
def forward(self, user_id, seq, item_id):

item_embs = np.expand_dims(self.Q(seq), 1)
user_emb = self.P(user_id)
out, out_h, out_v, out_hs = None, None, None, []
if self.d_prime:
out_v = self.conv_v(item_embs)
out_v = out_v.reshape(out_v.shape[0], self.fc1_dim_v)
if self.d:
for conv, maxp in zip(self.conv_h, self.max_pool):
conv_out = np.squeeze(npx.relu(conv(item_embs)), axis=3)
t = maxp(conv_out)
pool_out = np.squeeze(t, axis=2)
out_hs.append(pool_out)
out_h = np.concatenate(out_hs, axis=1)
out = np.concatenate([out_v, out_h], axis=1)
z = self.fc(self.dropout(out))
x = np.concatenate([z, user_emb], axis=1)
q_prime_i = np.squeeze(self.Q_prime(item_id))
b = np.squeeze(self.b(item_id))
res = (x * q_prime_i).sum(1) + b
return res

16.7.3 Sequential Dataset with Negative Sampling
To process the sequential interaction data, we need to reimplement the Dataset class. The fol-
lowing code creates a new dataset class named SeqDataset. In each sample, it outputs the user
identity, his previous L interacted items as a sequence and the next item he interacts as the target.
The following figure demonstrates the data loading process for one user. Suppose that this user
liked 8 movies, we organize these eight movies in chronological order. The latest movie is left out
as the test item. For the remaining seven movies, we can get three training samples, with each
sample containing a sequence of five (L = 5) movies and its subsequent item as the target item.
Negative samples are also included in the Customized dataset.
Fig. 16.7.2: Illustration of the data generation process
class SeqDataset(gluon.data.Dataset):
def __init__(self, user_ids, item_ids, L, num_users, num_items,
candidates):
user_ids, item_ids = np.array(user_ids), np.array(item_ids)
sort_idx = np.array(sorted(range(len(user_ids)),
key=lambda k: user_ids[k]))
u_ids, i_ids = user_ids[sort_idx], item_ids[sort_idx]
temp, u_ids, self.cand = {}, u_ids.asnumpy(), candidates
self.all_items = set([i for i in range(num_items)])
[temp.setdefault(u_ids[i], []).append(i) for i, _ in enumerate(u_ids)]
temp = sorted(temp.items(), key=lambda x: x[0])
u_ids = np.array([i[0] for i in temp])
idx = np.array([i[1][0] for i in temp])
self.ns = ns = int(sum([c - L if c >= L + 1 else 1 for c
in np.array([len(i[1]) for i in temp])]))
self.seq_items = np.zeros((ns, L))
self.seq_users = np.zeros(ns, dtype='int32')
self.seq_tgt = np.zeros((ns, 1))
self.test_seq = np.zeros((num_users, L))
test_users, _uid = np.empty(num_users), None
for i, (uid, i_seq) in enumerate(self._seq(u_ids, i_ids, idx, L + 1)):
if uid != _uid:
self.test_seq[uid][:] = i_seq[-L:]
test_users[uid], _uid = uid, uid
self.seq_tgt[i][:] = i_seq[-1:]
self.seq_items[i][:], self.seq_users[i] = i_seq[:L], uid
def _win(self, tensor, window_size, step_size=1):


if len(tensor) - window_size >= 0:
for i in range(len(tensor), 0, - step_size):
if i - window_size >= 0:
yield tensor[i - window_size:i]
else:
break
else:
yield tensor
def _seq(self, u_ids, i_ids, idx, max_len):

for i in range(len(idx)):
stop_idx = None if i >= len(idx) - 1 else int(idx[i + 1])
for s in self._win(i_ids[int(idx[i]):stop_idx], max_len):
yield (int(u_ids[i]), s)
def __len__(self):
return self.ns

neg = list(self.all_items - set(self.cand[int(self.seq_users[idx])]))
i = random.randint(0, len(neg) - 1)
return (self.seq_users[idx], self.seq_items[idx], self.seq_tgt[idx],
neg[i])
16.7.4 Load the MovieLens 100K dataset
Afterwards, we read and split the MovieLens 100K dataset in sequence-aware mode and load the
training data with sequential dataloader implemented above.
TARGET_NUM, L, batch_size = 1, 3, 4096

train_data, test_data = d2l.split_data_ml100k(df, num_users, num_items,
'seq-aware')
users_train, items_train, ratings_train, candidates = d2l.load_data_ml100k(
train_data, num_users, num_items, feedback="implicit")
users_test, items_test, ratings_test, test_iter = d2l.load_data_ml100k(
test_data, num_users, num_items, feedback="implicit")
train_seq_data = SeqDataset(users_train, items_train, L, num_users,
num_items, candidates)
train_iter = gluon.data.DataLoader(train_seq_data, batch_size, True,
last_batch="rollover",
test_seq_iter = train_seq_data.test_seq
train_seq_data[0]
(array(0, dtype=int32), array([110., 255., 4.]), array([101.]), 1590)
The training data structure is shown above. The first element is the user identity, the next list
indicates the last five items this user liked, and the last element is the item this user liked after the
five items.

16.7.5 Train the Model
Now, let us train the model. We use the same setting as NeuMF, including learning rate, optimizer,
and k, in the last section so that the results are comparable.
net = Caser(10, num_users, num_items, L)
loss = d2l.BPRLoss()
d2l.train_ranking(net, train_iter, test_iter, loss, trainer, test_seq_iter,

num_users, num_items, num_epochs, devices,
d2l.evaluate_ranking, candidates, eval_step=1)
train loss 0.829, test hit rate 0.381, test AUC 0.756
Summary
• Inferring a userʼs short-term and long-term interests can make prediction of the next item
that he preferred more effectively.
• Convolutional neural networks can be utilized to capture usersʼ short-term interests from
sequential interactions.

Exercises
• Conduct an ablation study by removing one of the horizontal and vertical convolutional net-
works, which component is the more important ?
• Vary the hyperparameter L. Does longer historical interactions bring higher accuracy?
• Apart from the sequence-aware recommendation task we introduced above, there is another
type of sequence-aware recommendation task called session-based recommendation (Hi-
dasi et al., 2015). Can you explain the differences between these two tasks?
Discussions235
16.8 Feature-Rich Recommender Systems
Interaction data is the most basic indication of usersʼ preferences and interests. It plays a critical
role in former introduced models. Yet, interaction data is usually extremely sparse and can be
noisy at times. To address this issue, we can integrate side information such as features of items,
profiles of users, and even in which context that the interaction occurred into the recommenda-
tion model. Utilizing these features are helpful in making recommendations in that these features
can be an effective predictor of users interests especially when interaction data is lacking. As such,
it is essential for recommendation models also have the capability to deal with those features and
give the model some content/context awareness. To demonstrate this type of recommendation
models, we introduce another task on click-through rate (CTR) for online advertisement recom-
mendations (McMahan et al., 2013) and present an anonymous advertising data. Targeted adver-
tisement services have attracted widespread attention and are often framed as recommendation
engines. Recommending advertisements that match usersʼ personal taste and interest is impor-
tant for click-through rate improvement.
Digital marketers use online advertising to display advertisements to customers. Click-through
rate is a metric that measures the number of clicks advertisers receive on their ads per number of
impressions and it is expressed as a percentage calculated with the formula:
#Clicks
CTR = × 100%. (16.8.1)
#Impressions
Click-through rate is an important signal that indicates the effectiveness of prediction algorithms.
Click-through rate prediction is a task of predicting the likelihood that something on a website will
be clicked. Models on CTR prediction can not only be employed in targeted advertising systems
but also in general item (e.g., movies, news, products) recommender systems, email campaigns,
and even search engines. It is also closely related to user satisfaction, conversion rate, and can be
helpful in setting campaign goals as it can help advertisers to set realistic expectations.
from collections import defaultdict

import os
235

16.8.1 An Online Advertising Dataset
With the considerable advancements of Internet and mobile technology, online advertising has
become an important income resource and generates vast majority of revenue in the Internet
industry. It is important to display relevant advertisements or advertisements that pique usersʼ
interests so that casual visitors can be converted into paying customers. The dataset we intro-
duced is an online advertising dataset. It consists of 34 fields, with the first column representing
the target variable that indicates if an ad was clicked (1) or not (0). All the other columns are
categorical features. The columns might represent the advertisement id, site or application id,
device id, time, user profiles and so on. The real semantics of the features are undisclosed due to
anonymization and privacy concern.
The following code downloads the dataset from our server and saves it into the local data folder.
#@save
d2l.DATA_HUB['ctr'] = (d2l.DATA_URL + 'ctr.zip',
'e18327c48c8e8e5c23da714dd614e390d369843f')
data_dir = d2l.download_extract('ctr')
Downloading ../data/ctr.zip from http://d2l-data.s3-accelerate.amazonaws.com/ctr.zip...
There are a training set and a test set, consisting of 15000 and 3000 samples/lines, respectively.
16.8.2 Dataset Wrapper
For the convenience of data loading, we implement a CTRDataset which loads the advertising
dataset from the CSV file and can be used by DataLoader.
#@save
class CTRDataset(gluon.data.Dataset):
def __init__(self, data_path, feat_mapper=None, defaults=None,
min_threshold=4, num_feat=34):
self.NUM_FEATS, self.count, self.data = num_feat, 0, {}
feat_cnts = defaultdict(lambda: defaultdict(int))
self.feat_mapper, self.defaults = feat_mapper, defaults
self.field_dims = np.zeros(self.NUM_FEATS, dtype=np.int64)
with open(data_path) as f:
for line in f:
instance = {}
values = line.rstrip('\n').split('\t')
if len(values) != self.NUM_FEATS + 1:
continue
label = np.float32([0, 0])
label[int(values[0])] = 1
instance['y'] = [np.float32(values[0])]
for i in range(1, self.NUM_FEATS + 1):
feat_cnts[i][values[i]] += 1
instance.setdefault('x', []).append(values[i])
self.data[self.count] = instance
self.count = self.count + 1
if self.feat_mapper is None and self.defaults is None:
16.8. Feature-Rich Recommender Systems 753

feat_mapper = {i: {feat for feat, c in cnt.items() if c >=
min_threshold} for i, cnt in feat_cnts.items()}
self.feat_mapper = {i: {feat: idx for idx, feat in enumerate(cnt)}
for i, cnt in feat_mapper.items()}
self.defaults = {i: len(cnt) for i, cnt in feat_mapper.items()}
for i, fm in self.feat_mapper.items():
self.field_dims[i - 1] = len(fm) + 1
self.offsets = np.array((0, *np.cumsum(self.field_dims).asnumpy()
[:-1]))
def __len__(self):
return self.count

feat = np.array([self.feat_mapper[i + 1].get(v, self.defaults[i + 1])
for i, v in enumerate(self.data[idx]['x'])])
return feat + self.offsets, self.data[idx]['y']
The following example loads the training data and print out the first record.
train_data = CTRDataset(os.path.join(data_dir, 'train.csv'))

train_data[0]
(array([ 143., 146., 227., 246., 957., 1250., 1471., 1566., 1624.,
1985., 2008., 2061., 2236., 2304., 2305., 2360., 2745., 2746.,
2747., 2748., 2892., 2988., 3165., 3189., 3194., 3195., 3308.,
3645., 3687., 3704., 3723., 3752., 3780., 3805.]),
[1.0])
As can be seen, all the 34 fields are categorical features. Each value represents the one-hot index
of the corresponding entry. The label 0 means that it is not clicked. This CTRDataset can also
be used to load other datasets such as the Criteo display advertising challenge Dataset236 and the
Avazu click-through rate prediction Dataset237 .
Summary
• Click-through rate is an important metric that is used to measure the effectiveness of adver-
tising systems and recommender systems.
• Click-through rate prediction is usually converted to a binary classification problem. The
target is to predict whether an ad/item will be clicked or not based on given features.
236
https://labs.criteo.com/2014/02/kaggle-display-advertising-challenge-dataset/
237
https://www.kaggle.com/c/avazu-ctr-prediction

Exercises
• Can you load the Criteo and Avazu dataset with the provided CTRDataset. It is worth noting
that the Criteo dataset consisting of real-valued features so you may have to revise the code
a bit.
Discussions238
16.9 Factorization Machines
Factorization machines (FM) (Rendle, 2010), proposed by Steffen Rendle in 2010, is a supervised
algorithm that can be used for classification, regression, and ranking tasks. It quickly took notice
and became a popular and impactful method for making predictions and recommendations. Par-
ticularly, it is a generalization of the linear regression model and the matrix factorization model.
Moreover, it is reminiscent of support vector machines with a polynomial kernel. The strengths of
factorization machines over the linear regression and matrix factorization are: (1) it can model χ-
way variable interactions, where χ is the number of polynomial order and is usually set to two. (2)
A fast optimization algorithm associated with factorization machines can reduce the polynomial
computation time to linear complexity, making it extremely efficient especially for high dimen-
sional sparse inputs. For these reasons, factorization machines are widely employed in modern
advertisement and products recommendations. The technical details and implementations are
described below.
16.9.1 2-Way Factorization Machines
Formally, let x ∈ Rd denote the feature vectors of one sample, and y denote the corresponding
label which can be real-valued label or class label such as binary class “click/non-click”. The model
for a factorization machine of degree two is defined as:
∑
d ∑
d ∑
d
ŷ(x) = w0 + wi xi + ⟨vi , vj ⟩xi xj (16.9.1)
i=1 i=1 j=i+1
where w0 ∈ R is the global bias; w ∈ Rd denotes the weights of the i-th variable; V ∈ Rd×k rep-
resents the feature embeddings; vi represents the ith row of V; k is the dimensionality of latent
factors; ⟨·, ·⟩ is the dot product of two vectors. ⟨vi , vj ⟩ model the interaction between the ith and
j th feature. Some feature interactions can be easily understood so they can be designed by ex-
perts. However, most other feature interactions are hidden in data and difficult to identify. So
modeling feature interactions automatically can greatly reduce the efforts in feature engineering.
It is obvious that the first two terms correspond to the linear regression model and the last term is
an extension of the matrix factorization model. If the feature i represents a item and the feature
j represents a user, the third term is exactly the dot product between user and item embeddings.
It is worth noting that FM can also generalize to higher orders (degree > 2). Nevertheless, the
numerical stability might weaken the generalization.
238
16.9. Factorization Machines 755

16.9.2 An Efficient Optimization Criterion
Optimizing the factorization machines in a straight forward method leads to a complexity of

O(kd2 ) as all pairwise interactions require to be computed. To solve this inefficiency problem,
we can reorganize the third term of FM which could greatly reduce the computation cost, lead-
ing to a linear time complexity (O(kd)). The reformulation of the pairwise interaction term is as
follows:
∑
d ∑
d
⟨vi , vj ⟩xi xj
i=1 j=i+1
1 ∑∑ 1∑
d d d
= ⟨vi , vj ⟩xi xj − ⟨vi , vi ⟩xi xi
2 2
i=1 j=1 i=1
1( ∑
d ∑
d ∑
k ∑
d ∑
k
)
= vi,l vj,l xi xj − vi,l vi,l xi xi (16.9.2)
2
i=1 j=1 l=1 i=1 l=1
1 ∑
k
( ∑
d ∑
d ∑
d
)
= ( vi,l xi )( vj,l xj ) − v2i,l x2i
2
l=1 i=1 j=1 i=1
1 ∑( ∑ ∑
k d d
= ( vi,l xi )2 − v2i,l x2i )
2
l=1 i=1 i=1
With this reformulation, the model complexity are decreased greatly. Moreover, for sparse fea-
tures, only non-zero elements needs to be computed so that the overall complexity is linear to the
number of non-zero features.
To learn the FM model, we can use the MSE loss for regression task, the cross entropy loss for
classification tasks, and the BPR loss for ranking task. Standard optimizers such as SGD and Adam
are viable for optimization.

from mxnet import init, gluon, np, npx
import os
import sys
npx.set_np()
The following code implement the factorization machines. It is clear to see that FM consists a
linear regression block and an efficient feature interaction block. We apply a sigmoid function
over the final score since we treat the CTR prediction as a classification task.
class FM(nn.Block):
def __init__(self, field_dims, num_factors):
super(FM, self).__init__()
num_inputs = int(sum(field_dims))
self.embedding = nn.Embedding(num_inputs, num_factors)
self.fc = nn.Embedding(num_inputs, 1)
self.linear_layer = nn.Dense(1, use_bias=True)


square_of_sum = np.sum(self.embedding(x), axis=1) ** 2
sum_of_square = np.sum(self.embedding(x) ** 2, axis=1)
x = self.linear_layer(self.fc(x).sum(1)) \
+ 0.5 * (square_of_sum - sum_of_square).sum(1, keepdims=True)
x = npx.sigmoid(x)
return x
16.9.4 Load the Advertising Dataset
We use the CTR data wrapper from the last section to load the online advertising dataset.
batch_size = 2048
train_data = d2l.CTRDataset(os.path.join(data_dir, 'train.csv'))
test_data = d2l.CTRDataset(os.path.join(data_dir, 'test.csv'),
feat_mapper=train_data.feat_mapper,
defaults=train_data.defaults)
train_data, shuffle=True, last_batch='rollover', batch_size=batch_size,
test_data, shuffle=False, last_batch='rollover', batch_size=batch_size,
16.9.5 Train the Model
Afterwards, we train the model. The learning rate is set to 0.01 and the embedding size is set to 20
by default. The Adam optimizer and the SigmoidBinaryCrossEntropyLoss loss are used for model
training.
net = FM(train_data.field_dims, num_factors=20)
lr, num_epochs, optimizer = 0.02, 30, 'adam'

16.9. Factorization Machines 757

Summary
• FM is a general framework that can be applied on a variety of tasks such as regression, clas-
sification, and ranking.
• Feature interaction/crossing is important for prediction tasks and the 2-way interaction can
be efficiently modeled with FM.
Exercises
• Can you test FM on other dataset such as Avazu, MovieLens, and Criteo datasets?
• Vary the embedding size to check its impact on performance, can you observe a similar
pattern as that of matrix factorization?
Discussions239
16.10 Deep Factorization Machines
Learning effective feature combinations is critical to the success of click-through rate prediction
task. Factorization machines model feature interactions in a linear paradigm (e.g., bilinear in-
teractions). This is often insufficient for real-world data where inherent feature crossing struc-
tures are usually very complex and nonlinear. Whatʼs worse, second-order feature interactions
are generally used in factorization machines in practice. Modeling higher degrees of feature com-
binations with factorization machines is possible theoretically but it is usually not adopted due to
numerical instability and high computational complexity.
One effective solution is using deep neural networks. Deep neural networks are powerful in fea-
ture representation learning and have the potential to learn sophisticated feature interactions.
As such, it is natural to integrate deep neural networks to factorization machines. Adding non-
linear transformation layers to factorization machines gives it the capability to model both low-
order feature combinations and high-order feature combinations. Moreover, non-linear inherent
structures from inputs can also be captured with deep neural networks. In this section, we will
239

introduce a representative model named deep factorization machines (DeepFM) (Guo et al., 2017)
which combine FM and deep neural networks.
16.10.1 Model Architectures
DeepFM consists of an FM component and a deep component which are integrated in a paral-
lel structure. The FM component is the same as the 2-way factorization machines which is used
to model the low-order feature interactions. The deep component is a multi-layered perceptron
that is used to capture high-order feature interactions and nonlinearities. These two components
share the same inputs/embeddings and their outputs are summed up as the final prediction. It
is worth pointing out that the spirit of DeepFM resembles that of the Wide & Deep architecture
which can capture both memorization and generalization. The advantages of DeepFM over the
Wide & Deep model is that it reduces the effort of hand-crafted feature engineering by identifying
feature combinations automatically.
We omit the description of the FM component for brevity and denote the output as ŷ (F M ) . Readers
are referred to the last section for more details. Let ei ∈ Rk denote the latent feature vector of the
ith field. The input of the deep component is the concatenation of the dense embeddings of all
fields that are looked up with the sparse categorical feature input, denoted as:
z(0) = [e1 , e2 , ..., ef ], (16.10.1)
where f is the number of fields. It is then fed into the following neural network:
z(l) = α(W(l) z(l−1) + b(l) ), (16.10.2)
where α is the activation function. Wl and bl are the weight and bias at the lth layer. Let yDN N
denote the output of the prediction. The ultimate prediction of DeepFM is the summation of the
outputs from both FM and DNN. So we have:
ŷ = σ(ŷ (F M ) + ŷ (DN N ) ), (16.10.3)
where σ is the sigmoid function. The architecture of DeepFM is illustrated below.
16.10. Deep Factorization Machines 759

It is worth noting that DeepFM is not the only way to combine deep neural networks with FM. We
can also add nonlinear layers over the feature interactions (He & Chua, 2017).

from mxnet import init, gluon, np, npx
import os
import sys
npx.set_np()
16.10.2 Implemenation of DeepFM
The implementation of DeepFM is similar to that of FM. We keep the FM part unchanged and use
an MLP block with relu as the activation function. Dropout is also used to regularize the model.
The number of neurons of the MLP can be adjusted with the mlp_dims hyperparameter.
class DeepFM(nn.Block):
def __init__(self, field_dims, num_factors, mlp_dims, drop_rate=0.1):
super(DeepFM, self).__init__()
num_inputs = int(sum(field_dims))
self.embedding = nn.Embedding(num_inputs, num_factors)
self.fc = nn.Embedding(num_inputs, 1)
self.linear_layer = nn.Dense(1, use_bias=True)
input_dim = self.embed_output_dim = len(field_dims) * num_factors
for dim in mlp_dims:
self.mlp.add(nn.Dense(dim, 'relu', True, in_units=input_dim))
self.mlp.add(nn.Dropout(rate=drop_rate))

input_dim = dim
self.mlp.add(nn.Dense(in_units=input_dim, units=1))

embed_x = self.embedding(x)
square_of_sum = np.sum(embed_x, axis=1) ** 2
sum_of_square = np.sum(embed_x ** 2, axis=1)
inputs = np.reshape(embed_x, (-1, self.embed_output_dim))
x = self.linear_layer(self.fc(x).sum(1)) \
+ 0.5 * (square_of_sum - sum_of_square).sum(1, keepdims=True) \
+ self.mlp(inputs)
x = npx.sigmoid(x)
return x
The data loading process is the same as that of FM. We set the MLP component of DeepFM to a
three-layered dense network with the a pyramid structure (30-20-10). All other hyperparameters
remain the same as FM.
batch_size = 2048
train_data = d2l.CTRDataset(os.path.join(data_dir, 'train.csv'))
test_data = d2l.CTRDataset(os.path.join(data_dir, 'test.csv'),
feat_mapper=train_data.feat_mapper,
defaults=train_data.defaults)
field_dims = train_data.field_dims
train_data, shuffle=True, last_batch='rollover', batch_size=batch_size,
test_data, shuffle=False, last_batch='rollover', batch_size=batch_size,
net = DeepFM(field_dims, num_factors=10, mlp_dims=[30, 20, 10])
lr, num_epochs, optimizer = 0.01, 30, 'adam'

16.10. Deep Factorization Machines 761

Compared with FM, DeepFM converges faster and achieves better performance.
Summary
• Integrating neural networks to FM enables it to model complex and high-order interactions.

• DeepFM outperforms the original FM on the advertising dataset.
Exercises
• Vary the structure of the MLP to check its impact on model performance.
• Change the dataset to Criteo and compare it with the original FM model.
Discussions240
240

17 | Generative Adversarial Networks
17.1 Generative Adversarial Networks
Throughout most of this book, we have talked about how to make predictions. In some form or
another, we used deep neural networks learned mappings from data examples to labels. This kind
of learning is called discriminative learning, as in, weʼd like to be able to discriminate between
photos cats and photos of dogs. Classifiers and regressors are both examples of discriminative
learning. And neural networks trained by backpropagation have upended everything we thought
we knew about discriminative learning on large complicated datasets. Classification accuracies
on high-res images has gone from useless to human-level (with some caveats) in just 5-6 years. We
will spare you another spiel about all the other discriminative tasks where deep neural networks
do astoundingly well.
But there is more to machine learning than just solving discriminative tasks. For example, given
a large dataset, without any labels, we might want to learn a model that concisely captures the
characteristics of this data. Given such a model, we could sample synthetic data examples that
resemble the distribution of the training data. For example, given a large corpus of photographs
of faces, we might want to be able to generate a new photorealistic image that looks like it might
plausibly have come from the same dataset. This kind of learning is called generative modeling.
Until recently, we had no method that could synthesize novel photorealistic images. But the suc-
cess of deep neural networks for discriminative learning opened up new possibilities. One big
trend over the last three years has been the application of discriminative deep nets to overcome
challenges in problems that we do not generally think of as supervised learning problems. The
recurrent neural network language models are one example of using a discriminative network
(trained to predict the next character) that once trained can act as a generative model.
In 2014, a breakthrough paper introduced Generative adversarial networks (GANs) (Goodfellow
et al., 2014), a clever new way to leverage the power of discriminative models to get good gener-
ative models. At their heart, GANs rely on the idea that a data generator is good if we cannot tell
fake data apart from real data. In statistics, this is called a two-sample test - a test to answer the
question whether datasets X = {x1 , . . . , xn } and X ′ = {x′1 , . . . , x′n } were drawn from the same dis-
tribution. The main difference between most statistics papers and GANs is that the latter use this
idea in a constructive way. In other words, rather than just training a model to say “hey, these two
datasets do not look like they came from the same distribution”, they use the two-sample test241 to
provide training signals to a generative model. This allows us to improve the data generator until
it generates something that resembles the real data. At the very least, it needs to fool the classifier.
Even if our classifier is a state of the art deep neural network.
241
https://en.wikipedia.org/wiki/Two-sample_hypothesis_testing
763
Fig. 17.1.1: Generative Adversarial Networks
The GAN architecture is illustrated in Fig. 17.1.1. As you can see, there are two pieces in GAN
architecture - first off, we need a device (say, a deep network but it really could be anything, such
as a game rendering engine) that might potentially be able to generate data that looks just like the
real thing. If we are dealing with images, this needs to generate images. If we are dealing with
speech, it needs to generate audio sequences, and so on. We call this the generator network. The
second component is the discriminator network. It attempts to distinguish fake and real data from
each other. Both networks are in competition with each other. The generator network attempts
to fool the discriminator network. At that point, the discriminator network adapts to the new fake
data. This information, in turn is used to improve the generator network, and so on.
The discriminator is a binary classifier to distinguish if the input x is real (from real data) or fake
(from the generator). Typically, the discriminator outputs a scalar prediction o ∈ R for input x,
such as using a dense layer with hidden size 1, and then applies sigmoid function to obtain the
predicted probability D(x) = 1/(1 + e−o ). Assume the label y for the true data is 1 and 0 for the
fake data. We train the discriminator to minimize the cross-entropy loss, i.e.,
min{−y log D(x) − (1 − y) log(1 − D(x))}, (17.1.1)

D
For the generator, it first draws some parameter z ∈ Rd from a source of randomness, e.g., a
normal distribution z ∼ N (0, 1). We often call z as the latent variable. It then applies a function
to generate x′ = G(z). The goal of the generator is to fool the discriminator to classify x′ = G(z)
as true data, i.e., we want D(G(z)) ≈ 1. In other words, for a given discriminator D, we update
the parameters of the generator G to maximize the cross-entropy loss when y = 0, i.e.,
max{−(1 − y) log(1 − D(G(z)))} = max{− log(1 − D(G(z)))}. (17.1.2)

G G
If the generator does a perfect job, then D(x′ ) ≈ 1 so the above loss near 0, which results the
gradients are too small to make a good progress for the discriminator. So commonly we minimize
the following loss:
min{−y log(D(G(z)))} = min{− log(D(G(z)))}, (17.1.3)

G G
which is just feed x′ = G(z) into the discriminator but giving label y = 1.
To sum up, D and G are playing a “minimax” game with the comprehensive objective function:
minD maxG {−Ex∼Data logD(x) − Ez∼Noise log(1 − D(G(z)))}. (17.1.4)
764 Chapter 17. Generative Adversarial Networks

Many of the GANs applications are in the context of images. As a demonstration purpose, we
are going to content ourselves with fitting a much simpler distribution first. We will illustrate
what happens if we use GANs to build the worldʼs most inefficient estimator of parameters for a
Gaussian. Let us get started.
%matplotlib inline
npx.set_np()
17.1.1 Generate some “real” data
Since this is going to be the worldʼs lamest example, we simply generate data drawn from a Gaus-
sian.
X = np.random.normal(size=(1000, 2))
A = np.array([[1, 2], [-0.1, 0.5]])
b = np.array([1, 2])
data = X.dot(A) + b
Let us see what we got. This should be a Gaussian shifted in some rather arbitrary way with mean
b and covariance matrix AT A.
d2l.set_figsize()
d2l.plt.scatter(data[:100, 0].asnumpy(), data[:100, 1].asnumpy());
print(f'The covariance matrix is\n{np.dot(A.T, A)}')
The covariance matrix is

[[1.01 1.95]
[1.95 4.25]]
batch_size = 8
data_iter = d2l.load_array((data,), batch_size)
17.1. Generative Adversarial Networks 765

17.1.2 Generator
Our generator network will be the simplest network possible - a single layer linear model. This
is since we will be driving that linear network with a Gaussian data generator. Hence, it literally
only needs to learn the parameters to fake things perfectly.
net_G = nn.Sequential()
net_G.add(nn.Dense(2))
17.1.3 Discriminator
For the discriminator we will be a bit more discriminating: we will use an MLP with 3 layers to
make things a bit more interesting.
net_D = nn.Sequential()
net_D.add(nn.Dense(5, activation='tanh'),
nn.Dense(3, activation='tanh'),
nn.Dense(1))
17.1.4 Training
First we define a function to update the discriminator.
#@save
def update_D(X, Z, net_D, net_G, loss, trainer_D):
"""Update discriminator."""
ones = np.ones((batch_size,), ctx=X.ctx)
zeros = np.zeros((batch_size,), ctx=X.ctx)
real_Y = net_D(X)
fake_X = net_G(Z)
# Do not need to compute gradient for `net_G`, detach it from
# computing gradients.
fake_Y = net_D(fake_X.detach())
loss_D = (loss(real_Y, ones) + loss(fake_Y, zeros)) / 2
loss_D.backward()
trainer_D.step(batch_size)
return float(loss_D.sum())
The generator is updated similarly. Here we reuse the cross-entropy loss but change the label of
the fake data from 0 to 1.
def update_G(Z, net_D, net_G, loss, trainer_G): #@save

"""Update generator."""
batch_size = Z.shape[0]
ones = np.ones((batch_size,), ctx=Z.ctx)
# We could reuse `fake_X` from ùpdate_D` to save computation
fake_X = net_G(Z)
# Recomputing `fake_Y` is needed since `net_D` is changed

fake_Y = net_D(fake_X)
loss_G = loss(fake_Y, ones)
loss_G.backward()
trainer_G.step(batch_size)
return float(loss_G.sum())
Both the discriminator and the generator performs a binary logistic regression with the cross-
entropy loss. We use Adam to smooth the training process. In each iteration, we first update the
discriminator and then the generator. We visualize both losses and generated examples.
def train(net_D, net_G, data_iter, num_epochs, lr_D, lr_G, latent_dim, data):

loss = gluon.loss.SigmoidBCELoss()
net_D.initialize(init=init.Normal(0.02), force_reinit=True)
net_G.initialize(init=init.Normal(0.02), force_reinit=True)
trainer_D = gluon.Trainer(net_D.collect_params(),
'adam', {'learning_rate': lr_D})
trainer_G = gluon.Trainer(net_G.collect_params(),
'adam', {'learning_rate': lr_G})
xlim=[1, num_epochs], nrows=2, figsize=(5, 5),
legend=['generator', 'discriminator'])
animator.fig.subplots_adjust(hspace=0.3)
# Train one epoch
timer = d2l.Timer()
metric = d2l.Accumulator(3) # loss_D, loss_G, num_examples
for X in data_iter:
Z = np.random.normal(0, 1, size=(batch_size, latent_dim))
metric.add(update_D(X, Z, net_D, net_G, loss, trainer_D),
update_G(Z, net_D, net_G, loss, trainer_G),
batch_size)
# Visualize generated examples
Z = np.random.normal(0, 1, size=(100, latent_dim))
fake_X = net_G(Z).asnumpy()
animator.axes[1].cla()
animator.axes[1].scatter(data[:, 0], data[:, 1])
animator.axes[1].scatter(fake_X[:, 0], fake_X[:, 1])
animator.axes[1].legend(['real', 'generated'])
# Show the losses
loss_D, loss_G = metric[0]/metric[2], metric[1]/metric[2]
animator.add(epoch + 1, (loss_D, loss_G))
print(f'loss_D {loss_D:.3f}, loss_G {loss_G:.3f}, '
f'{metric[2] / timer.stop():.1f} examples/sec')
Now we specify the hyperparameters to fit the Gaussian distribution.
lr_D, lr_G, latent_dim, num_epochs = 0.05, 0.005, 2, 20

train(net_D, net_G, data_iter, num_epochs, lr_D, lr_G,
latent_dim, data[:100].asnumpy())
loss_D 0.693, loss_G 0.693, 616.3 examples/sec
17.1. Generative Adversarial Networks 767

Summary
• Generative adversarial networks (GANs) composes of two deep networks, the generator and
the discriminator.
• The generator generates the image as much closer to the true image as possible to fool the
discriminator, via maximizing the cross-entropy loss, i.e., max log(D(x′ )).
• The discriminator tries to distinguish the generated images from the true images, via mini-
mizing the cross-entropy loss, i.e., min −y log D(x) − (1 − y) log(1 − D(x)).
Exercises
• Does an equilibrium exist where the generator wins, i.e. the discriminator ends up unable
to distinguish the two distributions on finite samples?
Discussions242
242

17.2 Deep Convolutional Generative Adversarial Networks
In Section 17.1, we introduced the basic ideas behind how GANs work. We showed that they can
draw samples from some simple, easy-to-sample distribution, like a uniform or normal distribu-
tion, and transform them into samples that appear to match the distribution of some dataset. And
while our example of matching a 2D Gaussian distribution got the point across, it is not especially
exciting.
In this section, we will demonstrate how you can use GANs to generate photorealistic images. We
will be basing our models on the deep convolutional GANs (DCGAN) introduced in (Radford et
al., 2015). We will borrow the convolutional architecture that have proven so successful for dis-
criminative computer vision problems and show how via GANs, they can be leveraged to generate
photorealistic images.

npx.set_np()
17.2.1 The Pokemon Dataset
The dataset we will use is a collection of Pokemon sprites obtained from pokemondb243 . First
download, extract and load this dataset.
#@save
d2l.DATA_HUB['pokemon'] = (d2l.DATA_URL + 'pokemon.zip',
'c065c0e2593b8b161a2d7873e42418bf6a21106c')
data_dir = d2l.download_extract('pokemon')
pokemon = gluon.data.vision.datasets.ImageFolderDataset(data_dir)
Downloading ../data/pokemon.zip from http://d2l-data.s3-accelerate.amazonaws.com/pokemon.zip.

,→..
We resize each image into 64 × 64. The ToTensor transformation will project the pixel value into
[0, 1], while our generator will use the tanh function to obtain outputs in [−1, 1]. Therefore we
normalize the data with 0.5 mean and 0.5 standard deviation to match the value range.
batch_size = 256
transformer = gluon.data.vision.transforms.Compose([
gluon.data.vision.transforms.Normalize(0.5, 0.5)
])
data_iter = gluon.data.DataLoader(
pokemon.transform_first(transformer), batch_size=batch_size,
shuffle=True, num_workers=d2l.get_dataloader_workers())
Let us visualize the first 20 images.

243
https://pokemondb.net/sprites
17.2. Deep Convolutional Generative Adversarial Networks 769

imgs = X[0:20,:,:,:].transpose(0, 2, 3, 1)/2+0.5
d2l.show_images(imgs, num_rows=4, num_cols=5)
break
17.2.2 The Generator
The generator needs to map the noise variable z ∈ Rd , a length-d vector, to a RGB image with width
and height to be 64 × 64 . In Section 13.11 we introduced the fully convolutional network that uses
transposed convolution layer (refer to Section 13.10) to enlarge input size. The basic block of the
generator contains a transposed convolution layer followed by the batch normalization and ReLU
activation.
class G_block(nn.Block):
def __init__(self, channels, kernel_size=4,
strides=2, padding=1, **kwargs):
super(G_block, self).__init__(**kwargs)
self.conv2d_trans = nn.Conv2DTranspose(
channels, kernel_size, strides, padding, use_bias=False)
self.batch_norm = nn.BatchNorm()
self.activation = nn.Activation('relu')

return self.activation(self.batch_norm(self.conv2d_trans(X)))
In default, the transposed convolution layer uses a kh = kw = 4 kernel, a sh = sw = 2 strides, and

′ ′
a ph = pw = 1 padding. With a input shape of nh × nw = 16 × 16, the generator block will double
inputʼs width and height.
′ ′
nh × nw = [(nh kh − (nh − 1)(kh − sh ) − 2ph ] × [(nw kw − (nw − 1)(kw − sw ) − 2pw ]
= [(kh + sh (nh − 1) − 2ph ] × [(kw + sw (nw − 1) − 2pw ]
(17.2.1)
= [(4 + 2 × (16 − 1) − 2 × 1] × [(4 + 2 × (16 − 1) − 2 × 1]
= 32 × 32.
x = np.zeros((2, 3, 16, 16))

g_blk = G_block(20)
g_blk.initialize()
g_blk(x).shape
(2, 20, 32, 32)
If changing the transposed convolution layer to a 4 × 4 kernel, 1 × 1 strides and zero padding. With
a input size of 1 × 1, the output will have its width and height increased by 3 respectively.
x = np.zeros((2, 3, 1, 1))
g_blk = G_block(20, strides=1, padding=0)
g_blk.initialize()
g_blk(x).shape
(2, 20, 4, 4)
The generator consists of four basic blocks that increase inputʼs both width and height from 1
to 32. At the same time, it first projects the latent variable into 64 × 8 channels, and then halve
the channels each time. At last, a transposed convolution layer is used to generate the output. It
further doubles the width and height to match the desired 64 × 64 shape, and reduces the channel
size to 3. The tanh activation function is applied to project output values into the (−1, 1) range.
n_G = 64
net_G = nn.Sequential()
net_G.add(G_block(n_G*8, strides=1, padding=0), # Output: (64 * 8, 4, 4)
G_block(n_G*4), # Output: (64 * 4, 8, 8)
G_block(n_G*2), # Output: (64 * 2, 16, 16)
G_block(n_G), # Output: (64, 32, 32)
nn.Conv2DTranspose(
3, kernel_size=4, strides=2, padding=1, use_bias=False,
activation='tanh')) # Output: (3, 64, 64)
Generate a 100 dimensional latent variable to verify the generatorʼs output shape.
x = np.zeros((1, 100, 1, 1))

net_G.initialize()
net_G(x).shape

(1, 3, 64, 64)
17.2.3 Discriminator
The discriminator is a normal convolutional network network except that it uses a leaky ReLU as
its activation function. Given α ∈ [0, 1], its definition is
{
x if x > 0
leaky ReLU(x) = . (17.2.2)
αx otherwise
As it can be seen, it is normal ReLU if α = 0, and an identity function if α = 1. For α ∈ (0, 1), leaky
ReLU is a nonlinear function that give a non-zero output for a negative input. It aims to fix the
“dying ReLU” problem that a neuron might always output a negative value and therefore cannot
make any progress since the gradient of ReLU is 0.
alphas = [0, 0.2, 0.4, .6, .8, 1]

x = np.arange(-2, 1, 0.1)
Y = [nn.LeakyReLU(alpha)(x).asnumpy() for alpha in alphas]
d2l.plot(x.asnumpy(), Y, 'x', 'y', alphas)
The basic block of the discriminator is a convolution layer followed by a batch normalization layer
and a leaky ReLU activation. The hyperparameters of the convolution layer are similar to the
transpose convolution layer in the generator block.
class D_block(nn.Block):
def __init__(self, channels, kernel_size=4, strides=2,
padding=1, alpha=0.2, **kwargs):
super(D_block, self).__init__(**kwargs)
self.conv2d = nn.Conv2D(
channels, kernel_size, strides, padding, use_bias=False)
self.batch_norm = nn.BatchNorm()
self.activation = nn.LeakyReLU(alpha)

return self.activation(self.batch_norm(self.conv2d(X)))

A basic block with default settings will halve the width and height of the inputs, as we demon-
strated in Section 6.3. For example, given a input shape nh = nw = 16, with a kernel shape
kh = kw = 4, a stride shape sh = sw = 2, and a padding shape ph = pw = 1, the output shape will
be:
′ ′
nh × nw = ⌊(nh − kh + 2ph + sh )/sh ⌋ × ⌊(nw − kw + 2pw + sw )/sw ⌋
= ⌊(16 − 4 + 2 × 1 + 2)/2⌋ × ⌊(16 − 4 + 2 × 1 + 2)/2⌋ (17.2.3)
= 8 × 8.
x = np.zeros((2, 3, 16, 16))

d_blk = D_block(20)
d_blk.initialize()
d_blk(x).shape
(2, 20, 8, 8)
The discriminator is a mirror of the generator.
n_D = 64
net_D = nn.Sequential()
net_D.add(D_block(n_D), # Output: (64, 32, 32)
D_block(n_D*2), # Output: (64 * 2, 16, 16)
D_block(n_D*4), # Output: (64 * 4, 8, 8)
D_block(n_D*8), # Output: (64 * 8, 4, 4)
nn.Conv2D(1, kernel_size=4, use_bias=False)) # Output: (1, 1, 1)
It uses a convolution layer with output channel 1 as the last layer to obtain a single prediction
value.
x = np.zeros((1, 3, 64, 64))

net_D.initialize()
net_D(x).shape
(1, 1, 1, 1)
17.2.4 Training
Compared to the basic GAN in Section 17.1, we use the same learning rate for both generator and
discriminator since they are similar to each other. In addition, we change β1 in Adam (Section
11.10) from 0.9 to 0.5. It decreases the smoothness of the momentum, the exponentially weighted
moving average of past gradients, to take care of the rapid changing gradients because the gener-
ator and the discriminator fight with each other. Besides, the random generated noise Z, is a 4-D
tensor and we are using GPU to accelerate the computation.
def train(net_D, net_G, data_iter, num_epochs, lr, latent_dim,

device=d2l.try_gpu()):
loss = gluon.loss.SigmoidBCELoss()
net_D.initialize(init=init.Normal(0.02), force_reinit=True, ctx=device)
net_G.initialize(init=init.Normal(0.02), force_reinit=True, ctx=device)

trainer_hp = {'learning_rate': lr, 'beta1': 0.5}
trainer_D = gluon.Trainer(net_D.collect_params(), 'adam', trainer_hp)
trainer_G = gluon.Trainer(net_G.collect_params(), 'adam', trainer_hp)
xlim=[1, num_epochs], nrows=2, figsize=(5, 5),
legend=['discriminator', 'generator'])
animator.fig.subplots_adjust(hspace=0.3)
for epoch in range(1, num_epochs + 1):
# Train one epoch
timer = d2l.Timer()
metric = d2l.Accumulator(3) # loss_D, loss_G, num_examples
for X, _ in data_iter:
Z = np.random.normal(0, 1, size=(batch_size, latent_dim, 1, 1))
X, Z = X.as_in_ctx(device), Z.as_in_ctx(device),
metric.add(d2l.update_D(X, Z, net_D, net_G, loss, trainer_D),
d2l.update_G(Z, net_D, net_G, loss, trainer_G),
batch_size)
# Show generated examples
Z = np.random.normal(0, 1, size=(21, latent_dim, 1, 1), ctx=device)
# Normalize the synthetic data to N(0, 1)
fake_x = net_G(Z).transpose(0, 2, 3, 1) / 2 + 0.5
imgs = np.concatenate(
[np.concatenate([fake_x[i * 7 + j] for j in range(7)], axis=1)
for i in range(len(fake_x)//7)], axis=0)
animator.axes[1].cla()
animator.axes[1].imshow(imgs.asnumpy())
# Show the losses
loss_D, loss_G = metric[0] / metric[2], metric[1] / metric[2]
animator.add(epoch, (loss_D, loss_G))
print(f'loss_D {loss_D:.3f}, loss_G {loss_G:.3f}, '
f'{metric[2] / timer.stop():.1f} examples/sec on {str(device)}')
We train the model with a small number of epochs just for demonstration. For better perfor-
mance, the variable num_epochs can be set to a larger number.
latent_dim, lr, num_epochs = 100, 0.005, 20

train(net_D, net_G, data_iter, num_epochs, lr, latent_dim)
loss_D 0.060, loss_G 5.690, 2661.8 examples/sec on gpu(0)

Summary
• DCGAN architecture has four convolutional layers for the Discriminator and four
“fractionally-strided” convolutional layers for the Generator.
• The Discriminator is a 4-layer strided convolutions with batch normalization (except its input
layer) and leaky ReLU activations.
• Leaky ReLU is a nonlinear function that give a non-zero output for a negative input. It aims
to fix the “dying ReLU” problem and helps the gradients flow easier through the architecture.
Exercises
1. What will happen if we use standard ReLU activation rather than leaky ReLU?
2. Apply DCGAN on Fashion-MNIST and see which category works well and which does not.
Discussions244
244

18 | Appendix: Mathematics for Deep
Learning
Brent Werness (Amazon), Rachel Hu (Amazon), and authors of this book

One of the wonderful parts of modern deep learning is the fact that much of it can be understood
and used without a full understanding of the mathematics below it. This is a sign that the field
is maturing. Just as most software developers no longer need to worry about the theory of com-
putable functions, neither should deep learning practitioners need to worry about the theoretical
foundations of maximum likelihood learning.
But, we are not quite there yet.
In practice, you will sometimes need to understand how architectural choices influence gradient
flow, or the implicit assumptions you make by training with a certain loss function. You might
need to know what in the world entropy measures, and how it can help you understand exactly
what bits-per-character means in your model. These all require deeper mathematical understand-
ing.
This appendix aims to provide you the mathematical background you need to understand the core
theory of modern deep learning, but it is not exhaustive. We will begin with examining linear al-
gebra in greater depth. We develop a geometric understanding of all the common linear algebraic
objects and operations that will enable us to visualize the effects of various transformations on our
data. A key element is the development of the basics of eigen-decompositions.
We next develop the theory of differential calculus to the point that we can fully understand why
the gradient is the direction of steepest descent, and why back-propagation takes the form it does.
Integral calculus is then discussed to the degree needed to support our next topic, probability
theory.
Problems encountered in practice frequently are not certain, and thus we need a language to speak
about uncertain things. We review the theory of random variables and the most commonly en-
countered distributions so we may discuss models probabilistically. This provides the foundation
for the naive Bayes classifier, a probabilistic classification technique.
Closely related to probability theory is the study of statistics. While statistics is far too large a field
to do justice in a short section, we will introduce fundamental concepts that all machine learning
practitioners should be aware of, in particular: evaluating and comparing estimators, conducting
hypothesis tests, and constructing confidence intervals.
Last, we turn to the topic of information theory, which is the mathematical study of information
storage and transmission. This provides the core language by which we may discuss quantitatively
how much information a model holds on a domain of discourse.
777
Taken together, these form the core of the mathematical concepts needed to begin down the path
towards a deep understanding of deep learning.
18.1 Geometry and Linear Algebraic Operations
In Section 2.3, we encountered the basics of linear algebra and saw how it could be used to express
common operations for transforming our data. Linear algebra is one of the key mathematical pil-
lars underlying much of the work that we do deep learning and in machine learning more broadly.
While Section 2.3 contained enough machinery to communicate the mechanics of modern deep
learning models, there is a lot more to the subject. In this section, we will go deeper, highlighting
some geometric interpretations of linear algebra operations, and introducing a few fundamental
concepts, including of eigenvalues and eigenvectors.
18.1.1 Geometry of Vectors
First, we need to discuss the two common geometric interpretations of vectors, as either points
or directions in space. Fundamentally, a vector is a list of numbers such as the Python list below.
v = [1, 7, 0, 1]
Mathematicians most often write this as either a column or row vector, which is to say either as
 
1
7
x= 
0 , (18.1.1)
1
or
[ ]
x⊤ = 1 7 0 1 . (18.1.2)
These often have different interpretations, where data examples are column vectors and weights
used to form weighted sums are row vectors. However, it can be beneficial to be flexible. As we
have described in Section 2.3, though a single vectorʼs default orientation is a column vector, for
any matrix representing a tabular dataset, treating each data example as a row vector in the matrix
is more conventional.
Given a vector, the first interpretation that we should give it is as a point in space. In two or three
dimensions, we can visualize these points by using the components of the vectors to define the
location of the points in space compared to a fixed reference called the origin. This can be seen
in Fig. 18.1.1.
778 Chapter 18. Appendix: Mathematics for Deep Learning

Fig. 18.1.1: An illustration of visualizing vectors as points in the plane. The first component of the
vector gives the x-coordinate, the second component gives the y-coordinate. Higher dimensions
are analogous, although much harder to visualize.
This geometric point of view allows us to consider the problem on a more abstract level. No longer
faced with some insurmountable seeming problem like classifying pictures as either cats or dogs,
we can start considering tasks abstractly as collections of points in space and picturing the task as
discovering how to separate two distinct clusters of points.
In parallel, there is a second point of view that people often take of vectors: as directions in space.
Not only can we think of the vector v = [3, 2]⊤ as the location 3 units to the right and 2 units up
from the origin, we can also think of it as the direction itself to take 3 steps to the right and 2 steps
up. In this way, we consider all the vectors in figure Fig. 18.1.2 the same.
Fig. 18.1.2: Any vector can be visualized as an arrow in the plane. In this case, every vector drawn
is a representation of the vector (3, 2)⊤ .
One of the benefits of this shift is that we can make visual sense of the act of vector addition. In
particular, we follow the directions given by one vector, and then follow the directions given by
the other, as is seen in Fig. 18.1.3.
18.1. Geometry and Linear Algebraic Operations 779

Fig. 18.1.3: We can visualize vector addition by first following one vector, and then another.
Vector subtraction has a similar interpretation. By considering the identity that u = v + (u − v),
we see that the vector u − v is the direction that takes us from the point v to the point u.
18.1.2 Dot Products and Angles
As we saw in Section 2.3, if we take two column vectors u and v, we can form their dot product by
computing:
∑
u⊤ v = ui · vi . (18.1.3)
i
Because (18.1.3) is symmetric, we will mirror the notation of classical multiplication and write
u · v = u⊤ v = v⊤ u, (18.1.4)
to highlight the fact that exchanging the order of the vectors will yield the same answer.
The dot product (18.1.3) also admits a geometric interpretation: it is closely related to the angle
between two vectors. Consider the angle shown in Fig. 18.1.4.
Fig. 18.1.4: Between any two vectors in the plane there is a well defined angle θ. We will see this
angle is intimately tied to the dot product.
To start, let us consider two specific vectors:

v = (r, 0) and w = (s cos(θ), s sin(θ)). (18.1.5)

The vector v is length r and runs parallel to the x-axis, and the vector w is of length s and at angle
θ with the x-axis.
If we compute the dot product of these two vectors, we see that
v · w = rs cos(θ) = ∥v∥∥w∥ cos(θ). (18.1.6)
With some simple algebraic manipulation, we can rearrange terms to obtain
( )
v·w
θ = arccos . (18.1.7)
∥v∥∥w∥
In short, for these two specific vectors, the dot product combined with the norms tell us the angle
between the two vectors. This same fact is true in general. We will not derive the expression here,
however, if we consider writing ∥v − w∥2 in two ways: one with the dot product, and the other
geometrically using the law of cosines, we can obtain the full relationship. Indeed, for any two
vectors v and w, the angle between the two vectors is
( )
v·w
θ = arccos . (18.1.8)
∥v∥∥w∥
This is a nice result since nothing in the computation references two-dimensions. Indeed, we can
use this in three or three million dimensions without issue.
As a simple example, let us see how to compute the angle between a pair of vectors:
%matplotlib inline
npx.set_np()
def angle(v, w):

return np.arccos(v.dot(w) / (np.linalg.norm(v) * np.linalg.norm(w)))
angle(np.array([0, 1, 2]), np.array([2, 3, 4]))
array(0.41899002)
We will not use it right now, but it is useful to know that we will refer to vectors for which the angle
is π/2 (or equivalently 90◦ ) as being orthogonal. By examining the equation above, we see that this
happens when θ = π/2, which is the same thing as cos(θ) = 0. The only way this can happen is
if the dot product itself is zero, and two vectors are orthogonal if and only if v · w = 0. This will
prove to be a helpful formula when understanding objects geometrically.
It is reasonable to ask: why is computing the angle useful? The answer comes in the kind of
invariance we expect data to have. Consider an image, and a duplicate image, where every pixel
value is the same but 10% the brightness. The values of the individual pixels are in general far
from the original values. Thus, if one computed the distance between the original image and the
darker one, the distance can be large.
However, for most ML applications, the content is the same—it is still an image of a cat as far as a
cat/dog classifier is concerned. However, if we consider the angle, it is not hard to see that for

any vector v, the angle between v and 0.1 · v is zero. This corresponds to the fact that scaling
vectors keeps the same direction and just changes the length. The angle considers the darker
image identical.
Examples like this are everywhere. In text, we might want the topic being discussed to not change
if we write twice as long of document that says the same thing. For some encoding (such as count-
ing the number of occurrences of words in some vocabulary), this corresponds to a doubling of
the vector encoding the document, so again we can use the angle.
Cosine Similarity
In ML contexts where the angle is employed to measure the closeness of two vectors, practitioners
adopt the term cosine similarity to refer to the portion
v·w
cos(θ) = . (18.1.9)
∥v∥∥w∥
The cosine takes a maximum value of 1 when the two vectors point in the same direction, a min-
imum value of −1 when they point in opposite directions, and a value of 0 when the two vectors
are orthogonal. Note that if the components of high-dimensional vectors are sampled randomly
with mean 0, their cosine will nearly always be close to 0.
18.1.3 Hyperplanes
In addition to working with vectors, another key object that you must understand to go far in linear
algebra is the hyperplane, a generalization to higher dimensions of a line (two dimensions) or of a
plane (three dimensions). In an d-dimensional vector space, a hyperplane has d − 1 dimensions
and divides the space into two half-spaces.
Let us start with an example. Suppose that we have a column vector w = [2, 1]⊤ . We want to know,
“what are the points v with w · v = 1?” By recalling the connection between dot products and
angles above (18.1.8), we can see that this is equivalent to
1 1
∥v∥∥w∥ cos(θ) = 1 ⇐⇒ ∥v∥ cos(θ) = =√ . (18.1.10)
∥w∥ 5
Fig. 18.1.5: Recalling trigonometry, we see the formula ∥v∥ cos(θ) is the length of the projection
of the vector v onto the direction of w

If we consider the geometric meaning of this expression, we see that this is equivalent to saying
that the length of the projection of v onto the direction of w is exactly 1/∥w∥, as is shown in :num-
ref:fig_vector-project. The set of all points where this is true is a line at right angles to the vector
w. If we wanted, we could find the equation for this line and see that it is 2x+y = 1 or equivalently
y = 1 − 2x.
If we now look at what happens when we ask about the set of points with w · v > 1 or w · v < 1, we
can see that these are cases where the projections are longer or shorter than 1/∥w∥, respectively.
Thus, those two inequalities define either side of the line. In this way, we have found a way to cut
our space into two halves, where all the points on one side have dot product below a threshold,
and the other side above as we see in Fig. 18.1.6.
Fig. 18.1.6: If we now consider the inequality version of the expression, we see that our hyperplane
(in this case: just a line) separates the space into two halves.
The story in higher dimension is much the same. If we now take w = [1, 2, 3]⊤ and ask about the
points in three dimensions with w · v = 1, we obtain a plane at right angles to the given vector w.
The two inequalities again define the two sides of the plane as is shown in Fig. 18.1.7.
Fig. 18.1.7: Hyperplanes in any dimension separate the space into two halves.
While our ability to visualize runs out at this point, nothing stops us from doing this in tens, hun-
dreds, or billions of dimensions. This occurs often when thinking about machine learned models.
For instance, we can understand linear classification models like those from Section 3.4, as meth-
ods to find hyperplanes that separate the different target classes. In this context, such hyperplanes
are often referred to as decision planes. The majority of deep learned classification models end with
a linear layer fed into a softmax, so one can interpret the role of the deep neural network to be to
find a non-linear embedding such that the target classes can be separated cleanly by hyperplanes.

To give a hand-built example, notice that we can produce a reasonable model to classify tiny im-
ages of t-shirts and trousers from the Fashion MNIST dataset (seen in Section 3.5) by just taking
the vector between their means to define the decision plane and eyeball a crude threshold. First
we will load the data and compute the averages.
# Load in the dataset

train = gluon.data.vision.FashionMNIST(train=True)
test = gluon.data.vision.FashionMNIST(train=False)
X_train_0 = np.stack([x[0] for x in train if x[1] == 0]).astype(float)

X_train_1 = np.stack([x[0] for x in train if x[1] == 1]).astype(float)
X_test = np.stack(
[x[0] for x in test if x[1] == 0 or x[1] == 1]).astype(float)
y_test = np.stack(
[x[1] for x in test if x[1] == 0 or x[1] == 1]).astype(float)
# Compute averages
ave_0 = np.mean(X_train_0, axis=0)
ave_1 = np.mean(X_train_1, axis=0)
It can be informative to examine these averages in detail, so let us plot what they look like. In this
case, we see that the average indeed resembles a blurry image of a t-shirt.
# Plot average t-shirt

d2l.set_figsize()
d2l.plt.imshow(ave_0.reshape(28, 28).tolist(), cmap='Greys')
d2l.plt.show()
In the second case, we again see that the average resembles a blurry image of trousers.
# Plot average trousers

d2l.plt.imshow(ave_1.reshape(28, 28).tolist(), cmap='Greys')
d2l.plt.show()

In a fully machine learned solution, we would learn the threshold from the dataset. In this case,
I simply eyeballed a threshold that looked good on the training data by hand.
# Print test set accuracy with eyeballed threshold

w = (ave_1 - ave_0).T
predictions = X_test.reshape(2000, -1).dot(w.flatten()) > -1500000
# Accuracy
np.mean(predictions.astype(y_test.dtype) == y_test, dtype=np.float64)
array(0.801, dtype=float64)
18.1.4 Geometry of Linear Transformations
Through Section 2.3 and the above discussions, we have a solid understanding of the geometry of
vectors, lengths, and angles. However, there is one important object we have omitted discussing,
and that is a geometric understanding of linear transformations represented by matrices. Fully
internalizing what matrices can do to transform data between two potentially different high di-
mensional spaces takes significant practice, and is beyond the scope of this appendix. However,
we can start building up intuition in two dimensions.
Suppose that we have some matrix:
[ ]
a b
A= . (18.1.11)
c d
If we want to apply this to an arbitrary vector v = [x, y]⊤ , we multiply and see that
[ ][ ]
a b x
Av =
c d y
[ ]
ax + by
=
cx + dy
[ ] [ ] (18.1.12)
a b
=x +y
c d
{ [ ]} { [ ]}
1 0
=x A +y A .
0 1

This may seem like an odd computation, where something clear became somewhat impenetrable.
However, it tells us that we can write the way that a matrix transforms any vector in terms of how
it transforms two specific vectors: [1, 0]⊤ and [0, 1]⊤ . This is worth considering for a moment. We
have essentially reduced an infinite problem (what happens to any pair of real numbers) to a finite
one (what happens to these specific vectors). These vectors are an example a basis, where we can
write any vector in our space as a weighted sum of these basis vectors.
Let us draw what happens when we use the specific matrix
[ ]
1 2
A= . (18.1.13)
−1 3
If we look at the specific vector v = [2, −1]⊤ , we see this is 2 · [1, 0]⊤ + −1 · [0, 1]⊤ , and thus we
know that the matrix A will send this to 2(A[1, 0]⊤ ) + −1(A[0, 1])⊤ = 2[1, −1]⊤ − [2, 3]⊤ = [0, −5]⊤ .
If we follow this logic through carefully, say by considering the grid of all integer pairs of points,
we see that what happens is that the matrix multiplication can skew, rotate, and scale the grid, but
the grid structure must remain as you see in Fig. 18.1.8.
Fig. 18.1.8: The matrix A acting on the given basis vectors. Notice how the entire grid is trans-
ported along with it.
This is the most important intuitive point to internalize about linear transformations represented
by matrices. Matrices are incapable of distorting some parts of space differently than others. All
they can do is take the original coordinates on our space and skew, rotate, and scale them.
Some distortions can be severe. For instance the matrix
[ ]
2 −1
B= , (18.1.14)
4 −2
compresses the entire two-dimensional plane down to a single line. Identifying and working with
such transformations are the topic of a later section, but geometrically we can see that this is
fundamentally different from the types of transformations we saw above. For instance, the result
from matrix A can be “bent back” to the original grid. The results from matrix B cannot because
we will never know where the vector [1, 2]⊤ came from—was it [1, 1]⊤ or [0, −1]⊤ ?
While this picture was for a 2 × 2 matrix, nothing prevents us from taking the lessons learned
into higher dimensions. If we take similar basis vectors like [1, 0, . . . , 0] and see where our matrix
sends them, we can start to get a feeling for how the matrix multiplication distorts the entire space
in whatever dimension space we are dealing with.

18.1.5 Linear Dependence
Consider again the matrix

[]
2 −1
B= . (18.1.15)
4 −2
This compresses the entire plane down to live on the single line y = 2x. The question now arises:
is there some way we can detect this just looking at the matrix itself? The answer is that indeed
we can. Let us take b1 = [2, 4]⊤ and b2 = [−1, −2]⊤ be the two columns of B. Remember that we
can write everything transformed by the matrix B as a weighted sum of the columns of the matrix:
like a1 b1 + a2 b2 . We call this a linear combination. The fact that b1 = −2 · b2 means that we can
write any linear combination of those two columns entirely in terms of say b2 since
a1 b1 + a2 b2 = −2a1 b2 + a2 b2 = (a2 − 2a1 )b2 . (18.1.16)
This means that one of the columns is, in a sense, redundant because it does not define a unique
direction in space. This should not surprise us too much since we already saw that this matrix
collapses the entire plane down into a single line. Moreover, we see that the linear dependence
b1 = −2 · b2 captures this. To make this more symmetrical between the two vectors, we will write
this as
b1 + 2 · b2 = 0. (18.1.17)
In general, we will say that a collection of vectors v1 , . . . , vk are linearly dependent if there exist
coefficients a1 , . . . , ak not all equal to zero so that
∑
k
ai vi = 0. (18.1.18)
i=1
In this case, we can solve for one of the vectors in terms of some combination of the others, and
effectively render it redundant. Thus, a linear dependence in the columns of a matrix is a witness
to the fact that our matrix is compressing the space down to some lower dimension. If there is
no linear dependence we say the vectors are linearly independent. If the columns of a matrix are
linearly independent, no compression occurs and the operation can be undone.
18.1.6 Rank
If we have a general n × m matrix, it is reasonable to ask what dimension space the matrix maps
into. A concept known as the rank will be our answer. In the previous section, we noted that a
linear dependence bears witness to compression of space into a lower dimension and so we will
be able to use this to define the notion of rank. In particular, the rank of a matrix A is the largest
number of linearly independent columns amongst all subsets of columns. For example, the matrix
[ ]
2 4
B= , (18.1.19)
−1 −2
has rank(B) = 1, since the two columns are linearly dependent, but either column by itself is not
linearly dependent. For a more challenging example, we can consider
 
1 3 0 −1 0
−1 0 1 1 −1
C= 0 3 1
, (18.1.20)
0 −1
2 3 −1 −2 1

and show that C has rank two since, for instance, the first two columns are linearly independent,
however any of the four collections of three columns are dependent.
This procedure, as described, is very inefficient. It requires looking at every subset of the columns
of our given matrix, and thus is potentially exponential in the number of columns. Later we will
see a more computationally efficient way to compute the rank of a matrix, but for now, this is
sufficient to see that the concept is well defined and understand the meaning.
18.1.7 Invertibility
We have seen above that multiplication by a matrix with linearly dependent columns cannot be
undone, i.e., there is no inverse operation that can always recover the input. However, multipli-
cation by a full-rank matrix (i.e., some A that is n × n matrix with rank n), we should always be
able to undo it. Consider the matrix
 
1 0 ··· 0
0 1 · · · 0
 
I = . . . ..  . (18.1.21)
. .
. . . . .
0 0 ··· 1
which is the matrix with ones along the diagonal, and zeros elsewhere. We call this the identity
matrix. It is the matrix which leaves our data unchanged when applied. To find a matrix which
undoes what our matrix A has done, we want to find a matrix A−1 such that
A−1 A = AA−1 = I. (18.1.22)
If we look at this as a system, we have n × n unknowns (the entries of A−1 ) and n × n equations
(the equality that needs to hold between every entry of the product A−1 A and every entry of I) so
we should generically expect a solution to exist. Indeed, in the next section we will see a quantity
called the determinant, which has the property that as long as the determinant is not zero, we can
find a solution. We call such a matrix A−1 the inverse matrix. As an example, if A is the general
2 × 2 matrix
[ ]
a b
A= , (18.1.23)
c d
then we can see that the inverse is

[ ]
1 d −b
. (18.1.24)
ad − bc −c a
We can test to see this by seeing that multiplying by the inverse given by the formula above works
in practice.
M = np.array([[1, 2], [1, 4]])

M_inv = np.array([[2, -1], [-0.5, 0.5]])
M_inv.dot(M)
array([[1., 0.],
[0., 1.]])

Numerical Issues
While the inverse of a matrix is useful in theory, we must say that most of the time we do not wish
to use the matrix inverse to solve a problem in practice. In general, there are far more numerically
stable algorithms for solving linear equations like
Ax = b, (18.1.25)
than computing the inverse and multiplying to get
x = A−1 b. (18.1.26)
Just as division by a small number can lead to numerical instability, so can inversion of a matrix
which is close to having low rank.
Moreover, it is common that the matrix A is sparse, which is to say that it contains only a small
number of non-zero values. If we were to explore examples, we would see that this does not mean
the inverse is sparse. Even if A was a 1 million by 1 million matrix with only 5 million non-zero
entries (and thus we need only store those 5 million), the inverse will typically have almost every
entry non-negative, requiring us to store all 1M2 entries—that is 1 trillion entries!
While we do not have time to dive all the way into the thorny numerical issues frequently encoun-
tered when working with linear algebra, we want to provide you with some intuition about when
to proceed with caution, and generally avoiding inversion in practice is a good rule of thumb.
18.1.8 Determinant
The geometric view of linear algebra gives an intuitive way to interpret a fundamental quantity
known as the determinant. Consider the grid image from before, but now with a highlighted region
(Fig. 18.1.9).
Fig. 18.1.9: The matrix A again distorting the grid. This time, I want to draw particular attention
to what happens to the highlighted square.
Look at the highlighted square. This is a square with edges given by (0, 1) and (1, 0) and thus it
has area one. After A transforms this square, we see that it becomes a parallelogram. There is
no reason this parallelogram should have the same area that we started with, and indeed in the
specific case shown here of
[ ]
1 2
A= , (18.1.27)
−1 3

it is an exercise in coordinate geometry to compute the area of this parallelogram and obtain that
the area is 5.
In general, if we have a matrix
[
]
a b
A= , (18.1.28)
c d
we can see with some computation that the area of the resulting parallelogram is ad − bc. This
area is referred to as the determinant.
Let us check this quickly with some example code.
import numpy as np
np.linalg.det(np.array([[1, -1], [2, 3]]))
5.000000000000001
The eagle-eyed amongst us will notice that this expression can be zero or even negative. For the
negative term, this is a matter of convention taken generally in mathematics: if the matrix flips
the figure, we say the area is negated. Let us see now that when the determinant is zero, we learn
more.
Let us consider
[ ]
2 4
B= . (18.1.29)
−1 −2
If we compute the determinant of this matrix, we get 2·(−2)−4·(−1) = 0. Given our understanding
above, this makes sense. B compresses the square from the original image down to a line segment,
which has zero area. And indeed, being compressed into a lower dimensional space is the only
way to have zero area after the transformation. Thus we see the following result is true: a matrix
A is invertible if and only if the determinant is not equal to zero.
As a final comment, imagine that we have any figure drawn on the plane. Thinking like computer
scientists, we can decompose that figure into a collection of little squares so that the area of the
figure is in essence just the number of squares in the decomposition. If we now transform that
figure by a matrix, we send each of these squares to parallelograms, each one of which has area
given by the determinant. We see that for any figure, the determinant gives the (signed) number
that a matrix scales the area of any figure.
Computing determinants for larger matrices can be laborious, but the intuition is the same. The
determinant remains the factor that n × n matrices scale n-dimensional volumes.
18.1.9 Tensors and Common Linear Algebra Operations
In Section 2.3 the concept of tensors was introduced. In this section, we will dive more deeply into
tensor contractions (the tensor equivalent of matrix multiplication), and see how it can provide a
unified view on a number of matrix and vector operations.
With matrices and vectors we knew how to multiply them to transform data. We need to have a
similar definition for tensors if they are to be useful to us. Think about matrix multiplication:
C = AB, (18.1.30)

or equivalently
∑
ci,j = ai,k bk,j . (18.1.31)
k
This pattern is one we can repeat for tensors. For tensors, there is no one case of what to sum over
that can be universally chosen, so we need specify exactly which indices we want to sum over. For
instance we could consider
∑
yil = xijkl ajk . (18.1.32)
jk
Such a transformation is called a tensor contraction. It can represent a far more flexible family of
transformations that matrix multiplication alone.
As a often-used notational simplification, we can notice that the sum is over exactly those indices
that occur more than once in the expression, thus people often work with Einstein notation, where
the summation is implicitly taken over all repeated indices. This gives the compact expression:
yil = xijkl ajk . (18.1.33)
Common Examples from Linear Algebra
Let us see how many of the linear algebraic definitions we have seen before can be expressed in
this compressed tensor notation:
∑
• v · w = i v i wi
∑
• ∥v∥22 = i vi vi
∑
• (Av)i = j aij vj
∑
• (AB)ik = j aij bjk
∑
• tr(A) = i aii
In this way, we can replace a myriad of specialized notations with short tensor expressions.
Expressing in Code
Tensors may flexibly be operated on in code as well. As seen in Section 2.3, we can create tensors
as is shown below.
# Define tensors
B = np.array([[[1, 2, 3], [4, 5, 6]], [[7, 8, 9], [10, 11, 12]]])
A = np.array([[1, 2], [3, 4]])
v = np.array([1, 2])
# Print out the shapes

A.shape, B.shape, v.shape
((2, 2), (2, 2, 3), (2,))

Einstein summation has been implemented directly. The indices that occurs in the Einstein sum-
mation can be passed as a string, followed by the tensors that are being acted upon. For in-
stance, to implement matrix multiplication, we can consider the Einstein summation seen above
(Av = aij vj ) and strip out the indices themselves to get the implementation:
# Reimplement matrix multiplication

np.einsum("ij, j -> i", A, v), A.dot(v)
(array([ 5, 11]), array([ 5, 11]))
This is a highly flexible notation. For instance if we want to compute what would be traditionally
written as
∑
ckl = bijk ail vj . (18.1.34)
ij
it can be implemented via Einstein summation as:
np.einsum("ijk, il, j -> kl", B, A, v)
array([[ 90, 126],

[102, 144],
[114, 162]])
This notation is readable and efficient for humans, however bulky if for whatever reason we need
to generate a tensor contraction programmatically. For this reason, einsum provides an alternative
notation by providing integer indices for each tensor. For example, the same tensor contraction
can also be written as:
np.einsum(B, [0, 1, 2], A, [0, 3], v, [1], [2, 3])
array([[ 90, 126],

[102, 144],
[114, 162]])
Either notation allows for concise and efficient representation of tensor contractions in code.
Summary
• Vectors can be interpreted geometrically as either points or directions in space.

• Dot products define the notion of angle to arbitrarily high-dimensional spaces.
• Hyperplanes are high-dimensional generalizations of lines and planes. They can be used to
define decision planes that are often used as the last step in a classification task.
• Matrix multiplication can be geometrically interpreted as uniform distortions of the under-
lying coordinates. They represent a very restricted, but mathematically clean, way to trans-
form vectors.
• Linear dependence is a way to tell when a collection of vectors are in a lower dimensional
space than we would expect (say you have 3 vectors living in a 2-dimensional space). The
rank of a matrix is the size of the largest subset of its columns that are linearly independent.

• When a matrixʼs inverse is defined, matrix inversion allows us to find another matrix that un-
does the action of the first. Matrix inversion is useful in theory, but requires care in practice
owing to numerical instability.
• Determinants allow us to measure how much a matrix expands or contracts a space. A
nonzero determinant implies an invertible (non-singular) matrix and a zero-valued deter-
minant means that the matrix is non-invertible (singular).
• Tensor contractions and Einstein summation provide for a neat and clean notation for ex-
pressing many of the computations that are seen in machine learning.
Exercises
1. What is the angle between

   
1 3
0 1
⃗v1 =  
−1 , ⃗v2 =  
0? (18.1.35)
2 1
[ ] [ ]
1 2 1 −2
2. True or false: and are inverses of one another?
0 1 0 1
3. Suppose that we draw a shape in the plane with area 100m2 . What is the area after trans-
forming the figure by the matrix
[ ]
2 3
. (18.1.36)
1 2
4. Which of the following sets of vectors are linearly independent?

      
 1 2 3 
•  0  ,  1  , 1
 
−1 −1 1
      
 3 1 0 
•    
1 , 1 , 0 
 
1 1 0
     
 1 0 1 
•   
1 , 1 , 0 
 
0 −1 1
[ ]
c [ ]
5. Suppose that you have a matrix written as A = · a b for some choice of values a, b, c,
d
and d. True or false: the determinant of such a matrix is always 0?
[ ] [ ]
1 0
6. The vectors e1 = and e2 = are orthogonal. What is the condition on a matrix A so
0 1
that Ae1 and Ae2 are orthogonal?
7. How can you write tr(A4 ) in Einstein notation for an arbitrary matrix A?
Discussions245
245

18.2 Eigendecompositions
Eigenvalues are often one of the most useful notions we will encounter when studying linear alge-
bra, however, as a beginner, it is easy to overlook their importance. Below, we introduce eigen-
decomposition and try to convey some sense of just why it is so important.
Suppose that we have a matrix A with the following entries:
[ ]
2 0
A= . (18.2.1)
0 −1
If we apply A to any vector v = [x, y]⊤ , we obtain a vector Av = [2x, −y]⊤ . This has an intuitive
interpretation: stretch the vector to be twice as wide in the x-direction, and then flip it in the
y-direction.
However, there are some vectors for which something remains unchanged. Namely [1, 0]⊤ gets
sent to [2, 0]⊤ and [0, 1]⊤ gets sent to [0, −1]⊤ . These vectors are still in the same line, and the only
modification is that the matrix stretches them by a factor of 2 and −1 respectively. We call such
vectors eigenvectors and the factor they are stretched by eigenvalues.
In general, if we can find a number λ and a vector v such that
Av = λv. (18.2.2)
We say that v is an eigenvector for A and λ is an eigenvalue.
18.2.1 Finding Eigenvalues
Let us figure out how to find them. By subtracting off the λv from both sides, and then factoring
out the vector, we see the above is equivalent to:
(A − λI)v = 0. (18.2.3)
For (18.2.3) to happen, we see that (A − λI) must compress some direction down to zero, hence
it is not invertible, and thus the determinant is zero. Thus, we can find the eigenvalues by finding
for what λ is det(A − λI) = 0. Once we find the eigenvalues, we can solve Av = λv to find the
associated eigenvector(s).
An Example
Let us see this with a more challenging matrix

[ ]
2 1
A= . (18.2.4)
2 3
If we consider det(A − λI) = 0, we see this is equivalent to the polynomial equation 0 = (2 − λ)(3 −
λ) − 2 = (4 − λ)(1 − λ). Thus, two eigenvalues are 4 and 1. To find the associated vectors, we then
need to solve
[ ][ ] [ ] [ ][ ] [ ]
2 1 x x 2 1 x 4x
= and = . (18.2.5)
2 3 y y 2 3 y 4y
We can solve this with the vectors [1, −1]⊤ and [1, 2]⊤ respectively.
We can check this in code using the built-in numpy.linalg.eig routine.

%matplotlib inline
import numpy as np
np.linalg.eig(np.array([[2, 1], [2, 3]]))
(array([1., 4.]),
array([[-0.70710678, -0.4472136 ],
[ 0.70710678, -0.89442719]]))
Note that numpy normalizes the eigenvectors to be of length one, whereas we took ours to be of
arbitrary length. Additionally, the choice of sign is arbitrary. However, the vectors computed are
parallel to the ones we found by hand with the same eigenvalues.
18.2.2 Decomposing Matrices
Let us continue the previous example one step further. Let

[ ]
1 1
W= , (18.2.6)
−1 2
be the matrix where the columns are the eigenvectors of the matrix A. Let
[ ]
1 0
Σ= , (18.2.7)
0 4
be the matrix with the associated eigenvalues on the diagonal. Then the definition of eigenvalues
and eigenvectors tells us that
AW = WΣ. (18.2.8)
The matrix W is invertible, so we may multiply both sides by W −1 on the right, we see that we
may write
A = WΣW−1 . (18.2.9)
In the next section we will see some nice consequences of this, but for now we need only know
that such a decomposition will exist as long as we can find a full collection of linearly independent
eigenvectors (so that W is invertible).
18.2.3 Operations on Eigendecompositions
One nice thing about eigendecompositions (18.2.9) is that we can write many operations we usu-
ally encounter cleanly in terms of the eigendecomposition. As a first example, consider:
n times
z }| { z }| {
n times n times
z }| { (18.2.10)
A = A · · · A = (WΣW ) · · · (WΣW ) = W Σ · · · Σ W−1 = WΣn W−1 .
n −1 −1
This tells us that for any positive power of a matrix, the eigendecomposition is obtained by just
raising the eigenvalues to the same power. The same can be shown for negative powers, so if we
want to invert a matrix we need only consider
A−1 = WΣ−1 W−1 , (18.2.11)
18.2. Eigendecompositions 795

or in other words, just invert each eigenvalue. This will work as long as each eigenvalue is non-
zero, so we see that invertible is the same as having no zero eigenvalues.
Indeed, additional work can show that if λ1 , . . . , λn are the eigenvalues of a matrix, then the de-
terminant of that matrix is
det(A) = λ1 · · · λn , (18.2.12)
or the product of all the eigenvalues. This makes sense intuitively because whatever stretching
W does, W −1 undoes it, so in the end the only stretching that happens is by multiplication by the
diagonal matrix Σ, which stretches volumes by the product of the diagonal elements.
Finally, recall that the rank was the maximum number of linearly independent columns of your
matrix. By examining the eigendecomposition closely, we can see that the rank is the same as the
number of non-zero eigenvalues of A.
The examples could continue, but hopefully the point is clear: eigendecomposition can simplify
many linear-algebraic computations and is a fundamental operation underlying many numerical
algorithms and much of the analysis that we do in linear algebra.
18.2.4 Eigendecompositions of Symmetric Matrices
It is not always possible to find enough linearly independent eigenvectors for the above process
to work. For instance the matrix
[ ]
1 1
A= , (18.2.13)
0 1
has only a single eigenvector, namely (1, 0)⊤ . To handle such matrices, we require more advanced
techniques than we can cover (such as the Jordan Normal Form, or Singular Value Decomposition).
We will often need to restrict our attention to those matrices where we can guarantee the existence
of a full set of eigenvectors.
The most commonly encountered family are the symmetric matrices, which are those matrices
where A = A⊤ . In this case, we may take W to be an orthogonal matrix—a matrix whose columns
are all length one vectors that are at right angles to one another, where W⊤ = W−1 —and all the
eigenvalues will be real.
Thus, in this special case, we can write (18.2.9) as
A = WΣW⊤ . (18.2.14)
18.2.5 Gershgorin Circle Theorem
Eigenvalues are often difficult to reason with intuitively. If presented an arbitrary matrix, there is
little that can be said about what the eigenvalues are without computing them. There is, however,
one theorem that can make it easy to approximate well if the largest values are on the diagonal.
∑
Let A = (aij ) be any square matrix (n × n). We will define ri = j̸=i |aij |. Let Di represent the disc
in the complex plane with center aii radius ri . Then, every eigenvalue of A is contained in one of
the Di .

This can be a bit to unpack, so let us look at an example.
Consider the matrix:
 
1.0 0.1 0.1 0.1
0.1 3.0 0.2 0.3
A=
0.1
. (18.2.15)
0.2 5.0 0.5
0.1 0.3 0.5 9.0
We have r1 = 0.3, r2 = 0.6, r3 = 0.8 and r4 = 0.9. The matrix is symmetric, so all eigenvalues are
real. This means that all of our eigenvalues will be in one of the ranges of
[a11 − r1 , a11 + r1 ] = [0.7, 1.3], (18.2.16)
[a22 − r2 , a22 + r2 ] = [2.4, 3.6], (18.2.17)
[a33 − r3 , a33 + r3 ] = [4.2, 5.8], (18.2.18)
[a44 − r4 , a44 + r4 ] = [8.1, 9.9]. (18.2.19)

Performing the numerical computation shows that the eigenvalues are approximately 0.99, 2.97,
4.95, 9.08, all comfortably inside the ranges provided.
A = np.array([[1.0, 0.1, 0.1, 0.1],

[0.1, 3.0, 0.2, 0.3],
[0.1, 0.2, 5.0, 0.5],
[0.1, 0.3, 0.5, 9.0]])
v, _ = np.linalg.eig(A)
v
array([9.08033648, 0.99228545, 4.95394089, 2.97343718])
In this way, eigenvalues can be approximated, and the approximations will be fairly accurate in
the case that the diagonal is significantly larger than all the other elements.
It is a small thing, but with a complex and subtle topic like eigendecomposition, it is good to get
any intuitive grasp we can.
18.2.6 A Useful Application: The Growth of Iterated Maps
Now that we understand what eigenvectors are in principle, let us see how they can be used to
provide a deep understanding of a problem central to neural network behavior: proper weight
initialization.

Eigenvectors as Long Term Behavior
The full mathematical investigation of the initialization of deep neural networks is beyond the
scope of the text, but we can see a toy version here to understand how eigenvalues can help us see
how these models work. As we know, neural networks operate by interspersing layers of linear
transformations with non-linear operations. For simplicity here, we will assume that there is no
non-linearity, and that the transformation is a single repeated matrix operation A, so that the
output of our model is
vout = A · A · · · Avin = AN vin . (18.2.20)
When these models are initialized, A is taken to be a random matrix with Gaussian entries, so let
us make one of those. To be concrete, we start with a mean zero, variance one Gaussian distributed
5 × 5 matrix.
np.random.seed(8675309)
k = 5
A = np.random.randn(k, k)
A
array([[ 0.58902366, 0.73311856, -1.1621888 , -0.55681601, -0.77248843],

[-0.16822143, -0.41650391, -1.37843129, 0.74925588, 0.17888446],
[ 0.69401121, -1.9780535 , -0.83381434, 0.56437344, 0.31201299],
[-0.87334496, 0.15601291, -0.38710108, -0.23920821, 0.88850104],
[ 1.29385371, -0.76774106, 0.20131613, 0.91800842, 0.38974115]])
Behavior on Random Data
For simplicity in our toy model, we will assume that the data vector we feed in vin is a random five
dimensional Gaussian vector. Let us think about what we want to have happen. For context, lets
think of a generic ML problem, where we are trying to turn input data, like an image, into a pre-
diction, like the probability the image is a picture of a cat. If repeated application of A stretches a
random vector out to be very long, then small changes in input will be amplified into large changes
in output—tiny modifications of the input image would lead to vastly different predictions. This
does not seem right!
On the flip side, if A shrinks random vectors to be shorter, then after running through many layers,
the vector will essentially shrink to nothing, and the output will not depend on the input. This is
also clearly not right either!
We need to walk the narrow line between growth and decay to make sure that our output changes
depending on our input, but not much!
Let us see what happens when we repeatedly multiply our matrix A against a random input vector,
and keep track of the norm.
# Calculate the sequence of norms after repeatedly applying À`

v_in = np.random.randn(k, 1)
norm_list = [np.linalg.norm(v_in)]

v_in = A.dot(v_in)
norm_list.append(np.linalg.norm(v_in))
d2l.plot(np.arange(0, 100), norm_list, 'Iteration', 'Value')
The norm is growing uncontrollably! Indeed if we take the list of quotients, we will see a pattern.
# Compute the scaling factor of the norms

norm_ratio_list = []
norm_ratio_list.append(norm_list[i]/norm_list[i - 1])
d2l.plot(np.arange(1, 100), norm_ratio_list, 'Iteration', 'Ratio')
If we look at the last portion of the above computation, we see that the random vector is stretched
by a factor of 1.974459321485[...], where the portion at the end shifts a little, but the stretching
factor is stable.

Relating Back to Eigenvectors
We have seen that eigenvectors and eigenvalues correspond to the amount something is stretched,
but that was for specific vectors, and specific stretches. Let us take a look at what they are for A. A
bit of a caveat here: it turns out that to see them all, we will need to go to complex numbers. You
can think of these as stretches and rotations. By taking the norm of the complex number (square
root of the sums of squares of real and imaginary parts) we can measure that stretching factor.
Let us also sort them.
# Compute the eigenvalues

eigs = np.linalg.eigvals(A).tolist()
norm_eigs = [np.absolute(x) for x in eigs]
norm_eigs.sort()
print(f'norms of eigenvalues: {norm_eigs}')
norms of eigenvalues: [0.8786205280381857, 1.2757952665062624, 1.4983381517710659, 1.

,→4983381517710659, 1.974459321485074]
An Observation
We see something a bit unexpected happening here: that number we identified before for the long
term stretching of our matrix A applied to a random vector is exactly (accurate to thirteen decimal
places!) the largest eigenvalue of A. This is clearly not a coincidence!
But, if we now think about what is happening geometrically, this starts to make sense. Consider a
random vector. This random vector points a little in every direction, so in particular, it points at
least a little bit in the same direction as the eigenvector of A associated with the largest eigenvalue.
This is so important that it is called the principle eigenvalue and principle eigenvector. After apply-
ing A, our random vector gets stretched in every possible direction, as is associated with every
possible eigenvector, but it is stretched most of all in the direction associated with this principle
eigenvector. What this means is that after apply in A, our random vector is longer, and points
in a direction closer to being aligned with the principle eigenvector. After applying the matrix
many times, the alignment with the principle eigenvector becomes closer and closer until, for
all practical purposes, our random vector has been transformed into the principle eigenvector!
Indeed this algorithm is the basis for what is known as the power iteration for finding the largest
eigenvalue and eigenvector of a matrix. For details see, for example, (VanLoan & Golub, 1983).
Fixing the Normalization
Now, from above discussions, we concluded that we do not want a random vector to be stretched
or squished at all, we would like random vectors to stay about the same size throughout the en-
tire process. To do so, we now rescale our matrix by this principle eigenvalue so that the largest
eigenvalue is instead now just one. Let us see what happens in this case.
# Rescale the matrix À`

A /= norm_eigs[-1]
# Do the same experiment again

v_in = np.random.randn(k, 1)

norm_list = [np.linalg.norm(v_in)]
v_in = A.dot(v_in)
norm_list.append(np.linalg.norm(v_in))
d2l.plot(np.arange(0, 100), norm_list, 'Iteration', 'Value')
We can also plot the ratio between consecutive norms as before and see that indeed it stabilizes.
# Also plot the ratio

norm_ratio_list = []
norm_ratio_list.append(norm_list[i]/norm_list[i-1])
d2l.plot(np.arange(1, 100), norm_ratio_list, 'Iteration', 'Ratio')

18.2.7 Conclusions
We now see exactly what we hoped for! After normalizing the matrices by the principle eigenvalue,
we see that the random data does not explode as before, but rather eventually equilibrates to a
specific value. It would be nice to be able to do these things from first principles, and it turns out
that if we look deeply at the mathematics of it, we can see that the largest eigenvalue of a large
random matrix with √independent mean zero, variance one Gaussian entries is on average about
√
n, or in our case 5 ≈ 2.2, due to a fascinating fact known as the circular law (Ginibre, 1965).
The relationship between the eigenvalues (and a related object called singular values) of random
matrices has been shown to have deep connections to proper initialization of neural networks as
was discussed in (Pennington et al., 2017) and subsequent works.
Summary
• Eigenvectors are vectors which are stretched by a matrix without changing direction.
• Eigenvalues are the amount that the eigenvectors are stretched by the application of the
matrix.
• The eigendecomposition of a matrix can allow for many operations to be reduced to opera-
tions on the eigenvalues.
• The Gershgorin Circle Theorem can provide approximate values for the eigenvalues of a
matrix.
• The behavior of iterated matrix powers depends primarily on the size of the largest eigen-
value. This understanding has many applications in the theory of neural network initializa-
tion.
Exercises
1. What are the eigenvalues and eigenvectors of

[ ]
2 1
A= ? (18.2.21)
1 2
2. What are the eigenvalues and eigenvectors of the following matrix, and what is strange about
this example compared to the previous one?
[ ]
2 1
A= . (18.2.22)
0 2
3. Without computing the eigenvalues, is it possible that the smallest eigenvalue of the follow-
ing matrix is less that 0.5? Note: this problem can be done in your head.
 
3.0 0.1 0.3 1.0
0.1 1.0 0.1 0.2
A= 
0.3 0.1 5.0 0.0 . (18.2.23)
1.0 0.2 0.0 1.8
Discussions246
246

18.3 Single Variable Calculus
In Section 2.4, we saw the basic elements of differential calculus. This section takes a deeper
dive into the fundamentals of calculus and how we can understand and apply it in the context of
machine learning.
18.3.1 Differential Calculus
Differential calculus is fundamentally the study of how functions behave under small changes. To
see why this is so core to deep learning, let us consider an example.
Suppose that we have a deep neural network where the weights are, for convenience, concatenated
into a single vector w = (w1 , . . . , wn ). Given a training dataset, we consider the loss of our neural
network on this dataset, which we will write as L(w).
This function is extraordinarily complex, encoding the performance of all possible models of the
given architecture on this dataset, so it is nearly impossible to tell what set of weights w will min-
imize the loss. Thus, in practice, we often start by initializing our weights randomly, and then
iteratively take small steps in the direction which makes the loss decrease as rapidly as possible.
The question then becomes something that on the surface is no easier: how do we find the direc-
tion which makes the weights decrease as quickly as possible? To dig into this, let us first examine
the case with only a single weight: L(w) = L(x) for a single real value x.
Let us take x and try to understand what happens when we change it by a small amount to x + ϵ.
If you wish to be concrete, think a number like ϵ = 0.0000001. To help us visualize what happens,
let us graph an example function, f (x) = sin(xx ), over the [0, 3].
%matplotlib inline
npx.set_np()
# Plot a function in a normal range

x_big = np.arange(0.01, 3.01, 0.01)
ys = np.sin(x_big**x_big)
d2l.plot(x_big, ys, 'x', 'f(x)')
18.3. Single Variable Calculus 803

At this large scale, the functionʼs behavior is not simple. However, if we reduce our range to some-
thing smaller like [1.75, 2.25], we see that the graph becomes much simpler.
# Plot a the same function in a tiny range

x_med = np.arange(1.75, 2.25, 0.001)
ys = np.sin(x_med**x_med)
d2l.plot(x_med, ys, 'x', 'f(x)')
Taking this to an extreme, if we zoom into a tiny segment, the behavior becomes far simpler: it is
just a straight line.
# Plot a the same function in a tiny range

x_small = np.arange(2.0, 2.01, 0.0001)
ys = np.sin(x_small**x_small)
d2l.plot(x_small, ys, 'x', 'f(x)')
This is the key observation of single variable calculus: the behavior of familiar functions can be
modeled by a line in a small enough range. This means that for most functions, it is reasonable to
expect that as we shift the x value of the function by a little bit, the output f (x) will also be shifted
by a little bit. The only question we need to answer is, “How large is the change in the output
compared to the change in the input? Is it half as large? Twice as large?”

Thus, we can consider the ratio of the change in the output of a function for a small change in the
input of the function. We can write this formally as
L(x + ϵ) − L(x) L(x + ϵ) − L(x)

= . (18.3.1)
(x + ϵ) − x ϵ
This is already enough to start to play around with in code. For instance, suppose that we know
that L(x) = x2 + 1701(x − 4)3 , then we can see how large this value is at the point x = 4 as follows.
# Define our function

def L(x):
return x**2 + 1701*(x-4)**3
# Print the difference divided by epsilon for several epsilon

for epsilon in [0.1, 0.001, 0.0001, 0.00001]:
print(f'epsilon = {epsilon:.5f} -> {(L(4+epsilon) - L(4)) / epsilon:.5f}')
epsilon = 0.10000 -> 25.11000

epsilon = 0.00100 -> 8.00270
epsilon = 0.00010 -> 8.00012
epsilon = 0.00001 -> 8.00001
Now, if we are observant, we will notice that the output of this number is suspiciously close to
8. Indeed, if we decrease ϵ, we will see value becomes progressively closer to 8. Thus we may
conclude, correctly, that the value we seek (the degree a change in the input changes the output)
should be 8 at the point x = 4. The way that a mathematician encodes this fact is
L(4 + ϵ) − L(4)
lim = 8. (18.3.2)
ϵ→0 ϵ
As a bit of a historical digression: in the first few decades of neural network research, scientists
used this algorithm (the method of finite differences) to evaluate how a loss function changed under
small perturbation: just change the weights and see how the loss changed. This is computationally
inefficient, requiring two evaluations of the loss function to see how a single change of one vari-
able influenced the loss. If we tried to do this with even a paltry few thousand parameters, it would
require several thousand evaluations of the network over the entire dataset! It was not solved until
1986 that the backpropagation algorithm introduced in (Rumelhart et al., 1988) provided a way to
calculate how any change of the weights together would change the loss in the same computation
time as a single prediction of the network over the dataset.
Back in our example, this value 8 is different for different values of x, so it makes sense to define it
as a function of x. More formally, this value dependent rate of change is referred to as the derivative
which is written as
df f (x + ϵ) − f (x)
(x) = lim . (18.3.3)
dx ϵ→0 ϵ
Different texts will use different notations for the derivative. For instance, all of the below nota-
tions indicate the same thing:
df d
= f = f ′ = ∇x f = Dx f = fx . (18.3.4)
dx dx
Most authors will pick a single notation and stick with it, however even that is not guaranteed. It
df
is best to be familiar with all of these. We will use the notation dx throughout this text, unless

d
we want to take the derivative of a complex expression, in which case we will use dx f to write
expressions like
[ ( 2 )]
d 4 x +1
x + cos . (18.3.5)
dx 2x − 1
Oftentimes, it is intuitively useful to unravel the definition of derivative (18.3.3) again to see how
a function changes when we make a small change of x:
df f (x + ϵ) − f (x) df f (x + ϵ) − f (x)
(x) = lim =⇒ (x) ≈
dx ϵ→0 ϵ dx ϵ
df
=⇒ ϵ (x) ≈ f (x + ϵ) − f (x) (18.3.6)
dx
df
=⇒ f (x + ϵ) ≈ f (x) + ϵ (x).
dx
The last equation is worth explicitly calling out. It tells us that if you take any function and change
the input by a small amount, the output would change by that small amount scaled by the deriva-
tive.
In this way, we can understand the derivative as the scaling factor that tells us how large of change
we get in the output from a change in the input.
18.3.2 Rules of Calculus
We now turn to the task of understanding how to compute the derivative of an explicit function. A
full formal treatment of calculus would derive everything from first principles. We will not indulge
in this temptation here, but rather provide an understanding of the common rules encountered.
Common Derivatives
As was seen in Section 2.4, when computing derivatives one can oftentimes use a series of rules
to reduce the computation to a few core functions. We repeat them here for ease of reference.
d
• Derivative of constants. dx c = 0.
d
• Derivative of linear functions. dx (ax) = a.
d n
• Power rule. dx x = nxn−1 .
d x
• Derivative of exponentials. dx e = ex .
d
• Derivative of the logarithm. dx log(x) = x1 .
Derivative Rules
If every derivative needed to be separately computed and stored in a table, differential calculus
would be near impossible. It is a gift of mathematics that we can generalize the( above derivatives
)
and compute more complex derivatives like finding the derivative of f (x) = log 1 + (x − 1)10 . As
was mentioned in Section 2.4, the key to doing so is to codify what happens when we take functions
and combine them in various ways, most importantly: sums, products, and compositions.
d dg dh
• Sum rule. dx (g(x) + h(x)) = dx (x) + dx (x).

dg
• Product rule. d
dx (g(x) · h(x)) = g(x) dh
dx (x) + dx (x)h(x).
dg
• Chain rule. d
dx g(h(x)) = dh (h(x)) · dh
dx (x).
Let us see how we may use (18.3.6) to understand these rules. For the sum rule, consider following
chain of reasoning:
f (x + ϵ) = g(x + ϵ) + h(x + ϵ)
dg dh
≈ g(x) + ϵ (x) + h(x) + ϵ (x)
dx ( dx )
dg dh (18.3.7)
= g(x) + h(x) + ϵ (x) + (x)
dx dx
( )
dg dh
= f (x) + ϵ (x) + (x) .
dx dx
df df dg
By comparing this result with the fact that f (x + ϵ) ≈ f (x) + ϵ dx (x), we see that dx (x) = dx (x) +
dh
dx (x) as desired. The intuition here is: when we change the input x, g and h jointly contribute to
dg
the change of the output by dx (x) and dh
dx (x).
The product is more subtle, and will require a new observation about how to work with these
expressions. We will begin as before using (18.3.6):
f (x + ϵ) = g(x + ϵ) · h(x + ϵ)
( ) ( )
dg dh
≈ g(x) + ϵ (x) · h(x) + ϵ (x)
dx dx
( )
dh dg dg dh (18.3.8)
= g(x) · h(x) + ϵ g(x) (x) + (x)h(x) + ϵ2 (x) (x)
dx dx dx dx
( )
dh dg dg dh
= f (x) + ϵ g(x) (x) + (x)h(x) + ϵ2 (x) (x).
dx dx dx dx
df
This resembles the computation done above, and indeed we see our answer ( dx (x) = g(x) dh
dx (x) +
dg 2
dx (x)h(x)) sitting next to ϵ, but there is the issue of that term of size ϵ . We will refer to this as a
higher-order term, since the power of ϵ2 is higher than the power of ϵ1 . We will see in a later section
that we will sometimes want to keep track of these, however for now observe that if ϵ = 0.0000001,
then ϵ2 = 0.0000000000001, which is vastly smaller. As we send ϵ → 0, we may safely ignore the
higher order terms. As a general convention in this appendix, we will use “≈” to denote that the
two terms are equal up to higher order terms. However, if we wish to be more formal we may
examine the difference quotient
f (x + ϵ) − f (x) dh dg dg dh
= g(x) (x) + (x)h(x) + ϵ (x) (x), (18.3.9)
ϵ dx dx dx dx
and see that as we send ϵ → 0, the right hand term goes to zero as well.
Finally, with the chain rule, we can again progress as before using (18.3.6) and see that
f (x + ϵ) = g(h(x + ϵ))
( )
dh
≈ g h(x) + ϵ (x)
dx
dh dg (18.3.10)
≈ g(h(x)) + ϵ (x) (h(x))
dx dh
dg dh
= f (x) + ϵ (h(x)) (x),
dh dx

where in the second line we view the function g as having its input (h(x)) shifted by the tiny quan-
tity ϵ dh
dx (x).
These rule provide us with a flexible set of tools to compute essentially any expression desired.
For instance,
d [ ( )] ( )−1 d [ ]
log 1 + (x − 1)10 = 1 + (x − 1)10 1 + (x − 1)10
dx dx
( )
( ) −1 d d
= 1 + (x − 1) 10
[1] + [(x − 1) ]
10
dx dx
( )
( )−1 d
= 1 + (x − 1)10 0 + 10(x − 1)9 [x − 1] (18.3.11)
dx
( ) −1
= 10 1 + (x − 1)10 (x − 1)9
10(x − 1)9
= .
1 + (x − 1)10
Where each line has used the following rules:
1. The chain rule and derivative of logarithm.
2. The sum rule.
3. The derivative of constants, chain rule, and power rule.
4. The sum rule, derivative of linear functions, derivative of constants.
Two things should be clear after doing this example:
1. Any function we can write down using sums, products, constants, powers, exponentials, and
logarithms can have its derivate computed mechanically by following these rules.
2. Having a human follow these rules can be tedious and error prone!
Thankfully, these two facts together hint towards a way forward: this is a perfect candidate for
mechanization! Indeed backpropagation, which we will revisit later in this section, is exactly that.
Linear Approximation
When working with derivatives, it is often useful to geometrically interpret the approximation
used above. In particular, note that the equation
df
f (x + ϵ) ≈ f (x) + ϵ (x), (18.3.12)
dx
df
approximates the value of f by a line which passes through the point (x, f (x)) and has slope dx (x).
In this way we say that the derivative gives a linear approximation to the function f , as illustrated
below:
# Compute sin
xs = np.arange(-np.pi, np.pi, 0.01)
plots = [np.sin(xs)]
# Compute some linear approximations. Use d(sin(x)) / dx = cos(x)

for x0 in [-1.5, 0, 2]:
plots.append(np.sin(x0) + (xs - x0) * np.cos(x0))
d2l.plot(xs, plots, 'x', 'f(x)', ylim=[-1.5, 1.5])

Higher Order Derivatives
Let us now do something that may on the surface seem strange. Take a function f and compute
df
the derivative dx . This gives us the rate of change of f at any point.
df
However, the derivative, dx , can be viewed
( as
) a function itself, so nothing stops us from comput-
df 2
ing the derivative of dx to get ddxf2 = dx
df df
dx . We will call this the second derivative of f . This
function is the rate of change of the rate of change of f , or in other words, how the rate of change
is changing. We may apply the derivative any number of times to obtain what is called the n-th
derivative. To keep the notation clean, we will denote the n-th derivative as
( )n
(n) dn f d
f (x) = n = f. (18.3.13)
dx dx
Let us try to understand why this is a useful notion. Below, we visualize f (2) (x), f (1) (x), and f (x).
First, consider the case that the second derivative f (2) (x) is a positive constant. This means that
the slope of the first derivative is positive. As a result, the first derivative f (1) (x) may start out
negative, becomes zero at a point, and then becomes positive in the end. This tells us the slope of
our original function f and therefore, the function f itself decreases, flattens out, then increases.
In other words, the function f curves up, and has a single minimum as is shown in Fig. 18.3.1.
Fig. 18.3.1: If we assume the second derivative is a positive constant, then the fist derivative in
increasing, which implies the function itself has a minimum.
Second, if the second derivative is a negative constant, that means that the first derivative is de-
creasing. This implies the first derivative may start out positive, becomes zero at a point, and then

becomes negative. Hence, the function f itself increases, flattens out, then decreases. In other
words, the function f curves down, and has a single maximum as is shown in Fig. 18.3.2.
Fig. 18.3.2: If we assume the second derivative is a negative constant, then the fist derivative in
decreasing, which implies the function itself has a maximum.
Third, if the second derivative is a always zero, then the first derivative will never change—it is
constant! This means that f increases (or decreases) at a fixed rate, and f is itself a straight line
as is shown in Fig. 18.3.3.
Fig. 18.3.3: If we assume the second derivative is zero, then the fist derivative is constant, which
implies the function itself is a straight line.
To summarize, the second derivative can be interpreted as describing the way that the function f
curves. A positive second derivative leads to a upwards curve, while a negative second derivative
means that f curves downwards, and a zero second derivative means that f does not curve at all.
Let us take this one step further. Consider the function g(x) = ax2 + bx + c. We can then compute
that
dg
(x) = 2ax + b
dx
(18.3.14)
d2 g
(x) = 2a.
dx2
If we have some original function f (x) in mind, we may compute the first two derivatives and
find the values for a, b, and c that make them match this computation. Similarly to the previous
section where we saw that the first derivative gave the best approximation with a straight line,
this construction provides the best approximation by a quadratic. Let us visualize this for f (x) =
sin(x).

# Compute sin
xs = np.arange(-np.pi, np.pi, 0.01)
plots = [np.sin(xs)]
# Compute some quadratic approximations. Use d(sin(x)) / dx = cos(x)

for x0 in [-1.5, 0, 2]:
plots.append(np.sin(x0) + (xs - x0) * np.cos(x0) -
(xs - x0)**2 * np.sin(x0) / 2)
d2l.plot(xs, plots, 'x', 'f(x)', ylim=[-1.5, 1.5])
We will extend this idea to the idea of a Taylor series in the next section.
Taylor Series
The Taylor series provides a method{to approximate the function f (x) if we } are given values for the
first n derivatives at a point x0 , i.e., f (x0 ), f (1) (x0 ), f (2) (x0 ), . . . , f (n) (x0 ) . The idea will be to find
a degree n polynomial that matches all the given derivatives at x0 .
We saw the case of n = 2 in the previous section and a little algebra shows this is
1 d2 f df
f (x) ≈ 2
(x0 )(x − x0 )2 + (x0 )(x − x0 ) + f (x0 ). (18.3.15)
2 dx dx
As we can see above, the denominator of 2 is there to cancel out the 2 we get when we take two
derivatives of x2 , while the other terms are all zero. Same logic applies for the first derivative and
the value itself.
If we push the logic further to n = 3, we will conclude that
d3 f d2 f
dx3
(x0 ) dx2
(x0 )
df (18.3.16)
f (x) ≈ (x − x0 ) +
3
(x − x0 )2 +
(x0 )(x − x0 ) + f (x0 ).
6 2 dx
where the 6 = 3 × 2 = 3! comes from the constant we get in front if we take three derivatives of
x3 .
Furthermore, we can get a degree n polynomial by
∑
n
f (i) (x0 )
Pn (x) = (x − x0 )i . (18.3.17)
i!
i=0

where the notation
( )n
(n) dn f d
f (x) = n = f. (18.3.18)
dx dx
Indeed, Pn (x) can be viewed as the best n-th degree polynomial approximation to our function
f (x).
While we are not going to dive all the way into the error of the above approximations, it is worth
mentioning the infinite limit. In this case, for well behaved functions (known as real analytic
functions) like cos(x) or ex , we can write out the infinite number of terms and approximate the
exactly same function
∞
∑ f (n) (x0 )
f (x) = (x − x0 )n . (18.3.19)
n!
n=0
Take f (x) = ex as am example. Since ex is its own derivative, we know that f (n) (x) = ex . There-
fore, ex can be reconstructed by taking the Taylor series at x0 = 0, i.e.,
∞
∑ xn x2 x3
x
e = =1+x+ + + ··· . (18.3.20)
n! 2 6
n=0
Let us see how this works in code and observe how increasing the degree of the Taylor approxi-
mation brings us closer to the desired function ex .
# Compute the exponential function

xs = np.arange(0, 3, 0.01)
ys = np.exp(xs)
# Compute a few Taylor series approximations

P1 = 1 + xs
P2 = 1 + xs + xs**2 / 2
P5 = 1 + xs + xs**2 / 2 + xs**3 / 6 + xs**4 / 24 + xs**5 / 120
d2l.plot(xs, [ys, P1, P2, P5], 'x', 'f(x)', legend=[

"Exponential", "Degree 1 Taylor Series", "Degree 2 Taylor Series",
"Degree 5 Taylor Series"])
Taylor series have two primary applications:

1. Theoretical applications: Often when we try to understand a too complex function, using Tay-
lor series enables us to turn it into a polynomial that we can work with directly.
2. Numerical applications: Some functions like ex or cos(x) are difficult for machines to com-
pute. They can store tables of values at a fixed precision (and this is often done), but it still
leaves open questions like “What is the 1000-th digit of cos(1)?” Taylor series are often help-
ful to answer such questions.
Summary
• Derivatives can be used to express how functions change when we change the input by a
small amount.
• Elementary derivatives can be combined using derivative rules to create arbitrarily complex
derivatives.
• Derivatives can be iterated to get second or higher order derivatives. Each increase in order
provides more fine grained information on the behavior of the function.
• Using information in the derivatives of a single data example, we can approximate well be-
haved functions by polynomials obtained from the Taylor series.
Exercises
1. What is the derivative of x3 − 4x + 1?

2. What is the derivative of log( x1 )?
3. True or False: If f ′ (x) = 0 then f has a maximum or minimum at x?
4. Where is the minimum of f (x) = x log(x) for x ≥ 0 (where we assume that f takes the
limiting value of 0 at f (0))?
Discussions247
18.4 Multivariable Calculus
Now that we have a fairly strong understanding of derivatives of a function of a single variable, let
us return to our original question where we were considering a loss function of potentially billions
of weights.
247
18.4. Multivariable Calculus 813

18.4.1 Higher-Dimensional Differentiation
What Section 18.3 tells us is that if we change a single one of these billions of weights leaving
every other one fixed, we know what will happen! This is nothing more than a function of a single
variable, so we can write
d
L(w1 + ϵ1 , w2 , . . . , wN ) ≈ L(w1 , w2 , . . . , wN ) + ϵ1 L(w1 , w2 , . . . , wN ). (18.4.1)
dw1
We will call the derivative in one variable while fixing the other the partial derivative, and we will
∂
use the notation ∂w 1
for the derivative in (18.4.1).
Now, let us take this and change w2 a little bit to w2 + ϵ2 :
∂
L(w1 + ϵ1 , w2 + ϵ2 , . . . , wN ) ≈ L(w1 , w2 + ϵ2 , . . . , wN ) + ϵ1 L(w1 , w2 + ϵ2 , . . . , wN )
∂w1
≈ L(w1 , w2 , . . . , wN )
∂
+ ϵ2 L(w1 , w2 , . . . , wN )
∂w2
∂
+ ϵ1 L(w1 , w2 , . . . , wN )
∂w1
(18.4.2)
∂ ∂
+ ϵ1 ϵ2 L(w1 , w2 , . . . , wN )
∂w2 ∂w1
≈ L(w1 , w2 , . . . , wN )
∂
+ ϵ2 L(w1 , w2 , . . . , wN )
∂w2
∂
+ ϵ1 L(w1 , w2 , . . . , wN ).
∂w1
We have again used the idea that ϵ1 ϵ2 is a higher order term that we can discard in the same way
we could discard ϵ2 in the previous section, along with what we saw in (18.4.1). By continuing in
this manner, we may write that
∑ ∂
L(w1 + ϵ1 , w2 + ϵ2 , . . . , wN + ϵN ) ≈ L(w1 , w2 , . . . , wN ) + ϵi L(w1 , w2 , . . . , wN ). (18.4.3)
∂wi
i
This may look like a mess, but we can make this more familiar by noting that the sum on the right
looks exactly like a dot product, so if we let
[ ]⊤
∂L ∂L
ϵ = [ϵ1 , . . . , ϵN ]⊤ and ∇x L = ,..., , (18.4.4)
∂x1 ∂xN
then
L(w + ϵ) ≈ L(w) + ϵ · ∇w L(w). (18.4.5)
We will call the vector ∇w L the gradient of L.

Equation (18.4.5) is worth pondering for a moment. It has exactly the format that we encountered
in one dimension, just we have converted everything to vectors and dot products. It allows us to
tell approximately how the function L will change given any perturbation to the input. As we will
see in the next section, this will provide us with an important tool in understanding geometrically
how we can learn using information contained in the gradient.

But first, let us see this approximation at work with an example. Suppose that we are working with
the function
[ ]
ex ey
f (x, y) = log(e + e ) with gradient ∇f (x, y) = x
x y
, . (18.4.6)
e + ey ex + ey
If we look at a point like (0, log(2)), we see that

[ ]
1 2
f (x, y) = log(3) with gradient ∇f (x, y) = , . (18.4.7)
3 3
Thus, if we want to approximate f at (ϵ1 , log(2) + ϵ2 ), we see that we should have the specific
instance of (18.4.5):
1 2
f (ϵ1 , log(2) + ϵ2 ) ≈ log(3) + ϵ1 + ϵ2 . (18.4.8)
3 3
We can test this in code to see how good the approximation is.
%matplotlib inline
npx.set_np()
def f(x, y):

return np.log(np.exp(x) + np.exp(y))
def grad_f(x, y):
return np.array([np.exp(x) / (np.exp(x) + np.exp(y)),
np.exp(y) / (np.exp(x) + np.exp(y))])
epsilon = np.array([0.01, -0.03])

grad_approx = f(0, np.log(2)) + epsilon.dot(grad_f(0, np.log(2)))
true_value = f(0 + epsilon[0], np.log(2) + epsilon[1])
f'approximation: {grad_approx}, true Value: {true_value}'
'approximation: 1.0819457, true Value: 1.0821242'
18.4.2 Geometry of Gradients and Gradient Descent
Consider the again (18.4.5):
L(w + ϵ) ≈ L(w) + ϵ · ∇w L(w). (18.4.9)
Let us suppose that I want to use this to help minimize our loss L. Let us understand geometrically
the algorithm of gradient descent first described in Section 2.5. What we will do is the following:
1. Start with a random choice for the initial parameters w.
2. Find the direction v that makes L decrease the most rapidly at w.
3. Take a small step in that direction: w → w + ϵv.
4. Repeat.

The only thing we do not know exactly how to do is to compute the vector v in the second step.
We will call such a direction the direction of steepest descent. Using the geometric understanding of
dot products from Section 18.1, we see that we can rewrite (18.4.5) as
L(w + v) ≈ L(w) + v · ∇w L(w) = L(w) + ∥∇w L(w)∥ cos(θ). (18.4.10)
Note that we have taken our direction to have length one for convenience, and used θ for the angle
between v and ∇w L(w). If we want to find the direction that decreases L as rapidly as possible, we
want to make this expression as negative as possible. The only way the direction we pick enters
into this equation is through cos(θ), and thus we wish to make this cosine as negative as possible.
Now, recalling the shape of cosine, we can make this as negative as possible by making cos(θ) = −1
or equivalently making the angle between the gradient and our chosen direction to be π radians,
or equivalently 180 degrees. The only way to achieve this is to head in the exact opposite direction:
pick v to point in the exact opposite direction to ∇w L(w)!
This brings us to one of the most important mathematical concepts in machine learning: the
direction of steepest decent points in the direction of −∇w L(w). Thus our informal algorithm
can be rewritten as follows.
1. Start with a random choice for the initial parameters w.
2. Compute ∇w L(w).
3. Take a small step in the opposite of that direction: w → w − ϵ∇w L(w).
4. Repeat.
This basic algorithm has been modified and adapted many ways by many researchers, but the core
concept remains the same in all of them. Use the gradient to find the direction that decreases the
loss as rapidly as possible, and update the parameters to take a step in that direction.
18.4.3 A Note on Mathematical Optimization
Throughout this book, we focus squarely on numerical optimization techniques for the practical
reason that all functions we encounter in the deep learning setting are too complex to minimize
explicitly.
However, it is a useful exercise to consider what the geometric understanding we obtained above
tells us about optimizing functions directly.
Suppose that we wish to find the value of x0 which minimizes some function L(x). Let us suppose
that moreover someone gives us a value and tells us that it is the value that minimizes L. Is there
anything we can check to see if their answer is even plausible?
Again consider (18.4.5):
L(x0 + ϵ) ≈ L(x0 ) + ϵ · ∇x L(x0 ). (18.4.11)
If the gradient is not zero, we know that we can take a step in the direction −ϵ∇x L(x0 ) to find a
value of L that is smaller. Thus, if we truly are at a minimum, this cannot be the case! We can
conclude that if x0 is a minimum, then ∇x L(x0 ) = 0. We call points with ∇x L(x0 ) = 0 critical
points.
This is nice, because in some rare settings, we can explicitly find all the points where the gradient
is zero, and find the one with the smallest value.

For a concrete example, consider the function
f (x) = 3x4 − 4x3 − 12x2 . (18.4.12)
This function has derivative

df
= 12x3 − 12x2 − 24x = 12x(x − 2)(x + 1). (18.4.13)
dx
The only possible location of minima are at x = −1, 0, 2, where the function takes the values
−5, 0, −32 respectively, and thus we can conclude that we minimize our function when x = 2. A
quick plot confirms this.
x = np.arange(-2, 3, 0.01)
f = (3 * x**4) - (4 * x**3) - (12 * x**2)
d2l.plot(x, f, 'x', 'f(x)')
This highlights an important fact to know when working either theoretically or numerically: the
only possible points where we can minimize (or maximize) a function will have gradient equal to
zero, however, not every point with gradient zero is the true global minimum (or maximum).
18.4.4 Multivariate Chain Rule
Let us suppose that we have a function of four variables (w, x, y, and z) which we can make by
composing many terms:
f (u, v) = (u + v)2
u(a, b) = (a + b)2 , v(a, b) = (a − b)2 , (18.4.14)
2
a(w, x, y, z) = (w + x + y + z) , b(w, x, y, z) = (w + x − y − z) .
2
Such chains of equations are common when working with neural networks, so trying to under-
stand how to compute gradients of such functions is key. We can start to see visual hints of this
connection in Fig. 18.4.1 if we take a look at what variables directly relate to one another.

Fig. 18.4.1: The function relations above where nodes represent values and edges show functional
dependence.
Nothing stops us from just composing everything from (18.4.14) and writing out that
(( )2 ( )2 )2
f (w, x, y, z) = (w + x + y + z)2 + (w + x − y − z)2 + (w + x + y + z)2 − (w + x − y − z)2 .
(18.4.15)
We may then take the derivative by just using single variable derivatives, but if we did that we
would quickly find ourself swamped with terms, many of which are repeats! Indeed, one can see
that, for instance:
∂f ( ( )
= 2 2 (2(w + x + y + z) − 2(w + x − y − z)) (w + x + y + z)2 − (w + x − y − z)2 +
∂w ( ))
2 (2(w + x − y − z) + 2(w + x + y + z)) (w + x − y − z)2 + (w + x + y + z)2 ×
(( )2 ( )2 )
(w + x + y + z)2 − (w + x − y − z)2 + (w + x − y − z)2 + (w + x + y + z)2 .
(18.4.16)
If we then also wanted to compute ∂f

∂x , we would end up with a similar equation again with many
repeated terms, and many shared repeated terms between the two derivatives. This represents
a massive quantity of wasted work, and if we needed to compute derivatives this way, the whole
deep learning revolution would have stalled out before it began!
Let us break up the problem. We will start by trying to understand how f changes when we change
a, essentially assuming that w, x, y, and z all do not exist. We will reason as we did back when we
worked with the gradient for the first time. Let us take a and add a small amount ϵ to it.
f (u(a + ϵ, b), v(a + ϵ, b))

( )
∂u ∂v
≈f u(a, b) + ϵ (a, b), v(a, b) + ϵ (a, b)
∂a ∂a (18.4.17)
[ ]
∂f ∂u ∂f ∂v
≈f (u(a, b), v(a, b)) + ϵ (u(a, b), v(a, b)) (a, b) + (u(a, b), v(a, b)) (a, b) .
∂u ∂a ∂v ∂a
The first line follows from the definition of partial derivative, and the second follows from the
definition of gradient. It is notationally burdensome to track exactly where we evaluate every
derivative, as in the expression ∂f
∂u (u(a, b), v(a, b)), so we often abbreviate this to the much more
memorable
∂f ∂f ∂u ∂f ∂v
= + . (18.4.18)
∂a ∂u ∂a ∂v ∂a
It is useful to think about the meaning of the process. We are trying to understand how a function
of the form f (u(a, b), v(a, b)) changes its value with a change in a. There are two pathways this can

occur: there is the pathway where a → u → f and where a → v → f . We can compute both of
∂u · ∂x and ∂v · ∂x respectively, and added up.
these contributions via the chain rule: ∂w ∂u ∂w ∂v
Imagine we have a different network of functions where the functions on the right depend on
those that are connected to on the left as is shown in Fig. 18.4.2.
Fig. 18.4.2: Another more subtle example of the chain rule.
∂f
To compute something like ∂y , we need to sum over all (in this case 3) paths from y to f giving
∂f ∂f ∂a ∂u ∂f ∂u ∂f ∂b ∂v
= + + . (18.4.19)
∂y ∂a ∂u ∂y ∂u ∂y ∂b ∂v ∂y
Understanding the chain rule in this way will pay great dividends when trying to understand how
gradients flow through networks, and why various architectural choices like those in LSTMs (Sec-
tion 9.2) or residual layers (Section 7.6) can help shape the learning process by controlling gradient
flow.
18.4.5 The Backpropagation Algorithm
Let us return to the example of (18.4.14) the previous section where
f (u, v) = (u + v)2
u(a, b) = (a + b)2 , v(a, b) = (a − b)2 , (18.4.20)
a(w, x, y, z) = (w + x + y + z)2 , b(w, x, y, z) = (w + x − y − z)2 .
∂f
If we want to compute say ∂w we may apply the multi-variate chain rule to see:
∂f ∂f ∂u ∂f ∂v
= + ,
∂w ∂u ∂w ∂v ∂w
∂u ∂u ∂a ∂u ∂b
= + , (18.4.21)
∂w ∂a ∂w ∂b ∂w
∂v ∂v ∂a ∂v ∂b
= + .
∂w ∂a ∂w ∂b ∂w
∂f
Let us try using this decomposition to compute ∂w . Notice that all we need here are the various
single step partials:
∂f ∂f
= 2(u + v), = 2(u + v),
∂u ∂v
∂u ∂u
= 2(a + b), = 2(a + b),
∂a ∂b (18.4.22)
∂v ∂v
= 2(a − b), = −2(a − b),
∂a ∂b
∂a ∂b
= 2(w + x + y + z), = 2(w + x − y − z).
∂w ∂w
If we write this out into code this becomes a fairly manageable expression.

# Compute the value of the function from inputs to outputs
w, x, y, z = -1, 0, -2, 1
a, b = (w + x + y + z)**2, (w + x - y - z)**2
u, v = (a + b)**2, (a - b)**2
f = (u + v)**2
print(f' f at {w}, {x}, {y}, {z} is {f}')
# Compute the single step partials

df_du, df_dv = 2*(u + v), 2*(u + v)
du_da, du_db, dv_da, dv_db = 2*(a + b), 2*(a + b), 2*(a - b), -2*(a - b)
da_dw, db_dw = 2*(w + x + y + z), 2*(w + x - y - z)
# Compute the final result from inputs to outputs

du_dw, dv_dw = du_da*da_dw + du_db*db_dw, dv_da*da_dw + dv_db*db_dw
df_dw = df_du*du_dw + df_dv*dv_dw
print(f'df/dw at {w}, {x}, {y}, {z} is {df_dw}')
f at -1, 0, -2, 1 is 1024

df/dw at -1, 0, -2, 1 is -4096
However, note that this still does not make it easy to compute something like ∂f
∂x . The reason for
that is the way we chose to apply the chain rule. If we look at what we did above, we always kept
∂w in the denominator when we could. In this way, we chose to apply the chain rule seeing how
w changed every other variable. If that is what we wanted, this would be a good idea. However,
think back to our motivation from deep learning: we want to see how every parameter changes
the loss. In essence, we want to apply the chain rule keeping ∂f in the numerator whenever we
can!
To be more explicit, note that we can write
∂f ∂f ∂a ∂f ∂b
= + ,
∂w ∂a ∂w ∂b ∂w
∂f ∂f ∂u ∂f ∂v
= + , (18.4.23)
∂a ∂u ∂a ∂v ∂a
∂f ∂f ∂u ∂f ∂v
= + .
∂b ∂u ∂b ∂v ∂b
∂f ∂f ∂f ∂f ∂f
Note that this application of the chain rule has us explicitly compute ∂u , ∂v , ∂a , ∂b , and ∂w . Noth-
ing stops us from also including the equations:
∂f ∂f ∂a ∂f ∂b
= + ,
∂x ∂a ∂x ∂b ∂x
∂f ∂f ∂a ∂f ∂b
= + , (18.4.24)
∂y ∂a ∂y ∂b ∂y
∂f ∂f ∂a ∂f ∂b
= + .
∂z ∂a ∂z ∂b ∂z
and then keeping track of how f changes when we change any node in the entire network. Let us
implement it.
# Compute the value of the function from inputs to outputs

w, x, y, z = -1, 0, -2, 1
a, b = (w + x + y + z)**2, (w + x - y - z)**2

u, v = (a + b)**2, (a - b)**2
f = (u + v)**2
print(f'f at {w}, {x}, {y}, {z} is {f}')
# Compute the derivative using the decomposition above

# First compute the single step partials
df_du, df_dv = 2*(u + v), 2*(u + v)
du_da, du_db, dv_da, dv_db = 2*(a + b), 2*(a + b), 2*(a - b), -2*(a - b)
da_dw, db_dw = 2*(w + x + y + z), 2*(w + x - y - z)
da_dx, db_dx = 2*(w + x + y + z), 2*(w + x - y - z)
da_dy, db_dy = 2*(w + x + y + z), -2*(w + x - y - z)
da_dz, db_dz = 2*(w + x + y + z), -2*(w + x - y - z)
# Now compute how f changes when we change any value from output to input
df_da, df_db = df_du*du_da + df_dv*dv_da, df_du*du_db + df_dv*dv_db
df_dw, df_dx = df_da*da_dw + df_db*db_dw, df_da*da_dx + df_db*db_dx
df_dy, df_dz = df_da*da_dy + df_db*db_dy, df_da*da_dz + df_db*db_dz
print(f'df/dw at {w}, {x}, {y}, {z} is {df_dw}')

print(f'df/dx at {w}, {x}, {y}, {z} is {df_dx}')
print(f'df/dy at {w}, {x}, {y}, {z} is {df_dy}')
print(f'df/dz at {w}, {x}, {y}, {z} is {df_dz}')
f at -1, 0, -2, 1 is 1024

df/dw at -1, 0, -2, 1 is -4096
df/dx at -1, 0, -2, 1 is -4096
df/dy at -1, 0, -2, 1 is -4096
df/dz at -1, 0, -2, 1 is -4096
The fact that we compute derivatives from f back towards the inputs rather than from the inputs
forward to the outputs (as we did in the first code snippet above) is what gives this algorithm its
name: backpropagation. Note that there are two steps: 1. Compute the value of the function, and
the single step partials from front to back. While not done above, this can be combined into a
single forward pass. 2. Compute the gradient of f from back to front. We call this the backwards
pass.
This is precisely what every deep learning algorithm implements to allow the computation of the
gradient of the loss with respect to every weight in the network at one pass. It is an astonishing
fact that we have such a decomposition.
To see how to encapsulated this, let us take a quick look at this example.
# Initialize as ndarrays, then attach gradients

w, x, y, z = np.array(-1), np.array(0), np.array(-2), np.array(1)
w.attach_grad()
x.attach_grad()
y.attach_grad()
z.attach_grad()
# Do the computation like usual, tracking gradients

a, b = (w + x + y + z)**2, (w + x - y - z)**2

u, v = (a + b)**2, (a - b)**2
f = (u + v)**2
# Execute backward pass

f.backward()
print(f'df/dw at {w}, {x}, {y}, {z} is {w.grad}')

print(f'df/dx at {w}, {x}, {y}, {z} is {x.grad}')
print(f'df/dy at {w}, {x}, {y}, {z} is {y.grad}')
print(f'df/dz at {w}, {x}, {y}, {z} is {z.grad}')
df/dw at -1.0, 0.0, -2.0, 1.0 is -4096.0

df/dx at -1.0, 0.0, -2.0, 1.0 is -4096.0
df/dy at -1.0, 0.0, -2.0, 1.0 is -4096.0
df/dz at -1.0, 0.0, -2.0, 1.0 is -4096.0
All of what we did above can be done automatically by calling f.backwards().
18.4.6 Hessians
As with single variable calculus, it is useful to consider higher-order derivatives in order to get a
handle on how we can obtain a better approximation to a function than using the gradient alone.
There is one immediate problem one encounters when working with higher order derivatives of
functions of several variables, and that is there are a large number of them. If we have a function
f (x1 , . . . , xn ) of n variables, then we can take n2 many second derivatives, namely for any choice
of i and j:
( )
d2 f d d
= f . (18.4.25)
dxi dxj dxi dxj
This is traditionally assembled into a matrix called the Hessian:

 2 2f

d f
· · · dxd1 dx
 .
dx1 dx1
.. 
n
Hf = 
 . . ..
. . .
 (18.4.26)
d2 f d2 f
dxn dx1 ··· dxn dxn
Not every entry of this matrix is independent. Indeed, we can show that as long as both mixed
partials (partial derivatives with respect to more than one variable) exist and are continuous, we
can say that for any i, and j,
d2 f d2 f
= . (18.4.27)
dxi dxj dxj dxi
This follows by considering first perturbing a function in the direction of xi , and then perturbing
it in xj and then comparing the result of that with what happens if we perturb first xj and then xi ,
with the knowledge that both of these orders lead to the same final change in the output of f .
As with single variables, we can use these derivatives to get a far better idea of how the function
behaves near a point. In particular, we can use it to find the best fitting quadratic near a point x0 ,
as we saw in a single variable.

Let us see an example. Suppose that f (x1 , x2 ) = a+b1 x1 +b2 x2 +c11 x21 +c12 x1 x2 +c22 x22 . This is the
general form for a quadratic in two variables. If we look at the value of the function, its gradient,
and its Hessian (18.4.26), all at the point zero:
f (0, 0) = a,
[ ]
b
∇f (0, 0) = 1 ,
b2 (18.4.28)
[ ]
2c11 c12
Hf (0, 0) = ,
c12 2c22
we can get our original polynomial back by saying

1
f (x) = f (0) + ∇f (0) · x + x⊤ Hf (0)x. (18.4.29)
2
In general, if we computed this expansion any point x0 , we see that
1
f (x) = f (x0 ) + ∇f (x0 ) · (x − x0 ) + (x − x0 )⊤ Hf (x0 )(x − x0 ). (18.4.30)
2
This works for any dimensional input, and provides the best approximating quadratic to any func-
tion at a point. To give an example, let us plot the function
f (x, y) = xe−x
2 −y 2
. (18.4.31)
One can compute that the gradient and Hessian are

( ) ( 3 )
−x2 −y 2 1 − 2x 4x − 6x 4x2 y − 2y
2
−x2 −y 2
∇f (x, y) = e and Hf (x, y) = e . (18.4.32)
−2xy 4x2 y − 2y 4xy 2 − 2x
And thus, with a little algebra, see that the approximating quadratic at [−1, 0]⊤ is
( )
f (x, y) ≈ e−1 −1 − (x + 1) + 2(x + 1)2 + 2y 2 . (18.4.33)
# Construct grid and compute function

x, y = np.meshgrid(np.linspace(-2, 2, 101),
np.linspace(-2, 2, 101), indexing='ij')
z = x*np.exp(- x**2 - y**2)
# Compute approximating quadratic with gradient and Hessian at (1, 0)

w = np.exp(-1)*(-1 - (x + 1) + 2 * (x + 1)**2 + 2 * y**2)
# Plot function
ax.plot_wireframe(x, y, w, **{'rstride': 10, 'cstride': 10}, color='purple')
d2l.plt.xlabel('x')
d2l.plt.ylabel('y')
d2l.set_figsize()
ax.set_xlim(-2, 2)
ax.set_ylim(-2, 2)
ax.set_zlim(-1, 1)
ax.dist = 12

This forms the basis for Newtonʼs Algorithm discussed in Section 11.3, where we perform numer-
ical optimization iteratively finding the best fitting quadratic, and then exactly minimizing that
quadratic.
18.4.7 A Little Matrix Calculus
Derivatives of functions involving matrices turn out to be particularly nice. This section can be-
come notationally heavy, so may be skipped in a first reading, but it is useful to know how deriva-
tives of functions involving common matrix operations are often much cleaner than one might
initially anticipate, particularly given how central matrix operations are to deep learning applica-
tions.
Let us begin with an example. Suppose that we have some fixed column vector β, and we want
to take the product function f (x) = β ⊤ x, and understand how the dot product changes when we
change x.
A bit of notation that will be useful when working with matrix derivatives in ML is called the de-
nominator layout matrix derivative where we assemble our partial derivatives into the shape of
whatever vector, matrix, or tensor is in the denominator of the differential. In this case, we will
write
 df 
dx
1
df  
=  ...  , (18.4.34)
dx df
dxn
where we matched the shape of the column vector x.

If we write out our function into components this is
∑
n
f (x) = βi xi = β1 x1 + · · · + βn xn . (18.4.35)
i=1
If we now take the partial derivative with respect to say β1 , note that everything is zero but the first
term, which is just x1 multiplied by β1 , so the we obtain that
df
= β1 , (18.4.36)
dx1
or more generally that
df
= βi . (18.4.37)
dxi

We can now reassemble this into a matrix to see
 df   
dx1 β1
df    
=  ...  =  ...  = β. (18.4.38)
dx df
dx
βn
n
This illustrates a few factors about matrix calculus that we will often counter throughout this sec-
tion:
• First, The computations will get rather involved.
• Second, The final results are much cleaner than the intermediate process, and will always
d
look similar to the single variable case. In this case, note that dx d
(bx) = b and dx (β ⊤ x) = β
are both similar.
• Third, transposes can often appear seemingly from nowhere. The core reason for this is the
convention that we match the shape of the denominator, thus when we multiply matrices,
we will need to take transposes to match back to the shape of the original term.
To keep building intuition, let us try a computation that is a little harder. Suppose that we have a
column vector x, and a square matrix A and we want to compute
d ⊤
(x Ax). (18.4.39)
dx
To drive towards easier to manipulate notation, let us consider this problem using Einstein nota-
tion. In this case we can write the function as
x⊤ Ax = xi aij xj . (18.4.40)
To compute our derivative, we need to understand for every k, what the value of
d d
(x⊤ Ax) = xi aij xj . (18.4.41)
dxk dxk
By the product rule, this is
d dxi dxj
xi aij xj = aij xj + xi aij . (18.4.42)
dxk dxk dxk
dxi
For a term like dxk
, it is not hard to see that this is one when i = k and zero otherwise. This means
that every term where i and k are different vanish from this sum, so the only terms that remain in
that first sum are the ones where i = k. The same reasoning holds for the second term where we
need j = k. This gives
d
xi aij xj = akj xj + xi aik . (18.4.43)
dxk
Now, the names of the indices in Einstein notation are arbitrary—the fact that i and j are different
is immaterial to this computation at this point, so we can re-index so that they both use i to see
that
d
xi aij xj = aki xi + xi aik = (aki + aik )xi . (18.4.44)
dxk
Now, here is where we start to need some practice to go further. Let us try and identify this out-
come in terms of matrix operations. aki + aik is the k, i-th component of A + A⊤ . This gives
d
xi aij xj = [A + A⊤ ]ki xi . (18.4.45)
dxk

Similarly, this term is now the product of the matrix A + A⊤ by the vector x, so we see that
[ ]
d ⊤ d
(x Ax) = xi aij xj = [(A + A⊤ )x]k . (18.4.46)
dx k dx k
Thus, we see that the k-th entry of the desired derivative from (18.4.39) is just the k-th entry of the
vector on the right, and thus the two are the same. Thus yields
d ⊤
(x Ax) = (A + A⊤ )x. (18.4.47)
dx
This required significantly more work than our last one, but the final result is small. More than
that, consider the following computation for traditional single variable derivatives:
d dx dx
(xax) = ax + xa = (a + a)x. (18.4.48)
dx dx dx
d
Equivalently dx (ax2 ) = 2ax = (a + a)x. Again, we get a result that looks rather like the single
variable result but with a transpose tossed in.
At this point, the pattern should be looking rather suspicious, so let us try to figure out why. When
we take matrix derivatives like this, let us first assume that the expression we get will be another
matrix expression: an expression we can write it in terms of products and sums of matrices and
their transposes. If such an expression exists, it will need to be true for all matrices. In particular,
it will need to be true of 1 × 1 matrices, in which case the matrix product is just the product of the
numbers, the matrix sum is just the sum, and the transpose does nothing at all! In other words,
whatever expression we get must match the single variable expression. This means that, with
some practice, one can often guess matrix derivatives just by knowing what the associated single
variable expression must look like!
Let us try this out. Suppose that X is a n × m matrix, U is an n × r and V is an r × m. Let us try to
compute
d
∥X − UV∥22 = ? (18.4.49)
dV
This computation is important in an area called matrix factorization. For us, however, it is just a
derivative to compute. Let us try to imaging what this would be for 1 × 1 matrices. In that case,
we get the expression
d
(x − uv)2 = −2(x − uv)u, (18.4.50)
dv
where, the derivative is rather standard. If we try to convert this back into a matrix expression we
get
d
∥X − UV∥22 = −2(X − UV)U. (18.4.51)
dV
However, if we look at this it does not quite work. Recall that X is n × m, as is UV, so the matrix
2(X − UV) is n × m. On the other hand U is n × r, and we cannot multiply a n × m and a n × r
matrix since the dimensions do not match!
d
We want to get dV , which is the same shape of V, which is r × m. So somehow we need to take a
n × m matrix and a n × r matrix, multiply them together (perhaps with some transposes) to get a
r × m. We can do this by multiplying U ⊤ by (X − UV). Thus, we can guess the solution to (18.4.49)
is
d
∥X − UV∥22 = −2U⊤ (X − UV). (18.4.52)
dV

To show that this works, we would be remiss to not provide a detailed computation. If we already
believe that this rule-of-thumb works, feel free to skip past this derivation. To compute
d
∥X − UV∥22 , (18.4.53)
dV
we must find for every a, and b
( )2
d d ∑ ∑
∥X − UV∥2 =
2
xij − uik vkj . (18.4.54)
dvab dvab
i,j k
Recalling that all entries of X and U are constants as far as dvdab is concerned, we may push the
derivative inside the sum, and apply the chain rule to the square to get
( )( )
d ∑ ∑ ∑ dvkj
∥X − UV∥2 =
2
2 xij − uik vkj − uik . (18.4.55)
dvab dvab
i,j k k
dv
As in the previous derivation, we may note that dvkjab
is only non-zero if the k = a and j = b. If
either of those conditions do not hold, the term in the sum is zero, and we may freely discard it.
We see that
( )
d ∑ ∑
∥X − UV∥22 = −2 xib − uik vkb uia . (18.4.56)
dvab
i k
An important subtlety here is that the requirement that k = a does not occur inside the inner
sum since that k is a dummy variable which we are summing over inside the inner term. For a
notationally cleaner example, consider why
( )2 ( )
d ∑ ∑
xi =2 xi . (18.4.57)
dx1
i i
From this point, we may start identifying components of the sum. First,
∑
uik vkb = [UV]ib . (18.4.58)
k
So the entire expression in the inside of the sum is

∑
xib − uik vkb = [X − UV]ib . (18.4.59)
k
This means we may now write our derivative as
d ∑
∥X − UV∥22 = −2 [X − UV]ib uia . (18.4.60)
dvab
i
We want this to look like the a, b element of a matrix so we can use the technique as in the previous
example to arrive at a matrix expression, which means that we need to exchange the order of the
indices on uia . If we notice that uia = [U⊤ ]ai , we can then write
d ∑
∥X − UV∥22 = −2 [U⊤ ]ai [X − UV]ib . (18.4.61)
dvab
i

This is a matrix product, and thus we can conclude that
d
∥X − UV∥22 = −2[U⊤ (X − UV)]ab . (18.4.62)
dvab
and thus we may write the solution to (18.4.49)
d
∥X − UV∥22 = −2U⊤ (X − UV). (18.4.63)
dV
This matches the solution we guessed above!
It is reasonable to ask at this point, “Why can I not just write down matrix versions of all the cal-
culus rules I have learned? It is clear this is still mechanical. Why do we not just get it over with!”
And indeed there are such rules and (Petersen et al., 2008) provides an excellent summary. How-
ever, due to the plethora of ways matrix operations can be combined compared to single values,
there are many more matrix derivative rules than single variable ones. It is often the case that it
is best to work with the indices, or leave it up to automatic differentiation when appropriate.
Summary
• In higher dimensions, we can define gradients which serve the same purpose as derivatives
in one dimension. These allow us to see how a multi-variable function changes when we
make an arbitrary small change to the inputs.
• The backpropagation algorithm can be seen to be a method of organizing the multi-variable
chain rule to allow for the efficient computation of many partial derivatives.
• Matrix calculus allows us to write the derivatives of matrix expressions in concise ways.
Exercises
1. Given a column vector β, compute the derivatives of both f (x) = β ⊤ x and g(x) = x⊤ β. Why
do you get the same answer?
∂v ∥v∥2 ?
∂
2. Let v be an n dimension vector. What is
3. Let L(x, y) = log(ex + ey ). Compute the gradient. What is the sum of the components of the
gradient?
4. Let f (x, y) = x2 y + xy 2 . Show that the only critical point is (0, 0). By considering f (x, x),
determine if (0, 0) is a maximum, minimum, or neither.
5. Suppose that we are minimizing a function f (x) = g(x) + h(x). How can we geometrically
interpret the condition of ∇f = 0 in terms of g and h?
Discussions248
248

18.5 Integral Calculus
Differentiation only makes up half of the content of a traditional calculus education. The other
pillar, integration, starts out seeming a rather disjoint question, “What is the area underneath this
curve?” While seemingly unrelated, integration is tightly intertwined with the differentiation via
what is known as the fundamental theorem of calculus.
At the level of machine learning we discuss in this book, we will not need a deep understanding of
integration. However, we will provide a brief introduction to lay the groundwork for any further
applications we will encounter later on.
18.5.1 Geometric Interpretation
Suppose that we have a function f (x). For simplicity, let us assume that f (x) is non-negative (never
takes a value less than zero). What we want to try and understand is: what is the area contained
between f (x) and the x-axis?
%matplotlib inline
npx.set_np()
x = np.arange(-2, 2, 0.01)
f = np.exp(-x**2)
d2l.set_figsize()
d2l.plt.plot(x, f, color='black')
d2l.plt.fill_between(x.tolist(), f.tolist())
d2l.plt.show()
In most cases, this area will be infinite or undefined (consider the area under f (x) = x2 ), so people
will often talk about the area between a pair of ends, say a and b.
x = np.arange(-2, 2, 0.01)
f = np.exp(-x**2)
18.5. Integral Calculus 829

d2l.set_figsize()
d2l.plt.fill_between(x.tolist()[50:250], f.tolist()[50:250])
d2l.plt.show()
We will denote this area by the integral symbol below:

∫ b
Area(A) = f (x) dx. (18.5.1)
a
∑
The inner variable is a dummy variable, much like the index of a sum in a , and so this can be
equivalently written with any inner value we like:
∫ b ∫ b
f (x) dx = f (z) dz. (18.5.2)
a a
There is a traditional way to try and understand how we might try to approximate such integrals:
we can imagine taking the region in-between a and b and chopping it into N vertical slices. If N
is large, we can approximate the area of each slice by a rectangle, and then add up the areas to get
the total area under the curve. Let us take a look at an example doing this in code. We will see
how to get the true value in a later section.
epsilon = 0.05
a = 0
b = 2
x = np.arange(a, b, epsilon)
f = x / (1 + x**2)
approx = np.sum(epsilon*f)
true = np.log(2) / 2
d2l.set_figsize()
d2l.plt.bar(x.asnumpy(), f.asnumpy(), width=epsilon, align='edge')
d2l.plt.ylim([0, 1])
d2l.plt.show()
f'approximation: {approx}, truth: {true}'

'approximation: 0.79448557, truth: 0.34657359027997264'
The issue is that while it can be done numerically, we can do this approach analytically for only
the simplest functions like
∫ b
x dx. (18.5.3)
a
Anything somewhat more complex like our example from the code above
∫ b
x
dx. (18.5.4)
a 1 + x2
is beyond what we can solve with such a direct method.

We will instead take a different approach. We will work intuitively with the notion of the area, and
learn the main computational tool used to find integrals: the fundamental theorem of calculus. This
will be the basis for our study of integration.
18.5.2 The Fundamental Theorem of Calculus
To dive deeper into the theory of integration, let us introduce a function

∫ x
F (x) = f (y)dy. (18.5.5)
0
This function measures the area between 0 and x depending on how we change x. Notice that this
is everything we need since
∫ b
f (x) dx = F (b) − F (a). (18.5.6)
a
This is a mathematical encoding of the fact that we can measure the area out to the far end-point
and then subtract off the area to the near end point as indicated in Fig. 18.5.1.

Fig. 18.5.1: Visualizing why we may reduce the problem of computing the area under a curve
between two points to computing the area to the left of a point.
Thus, we can figure out what the integral over any interval is by figuring out what F (x) is.
To do so, let us consider an experiment. As we often do in calculus, let us imagine what happens
when we shift the value by a tiny bit. From the comment above, we know that
∫ x+ϵ
F (x + ϵ) − F (x) = f (y) dy. (18.5.7)
x
This tells us that the function changes by the area under a tiny sliver of a function.
This is the point at which we make an approximation. If we look at a tiny sliver of area like this, it
looks like this area is close to the rectangular area with height the value of f (x) and the base width
ϵ. Indeed, one can show that as ϵ → 0 this approximation becomes better and better. Thus we can
conclude:
F (x + ϵ) − F (x) ≈ ϵf (x). (18.5.8)
However, we can now notice: this is exactly the pattern we expect if we were computing the deriva-
tive of F ! Thus we see the following rather surprising fact:
dF
(x) = f (x). (18.5.9)
dx
This is the fundamental theorem of calculus. We may write it in expanded form as
∫ x
d
f (y) dy = f (x). (18.5.10)
dx −∞
It takes the concept of finding areas (a priori rather hard), and reduces it to a statement derivatives
(something much more completely understood). One last comment that we must make is that this
does not tell us exactly what F (x) is. Indeed F (x) + C for any C has the same derivative. This
is a fact-of-life in the theory of integration. Thankfully, notice that when working with definite
integrals, the constants drop out, and thus are irrelevant to the outcome.
∫ b
f (x) dx = (F (b) + C) − (F (a) + C) = F (b) − F (a). (18.5.11)
a
This may seem like abstract non-sense, but let us take a moment to appreciate that it has given us
a whole new perspective on computing integrals. Our goal is no-longer to do some sort of chop-
and-sum process to try and recover the area, rather we need only find a function whose derivative
is the function we have! This is incredible since we can now list many rather difficult integrals
by just reversing the table from Section 18.3.2. For instance, we know that the derivative of xn is
nxn−1 . Thus, we can say using the fundamental theorem (18.5.10) that
∫ x
ny n−1 dy = xn − 0n = xn . (18.5.12)
0

Similarly, we know that the derivative of ex is itself, so that means
∫ x
ex dx = ex − e0 = ex − 1. (18.5.13)
0
In this way, we can develop the entire theory of integration leveraging ideas from differential
calculus freely. Every integration rule derives from this one fact.
18.5.3 Change of Variables
Just as with differentiation, there are a number of rules which make the computation of integrals
more tractable. In fact, every rule of differential calculus (like the product rule, sum rule, and
chain rule) has a corresponding rule for integral calculus (integration by parts, linearity of inte-
gration, and the change of variables formula respectively). In this section, we will dive into what
is arguably the most important from the list: the change of variables formula.
First, suppose that we have a function which is itself an integral:
∫ x
F (x) = f (y) dy. (18.5.14)
0
Let us suppose that we want to know how this function looks when we compose it with another to
obtain F (u(x)). By the chain rule, we know
d dF du
F (u(x)) = (u(x)) · . (18.5.15)
dx du dx
We can turn this into a statement about integration by using the fundamental theorem (18.5.10)
as above. This gives
∫ x
dF du
F (u(x)) − F (u(0)) = (u(y)) · dy. (18.5.16)
0 du dy
Recalling that F is itself an integral gives that the left hand side may be rewritten to be
∫ u(x) ∫ x
dF du
f (y) dy = (u(y)) · dy. (18.5.17)
u(0) 0 du dy
dF
Similarly, recalling that F is an integral allows us to recognize that dx = f using the fundamental
theorem (18.5.10), and thus we may conclude
∫ u(x) ∫ x
du
f (y) dy = f (u(y)) · dy. (18.5.18)
u(0) 0 dy
This is the change of variables formula.

For a more intuitive derivation, consider what happens when we take an integral of f (u(x)) be-
tween x and x + ϵ. For a small ϵ, this integral is approximately ϵf (u(x)), the area of the associated
rectangle. Now, let us compare this with the integral of f (y) from u(x) to u(x + ϵ). We know that
u(x + ϵ) ≈ u(x) + ϵ du du
dx (x), so the area of this rectangle is approximately ϵ dx (x)f (u(x)). Thus, to
make the area of these two rectangles to agree, we need to multiply the first one by du dx (x) as is
illustrated in Fig. 18.5.2.

Fig. 18.5.2: Visualizing the transformation of a single thin rectangle under the change of variables.
This tells us that

∫ x+ϵ ∫ u(x+ϵ)
du
f (u(y)) (y) dy = f (y) dy. (18.5.19)
x dy u(x)
This is the change of variables formula expressed for a single small rectangle.
If u(x) and f (x) are properly chosen, this can allow for the computation of incredibly complex in-
tegrals. For instance, if we even chose f (y) = 1 and u(x) = e−x (which means du −x2 ),
2
dx (x) = −2xe
this can show for instance that
∫ e−1 ∫ 1
e−1 − 1 = ye−y dy,
2
1 dy = −2 (18.5.20)
e−0 0
and thus by rearranging that

∫ 1
1 − e−1
ye−y dy =
2
. (18.5.21)
0 2
18.5.4 A Comment on Sign Conventions
Keen-eyed readers will observe something strange about the computations above. Namely, com-
putations like
∫ e−1
1 dy = e−1 − 1 < 0, (18.5.22)
e−0
can produce negative numbers. When thinking about areas, it can be strange to see a negative
value, and so it is worth digging into what the convention is.
Mathematicians take the notion of signed areas. This manifests itself in two ways. First, if we
consider a function f (x) which is sometimes less than zero, then the area will also be negative. So
for instance
∫ 1
(−1) dx = −1. (18.5.23)
0
Similarly, integrals which progress from right to left, rather than left to right are also taken to be
negative areas
∫ −1
1 dx = −1. (18.5.24)
0
The standard area (from left to right of a positive function) is always positive. Anything obtained
by flipping it (say flipping over the x-axis to get the integral of a negative number, or flipping over

the y-axis to get an integral in the wrong order) will produce a negative area. And indeed, flipping
twice will give a pair of negative signs that cancel out to have positive area
∫ −1
(−1) dx = 1. (18.5.25)
0
If this discussion sounds familiar, it is! In Section 18.1 we discussed how the determinant repre-
sented the signed area in much the same way.
18.5.5 Multiple Integrals
In some cases, we will need to work in higher dimensions. For instance, suppose that we have
a function of two variables, like f (x, y) and we want to know the volume under f when x ranges
over [a, b] and y ranges over [c, d].
# Construct grid and compute function

x, y = np.meshgrid(np.linspace(-2, 2, 101), np.linspace(-2, 2, 101),
indexing='ij')
z = np.exp(- x**2 - y**2)
# Plot function
ax.plot_wireframe(x, y, z)
d2l.plt.xlabel('x')
d2l.plt.ylabel('y')
d2l.plt.xticks([-2, -1, 0, 1, 2])
d2l.plt.yticks([-2, -1, 0, 1, 2])
d2l.set_figsize()
ax.set_xlim(-2, 2)
ax.set_ylim(-2, 2)
ax.set_zlim(0, 1)
ax.dist = 12
We write this as
∫
f (x, y) dx dy. (18.5.26)
[a,b]×[c,d]
Suppose that we wish to compute this integral. My claim is that we can do this by iteratively com-
puting first the integral in x and then shifting to the integral in y, that is to say
∫ ∫ d (∫ b )
f (x, y) dx dy = f (x, y) dx dy. (18.5.27)
[a,b]×[c,d] c a

Let us see why this is.
Consider the figure above where we have split the function into ϵ × ϵ squares which we will index
with integer coordinates i, j. In this case, our integral is approximately
∑
ϵ2 f (ϵi, ϵj). (18.5.28)
i,j
Once we discretize the problem, we may add up the values on these squares in whatever order we
like, and not worry about changing the values. This is illustrated in Fig. 18.5.3. In particular, we
can say that
( )
∑ ∑
ϵ ϵf (ϵi, ϵj) . (18.5.29)
j i
Fig. 18.5.3: Illustrating how to decompose a sum over many squares as a sum over first the columns
(1), then adding the column sums together (2).
The sum on the inside is precisely the discretization of the integral

∫ b
G(ϵj) = f (x, ϵj) dx. (18.5.30)
a
Finally, notice that if we combine these two expressions we get

∑ ∫ d ∫
ϵG(ϵj) ≈ G(y) dy = f (x, y) dx dy. (18.5.31)
j c [a,b]×[c,d]
Thus putting it all together, we have that

∫ ∫ d (∫ b )
f (x, y) dx dy = f (x, y) dx dy. (18.5.32)
[a,b]×[c,d] c a
Notice that, once discretized, all we did was rearrange the order in which we added a list of num-
bers. This may make it seem like it is nothing, however this result (called Fubini’s Theorem) is not
always true! For the type of mathematics encountered when doing machine learning (continu-
ous functions), there is no concern, however it is possible to create examples where it fails (for
example the function f (x, y) = xy(x2 − y 2 )/(x2 + y 2 )3 over the rectangle [0, 2] × [0, 1]).
Note that the choice to do the integral in x first, and then the integral in y was arbitrary. We could
have equally well chosen to do y first and then x to see
∫ ∫ b (∫ d )
f (x, y) dx dy = f (x, y) dy dx. (18.5.33)
[a,b]×[c,d] a c

Often times, we will condense down to vector notation, and say that for U = [a, b] × [c, d] this is
∫
f (x) dx. (18.5.34)
U
18.5.6 Change of Variables in Multiple Integrals
As with single variables in (18.5.18), the ability to change variables inside a higher dimensional
integral is a key tool. Let us summarize the result without derivation.
We need a function that reparameterizes our domain of integration. We can take this to be ϕ :
Rn → Rn , that is any function which takes in n real variables and returns another n. To keep
the expressions clean, we will assume that ϕ is injective which is to say it never folds over itself
(ϕ(x) = ϕ(y) =⇒ x = y).
In this case, we can say that
∫ ∫
f (x) dx = f (ϕ(x)) |det(Dϕ(x))| dx. (18.5.35)
ϕ(U ) U
where Dϕ is the Jacobian of ϕ, which is the matrix of partial derivatives of ϕ =

(ϕ1 (x1 , . . . , xn ), . . . , ϕn (x1 , . . . , xn )),
 ∂ϕ1 ∂ϕ1 
∂x1 · · · ∂x n
 ..  .
Dϕ =  ... ..
. .  (18.5.36)
∂ϕn ∂ϕn
∂x1 · · · ∂xn
Looking closely, we see that this is similar to the single variable chain rule (18.5.18), except we
have replaced the term du dx (x) with |det(Dϕ(x))|. Let us see how we can to interpret this term.
du
Recall that the dx (x) term existed to say how much we stretched our x-axis by applying u. The
same process in higher dimensions is to determine how much we stretch the area (or volume, or
hyper-volume) of a little square (or little hyper-cube) by applying ϕ. If ϕ was the multiplication by
a matrix, then we know how the determinant already gives the answer.
With some work, one can show that the Jacobian provides the best approximation to a multivari-
able function ϕ at a point by a matrix in the same way we could approximate by lines or planes
with derivatives and gradients. Thus the determinant of the Jacobian exactly mirrors the scaling
factor we identified in one dimension.
It takes some work to fill in the details to this, so do not worry if they are not clear now. Let us see
at least one example we will make use of later on. Consider the integral
∫ ∞∫ ∞
e−x −y dx dy.
2 2
(18.5.37)
−∞ −∞
Playing with this integral directly will get us no-where, but if we change variables, we can make
significant progress. If we let ϕ(r, θ) = (r cos(θ), r sin(θ)) (which is to say that x = r cos(θ), y =
r sin(θ)), then we can apply the change of variable formula to see that this is the same thing as
∫ ∞ ∫ 2π
e−r |det(Dϕ(x))| dθ dr,
2
(18.5.38)
0 0
where
[ ]
cos(θ) −r sin(θ)

|det(Dϕ(x))| = det = r(cos2 (θ) + sin2 (θ)) = r. (18.5.39)
sin(θ) r cos(θ)

Thus, the integral is
∫ ∞ ∫ 2π ∫ ∞
−r2
re−r dr = π,
2
re dθ dr = 2π (18.5.40)
0 0 0
where the final equality follows by the same computation that we used in section Section 18.5.3.
We will meet this integral again when we study continuous random variables in Section 18.6.
Summary
• The theory of integration allows us to answer questions about areas or volumes.

• The fundamental theorem of calculus allows us to leverage knowledge about derivatives to
compute areas via the observation that the derivative of the area up to some point is given
by the value of the function being integrated.
• Integrals in higher dimensions can be computed by iterating single variable integrals.
Exercises
∫2 1
1. What is 1 x dx?
∫ √π
2. Use the change of variables formula to integrate 0 x sin(x2 ) dx.
∫
3. What is [0,1]2 xy dx dy?
∫2∫1
4. Use the change of variables formula to compute 0 0 xy(x2 − y 2 )/(x2 + y 2 )3 dy dx and
∫1∫2
0 0 f (x, y) = xy(x − y )/(x + y ) dx dy to see they are different.
2 2 2 2 3
Discussions249
18.6 Random Variables
In Section 2.6 we saw the basics of how to work with discrete random variables, which in our case
refer to those random variables which take either a finite set of possible values, or the integers.
In this section, we develop the theory of continuous random variables, which are random variables
which can take on any real value.
18.6.1 Continuous Random Variables
Continuous random variables are a significantly more subtle topic than discrete random variables.
A fair analogy to make is that the technical jump is comparable to the jump between adding lists
of numbers and integrating functions. As such, we will need to take some time to develop the
theory.
249

From Discrete to Continuous
To understand the additional technical challenges encountered when working with continuous
random variables, let us perform a thought experiment. Suppose that we are throwing a dart at
the dart board, and we want to know the probability that it hits exactly 2cm from the center of the
board.
To start with, we imagine measuring a single digit of accuracy, that is to say with bins for 0cm,
1cm, 2cm, and so on. We throw say 100 darts at the dart board, and if 20 of them fall into the bin
for 2cm we conclude that 20% of the darts we throw hit the board 2cm away from the center.
However, when we look closer, this does not match our question! We wanted exact equality,
whereas these bins hold all that fell between say 1.5cm and 2.5cm.
Undeterred, we continue further. We measure even more precisely, say 1.9cm, 2.0cm, 2.1cm, and
now see that perhaps 3 of the 100 darts hit the board in the 2.0cm bucket. Thus we conclude the
probability is 3%.
However, this does not solve anything! We have just pushed the issue down one digit further. Let
us abstract a bit. Imagine we know the probability that the first k digits match with 2.00000 . . . and
we want to know the probability it matches for the first k+1 digits. It is fairly reasonable to assume
that the k + 1th digit is essentially a random choice from the set {0, 1, 2, . . . , 9}. At least, we cannot
conceive of a physically meaningful process which would force the number of micrometers away
form the center to prefer to end in a 7 vs a 3.
What this means is that in essence each additional digit of accuracy we require should decrease
probability of matching by a factor of 10. Or put another way, we would expect that
P (distance is 2.00 . . . , to k digits) ≈ p · 10−k . (18.6.1)
The value p essentially encodes what happens with the first few digits, and the 10−k handles the
rest.
Notice that if we know the position accurate to k = 4 digits after the decimal. that means we know
the value falls within the interval say [(1.99995, 2.00005] which is an interval of length 2.00005 −
1.99995 = 10−4 . Thus, if we call the length of this interval ϵ, we can say
P (distance is in an ϵ-sized interval around 2) ≈ ϵ · p. (18.6.2)
Let us take this one final step further. We have been thinking about the point 2 the entire time,
but never thinking about other points. Nothing is different there fundamentally, but it is the case
that the value p will likely be different. We would at least hope that a dart thrower was more likely
to hit a point near the center, like 2cm rather than 20cm. Thus, the value p is not fixed, but rather
should depend on the point x. This tells us that we should expect
P (distance is in an ϵ-sized interval around x) ≈ ϵ · p(x). (18.6.3)
Indeed, (18.6.3) precisely defines the probability density function. It is a function p(x) which en-
codes the relative probability of hitting near one point vs. another. Let us visualize what such a
function might look like.
%matplotlib inline
18.6. Random Variables 839

npx.set_np()
# Plot the probability density function for some random variable

x = np.arange(-5, 5, 0.01)
p = 0.2*np.exp(-(x - 3)**2 / 2)/np.sqrt(2 * np.pi) + \
0.8*np.exp(-(x + 1)**2 / 2)/np.sqrt(2 * np.pi)
d2l.plot(x, p, 'x', 'Density')
The locations where the function value is large indicates regions where we are more likely to find
the random value. The low portions are areas where we are unlikely to find the random value.
Probability Density Functions
Let us now investigate this further. We have already seen what a probability density function is
intuitively for a random variable X, namely the density function is a function p(x) so that
P (X is in an ϵ-sized interval around x) ≈ ϵ · p(x). (18.6.4)
But what does this imply for the properties of p(x)?

First, probabilities are never negative, thus we should expect that p(x) ≥ 0 as well.
Second, let us imagine that we slice up the R into an infinite number of slices which are ϵ wide, say
with slices (ϵ·i, ϵ·(i+1)]. For each of these, we know from (18.6.4) the probability is approximately
P (X is in an ϵ-sized interval around x) ≈ ϵ · p(ϵ · i), (18.6.5)
so summed over all of them it should be

∑
P (X ∈ R) ≈ ϵ · p(ϵ · i). (18.6.6)
i
This is nothing more than the approximation of an integral discussed in Section 18.5, thus we can
say that
∫ ∞
P (X ∈ R) = p(x) dx. (18.6.7)
−∞

We know that P (X ∈ R) = 1, since the random variable must take on some number, we can
conclude that for any density
∫ ∞
p(x) dx = 1. (18.6.8)
−∞
Indeed, digging into this further shows that for any a, and b, we see that
∫ b
P (X ∈ (a, b]) = p(x) dx. (18.6.9)
a
We may approximate this in code by using the same discrete approximation methods as before.
In this case we can approximate the probability of falling in the blue region.
# Approximate probability using numerical integration

epsilon = 0.01
x = np.arange(-5, 5, 0.01)
p = 0.2*np.exp(-(x - 3)**2 / 2) / np.sqrt(2 * np.pi) + \
0.8*np.exp(-(x + 1)**2 / 2) / np.sqrt(2 * np.pi)
d2l.set_figsize()
d2l.plt.plot(x, p, color='black')
d2l.plt.fill_between(x.tolist()[300:800], p.tolist()[300:800])
d2l.plt.show()
f'approximate Probability: {np.sum(epsilon*p[300:800])}'
'approximate Probability: 0.7736172'
It turns out that these two properties describe exactly the space of possible probability density
functions (or p.d.f.ʼs for the commonly encountered abbreviation). They are non-negative func-
tions p(x) ≥ 0 such that
∫ ∞
p(x) dx = 1. (18.6.10)
−∞
We interpret this function by using integration to obtain the probability our random variable is in
a specific interval:
∫ b
P (X ∈ (a, b]) = p(x) dx. (18.6.11)
a

In Section 18.8 we will see a number of common distributions, but let us continue working in the
abstract.
Cumulative Distribution Functions
In the previous section, we saw the notion of the p.d.f. In practice, this is a commonly encountered
method to discuss continuous random variables, but it has one significant pitfall: that the values
of the p.d.f. are not themselves probabilities, but rather a function that we must integrate to yield
probabilities. There is nothing wrong with a density being larger than 10, as long as it is not larger
than 10 for more than an interval of length 1/10. This can be counter-intuitive, so people often
also think in terms of the cumulative distribution function, or c.d.f., which is a probability.
In particular, by using (18.6.11), we define the c.d.f. for a random variable X with density p(x) by
∫ x
F (x) = p(x) dx = P (X ≤ x). (18.6.12)
−∞
Let us observe a few properties.

• F (x) → 0 as x → −∞.
• F (x) → 1 as x → ∞.
• F (x) is non-decreasing (y > x =⇒ F (y) ≥ F (x)).
• F (x) is continuous (has no jumps) if X is a continuous random variable.
With the fourth bullet point, note that this would not be true if X were discrete, say taking the
values 0 and 1 both with probability 1/2. In that case


0 x < 0,
F (x) = 12 x < 1, (18.6.13)


1 x ≥ 1.
In this example, we see one of the benefits of working with the c.d.f., the ability to deal with con-
tinuous or discrete random variables in the same framework, or indeed mixtures of the two (flip
a coin: if heads return the roll of a die, if tails return the distance of a dart throw from the center
of a dart board).
Means
Suppose that we are dealing with a random variables X. The distribution itself can be hard to
interpret. It is often useful to be able to summarize the behavior of a random variable concisely.
Numbers that help us capture the behavior of a random variable are called summary statistics. The
most commonly encountered ones are the mean, the variance, and the standard deviation.
The mean encodes the average value of a random variable. If we have a discrete random variable
X, which takes the values xi with probabilities pi , then the mean is given by the weighted average:
sum the values times the probability that the random variable takes on that value:
∑
µX = E[X] = x i pi . (18.6.14)
i
The way we should interpret the mean (albeit with caution) is that it tells us essentially where the
random variable tends to be located.

As a minimalistic example that we will examine throughout this section, let us take X to be the
random variable which takes the value a − 2 with probability p, a + 2 with probability p and a with
probability 1 − 2p. We can compute using (18.6.14) that, for any possible choice of a and p, the
mean is
∑
µX = E[X] = xi pi = (a − 2)p + a(1 − 2p) + (a + 2)p = a. (18.6.15)
i
Thus we see that the mean is a. This matches the intuition since a is the location around which
we centered our random variable.
Because they are helpful, let us summarize a few properties.
• For any random variable X and numbers a and b, we have that µaX+b = aµX + b.
• If we have two random variables X and Y , we have µX+Y = µX + µY .
Means are useful for understanding the average behavior of a random variable, however the mean
is not sufficient to even have a full intuitive understanding. Making a profit of $10 ± $1 per sale is
very different from making $10 ± $15 per sale despite having the same average value. The second
one has a much larger degree of fluctuation, and thus represents a much larger risk. Thus, to
understand the behavior of a random variable, we will need at minimum one more measure: some
measure of how widely a random variable fluctuates.
Variances
This leads us to consider the variance of a random variable. This is a quantitative measure of
how far a random variable deviates from the mean. Consider the expression X − µX . This is the
deviation of the random variable from its mean. This value can be positive or negative, so we need
to do something to make it positive so that we are measuring the magnitude of the deviation.
A reasonable thing to try is to look at |X − µX |, and indeed this leads to a useful quantity called the
mean absolute deviation, however due to connections with other areas of mathematics and statis-
tics, people often use a different solution.
In particular, they look at (X − µX )2 . If we look at the typical size of this quantity by taking the
mean, we arrive at the variance
[ ]
2
σX = Var(X) = E (X − µX )2 = E[X 2 ] − µ2X . (18.6.16)
The last equality in (18.6.16) holds by expanding out the definition in the middle, and applying the
properties of expectation.
Let us look at our example where X is the random variable which takes the value a − 2 with prob-
ability p, a + 2 with
[ 2probability
] p and a with probability 1 − 2p. In this case µX = a, so all we need
to compute is E X . This can readily be done:
[ ]
E X 2 = (a − 2)2 p + a2 (1 − 2p) + (a + 2)2 p = a2 + 8p. (18.6.17)
Thus, we see that by (18.6.16) our variance is

2
σX = Var(X) = E[X 2 ] − µ2X = a2 + 8p − a2 = 8p. (18.6.18)
This result again makes sense. The largest p can be is 1/2 which corresponds to picking a − 2 or
a + 2 with a coin flip. The variance of this being 4 corresponds to the fact that both a − 2 and

a + 2 are 2 units away from the mean, and 22 = 4. On the other end of the spectrum, if p = 0, this
random variable always takes the value 0 and so it has no variance at all.
We will list a few properties of variance below:
• For any random variable X, Var(X) ≥ 0, with Var(X) = 0 if and only if X is a constant.
• For any random variable X and numbers a and b, we have that Var(aX + b) = a2 Var(X).
• If we have two independent random variables X and Y , we have Var(X + Y ) = Var(X) +
Var(Y ).
When interpreting these values, there can be a bit of a hiccup. In particular, let us try imagining
what happens if we keep track of units through this computation. Suppose that we are working
with the star rating assigned to a product on the web page. Then a, a−2, and a+2 are all measured
in units of stars. Similarly, the mean µX is then also measured in stars (being a weighted average).
However, if we get to the variance, we immediately encounter an issue, which is we want to look at
(X − µX )2 , which is in units of squared stars. This means that the variance itself is not comparable
to the original measurements. To make it interpretable, we will need to return to our original
units.
Standard Deviations
This summary statistics can always be deduced from the variance by taking the square root! Thus
we define the standard deviation to be
√
σX = Var(X). (18.6.19)
√
In our example, this means we now have the standard deviation is σX = 2 2p. If we are dealing
with units of stars for our review example, σX is again in units of stars.
The properties we had for the variance can be restated for the standard deviation.
• For any random variable X, σX ≥ 0.
• For any random variable X and numbers a and b, we have that σaX+b = |a|σX
√
• If we have two independent random variables X and Y , we have σX+Y = σX 2 + σ2 .
Y
It is natural at this moment to ask, “If the standard deviation is in the units of our original random
variable, does it represent something we can draw with regards to that random variable?” The
answer is a resounding yes! Indeed much like the mean told we the typical location of our random
variable, the standard deviation gives the typical range of variation of that random variable. We
can make this rigorous with what is known as Chebyshevʼs inequality:
1
P (X ̸∈ [µX − ασX , µX + ασX ]) ≤ . (18.6.20)
α2
Or to state it verbally in the case of α = 10, 99% of the samples from any random variable fall
within 10 standard deviations of the mean. This gives an immediate interpretation to our standard
summary statistics.
To see how this statement is rather subtle, let us take a look at our running example again where
X is the random variable which takes the value a − 2 with probability p, a + 2 with probability p
√
and a with probability 1 − 2p. We saw that the mean was a and the standard deviation was 2 2p.
This means, if we take Chebyshevʼs inequality (18.6.20) with α = 2, we see that the expression is
( √ √ ) 1
P X ̸∈ [a − 4 2p, a + 4 2p] ≤ . (18.6.21)
4

This means that 75% of the time, this random variable will fall within this interval for any value of
p. Now, notice that as p → 0, this interval also converges to the single point a. But we know that
our random variable takes the values a − 2, a, and a + 2 only so eventually we can be certain a − 2
and a + 2 will fall outside the interval! The question is, at what p does that happen. So we want to
√
solve: for what p does a + 4 2p = a + 2, which is solved when p = 1/8, which is exactly the first p
where it could possibly happen without violating our claim that no more than 1/4 of samples from
the distribution would fall outside the interval (1/8 to the left, and 1/8 to the right).
Let us visualize this. We will show the probability of getting the three values as three vertical bars
with height proportional to the probability. The interval will be drawn as a horizontal line in the
middle. The first plot shows what happens for p > 1/8 where the interval safely contains all points.
# Define a helper to plot these figures

def plot_chebyshev(a, p):
d2l.set_figsize()
d2l.plt.stem([a-2, a, a+2], [p, 1-2*p, p], use_line_collection=True)
d2l.plt.xlim([-4, 4])
d2l.plt.xlabel('x')
d2l.plt.ylabel('p.m.f.')
d2l.plt.hlines(0.5, a - 4 * np.sqrt(2 * p),

a + 4 * np.sqrt(2 * p), 'black', lw=4)
d2l.plt.vlines(a - 4 * np.sqrt(2 * p), 0.53, 0.47, 'black', lw=1)
d2l.plt.vlines(a + 4 * np.sqrt(2 * p), 0.53, 0.47, 'black', lw=1)
d2l.plt.title(f'p = {p:.3f}')
d2l.plt.show()
# Plot interval when p > 1/8

plot_chebyshev(0.0, 0.2)
The second shows that at p = 1/8, the interval exactly touches the two points. This shows that the
inequality is sharp, since no smaller interval could be taken while keeping the inequality true.
# Plot interval when p = 1/8


The third shows that for p < 1/8 the interval only contains the center. This does not invalidate the
inequality since we only needed to ensure that no more than 1/4 of the probability falls outside
the interval, which means that once p < 1/8, the two points at a − 2 and a + 2 can be discarded.
# Plot interval when p < 1/8

Means and Variances in the Continuum
This has all been in terms of discrete random variables, but the case of continuous random vari-
ables is similar. To intuitively understand how this works, imagine that we split the real number
line into intervals of length ϵ given by (ϵi, ϵ(i + 1)]. Once we do this, our continuous random vari-
able has been made discrete and we can use (18.6.14) say that
∑
µX ≈ (ϵi)P (X ∈ (ϵi, ϵ(i + 1)])
i
∑ (18.6.22)
≈ (ϵi)pX (ϵi)ϵ,
i

where pX is the density of X. This is an approximation to the integral of xpX (x), so we can con-
clude that
∫ ∞
µX = xpX (x) dx. (18.6.23)
−∞
Similarly, using (18.6.16) the variance can be written as

∫ ∞ (∫ ∞ )2
2
σX = E[X ] −
2
µ2X = x pX (x) dx −
2
xpX (x) dx . (18.6.24)
−∞ −∞
Everything stated above about the mean, the variance, and the standard deviation still applies in
this case. For instance, if we consider the random variable with density
{
1 x ∈ [0, 1],
p(x) = (18.6.25)
0 otherwise.
we can compute
∫ ∞ ∫ 1
1
µX = xp(x) dx = x dx = . (18.6.26)
−∞ 0 2
and
∫ ∞ ( )2
1 1 1 1
2
σX = x p(x) dx −
2
= − = . (18.6.27)
−∞ 2 3 4 12
As a warning, let us examine one more example, known as the Cauchy distribution. This is the
distribution with p.d.f. given by
1
p(x) = . (18.6.28)
1 + x2
# Plot the Cauchy distribution p.d.f.

x = np.arange(-5, 5, 0.01)
p = 1 / (1 + x**2)
d2l.plot(x, p, 'x', 'p.d.f.')

This function looks innocent, and indeed consulting a table of integrals will show it has area one
under it, and thus it defines a continuous random variable.
To see what goes astray, let us try to compute the variance of this. This would involve using
(18.6.16) computing
∫ ∞
x2
2
dx. (18.6.29)
−∞ 1 + x
The function on the inside looks like this:
# Plot the integrand needed to compute the variance

x = np.arange(-20, 20, 0.01)
p = x**2 / (1 + x**2)
d2l.plot(x, p, 'x', 'integrand')
This function clearly has infinite area under it since it is essentially the constant one with a small
dip near zero, and indeed we could show that
∫ ∞
x2
2
dx = ∞. (18.6.30)
−∞ 1 + x
This means it does not have a well-defined finite variance.

However, looking deeper shows an even more disturbing result. Let us try to compute the mean
using (18.6.14). Using the change of variables formula, we see
∫ ∞ ∫
x 1 ∞1
µX = 2
dx = du. (18.6.31)
−∞ 1 + x 2 1 u
The integral inside is the definition of the logarithm, so this is in essence log(∞) = ∞, so there is
no well-defined average value either!
Machine learning scientists define their models so that we most often do not need to deal with
these issues, and will in the vast majority of cases deal with random variables with well-defined
means and variances. However, every so often random variables with heavy tails (that is those
random variables where the probabilities of getting large values are large enough to make things
like the mean or variance undefined) are helpful in modeling physical systems, thus it is worth
knowing that they exist.

Joint Density Functions
The above work all assumes we are working with a single real valued random variable. But what
if we are dealing with two or more potentially highly correlated random variables? This circum-
stance is the norm in machine learning: imagine random variables like Ri,j which encode the red
value of the pixel at the (i, j) coordinate in an image, or Pt which is a random variable given by
a stock price at time t. Nearby pixels tend to have similar color, and nearby times tend to have
similar prices. We cannot treat them as separate random variables, and expect to create a success-
ful model (we will see in Section 18.9 a model that under-performs due to such an assumption).
We need to develop the mathematical language to handle these correlated continuous random
variables.
Thankfully, with the multiple integrals in Section 18.5 we can develop such a language. Suppose
that we have, for simplicity, two random variables X, Y which can be correlated. Then, similar to
the case of a single variable, we can ask the question:
P (X is in an ϵ-sized interval around x and Y is in an ϵ-sized interval around y). (18.6.32)
Similar reasoning to the single variable case shows that this should be approximately
P (X is in an ϵ-sized interval around x and Y is in an ϵ-sized interval around y) ≈ ϵ2 p(x, y),

(18.6.33)
for some function p(x, y). This is referred to as the joint density of X and Y . Similar properties
are true for this as we saw in the single variable case. Namely:
• p(x, y) ≥ 0;
∫
• R2 p(x, y) dx dy = 1;
∫
• P ((X, Y ) ∈ D) = D p(x, y) dx dy.
In this way, we can deal with multiple, potentially correlated random variables. If we wish to
work with more than two random variables, we can extend the multivariate density to as many
coordinates as desired by considering p(x) = p(x1 , . . . , xn ). The same properties of being non-
negative, and having total integral of one still hold.
Marginal Distributions
When dealing with multiple variables, we oftentimes want to be able to ignore the relationships
and ask, “how is this one variable distributed?” Such a distribution is called a marginal distribution.
To be concrete, let us suppose that we have two random variables X, Y with joint density given by
pX,Y (x, y). We will be using the subscript to indicate what random variables the density is for. The
question of finding the marginal distribution is taking this function, and using it to find pX (x).
As with most things, it is best to return to the intuitive picture to figure out what should be true.
Recall that the density is the function pX so that
P (X ∈ [x, x + ϵ]) ≈ ϵ · pX (x). (18.6.34)
There is no mention of Y , but if all we are given is pX,Y , we need to include Y somehow. We can
first observe that this is the same as
P (X ∈ [x, x + ϵ], and Y ∈ R) ≈ ϵ · pX (x). (18.6.35)

Our density does not directly tell us about what happens in this case, we need to split into small
intervals in y as well, so we can write this as
∑
ϵ · pX (x) ≈ P (X ∈ [x, x + ϵ], and Y ∈ [ϵ · i, ϵ · (i + 1)])
i
∑ (18.6.36)
≈ ϵ2 pX,Y (x, ϵ · i).
i
Fig. 18.6.1: By summing along the columns of our array of probabilities, we are able to obtain the
marginal distribution for just the random variable represented along the x-axis.
This tells us to add up the value of the density along a series of squares in a line as is shown in Fig.
18.6.1. Indeed, after canceling one factor of epsilon from both sides, and recognizing the sum on
the right is the integral over y, we can conclude that
∑
pX (x) ≈ ϵpX,Y (x, ϵ · i)
∫i ∞ (18.6.37)
≈ pX,Y (x, y) dy.
−∞
Thus we see
∫ ∞
pX (x) = pX,Y (x, y) dy. (18.6.38)
−∞
This tells us that to get a marginal distribution, we integrate over the variables we do not care
about. This process is often referred to as integrating out or marginalized out the unneeded vari-
ables.
Covariance
When dealing with multiple random variables, there is one additional summary statistic which
is helpful to know: the covariance. This measures the degree that two random variable fluctuate
together.
Suppose that we have two random variables X and Y , to begin with, let us suppose they are dis-
crete, taking on values (xi , yj ) with probability pij . In this case, the covariance is defined as
∑
σXY = Cov(X, Y ) = (xi − µX )(yj − µY )pij . = E[XY ] − E[X]E[Y ]. (18.6.39)
i,j

To think about this intuitively: consider the following pair of random variables. Suppose that X
takes the values 1 and 3, and Y takes the values −1 and 3. Suppose that we have the following
probabilities
p
P (X = 1 and Y = −1) = ,
2
1−p
P (X = 1 and Y = 3) = ,
2 (18.6.40)
1−p
P (X = 3 and Y = −1) = ,
2
p
P (X = 3 and Y = 3) = ,
2
where p is a parameter in [0, 1] we get to pick. Notice that if p = 1 then they are both always
their minimum or maximum values simultaneously, and if p = 0 they are guaranteed to take their
flipped values simultaneously (one is large when the other is small and vice versa). If p = 1/2,
then the four possibilities are all equally likely, and neither should be related. Let us compute the
covariance. First, note µX = 2 and µY = 1, so we may compute using (18.6.39):
∑
Cov(X, Y ) = (xi − µX )(yj − µY )pij
i,j
p 1−p 1−p p
= (1 − 2)(−1 − 1) + (1 − 2)(3 − 1) + (3 − 2)(−1 − 1) + (3 − 2)(3 − 1)
2 2 2 2
= 4p − 2.
(18.6.41)
When p = 1 (the case where they are both maximally positive or negative at the same time) has a
covariance of 2. When p = 0 (the case where they are flipped) the covariance is −2. Finally, when
p = 1/2 (the case where they are unrelated), the covariance is 0. Thus we see that the covariance
measures how these two random variables are related.
A quick note on the covariance is that it only measures these linear relationships. More complex
relationships like X = Y 2 where Y is randomly chosen from {−2, −1, 0, 1, 2} with equal probabil-
ity can be missed. Indeed a quick computation shows that these random variables have covariance
zero, despite one being a deterministic function of the other.
For continuous random variables, much the same story holds. At this point, we are pretty comfort-
able with doing the transition between discrete and continuous, so we will provide the continuous
analogue of (18.6.39) without any derivation.
∫
σXY = (x − µX )(y − µY )p(x, y) dx dy. (18.6.42)
R2
For visualization, let us take a look at a collection of random variables with tunable covariance.
# Plot a few random variables adjustable covariance

covs = [-0.9, 0.0, 1.2]
d2l.plt.figure(figsize=(12, 3))
for i in range(3):
X = np.random.normal(0, 1, 500)
Y = covs[i]*X + np.random.normal(0, 1, (500))
d2l.plt.subplot(1, 4, i+1)
d2l.plt.scatter(X.asnumpy(), Y.asnumpy())
d2l.plt.xlabel('X')

d2l.plt.ylabel('Y')
d2l.plt.title(f'cov = {covs[i]}')
d2l.plt.show()
Let us see some properties of covariances:

• For any random variable X, Cov(X, X) = Var(X).
• For any random variables X, Y and numbers a and b, Cov(aX + b, Y ) = Cov(X, aY + b) =
aCov(X, Y ).
• If X and Y are independent then Cov(X, Y ) = 0.
In addition, we can use the covariance to expand a relationship we saw before. Recall that is X
and Y are two independent random variables then
Var(X + Y ) = Var(X) + Var(Y ). (18.6.43)
With knowledge of covariances, we can expand this relationship. Indeed, some algebra can show
that in general,
Var(X + Y ) = Var(X) + Var(Y ) + 2Cov(X, Y ). (18.6.44)
This allows us to generalize the variance summation rule for correlated random variables.
Correlation
As we did in the case of means and variances, let us now consider units. If X is measured in one
unit (say inches), and Y is measured in another (say dollars), the covariance is measured in the
product of these two units inches × dollars. These units can be hard to interpret. What we will
often want in this case is a unit-less measurement of relatedness. Indeed, often we do not care
about exact quantitative correlation, but rather ask if the correlation is in the same direction, and
how strong the relationship is.
To see what makes sense, let us perform a thought experiment. Suppose that we convert our ran-
dom variables in inches and dollars to be in inches and cents. In this case the random variable Y is

multiplied by 100. If we work through the definition, this means that Cov(X, Y ) will be multiplied
by 100. Thus we see that in this case a change of units change the covariance by a factor of 100.
Thus, to find our unit-invariant measure of correlation, we will need to divide by something else
that also gets scaled by 100. Indeed we have a clear candidate, the standard deviation! Indeed if
we define the correlation coefficient to be
Cov(X, Y )
ρ(X, Y ) = , (18.6.45)
σX σY
we see that this is a unit-less value. A little mathematics can show that this number is between −1
and 1 with 1 meaning maximally positively correlated, whereas −1 means maximally negatively
correlated.
Returning to our explicit discrete example above, we can see that σX = 1 and σY = 2, so we can
compute the correlation between the two random variables using (18.6.45) to see that
4p − 2
ρ(X, Y ) = = 2p − 1. (18.6.46)
1·2
This now ranges between −1 and 1 with the expected behavior of 1 meaning most correlated, and
−1 meaning minimally correlated.
As another example, consider X as any random variable, and Y = aX + b as any linear determin-
istic function of X. Then, one can compute that
σY = σaX+b = |a|σX , (18.6.47)
Cov(X, Y ) = Cov(X, aX + b) = aCov(X, X) = aVar(X), (18.6.48)

and thus by (18.6.45) that
aVar(X) a
ρ(X, Y ) = = = sign(a). (18.6.49)
|a|σX
2 |a|
Thus we see that the correlation is +1 for any a > 0, and −1 for any a < 0 illustrating that corre-
lation measures the degree and directionality the two random variables are related, not the scale
that the variation takes.
Let us again plot a collection of random variables with tunable correlation.
# Plot a few random variables adjustable correlations

cors = [-0.9, 0.0, 1.0]
for i in range(3):
X = np.random.normal(0, 1, 500)
Y = cors[i] * X + np.sqrt(1 - cors[i]**2) * np.random.normal(0, 1, 500)
d2l.plt.subplot(1, 4, i + 1)
d2l.plt.scatter(X.asnumpy(), Y.asnumpy())
d2l.plt.xlabel('X')
d2l.plt.ylabel('Y')
d2l.plt.title(f'cor = {cors[i]}')
d2l.plt.show()

Let us list a few properties of the correlation below.
• For any random variable X, ρ(X, X) = 1.
• For any random variables X, Y and numbers a and b, ρ(aX + b, Y ) = ρ(X, aY + b) = ρ(X, Y ).
• If X and Y are independent with non-zero variance then ρ(X, Y ) = 0.
As a final note, you may feel like some of these formulae are familiar. Indeed, if we expand ev-
erything out assuming that µX = µY = 0, we see that this is
∑
i,j xi yi pij
ρ(X, Y ) = √∑ √∑ . (18.6.50)
x 2p y 2p
i,j i ij i,j j ij
This looks like a sum of a product of terms divided by the square root of sums of terms. This is
exactly the formula for the cosine of the angle between two vectors v, w with the different coordi-
nates weighted by pij :
∑
v·w v i wi
cos(θ) = = √∑ i √∑ . (18.6.51)
∥v∥∥w∥ v 2 w 2
i i i i
Indeed if we think of norms as being related to standard deviations, and correlations as being
cosines of angles, much of the intuition we have from geometry can be applied to thinking about
random variables.
Summary
• Continuous random variables are random variables that can take on a continuum of values.
They have some technical difficulties that make them more challenging to work with com-
pared to discrete random variables.
• The probability density function allows us to work with continuous random variables by
giving a function where the area under the curve on some interval gives the probability of
finding a sample point in that interval.
• The cumulative distribution function is the probability of observing the random variable to
be less than a given threshold. It can provide a useful alternate viewpoint which unifies
discrete and continuous variables.

• The mean is the average value of a random variable.
• The variance is the expected square of the difference between the random variable and its
mean.
• The standard deviation is the square root of the variance. It can be thought of as measuring
the range of values the random variable may take.
• Chebyshevʼs inequality allows us to make this intuition rigorous by giving an explicit interval
that contains the random variable most of the time.
• Joint densities allow us to work with correlated random variables. We may marginalize joint
densities by integrating over unwanted random variables to get the distribution of the de-
sired random variable.
• The covariance and correlation coefficient provide a way to measure any linear relationship
between two correlated random variables.
Exercises
1. Suppose that we have the random variable with density given by p(x) = 1
x2
for x ≥ 1 and
p(x) = 0 otherwise. What is P (X > 2)?
2. The Laplace distribution is a random variable whose density is given by = 12 e−|x| . What
∫ ∞p(x −x
∫is∞the2 mean and the standard deviation of this function? As a hint, 0 xe
−x dx = 2.
dx = 1 and
0 x e
3. I walk up to you on the street and say “I have a random variable with mean 1, standard devi-
ation 2, and I observed 25% of my samples taking a value larger than 9.” Do you believe me?
Why or why not?
4. Suppose that you have two random variables X, Y , with joint density given by pXY (x, y) =
4xy for x, y ∈ [0, 1] and pXY (x, y) = 0 otherwise. What is the covariance of X and Y ?
Discussions250
18.7 Maximum Likelihood
One of the most commonly encountered way of thinking in machine learning is the maximum
likelihood point of view. This is the concept that when working with a probabilistic model with
unknown parameters, the parameters which make the data have the highest probability are the
most likely ones.
250
18.7. Maximum Likelihood 855

18.7.1 The Maximum Likelihood Principle
This has a Bayesian interpretation which can be helpful to think about. Suppose that we have a
model with parameters θ and a collection of data examples X. For concreteness, we can imagine
that θ is a single value representing the probability that a coin comes up heads when flipped, and
X is a sequence of independent coin flips. We will look at this example in depth later.
If we want to find the most likely value for the parameters of our model, that means we want to
find
argmax P (θ | X). (18.7.1)
By Bayesʼ rule, this is the same thing as
P (X | θ)P (θ)
argmax . (18.7.2)
P (X)
The expression P (X), a parameter agnostic probability of generating the data, does not depend
on θ at all, and so can be dropped without changing the best choice of θ. Similarly, we may now
posit that we have no prior assumption on which set of parameters are better than any others, so
we may declare that P (θ) does not depend on theta either! This, for instance, makes sense in our
coin flipping example where the probability it comes up heads could be any value in [0, 1] without
any prior belief it is fair or not (often referred to as an uninformative prior). Thus we see that our
application of Bayesʼ rule shows that our best choice of θ is the maximum likelihood estimate for
θ:
θ̂ = argmax P (X | θ). (18.7.3)

θ
As a matter of common terminology, the probability of the data given the parameters (P (X | θ))
is referred to as the likelihood.
A Concrete Example
Let us see how this works in a concrete example. Suppose that we have a single parameter θ
representing the probability that a coin flip is heads. Then the probability of getting a tails is 1 − θ,
and so if our observed data X is a sequence with nH heads and nT tails, we can use the fact that
independent probabilities multiply to see that
P (X | θ) = θnH (1 − θ)nT . (18.7.4)
If we flip 13 coins and get the sequence “HHHTHTTHHHHHT”, which has nH = 9 and nT = 4, we
see that this is
P (X | θ) = θ9 (1 − θ)4 . (18.7.5)
One nice thing about this example will be that we know the answer going in. Indeed, if we said
verbally, “I flipped 13 coins, and 9 came up heads, what is our best guess for the probability that
the coin comes us heads?,” everyone would correctly guess 9/13. What this maximum likelihood
method will give us is a way to get that number from first principals in a way that will generalize
to vastly more complex situations.
For our example, the plot of P (X | θ) is as follows:

%matplotlib inline
npx.set_np()
theta = np.arange(0, 1, 0.001)

p = theta**9 * (1 - theta)**4.
d2l.plot(theta, p, 'theta', 'likelihood')
This has its maximum value somewhere near our expected 9/13 ≈ 0.7 . . .. To see if it is exactly
there, we can turn to calculus. Notice that at the maximum, the function is flat. Thus, we could
find the maximum likelihood estimate (18.7.1) by finding the values of θ where the derivative is
zero, and finding the one that gives the highest probability. We compute:
d
0= P (X | θ)
dθ
d 9
= θ (1 − θ)4 (18.7.6)
dθ
= 9θ8 (1 − θ)4 − 4θ9 (1 − θ)3
= θ8 (1 − θ)3 (9 − 13θ).
This has three solutions: 0, 1 and 9/13. The first two are clearly minima, not maxima as they assign
probability 0 to our sequence. The final value does not assign zero probability to our sequence,
and thus must be the maximum likelihood estimate θ̂ = 9/13.
18.7.2 Numerical Optimization and the Negative Log-Likelihood
The previous example is nice, but what if we have billions of parameters and data examples.
First notice that, if we make the assumption that all the data examples are independent, we can
no longer practically consider the likelihood itself as it is a product of many probabilities. Indeed,
each probability is in [0, 1], say typically of value about 1/2, and the product of (1/2)1000000000 is far
below machine precision. We cannot work with that directly.
However, recall that the logarithm turns products to sums, in which case
log((1/2)1000000000 ) = 1000000000 · log(1/2) ≈ −301029995.6 . . . (18.7.7)

This number fits perfectly within even a single precision 32-bit float. Thus, we should consider
the log-likelihood, which is
log(P (X | θ)). (18.7.8)
Since the function x 7→ log(x) is increasing, maximizing the likelihood is the same thing as maxi-
mizing the log-likelihood. Indeed in Section 18.9 we will see this reasoning applied when working
with the specific example of the naive Bayes classifier.
We often work with loss functions, where we wish to minimize the loss. We may turn maximum
likelihood into the minimization of a loss by taking − log(P (X | θ)), which is the negative log-
likelihood.
To illustrate this, consider the coin flipping problem from before, and pretend that we do not know
the closed form solution. We may compute that
− log(P (X | θ)) = − log(θnH (1 − θ)nT ) = −(nH log(θ) + nT log(1 − θ)). (18.7.9)
This can be written into code, and freely optimized even for billions of coin flips.
# Set up our data

n_H = 8675309
n_T = 25624
# Initialize our paramteres

theta = np.array(0.5)
theta.attach_grad()
# Perform gradient descent

lr = 0.00000000001
for iter in range(10):
loss = -(n_H * np.log(theta) + n_T * np.log(1 - theta))
loss.backward()
theta -= lr * theta.grad
# Check output
theta, n_H / (n_H + n_T)
(array(0.50172704), 0.9970550284664874)
Numerical convenience is only one reason people like to use negative log-likelihoods. Indeed,
there are a several reasons that it can be preferable.
The second reason we consider the log-likelihood is the simplified application of calculus rules. As
discussed above, due to independence assumptions, most probabilities we encounter in machine
learning are products of individual probabilities.
P (X | θ) = p(x1 | θ) · p(x2 | θ) · · · p(xn | θ). (18.7.10)

This means that if we directly apply the product rule to compute a derivative we get
( )
∂ ∂
P (X | θ) = P (x1 | θ) · P (x2 | θ) · · · P (xn | θ)
∂θ ∂θ
( )
∂
+ P (x1 | θ) · P (x2 | θ) · · · P (xn | θ)
∂θ (18.7.11)
..
.
( )
∂
+ P (x1 | θ) · P (x2 | θ) · · · P (xn | θ) .
∂θ
This requires n(n − 1) multiplications, along with (n − 1) additions, so it is total of quadratic time
in the inputs! Sufficient cleverness in grouping terms will reduce this to linear time, but it requires
some thought. For the negative log-likelihood we have instead
− log (P (X | θ)) = − log(P (x1 | θ)) − log(P (x2 | θ)) · · · − log(P (xn | θ)), (18.7.12)
which then gives

( ) ( )
∂ 1 ∂ 1 ∂
− log (P (X | θ)) = P (x1 | θ) + · · · + P (xn | θ) . (18.7.13)
∂θ P (x1 | θ) ∂θ P (xn | θ) ∂θ
This requires only n divides and n − 1 sums, and thus is linear time in the inputs.
The third and final reason to consider the negative log-likelihood is the relationship to information
theory, which we will discuss in detail in Section 18.11. This is a rigorous mathematical theory
which gives a way to measure the degree of information or randomness in a random variable.
The key object of study in that field is the entropy which is
∑
H(p) = − pi log2 (pi ), (18.7.14)
i
which measures the randomness of a source. Notice that this is nothing more than the average
− log probability, and thus if we take our negative log-likelihood and divide by the number of data
examples, we get a relative of entropy known as cross-entropy. This theoretical interpretation
alone would be sufficiently compelling to motivate reporting the average negative log-likelihood
over the dataset as a way of measuring model performance.
18.7.3 Maximum Likelihood for Continuous Variables
Everything that we have done so far assumes we are working with discrete random variables, but
what if we want to work with continuous ones?
The short summary is that nothing at all changes, except we replace all the instances of the proba-
bility with the probability density. Recalling that we write densities with lower case p, this means
that for example we now say
∑
− log (p(X | θ)) = − log(p(x1 | θ)) − log(p(x2 | θ)) · · · − log(p(xn | θ)) = − log(p(xi | θ)).
i
(18.7.15)
The question becomes, “Why is this OK?” After all, the reason we introduced densities was because
probabilities of getting specific outcomes themselves was zero, and thus is not the probability of
generating our data for any set of parameters zero?

Indeed, this is the case, and understanding why we can shift to densities is an exercise in tracing
what happens to the epsilons.
Let us first re-define our goal. Suppose that for continuous random variables we no longer want
to compute the probability of getting exactly the right value, but instead matching to within some
range ϵ. For simplicity, we assume our data is repeated observations x1 , . . . , xN of identically dis-
tributed random variables X1 , . . . , XN . As we have seen previously, this can be written as
P (X1 ∈ [x1 , x1 + ϵ], X2 ∈ [x2 , x2 + ϵ], . . . , XN ∈ [xN , xN + ϵ] | θ)

(18.7.16)
≈ϵN p(x1 | θ) · p(x2 | θ) · · · p(xn | θ).
Thus, if we take negative logarithms of this we obtain
− log(P (X1 ∈ [x1 , x1 + ϵ], X2 ∈ [x2 , x2 + ϵ], . . . , XN ∈ [xN , xN + ϵ] | θ))

∑
≈ − N log(ϵ) − log(p(xi | θ)). (18.7.17)
i
If we examine this expression, the only place that the ϵ occurs is in the additive constant −N log(ϵ).
This does not depend on the parameters θ at all, so the optimal choice of θ does not depend on
our choice of ϵ! If we demand four digits or four-hundred, the best choice of θ remains the same,
thus we may freely drop the epsilon to see that what we want to optimize is
∑
− log(p(xi | θ)). (18.7.18)
i
Thus, we see that the maximum likelihood point of view can operate with continuous random
variables as easily as with discrete ones by replacing the probabilities with probability densities.
Summary
• The maximum likelihood principle tells us that the best fit model for a given dataset is the
one that generates the data with the highest probability.
• Often people work with the negative log-likelihood instead for a variety of reasons: numer-
ical stability, conversion of products to sums (and the resulting simplification of gradient
computations), and theoretical ties to information theory.
• While simplest to motivate in the discrete setting, it may be freely generalized to the contin-
uous setting as well by maximizing the probability density assigned to the datapoints.
Exercises
1. Suppose that you know that a random variable has density α1 e−αx for some value α. You
obtain a single observation from the random variable which is the number 3. What is the
maximum likelihood estimate for α?
2. Suppose that you have a dataset of samples {xi }N
i=1 drawn from a Gaussian with unknown
mean, but variance 1. What is the maximum likelihood estimate for the mean?
Discussions251
251

18.8 Distributions
Now that we have learned how to work with probability in both the discrete and the continuous
setting, let us get to know some of the common distributions encountered. Depending on the area
of machine learning, we may need to be familiar with vastly more of these, or for some areas of
deep learning potentially none at all. This is, however, a good basic list to be familiar with. Let us
first import some common libraries.
%matplotlib inline
from math import erf, factorial
import numpy as np
18.8.1 Bernoulli
This is the simplest random variable usually encountered. This random variable encodes a coin
flip which comes up 1 with probability p and 0 with probability 1 − p. If we have a random variable
X with this distribution, we will write
X ∼ Bernoulli(p). (18.8.1)
The cumulative distribution function is



0 x < 0,
F (x) = 1 − p 0 ≤ x < 1, (18.8.2)


1 x >= 1.
The probability mass function is plotted below.
p = 0.3
d2l.set_figsize()
d2l.plt.stem([0, 1], [1 - p, p], use_line_collection=True)
d2l.plt.xlabel('x')
d2l.plt.show()
18.8. Distributions 861

Now, let us plot the cumulative distribution function (18.8.2).
x = np.arange(-1, 2, 0.01)
def F(x):
return 0 if x < 0 else 1 if x > 1 else 1 - p
d2l.plot(x, np.array([F(y) for y in x]), 'x', 'c.d.f.')
If X ∼ Bernoulli(p), then:
• µX = p,
2 = p(1 − p).
• σX
We can sample an array of arbitrary shape from a Bernoulli random variable as follows.
1*(np.random.rand(10, 10) < p)
array([[0, 0, 0, 0, 1, 1, 1, 0, 0, 0],
[1, 0, 0, 1, 0, 0, 0, 1, 0, 0],
[0, 0, 1, 1, 0, 1, 1, 0, 1, 1],
[0, 1, 0, 0, 1, 0, 0, 1, 1, 0],
[0, 1, 0, 1, 0, 1, 0, 0, 1, 1],
[1, 0, 1, 0, 0, 1, 1, 1, 0, 0],
[1, 0, 0, 1, 0, 0, 0, 1, 0, 0],
[0, 0, 0, 0, 0, 0, 1, 1, 0, 0],
[1, 1, 0, 1, 0, 0, 0, 0, 0, 0],
[1, 1, 0, 0, 1, 0, 0, 1, 1, 1]])

18.8.2 Discrete Uniform
The next commonly encountered random variable is a discrete uniform. For our discussion here,
we will assume that it is supported on the integers {1, 2, . . . , n}, however any other set of values
can be freely chosen. The meaning of the word uniform in this context is that every possible value
is equally likely. The probability for each value i ∈ {1, 2, 3, . . . , n} is pi = n1 . We will denote a
random variable X with this distribution as
X ∼ U (n). (18.8.3)



0 x < 1,
F (x) = nk k ≤ x < k + 1 with 1 ≤ k < n, (18.8.4)


1 x >= n.
Let us first plot the probability mass function.
n = 5
d2l.plt.stem([i+1 for i in range(n)], n*[1 / n], use_line_collection=True)

d2l.plt.xlabel('x')
d2l.plt.show()
x = np.arange(-1, 6, 0.01)
def F(x):
return 0 if x < 1 else 1 if x > n else np.floor(x) / n

If X ∼ U (n), then:
1+n
• µX = 2 ,
2 = n2 −1
• σX 12 .
We can sample an array of arbitrary shape from a discrete uniform random variable as follows.
np.random.randint(1, n, size=(10, 10))
array([[1, 4, 3, 2, 1, 1, 3, 1, 2, 1],
[1, 3, 2, 3, 3, 2, 1, 4, 2, 3],
[4, 4, 1, 4, 3, 2, 1, 4, 3, 3],
[4, 4, 1, 4, 3, 3, 3, 2, 1, 4],
[3, 4, 3, 1, 2, 2, 2, 4, 2, 1],
[2, 1, 1, 2, 3, 3, 4, 1, 1, 4],
[1, 1, 1, 4, 2, 1, 2, 1, 3, 1],
[2, 2, 3, 2, 1, 4, 1, 3, 4, 2],
[2, 4, 1, 3, 4, 3, 1, 4, 4, 3],
[3, 4, 3, 3, 4, 4, 1, 4, 2, 2]])
18.8.3 Continuous Uniform
Next, let us discuss the continuous uniform distribution. The idea behind this random variable
is that if we increase the n in the discrete uniform distribution, and then scale it to fit within the
interval [a, b], we will approach a continuous random variable that just picks an arbitrary value in
[a, b] all with equal probability. We will denote this distribution as
X ∼ U (a, b). (18.8.5)
The probability density function is

{
1
b−a x ∈ [a, b],
p(x) = (18.8.6)
0 x ̸∈ [a, b].



0 x < a,
F (x) = x−a
x ∈ [a, b], (18.8.7)
 b−a

1 x >= b.
Let us first plot the probability density function (18.8.6).
a, b = 1, 3
x = np.arange(0, 4, 0.01)
p = (x > a)*(x < b)/(b - a)
def F(x):
return 0 if x < a else 1 if x > b else (x - a) / (b - a)
If X ∼ U (a, b), then:

a+b
• µX = 2 ,
2 = (b−a)2
• σX 12 .
We can sample an array of arbitrary shape from a uniform random variable as follows. Note that
it by default samples from a U (0, 1), so if we want a different range we need to scale it.
(b - a) * np.random.rand(10, 10) + a
array([[1.4263995 , 1.32234442, 1.98789028, 2.62305405, 1.81802665,

1.7479621 , 1.86588587, 1.43083642, 2.0871248 , 2.16239639],
[1.18332507, 2.67369682, 1.507526 , 2.41555196, 1.02998459,
2.33625933, 1.20203081, 2.08790005, 2.01575794, 1.82344995],
[2.89199317, 2.14589118, 1.85481229, 1.33449574, 2.93617602,
2.98310547, 2.96235126, 2.14785852, 2.77316042, 2.62551597],
[2.14243349, 2.9940622 , 1.19788307, 1.71621293, 1.32188867,
2.61265492, 2.15860731, 2.55533404, 1.75299553, 1.83622521],
[2.92449044, 2.28884528, 1.94550309, 2.76315966, 2.10196016,
2.51960426, 1.56423267, 1.5574463 , 1.70708128, 2.53040129],
[2.22330704, 1.77008105, 2.36421772, 2.7777458 , 2.25120004,
1.35396632, 1.9726549 , 2.27340037, 1.74838093, 2.92967289],
[2.78088808, 1.12106631, 1.93326138, 1.34366358, 2.78746558,
1.73978549, 1.59001827, 2.29027882, 2.25947092, 1.37424085],
[2.22085405, 1.0699301 , 1.46896539, 1.91263425, 1.97568991,
2.16305974, 1.88500816, 2.56292776, 2.62649923, 2.2548305 ],
[2.87977449, 2.584483 , 2.74738304, 1.64931996, 1.90518891,
2.04801046, 1.67777837, 2.35125907, 2.55353478, 1.69605429],
[2.97503039, 2.78367052, 1.48010665, 1.36524376, 2.45156427,
2.15156029, 2.13933111, 2.12242048, 1.04232867, 2.92524385]])
18.8.4 Binomial
Let us make things a little more complex and examine the binomial random variable. This random
variable originates from performing a sequence of n independent experiments, each of which has
probability p of succeeding, and asking how many successes we expect to see.
Let us express this mathematically. Each experiment is an independent random variable Xi where
we will use 1 to encode success, and 0 to encode failure. Since each is an independent coin flip
which is successful with probability p, we can say that Xi ∼ Bernoulli(p). Then, the binomial
random variable is
∑n
X= Xi . (18.8.8)
i=1
In this case, we will write
X ∼ Binomial(n, p). (18.8.9)
To get the( cumulative

) distribution function, we need to notice that getting exactly k successes can
n
occur in k = k!(n−k)! ways each of which has a probability of pk (1 − p)n−k of occurring. Thus the
n!
cumulative distribution function is



0∑ x < 0,
(n) m
m≤k m p (1 − p) k ≤ x < k + 1 with 0 ≤ k < n,
F (x) = n−m (18.8.10)


1 x >= n.

Let us first plot the probability mass function.
n, p = 10, 0.2
# Compute binomial coefficient

def binom(n, k):
comb = 1
for i in range(min(k, n - k)):
comb = comb * (n - i) // (i + 1)
return comb
pmf = np.array([p**i * (1-p)**(n - i) * binom(n, i) for i in range(n + 1)])
d2l.plt.stem([i for i in range(n + 1)], pmf, use_line_collection=True)

d2l.plt.xlabel('x')
d2l.plt.show()
x = np.arange(-1, 11, 0.01)

cmf = np.cumsum(pmf)
def F(x):
return 0 if x < 0 else 1 if x > n else cmf[int(x)]
d2l.plot(x, np.array([F(y) for y in x.tolist()]), 'x', 'c.d.f.')

While this result is not simple, the means and variances are. If X ∼ Binomial(n, p), then:
• µX = np,
2 = np(1 − p).
• σX
This can be sampled as follows.
np.random.binomial(n, p, size=(10, 10))
array([[2, 2, 2, 3, 4, 2, 5, 1, 2, 2],
[1, 2, 2, 1, 3, 1, 5, 1, 2, 2],
[1, 2, 0, 0, 3, 2, 0, 1, 3, 0],
[0, 2, 2, 1, 6, 0, 1, 3, 3, 0],
[1, 3, 3, 6, 1, 0, 5, 0, 3, 2],
[1, 1, 1, 3, 3, 2, 3, 3, 2, 2],
[4, 2, 2, 0, 1, 3, 3, 2, 0, 2],
[3, 3, 1, 0, 1, 2, 3, 0, 3, 1],
[5, 1, 2, 0, 0, 1, 2, 1, 0, 2],
[0, 2, 2, 1, 0, 2, 2, 2, 1, 1]])
18.8.5 Poisson
Let us now perform a thought experiment. We are standing at a bus stop and we want to know how
many buses will arrive in the next minute. Let us start by considering X (1) ∼ Bernoulli(p) which
is simply the probability that a bus arrives in the one minute window. For bus stops far from an
urban center, this might be a pretty good approximation. We may never see more than one bus in
a minute.
However, if we are in a busy area, it is possible or even likely that two buses will arrive. We can
model this by splitting our random variable into two parts for the first 30 seconds, or the second
30 seconds. In this case we can write
(2) (2)
X (2) ∼ X1 + X2 , (18.8.11)
(2)
where X (2) is the total sum, and Xi ∼ Bernoulli(p/2). The total distribution is then X (2) ∼
Binomial(2, p/2).

Why stop here? Let us continue to split that minute into n parts. By the same reasoning as above,
we see that
X (n) ∼ Binomial(n, p/n). (18.8.12)
Consider these random variables. By the previous section, we know that (18.8.12) has mean
(n) = n(p/n)(1 − (p/n)) = p(1 − p/n). If we take n → ∞,
µX (n) = n(p/n) = p, and variance σX2
2
we can see that these numbers stabilize to µX (∞) = p, and variance σX (∞) = p. This indicates that
there could be some random variable we can define in this infinite subdivision limit.
This should not come as too much of a surprise, since in the real world we can just count the
number of bus arrivals, however it is nice to see that our mathematical model is well defined.
This discussion can be made formal as the law of rare events.
Following through this reasoning carefully, we can arrive at the following model. We will say that
X ∼ Poisson(λ) if it is a random variable which takes the values {0, 1, 2, . . .} with probability
λk e−λ (18.8.13)
pk = .
k!
The value λ > 0 is known as the rate (or the shape parameter), and denotes the average number of
arrivals we expect in one unit of time.
We may sum this probability mass function to get the cumulative distribution function.
{
0 x < 0,
F (x) = −λ
∑k λm
(18.8.14)
e m=0 m! k ≤ x < k + 1 with 0 ≤ k.
Let us first plot the probability mass function (18.8.13).
lam = 5.0
xs = [i for i in range(20)]
pmf = np.array([np.exp(-lam) * lam**k / factorial(k) for k in xs])
d2l.plt.stem(xs, pmf, use_line_collection=True)

d2l.plt.xlabel('x')
d2l.plt.show()

x = np.arange(-1, 21, 0.01)
cmf = np.cumsum(pmf)
def F(x):
return 0 if x < 0 else 1 if x > n else cmf[int(x)]
d2l.plot(x, np.array([F(y) for y in x.tolist()]), 'x', 'c.d.f.')
As we saw above, the means and variances are particularly concise. If X ∼ Poisson(λ), then:
• µX = λ,
2 = λ.
• σX
This can be sampled as follows.
np.random.poisson(lam, size=(10, 10))
array([[ 8, 4, 3, 6, 6, 11, 2, 4, 2, 8],

[ 6, 4, 7, 3, 4, 1, 5, 6, 4, 1],
[ 4, 6, 6, 3, 9, 3, 5, 6, 4, 5],
[ 5, 5, 6, 3, 9, 2, 4, 12, 5, 7],
[ 6, 5, 6, 3, 4, 4, 5, 2, 0, 7],
[ 4, 3, 1, 8, 2, 5, 7, 2, 8, 7],
[ 4, 6, 9, 10, 3, 5, 2, 3, 7, 5],
[ 5, 4, 3, 8, 4, 3, 4, 5, 2, 3],
[ 8, 7, 6, 2, 3, 1, 2, 5, 2, 7],
[ 8, 4, 7, 4, 6, 5, 4, 7, 5, 5]])

18.8.6 Gaussian
Now Let us try a different, but related experiment. Let us say we again are performing n indepen-
dent Bernoulli(p) measurements Xi . The distribution of the sum of these is X (n) ∼ Binomial(n, p).
Rather than taking a limit as n increases and p decreases, Let us fix p, and then send n → ∞. In
this case µX (n) = np → ∞ and σX 2
(n) = np(1 − p) → ∞, so there is no reason to think this limit
should be well defined.
However, not all hope is lost! Let us just make the mean and variance be well behaved by defining
X (n) − µX (n)
Y (n) = . (18.8.15)
σX (n)
This can be seen to have mean zero and variance one, and so it is plausible to believe that it will
converge to some limiting distribution. If we plot what these distributions look like, we will be-
come even more convinced that it will work.
p = 0.2
ns = [1, 10, 100, 1000]
for i in range(4):
n = ns[i]
pmf = np.array([p**i * (1-p)**(n-i) * binom(n, i) for i in range(n + 1)])
d2l.plt.subplot(1, 4, i + 1)
d2l.plt.stem([(i - n*p)/np.sqrt(n*p*(1 - p)) for i in range(n + 1)], pmf,
use_line_collection=True)
d2l.plt.xlim([-4, 4])
d2l.plt.xlabel('x')
d2l.plt.title("n = {}".format(n))
d2l.plt.show()
One thing to note: compared to the Poisson case, we are now dividing by the standard deviation
which means that we are squeezing the possible outcomes into smaller and smaller areas. This is
an indication that our limit will no longer be discrete, but rather a continuous.
A derivation of what occurs is beyond the scope of this document, but the central limit theorem
states that as n → ∞, this will yield the Gaussian Distribution (or sometimes normal distribution).
More explicitly, for any a, b:
lim P (Y (n) ∈ [a, b]) = P (N (0, 1) ∈ [a, b]), (18.8.16)
n→∞

where we say a random variable is normally distributed with given mean µ and variance σ 2 , writ-
ten X ∼ N (µ, σ 2 ) if X has density
1 (x−µ)2
pX (x) = √ e− 2σ2 . (18.8.17)
2πσ 2
Let us first plot the probability density function (18.8.17).
mu, sigma = 0, 1
x = np.arange(-3, 3, 0.01)
p = 1 / np.sqrt(2 * np.pi * sigma**2) * np.exp(-(x - mu)**2 / (2 * sigma**2))
Now, let us plot the cumulative distribution function. It is beyond the scope of this appendix, but
the Gaussian c.d.f. does not have a closed-form formula in terms of more elementary functions.
We will use erf which provides a way to compute this integral numerically.
def phi(x):
return (1.0 + erf((x - mu) / (sigma * np.sqrt(2)))) / 2.0
d2l.plot(x, np.array([phi(y) for y in x.tolist()]), 'x', 'c.d.f.')

Keen-eyed readers will recognize some of these terms. Indeed, we encountered this integral in
Section 18.5. Indeed we need exactly that computation to see that this pX (x) has total area one
and is thus a valid density.
Our choice of working with coin flips made computations shorter, but nothing about that choice
was fundamental. Indeed, if we take any collection of independent identically distributed random
variables Xi , and form
∑
N
X (N ) = Xi . (18.8.18)
i=1
Then
X (N ) − µX (N )
(18.8.19)
σX (N )
will be approximately Gaussian. There are additional requirements needed to make it work, most
commonly E[X 4 ] < ∞, but the philosophy is clear.
The central limit theorem is the reason that the Gaussian is fundamental to probability, statis-
tics, and machine learning. Whenever we can say that something we measured is a sum of many
small independent contributions, we can assume that the thing being measured will be close to
Gaussian.
There are many more fascinating properties of Gaussians, and we would like to discuss one more
here. The Gaussian is what is known as a maximum entropy distribution. We will get into entropy
more deeply in Section 18.11, however all we need to know at this point is that it is a measure of
randomness. In a rigorous mathematical sense, we can think of the Gaussian as the most ran-
dom choice of random variable with fixed mean and variance. Thus, if we know that our random
variable has some mean and variance, the Gaussian is in a sense the most conservative choice of
distribution we can make.
To close the section, Let us recall that if X ∼ N (µ, σ 2 ), then:
• µX = µ,
2 = σ2.
• σX
We can sample from the Gaussian (or standard normal) distribution as shown below.
np.random.normal(mu, sigma, size=(10, 10))
array([[-0.57855698, 0.92902142, -1.28065851, 0.62486226, 0.29431553,

0.48600285, 0.37483924, 0.27128912, -0.91282888, -1.51005878],
[ 2.21706625, 1.91460794, -0.78900758, -0.28104264, -0.20517255,
-0.09443792, -1.44056502, 0.84557181, 0.87875186, 0.11234779],
[ 0.87110348, -0.54684884, 0.84276367, -0.85837276, 0.51980555,
-0.87309186, -0.68363545, 0.11964016, 0.15145944, -0.91087196],
[ 0.49176061, 0.81555132, 0.45626283, 0.67039046, -0.3744689 ,
-1.85015505, -0.05151674, 0.02833729, -1.21424723, -0.00606646],
[ 0.09238888, 0.89545711, 0.45908982, -0.55859736, -0.09947204,
-0.54391891, -0.63874853, 1.64201487, 0.468613 , 0.20090474],
[ 0.03436811, -0.44335993, -0.2909628 , -1.39664818, 0.61456931,
-1.78495687, 0.89107593, 0.13847772, 1.8170571 , 1.65123838],
[-1.39054039, 1.37435581, 0.56844177, 0.84152274, -0.413917 ,

0.38519174, -1.82993419, 0.94526872, 1.4774141 , 0.70311165],
[-0.89720881, -1.73036467, -0.09530571, -0.44387358, 2.45545522,
0.06529793, 1.07548743, -0.12976989, -1.0027376 , -0.17622635],
[-0.5976888 , -1.36432736, -0.32955077, 0.12645766, 0.09316411,
0.07775715, 0.52900558, 0.39635442, -0.33988681, 1.11368078],
[ 0.62088067, 0.99038013, 1.64819413, -0.39226861, 0.60808831,
-0.82700987, -0.80826492, -0.13359222, -0.86161707, -0.76045281]])
18.8.7 Exponential Family
One shared property for all the distributions listed above is that they all belong to which is known
as the exponential family. The exponential family is a set of distributions whose density can be
expressed in the following form:
( )
p(x|η) = h(x) · exp η ⊤ · T (x) − A(η) (18.8.20)
As this definition can be a little subtle, let us examine it closely.

First, h(x) is known as the underlying measure or the base measure. This can be viewed as an original
choice of measure we are modifying with our exponential weight.
Second, we have the vector η = (η1 , η2 , ..., ηl ) ∈ Rl called the natural parameters or canonical pa-
rameters. These define how the base measure will be modified. The natural parameters enter
into the new measure by taking the dot product of these parameters against some function T (·)
of x = (x1 , x2 , ..., xn ) ∈ Rn and exponentiated. T (x) = (T1 (x), T2 (x), ..., Tl (x)) is called the suffi-
cient statistics for η. This name is used since the information represented by T (x) is sufficient to
calculate the probability density and no other information from the sample xʼs are required.
Third, we have A(η), which is referred to as the cumulant function, which ensures that the above
distribution (18.8.20) integrates to one, i.e.,
[∫ ]
( ⊤ )
A(η) = log h(x) · exp η · T (x) dx . (18.8.21)
To be concrete, let us consider the Gaussian. Assuming that x is an univariate variable, we saw
that it had a density of
1 { −(x − µ)2 }
p(x|µ, σ) = √ exp
2πσ 2 2σ 2
{ (18.8.22)
1 µ 1 ( 1 2 )}
= √ · exp 2 x − 2 x2 − µ + log(σ) .
2π σ 2σ 2σ 2
This matches the definition of the exponential family with:

• underlying measure: h(x) = √12π ,
[ ] [ µ ]
η
• natural parameters: η = 1 = σ12 ,
η2 2σ 2
[ ]
x
• sufficient statistics: T (x) = , and
−x2
η12
• cumulant function: A(η) = 1
2σ 2
µ2 + log(σ) = 4η2 − 12 log(2η2 ).

It is worth noting that the exact choice of each of above terms is somewhat arbitrary. Indeed, the
important feature is that the distribution can be expressed in this form, not the exact form itself.
As we allude to in Section 3.4.5, a widely used technique is to assume that the final output y follows
an exponential family distribution. The exponential family is a common and powerful family of
distributions encountered frequently in machine learning.
Summary
• Bernoulli random variables can be used to model events with a yes/no outcome.
• Discrete uniform distributions model selects from a finite set of possibilities.
• Continuous uniform distributions select from an interval.
• Binomial distributions model a series of Bernoulli random variables, and count the number
of successes.
• Poisson random variables model the arrival of rare events.
• Gaussian random variables model the result of adding a large number of independent ran-
dom variables together.
• All the above distributions belong to exponential family.
Exercises
1. What is the standard deviation of a random variable that is the difference X − Y of two
independent binomial random variables X, Y ∼ Binomial(16, 1/2).
√
2. If we take a Poisson random variable X ∼ Poisson(λ) and consider (X − λ)/ λ as λ → ∞,
we can show that this becomes approximately Gaussian. Why does this make sense?
3. What is the probability mass function for a sum of two discrete uniform random variables
on n elements?
Discussions252
18.9 Naive Bayes
Throughout the previous sections, we learned about the theory of probability and random vari-
ables. To put this theory to work, let us introduce the naive Bayes classifier. This uses nothing but
probabilistic fundamentals to allow us to perform classification of digits.
Learning is all about making assumptions. If we want to classify a new data example that we have
never seen before we have to make some assumptions about which data examples are similar to
each other. The naive Bayes classifier, a popular and remarkably clear algorithm, assumes all
features are independent from each other to simplify the computation. In this section, we will
apply this model to recognize characters in images.
252
18.9. Naive Bayes 875

%matplotlib inline
import math
npx.set_np()
18.9.1 Optical Character Recognition
MNIST (LeCun et al., 1998) is one of widely used datasets. It contains 60,000 images for training
and 10,000 images for validation. Each image contains a handwritten digit from 0 to 9. The task is
classifying each image into the corresponding digit.
Gluon provides a MNIST class in the data.vision module to automatically retrieve the dataset from
the Internet. Subsequently, Gluon will use the already-downloaded local copy. We specify whether
we are requesting the training set or the test set by setting the value of the parameter train to
True or False, respectively. Each image is a grayscale image with both width and height of 28 with
shape (28,28,1). We use a customized transformation to remove the last channel dimension. In
addition, the dataset represents each pixel by an unsigned 8-bit integer. We quantize them into
binary features to simplify the problem.
def transform(data, label):

return np.floor(data.astype('float32') / 128).squeeze(axis=-1), label
mnist_train = gluon.data.vision.MNIST(train=True, transform=transform)

mnist_test = gluon.data.vision.MNIST(train=False, transform=transform)
We can access a particular example, which contains the image and the corresponding label.
image, label = mnist_train[2]

image.shape, label
((28, 28), array(4, dtype=int32))
Our example, stored here in the variable image, corresponds to an image with a height and width
of 28 pixels.
image.shape, image.dtype
((28, 28), dtype('float32'))
Our code stores the label of each image as a scalar. Its type is a 32-bit integer.
label, type(label), label.dtype
(array(4, dtype=int32), mxnet.numpy.ndarray, dtype('int32'))
We can also access multiple examples at the same time.

images, labels = mnist_train[10:38]
images.shape, labels.shape
((28, 28, 28), (28,))
Let us visualize these examples.
d2l.show_images(images, 2, 9);
18.9.2 The Probabilistic Model for Classification
In a classification task, we map an example into a category. Here an example is a grayscale 28 × 28

image, and a category is a digit. (Refer to Section 3.4 for a more detailed explanation.) One natural
way to express the classification task is via the probabilistic question: what is the most likely label
given the features (i.e., image pixels)? Denote by x ∈ Rd the features of the example and y ∈ R the
label. Here features are image pixels, where we can reshape a 2-dimensional image to a vector so
that d = 282 = 784, and labels are digits. The probability of the label given the features is p(y | x).
If we are able to compute these probabilities, which are p(y | x) for y = 0, . . . , 9 in our example,
then the classifier will output the prediction ŷ given by the expression:
ŷ = argmax p(y | x). (18.9.1)
Unfortunately, this requires that we estimate p(y | x) for every value of x = x1 , ..., xd . Imagine
that each feature could take one of 2 values. For example, the feature x1 = 1 might signify that
the word apple appears in a given document and x1 = 0 would signify that it does not. If we had
30 such binary features, that would mean that we need to be prepared to classify any of 230 (over
1 billion!) possible values of the input vector x.
Moreover, where is the learning? If we need to see every single possible example in order to predict
the corresponding label then we are not really learning a pattern but just memorizing the dataset.
18.9.3 The Naive Bayes Classifier
Fortunately, by making some assumptions about conditional independence, we can introduce

some inductive bias and build a model capable of generalizing from a comparatively modest se-
lection of training examples. To begin, let us use Bayes theorem, to express the classifier as
p(x | y)p(y)
ŷ = argmaxy p(y | x) = argmaxy . (18.9.2)
p(x)

Note that the denominator is the normalizing term p(x) which does not depend on the value of
the label y. As a result, we only need to worry about comparing the numerator across different
values of y. Even if calculating the denominator turned out to be intractable, we could get away
with ignoring it, so long as we could evaluate the numerator. Fortunately, even if we wanted to
recover
∑ the normalizing constant, we could. We can always recover the normalization term since
y p(y | x) = 1.
Now, let us focus on p(x | y). Using the chain rule of probability, we can express the term p(x | y)
as
p(x1 | y) · p(x2 | x1 , y) · ... · p(xd | x1 , ..., xd−1 , y). (18.9.3)
By itself, this expression does not get us any further. We still must estimate roughly 2d parameters.
However, if we assume that the features are conditionally independent of∏ each other, given the label,
then suddenly we are in much better shape, as this term simplifies to i p(xi | y), giving us the
predictor
∏
d
ŷ = argmaxy p(xi | y)p(y). (18.9.4)
i=1
∏
If we can estimate i p(xi = 1 | y) for every i and y, and save its value in Pxy [i, y], here Pxy is a
d × n matrix with n being the number of classes and y ∈ {1, . . . , n}. In addition, we estimate p(y)
for every y and save it in Py [y], with Py a n-length vector. Then for any new example x, we could
compute
∏
d
ŷ = argmaxy Pxy [xi , y]Py [y], (18.9.5)
i=1
for any y. So our assumption of conditional independence has taken the complexity of our model
from an exponential dependence on the number of features O(2d n) to a linear dependence, which
is O(dn).
18.9.4 Training
The problem now is that we do not know Pxy and Py . So we need to estimate their values given
some training data first. This is training the model. Estimating Py is not too hard. Since we are
only dealing with 10 classes, we may count the number of occurrences ny for each of the digits
and divide it by the total amount of data n. For instance, if digit 8 occurs n8 = 5, 800 times and we
have a total of n = 60, 000 images, the probability estimate is p(y = 8) = 0.0967.
X, Y = mnist_train[:] # All training examples
n_y = np.zeros((10))
for y in range(10):
n_y[y] = (Y == y).sum()
P_y = n_y / n_y.sum()
P_y
array([0.09871667, 0.11236667, 0.0993 , 0.10218333, 0.09736667,

0.09035 , 0.09863333, 0.10441667, 0.09751666, 0.09915 ])

Now on to slightly more difficult things Pxy . Since we picked black and white images, p(xi | y)
denotes the probability that pixel i is switched on for class y. Just like before we can go and count
the number of times niy such that an event occurs and divide it by the total number of occurrences
of y, i.e., ny . But there is something slightly troubling: certain pixels may never be black (e.g., for
well cropped images the corner pixels might always be white). A convenient way for statisticians
to deal with this problem is to add pseudo counts to all occurrences. Hence, rather than niy we
use niy + 1 and instead of ny we use ny + 1. This is also called Laplace Smoothing. It may seem
ad-hoc, however it may be well motivated from a Bayesian point-of-view.
n_x = np.zeros((10, 28, 28))

for y in range(10):
n_x[y] = np.array(X.asnumpy()[Y.asnumpy() == y].sum(axis=0))
P_xy = (n_x + 1) / (n_y + 1).reshape(10, 1, 1)
d2l.show_images(P_xy, 2, 5);
By visualizing these 10 × 28 × 28 probabilities (for each pixel for each class) we could get some
mean looking digits.
Now we can use (18.9.5) to predict a new image. Given x, the following functions computes p(x |
y)p(y) for every y.
def bayes_pred(x):
x = np.expand_dims(x, axis=0) # (28, 28) -> (1, 28, 28)
p_xy = P_xy * x + (1 - P_xy)*(1 - x)
p_xy = p_xy.reshape(10, -1).prod(axis=1) # p(x|y)
return np.array(p_xy) * P_y
image, label = mnist_test[0]

bayes_pred(image)
array([0., 0., 0., 0., 0., 0., 0., 0., 0., 0.])
This went horribly wrong! To find out why, let us look at the per pixel probabilities. They are
typically numbers between 0.001 and 1. We are multiplying 784 of them. At this point it is worth
mentioning that we are calculating these numbers on a computer, hence with a fixed range for the
exponent. What happens is that we experience numerical underflow, i.e., multiplying all the small

numbers leads to something even smaller until it is rounded down to zero. We discussed this as a
theoretical issue in Section 18.7, but we see the phenomena clearly here in practice.
As discussed in that section, we fix this by use the fact that log ab = log a + log b, i.e., we switch to
summing logarithms. Even if both a and b are small numbers, the logarithm values should be in
a proper range.
a = 0.1
print('underflow:', a**784)
print('logarithm is normal:', 784*math.log(a))
underflow: 0.0
logarithm is normal: -1805.2267129073316
Since the logarithm is an increasing function, we can rewrite (18.9.5) as

∑
d
ŷ = argmaxy log Pxy [xi , y] + log Py [y]. (18.9.6)
i=1
We can implement the following stable version:
log_P_xy = np.log(P_xy)
log_P_xy_neg = np.log(1 - P_xy)
log_P_y = np.log(P_y)
def bayes_pred_stable(x):
x = np.expand_dims(x, axis=0) # (28, 28) -> (1, 28, 28)
p_xy = log_P_xy * x + log_P_xy_neg * (1 - x)
p_xy = p_xy.reshape(10, -1).sum(axis=1) # p(x|y)
return p_xy + log_P_y
py = bayes_pred_stable(image)
py
array([-269.00424, -301.73447, -245.21458, -218.8941 , -193.46907,

-206.10315, -292.54315, -114.62834, -220.35619, -163.18881])
We may now check if the prediction is correct.
# Convert label which is a scalar tensor of int32 dtype

# to a Python scalar integer for comparison
py.argmax(axis=0) == int(label)
array(True)
If we now predict a few validation examples, we can see the Bayes classifier works pretty well.
def predict(X):
return [bayes_pred_stable(x).argmax(axis=0).astype(np.int32) for x in X]
X, y = mnist_test[:18]
preds = predict(X)
d2l.show_images(X, 2, 9, titles=[str(d) for d in preds]);

Finally, let us compute the overall accuracy of the classifier.
X, y = mnist_test[:]
preds = np.array(predict(X), dtype=np.int32)
float((preds == y).sum()) / len(y) # Validation accuracy
0.8426
Modern deep networks achieve error rates of less than 0.01. The relatively poor performance is
due to the incorrect statistical assumptions that we made in our model: we assumed that each
and every pixel are independently generated, depending only on the label. This is clearly not how
humans write digits, and this wrong assumption led to the downfall of our overly naive (Bayes)
classifier.
Summary
• Using Bayesʼ rule, a classifier can be made by assuming all observed features are indepen-
dent.
• This classifier can be trained on a dataset by counting the number of occurrences of combi-
nations of labels and pixel values.
• This classifier was the gold standard for decades for tasks such as spam detection.
Exercises
1. Consider the dataset [[0, 0], [0, 1], [1, 0], [1, 1]] with labels given by the XOR of the two elements
[0, 1, 1, 0]. What are the probabilities for a Naive Bayes classifier built on this dataset. Does
it successfully classify our points? If not, what assumptions are violated?
2. Suppose that we did not use Laplace smoothing when estimating probabilities and a data
example arrived at testing time which contained a value never observed in training. What
would the model output?
3. The naive Bayes classifier is a specific example of a Bayesian network, where the dependence
of random variables are encoded with a graph structure. While the full theory is beyond the
scope of this section (see (Koller & Friedman, 2009) for full details), explain why allowing ex-
plicit dependence between the two input variables in the XOR model allows for the creation
of a successful classifier.

Discussions253
18.10 Statistics
Undoubtedly, to be a top deep learning practitioner, the ability to train the state-of-the-art and
high accurate models is crucial. However, it is often unclear when improvements are significant,
or only the result of random fluctuations in the training process. To be able to discuss uncertainty
in estimated values, we must learn some statistics.
The earliest reference of statistics can be traced back to an Arab scholar Al-Kindi in the 9th -century,
who gave a detailed description of how to use statistics and frequency analysis to decipher en-
crypted messages. After 800 years, the modern statistics arose from Germany in 1700s, when
the researchers focused on the demographic and economic data collection and analysis. Today,
statistics is the science subject that concerns the collection, processing, analysis, interpretation
and visualization of data. What is more, the core theory of statistics has been widely used in the
research within academia, industry, and government.
More specifically, statistics can be divided to descriptive statistics and statistical inference. The for-
mer focus on summarizing and illustrating the features of a collection of observed data, which is
referred to as a sample. The sample is drawn from a population, denotes the total set of similar
individuals, items, or events of our experiment interests. Contrary to descriptive statistics, sta-
tistical inference further deduces the characteristics of a population from the given samples, based
on the assumptions that the sample distribution can replicate the population distribution at some
degree.
You may wonder: “What is the essential difference between machine learning and statistics?”
Fundamentally speaking, statistics focuses on the inference problem. This type of problems in-
cludes modeling the relationship between the variables, such as causal inference, and testing the
statistically significance of model parameters, such as A/B testing. In contrast, machine learning
emphasizes on making accurate predictions, without explicitly programming and understanding
each parameterʼs functionality.
In this section, we will introduce three types of statistics inference methods: evaluating and com-
paring estimators, conducting hypothesis tests, and constructing confidence intervals. These
methods can help us infer the characteristics of a given population, i.e., the true parameter θ. For
brevity, we assume that the true parameter θ of a given population is a scalar value. It is straight-
forward to extend to the case where θ is a vector or a tensor, thus we omit it in our discussion.
18.10.1 Evaluating and Comparing Estimators
In statistics, an estimator is a function of given samples used to estimate the true parameter θ. We
will write θ̂n = fˆ(x1 , . . . , xn ) for the estimate of θ after observing the samples {x1 , x2 , . . . , xn }.
We have seen simple examples of estimators before in section Section 18.7. If you have a num-
ber of samples from a Bernoulli random variable, then the maximum likelihood estimate for the
probability the random variable is one can be obtained by counting the number of ones observed
and dividing by the total number of samples. Similarly, an exercise asked you to show that the
maximum likelihood estimate of the mean of a Gaussian given a number of samples is given by
the average value of all the samples. These estimators will almost never give the true value of the
parameter, but ideally for a large number of samples the estimate will be close.
253

As an example, we show below the true density of a Gaussian random variable with mean zero and
variance one, along with a collection samples from that Gaussian. We constructed the y coordinate
so every point is visible and the relationship to the original density is clearer.

import random
npx.set_np()
# Sample datapoints and create y coordinate

epsilon = 0.1
random.seed(8675309)
xs = np.random.normal(loc=0, scale=1, size=(300,))
ys = [np.sum(np.exp(-(xs[:i] - xs[i])**2 / (2 * epsilon**2))

/ np.sqrt(2*np.pi*epsilon**2)) / len(xs) for i in range(len(xs))]
# Compute true density

xd = np.arange(np.min(xs), np.max(xs), 0.01)
yd = np.exp(-xd**2/2) / np.sqrt(2 * np.pi)
# Plot the results

d2l.plot(xd, yd, 'x', 'density')
d2l.plt.scatter(xs, ys)
d2l.plt.axvline(x=0)
d2l.plt.axvline(x=np.mean(xs), linestyle='--', color='purple')
d2l.plt.title(f'sample mean: {float(np.mean(xs)):.2f}')
d2l.plt.show()
There can be many ways to compute an estimator of a parameter θ̂n . In this section, we intro-
duce three common methods to evaluate and compare estimators: the mean squared error, the
standard deviation, and statistical bias.
18.10. Statistics 883

Mean Squared Error
Perhaps the simplest metric used to evaluate estimators is the mean squared error (MSE) (or l2 loss)
of an estimator can be defined as
MSE(θ̂n , θ) = E[(θ̂n − θ)2 ]. (18.10.1)
This allows us to quantify the average squared deviation from the true value. MSE is always non-
negative. If you have read Section 3.1, you will recognize it as the most commonly used regression
loss function. As a measure to evaluate an estimator, the closer its value to zero, the closer the
estimator is close to the true parameter θ.
Statistical Bias
The MSE provides a natural metric, but we can easily imagine multiple different phenomena that
might make it large. Two fundamentally important are fluctuation in the estimator due to ran-
domness in the dataset, and systematic error in the estimator due to the estimation procedure.
First, let us measure the systematic error. For an estimator θ̂n , the mathematical illustration of
statistical bias can be defined as
bias(θ̂n ) = E(θ̂n − θ) = E(θ̂n ) − θ. (18.10.2)
Note that when bias(θ̂n ) = 0, the expectation of the estimator θ̂n is equal to the true value of
parameter. In this case, we say θ̂n is an unbiased estimator. In general, an unbiased estimator is
better than a biased estimator since its expectation is the same as the true parameter.
It is worth being aware, however, that biased estimators are frequently used in practice. There are
cases where unbiased estimators do not exist without further assumptions, or are intractable to
compute. This may seem like a significant flaw in an estimator, however the majority of estimators
encountered in practice are at least asymptotically unbiased in the sense that the bias tends to zero
as the number of available samples tends to infinity: limn→∞ bias(θ̂n ) = 0.
Variance and Standard Deviation
Second, let us measure the randomness in the estimator. Recall from Section 18.6, the standard
deviation (or standard error) is defined as the squared root of the variance. We may measure the
degree of fluctuation of an estimator by measuring the standard deviation or variance of that es-
timator.
√ √
σθ̂n = Var(θ̂n ) = E[(θ̂n − E(θ̂n ))2 ]. (18.10.3)
It is important to compare (18.10.3) to (18.10.1). In this equation we do not compare to the true
population value θ, but instead to E(θ̂n ), the expected sample mean. Thus we are not measuring
how far the estimator tends to be from the true value, but instead we measuring the fluctuation of
the estimator itself.

The Bias-Variance Trade-off
It is intuitively clear that these two main components contribute to the mean squared error. What
is somewhat shocking is that we can show that this is actually a decomposition of the mean squared
error into these two contributions plus a third one. That is to say that we can write the mean
squared error as the sum of the square of the bias, the variance and the irreducible error.
MSE(θ̂n , θ) = E[(θ̂n − θ)2 ]

= E[(θ̂n )2 ] + E[θ2 ] − 2E[θ̂n θ]
= Var[θ̂n ] + E[θ̂n ]2 + Var[θ] + E[θ]2 − 2E[θ̂n ]E[θ]
(18.10.4)
= (E[θ̂n ] − E[θ])2 + Var[θ̂n ] + Var[θ]
= (E[θ̂n − θ])2 + Var[θ̂n ] + Var[θ]
= (bias[θ̂n ])2 + Var(θ̂n ) + Var[θ].
We refer the above formula as bias-variance trade-off. The mean squared error can be divided into
three sources of error: the error from high bias, the error from high variance and the irreducible
error. The bias error is commonly seen in a simple model (such as a linear regression model),
which cannot extract high dimensional relations between the features and the outputs. If a model
suffers from high bias error, we often say it is underfitting or lack of generalization as introduced
in (Section 4.4). The high variance usually results from a too complex model, which overfits the
training data. As a result, an overfitting model is sensitive to small fluctuations in the data. If a
model suffers from high variance, we often say it is overfitting and lack of flexibility as introduced
in (Section 4.4). The irreducible error is the result from noise in the θ itself.
Evaluating Estimators in Code
Since the standard deviation of an estimator has been implementing by simply calling a.std() for
a tensor a, we will skip it but implement the statistical bias and the mean squared error.
# Statistical bias
def stat_bias(true_theta, est_theta):
return(np.mean(est_theta) - true_theta)
# Mean squared error

def mse(data, true_theta):
return(np.mean(np.square(data - true_theta)))
To illustrate the equation of the bias-variance trade-off, let us simulate of normal distribution
N (θ, σ 2 ) with 10, 000 samples. Here, we use a θ = 1 and σ = 4. As the estimator is a function
of the given samples, here we use the mean of the samples as an estimator for true θ in this nor-
mal distribution N (θ, σ 2 ) .
theta_true = 1
sigma = 4
sample_len = 10000
samples = np.random.normal(theta_true, sigma, sample_len)
theta_est = np.mean(samples)
theta_est

array(0.9503336)
Let us validate the trade-off equation by calculating the summation of the squared bias and the
variance of our estimator. First, calculate the MSE of our estimator.
mse(samples, theta_true)
array(15.781996)
Next, we calculate Var(θ̂n )+[bias(θ̂n )]2 as below. As you can see, the two values agree to numerical
precision.
bias = stat_bias(theta_true, theta_est)

np.square(samples.std()) + np.square(bias)
array(15.781995)
18.10.2 Conducting Hypothesis Tests
The most commonly encountered topic in statistical inference is hypothesis testing. While hy-
pothesis testing was popularized in the early 20th century, the first use can be traced back to John
Arbuthnot in the 1700s. John tracked 80-year birth records in London and concluded that more
men were born than women each year. Following that, the modern significance testing is the in-
telligence heritage by Karl Pearson who invented p-value and Pearsonʼs chi-squared test, William
Gosset who is the father of Studentʼs t-distribution, and Ronald Fisher who initialed the null hy-
pothesis and the significance test.
A hypothesis test is a way of evaluating some evidence against the default statement about a pop-
ulation. We refer the default statement as the null hypothesis H0 , which we try to reject using the
observed data. Here, we use H0 as a starting point for the statistical significance testing. The alter-
native hypothesis HA (or H1 ) is a statement that is contrary to the null hypothesis. A null hypothesis
is often stated in a declarative form which posits a relationship between variables. It should reflect
the brief as explicit as possible, and be testable by statistics theory.
Imagine you are a chemist. After spending thousands of hours in the lab, you develop a new
medicine which can dramatically improve oneʼs ability to understand math. To show its magic
power, you need to test it. Naturally, you may need some volunteers to take the medicine and see
whether it can help them learn math better. How do you get started?
First, you will need carefully random selected two groups of volunteers, so that there is no differ-
ence between their math understanding ability measured by some metrics. The two groups are
commonly referred to as the test group and the control group. The test group (or treatment group)
is a group of individuals who will experience the medicine, while the control group represents the
group of users who are set aside as a benchmark, i.e., identical environment setups except taking
this medicine. In this way, the influence of all the variables are minimized, except the impact of
the independent variable in the treatment.
Second, after a period of taking the medicine, you will need to measure the two groupsʼ math
understanding by the same metrics, such as letting the volunteers do the same tests after learning a
new math formula. Then, you can collect their performance and compare the results. In this case,

our null hypothesis will be that there is no difference between the two groups, and our alternate
will be that there is.
This is still not fully formal. There are many details you have to think of carefully. For example,
what is the suitable metrics to test their math understanding ability? How many volunteers for
your test so you can be confident to claim the effectiveness of your medicine? How long should
you run the test? How do you decide if there is a difference between the two groups? Do you care
about the average performance only, or also the range of variation of the scores? And so on.
In this way, hypothesis testing provides a framework for experimental design and reasoning about
certainty in observed results. If we can now show that the null hypothesis is very unlikely to be
true, we may reject it with confidence.
To complete the story of how to work with hypothesis testing, we need to now introduce some
additional terminology and make some of our concepts above formal.
Statistical Significance
The statistical significance measures the probability of erroneously rejecting the null hypothesis,
H0 , when it should not be rejected, i.e.,
statistical significance = 1 − α = 1 − P (reject H0 | H0 is true). (18.10.5)
It is also referred to as the type I error or false positive. The α, is called as the significance level and
its commonly used value is 5%, i.e., 1 − α = 95%. The significance level can be explained as the
level of risk that we are willing to take, when we reject a true null hypothesis.
Fig. 18.10.1 shows the observationsʼ values and probability of a given normal distribution in a two-
sample hypothesis test. If the observation data example is located outsides the 95% threshold, it
will be a very unlikely observation under the null hypothesis assumption. Hence, there might be
something wrong with the null hypothesis and we will reject it.
Fig. 18.10.1: Statistical significance.

Statistical Power
The statistical power (or sensitivity) measures the probability of reject the null hypothesis, H0 , when
it should be rejected, i.e.,
statistical power = 1 − β = 1 − P ( fail to reject H0 | H0 is false). (18.10.6)
Recall that a type I error is error caused by rejecting the null hypothesis when it is true, whereas a
type II error is resulted from failing to reject the null hypothesis when it is false. A type II error is
usually denoted as β, and hence the corresponding statistical power is 1 − β.
Intuitively, statistical power can be interpreted as how likely our test will detect a real discrepancy
of some minimum magnitude at a desired statistical significance level. 80% is a commonly used
statistical power threshold. The higher the statistical power, the more likely we are to detect true
differences.
One of the most common uses of statistical power is in determining the number of samples
needed. The probability you reject the null hypothesis when it is false depends on the degree
to which it is false (known as the effect size) and the number of samples you have. As you might
expect, small effect sizes will require a very large number of samples to be detectable with high
probability. While beyond the scope of this brief appendix to derive in detail, as an example, want
to be able to reject a null hypothesis that our sample came from a mean zero variance one Gaus-
sian, and we believe that our sampleʼs mean is actually close to one, we can do so with acceptable
error rates with a sample size of only 8. However, if we think our sample population true mean is
close to 0.01, then weʼd need a sample size of nearly 80000 to detect the difference.
We can imagine the power as a water filter. In this analogy, a high power hypothesis test is like a
high quality water filtration system that will reduce harmful substances in the water as much as
possible. On the other hand, a smaller discrepancy is like a low quality water filter, where some
relative small substances may easily escape from the gaps. Similarly, if the statistical power is not
of enough high power, then the test may not catch the smaller discrepancy.
Test Statistic
A test statistic T (x) is a scalar which summarizes some characteristic of the sample data. The goal
of defining such a statistic is that it should allow us to distinguish between different distributions
and conduct our hypothesis test. Thinking back to our chemist example, if we wish to show that
one population performs better than the other, it could be reasonable to take the mean as the
test statistic. Different choices of test statistic can lead to statistical test with drastically different
statistical power.
Often, T (X) (the distribution of the test statistic under our null hypothesis) will follow, at least
approximately, a common probability distribution such as a normal distribution when considered
under the null hypothesis. If we can derive explicitly such a distribution, and then measure our
test statistic on our dataset, we can safely reject the null hypothesis if our statistic is far outside
the range that we would expect. Making this quantitative leads us to the notion of p-values.

p-value
The p-value (or the probability value) is the probability that T (X) is at least as extreme as the ob-
served test statistic T (x) assuming that the null hypothesis is true, i.e.,
p-value = PH0 (T (X) ≥ T (x)). (18.10.7)
If the p-value is smaller than or equal to a predefined and fixed statistical significance level α, we
may reject the null hypothesis. Otherwise, we will conclude that we are lack of evidence to reject
the null hypothesis. For a given population distribution, the region of rejection will be the interval
contained of all the points which has a p-value smaller than the statistical significance level α.
One-side Test and Two-sided Test
Normally there are two kinds of significance test: the one-sided test and the two-sided test. The
one-sided test (or one-tailed test) is applicable when the null hypothesis and the alternative hypoth-
esis only have one direction. For example, the null hypothesis may state that the true parameter
θ is less than or equal to a value c. The alternative hypothesis would be that θ is greater than c.
That is, the region of rejection is on only one side of the sampling distribution. Contrary to the
one-sided test, the two-sided test (or two-tailed test) is applicable when the region of rejection is on
both sides of the sampling distribution. An example in this case may have a null hypothesis state
that the true parameter θ is equal to a value c. The alternative hypothesis would be that θ is not
equal to c.
General Steps of Hypothesis Testing
After getting familiar with the above concepts, let us go through the general steps of hypothesis
testing.
1. State the question and establish a null hypotheses H0 .
2. Set the statistical significance level α and a statistical power (1 − β).
3. Obtain samples through experiments. The number of samples needed will depend on the
statistical power, and the expected effect size.
4. Calculate the test statistic and the p-value.
5. Make the decision to keep or reject the null hypothesis based on the p-value and the statistical
significance level α.
To conduct a hypothesis test, we start by defining a null hypothesis and a level of risk that we
are willing to take. Then we calculate the test statistic of the sample, taking an extreme value of
the test statistic as evidence against the null hypothesis. If the test statistic falls within the reject
region, we may reject the null hypothesis in favor of the alternative.
Hypothesis testing is applicable in a variety of scenarios such as the clinical trails and A/B testing.

18.10.3 Constructing Confidence Intervals
When estimating the value of a parameter θ, point estimators like θ̂ are of limited utility since they
contain no notion of uncertainty. Rather, it would be far better if we could produce an interval
that would contain the true parameter θ with high probability. If you were interested in such
ideas a century ago, then you would have been excited to read “Outline of a Theory of Statistical
Estimation Based on the Classical Theory of Probability” by Jerzy Neyman (Neyman, 1937), who
first introduced the concept of confidence interval in 1937.
To be useful, a confidence interval should be as small as possible for a given degree of certainty.
Let us see how to derive it.
Definition
Mathematically, a confidence interval for the true parameter θ is an interval Cn that computed from
the sample data such that
Pθ (Cn ∋ θ) ≥ 1 − α, ∀θ. (18.10.8)
Here α ∈ (0, 1), and 1 − α is called the confidence level or coverage of the interval. This is the same
α as the significance level as we discussed about above.
Note that (18.10.8) is about variable Cn , not about the fixed θ. To emphasize this, we write Pθ (Cn ∋
θ) rather than Pθ (θ ∈ Cn ).
Interpretation
It is very tempting to interpret a 95% confidence interval as an interval where you can be 95% sure
the true parameter lies, however this is sadly not true. The true parameter is fixed, and it is the
interval that is random. Thus a better interpretation would be to say that if you generated a large
number of confidence intervals by this procedure, 95% of the generated intervals would contain
the true parameter.
This may seem pedantic, but it can have real implications for the interpretation of the results.
In particular, we may satisfy (18.10.8) by constructing intervals that we are almost certain do not
contain the true value, as long as we only do so rarely enough. We close this section by providing
three tempting but false statements. An in-depth discussion of these points can be found in (Morey
et al., 2016).
• Fallacy 1. Narrow confidence intervals mean we can estimate the parameter precisely.
• Fallacy 2. The values inside the confidence interval are more likely to be the true value than
those outside the interval.
• Fallacy 3. The probability that a particular observed 95% confidence interval contains the
true value is 95%.
Sufficed to say, confidence intervals are subtle objects. However, if you keep the interpretation
clear, they can be powerful tools.

A Gaussian Example
Let us discuss the most classical example, the confidence interval for the mean of a Gaussian of
unknown mean and variance. Suppose we collect n samples {xi }ni=1 from our Gaussian N (µ, σ 2 ).
We can compute estimators for the mean and standard deviation by taking
1∑ 1 ∑
n n
µ̂n = xi and σ̂n2 = (xi − µ̂)2 . (18.10.9)
n n−1
i=1 i=1
If we now consider the random variable

µ̂n − µ
T = √ , (18.10.10)
σ̂n / n
we obtain a random variable following a well-known distribution called the Student’s t-distribution
on n − 1 degrees of freedom.
This distribution is very well studied, and it is known, for instance, that as n → ∞, it is approx-
imately a standard Gaussian, and thus by looking up values of the Gaussian c.d.f. in a table, we
may conclude that the value of T is in the interval [−1.96, 1.96] at least 95% of the time. For finite
values of n, the interval needs to be somewhat larger, but are well known and precomputed in
tables.
Thus, we may conclude that for large n,
( )
µ̂n − µ
P √ ∈ [−1.96, 1.96] ≥ 0.95. (18.10.11)
σ̂n / n
√
Rearranging this by multiplying both sides by σ̂n / n and then adding µ̂n , we obtain
( [ ])
σ̂n σ̂n
P µ ∈ µ̂n − 1.96 √ , µ̂n + 1.96 √ ≥ 0.95. (18.10.12)
n n
Thus we know that we have found our 95% confidence interval:
[ ]
σ̂n σ̂n
µ̂n − 1.96 √ , µ̂n + 1.96 √ . (18.10.13)
n n
It is safe to say that (18.10.13) is one of the most used formula in statistics. Let us close our discus-
sion of statistics by implementing it. For simplicity, we assume we are in the asymptotic regime.
Small values of N should include the correct value of t_star obtained either programmatically or
from a t-table.
# Number of samples
N = 1000
# Sample dataset
samples = np.random.normal(loc=0, scale=1, size=(N,))
# Lookup Students's t-distribution c.d.f.

t_star = 1.96
# Construct interval
mu_hat = np.mean(samples)
sigma_hat = samples.std(ddof=1)
(mu_hat - t_star*sigma_hat/np.sqrt(N), mu_hat + t_star*sigma_hat/np.sqrt(N))

(array(-0.07853346), array(0.04412608))
Summary
• Statistics focuses on inference problems, whereas deep learning emphasizes on making ac-
curate predictions without explicitly programming and understanding.
• There are three common statistics inference methods: evaluating and comparing estima-
tors, conducting hypothesis tests, and constructing confidence intervals.
• There are three most common estimators: statistical bias, standard deviation, and mean
square error.
• A confidence interval is an estimated range of a true population parameter that we can con-
struct by given the samples.
• Hypothesis testing is a way of evaluating some evidence against the default statement about
a population.
Exercises
iid
1. Let X1 , X2 , . . . , Xn ∼ Unif(0, θ), where “iid” stands for independent and identically distributed.
Consider the following estimators of θ:
θ̂ = max{X1 , X2 , . . . , Xn }; (18.10.14)
2∑
n
θ̃ = 2X¯n = Xi . (18.10.15)
n
i=1
• Find the statistical bias, standard deviation, and mean square error of θ̂.
• Find the statistical bias, standard deviation, and mean square error of θ̃.
• Which estimator is better?
2. For our chemist example in introduction, can you derive the 5 steps to conduct a two-sided
hypothesis testing? Given the statistical significance level α = 0.05 and the statistical power
1 − β = 0.8.
3. Run the confidence interval code with N = 2 and α = 0.5 for 100 independently generated
dataset, and plot the resulting intervals (in this case t_star = 1.0). You will see several very
short intervals which are very far from containing the true mean 0. Does this contradict the
interpretation of the confidence interval? Do you feel comfortable using short intervals to
indicate high precision estimates?
Discussions254
254

18.11 Information Theory
The universe is overflowing with information. Information provides a common language across
disciplinary rifts: from Shakespeareʼs Sonnet to researchersʼ paper on Cornell ArXiv, from Van
Goghʼs printing Starry Night to Beethovenʼs music Symphony No. 5, from the first programming
language Plankalkül to the state-of-the-art machine learning algorithms. Everything must follow
the rules of information theory, no matter the format. With information theory, we can measure
and compare how much information is present in different signals. In this section, we will inves-
tigate the fundamental concepts of information theory and applications of information theory in
machine learning.
Before we get started, let us outline the relationship between machine learning and information
theory. Machine learning aims to extract interesting signals from data and make critical pre-
dictions. On the other hand, information theory studies encoding, decoding, transmitting, and
manipulating information. As a result, information theory provides fundamental language for
discussing the information processing in machine learned systems. For example, many machine
learning applications use the cross entropy loss as described in Section 3.4. This loss can be di-
rectly derived from information theoretic considerations.
18.11.1 Information
Let us start with the “soul” of information theory: information. Information can be encoded in
anything with a particular sequence of one or more encoding formats. Suppose that we task our-
selves with trying to define a notion of information. What could be our starting point?
Consider the following thought experiment. We have a friend with a deck of cards. They will
shuffle the deck, flip over some cards, and tell us statements about the cards. We will try to assess
the information content of each statement.
First, they flip over a card and tell us, “I see a card.” This provides us with no information at all.
We were already certain that this was the case so we hope the information should be zero.
Next, they flip over a card and say, “I see a heart.” This provides us some information, but in reality
there are only 4 different suits that were possible, each equally likely, so we are not surprised by
this outcome. We hope that whatever the measure of information, this event should have low
information content.
Next, they flip over a card and say, “This is the 3 of spades.” This is more information. Indeed there
were 52 equally likely possible outcomes, and our friend told us which one it was. This should be
a medium amount of information.
Let us take this to the logical extreme. Suppose that finally they flip over every card from the deck
and read off the entire sequence of the shuffled deck. There are 52! different orders to the deck,
again all equally likely, so we need a lot of information to know which one it is.
Any notion of information we develop must conform to this intuition. Indeed, in the next sec-
tions we will learn how to compute that these events have 0 bits, 2 bits, 5.7 bits, and 225.6 bits of
information respectively.
If we read through these thought experiments, we see a natural idea. As a starting point, rather
than caring about the knowledge, we may build off the idea that information represents the degree
of surprise or the abstract possibility of the event. For example, if we want to describe an unusual
event, we need a lot information. For a common event, we may not need much information.
18.11. Information Theory 893

In 1948, Claude E. Shannon published A Mathematical Theory of Communication (Shannon, 1948)
establishing the theory of information. In his article, Shannon introduced the concept of infor-
mation entropy for the first time. We will begin our journey here.
Self-information
Since information embodies the abstract possibility of an event, how do we map the possibility
to the number of bits? Shannon introduced the terminology bit as the unit of information, which
was originally created by John Tukey. So what is a “bit” and why do we use it to measure infor-
mation? Historically, an antique transmitter can only send or receive two types of code: 0 and
1. Indeed, binary encoding is still in common use on all modern digital computers. In this way,
any information is encoded by a series of 0 and 1. And hence, a series of binary digits of length n
contains n bits of information.
Now, suppose that for any series of codes, each 0 or 1 occurs with a probability of 12 . Hence, an
event X with a series of codes of length n, occurs with a probability of 21n . At the same time, as
we mentioned before, this series contains n bits of information. So, can we generalize to a math
function which can transfer the probability p to the number of bits? Shannon gave the answer by
defining self-information
I(X) = − log2 (p), (18.11.1)
as the bits of information we have received for this event X. Note that we will always use base-2
logarithms in this section. For the sake of simplicity, the rest of this section will omit the subscript
2 in the logarithm notation, i.e., log(.) always refers to log2 (.). For example, the code “0010” has a
self-information
( )
1
I("0010") = − log(p("0010")) = − log = 4 bits. (18.11.2)
24
We can calculate self information as shown below. Before that, let us first import all the necessary
packages in this section.

from mxnet.metric import NegativeLogLikelihood
from mxnet.ndarray import nansum
import random
def self_information(p):
return -np.log2(p)
self_information(1 / 64)
6.0

18.11.2 Entropy
As self-information only measures the information of a single discrete event, we need a more
generalized measure for any random variable of either discrete or continuous distribution.
Motivating Entropy
Let us try to get specific about what we want. This will be an informal statement of what are known
as the axioms of Shannon entropy. It will turn out that the following collection of common-sense
statements force us to a unique definition of information. A formal version of these axioms, along
with several others may be found in (Csiszar, 2008).
1. The information we gain by observing a random variable does not depend on what we call
the elements, or the presence of additional elements which have probability zero.
2. The information we gain by observing two random variables is no more than the sum of the
information we gain by observing them separately. If they are independent, then it is exactly
the sum.
3. The information gained when observing (nearly) certain events is (nearly) zero.
While proving this fact is beyond the scope of our text, it is important to know that this uniquely
determines the form that entropy must take. The only ambiguity that these allow is in the choice
of fundamental units, which is most often normalized by making the choice we saw before that
the information provided by a single fair coin flip is one bit.
Definition
For any random variable X that follows a probability distribution P with a probability density
function (p.d.f.) or a probability mass function (p.m.f.) p(x), we measure the expected amount of
information through entropy (or Shannon entropy)
H(X) = −Ex∼P [log p(x)]. (18.11.3)
To be specific, if X is discrete,
∑
H(X) = − pi log pi , where pi = P (Xi ). (18.11.4)
i
Otherwise, if X is continuous, we also refer entropy as differential entropy

∫
H(X) = − p(x) log p(x) dx. (18.11.5)
x
We can define entropy as below.
def entropy(p):
entropy = - p * np.log2(p)
# Operator nansum will sum up the non-nan number
out = nansum(entropy.as_nd_ndarray())
return out
entropy(np.array([0.1, 0.5, 0.1, 0.3]))

[1.6854753]
<NDArray 1 @cpu(0)>
Interpretations
You may be curious: in the entropy definition (18.11.3), why do we use an expectation of a negative
logarithm? Here are some intuitions.
First, why do we use a logarithm function log? Suppose that p(x) = f1 (x)f2 (x) . . . , fn (x), where
each component function fi (x) is independent from each other. This means that each fi (x) con-
tributes independently to the total information obtained from p(x). As discussed above, we want
the entropy formula to be additive over independent random variables. Luckily, log can naturally
turn a product of probability distributions to a summation of the individual terms.
Next, why do we use a negative log? Intuitively, more frequent events should contain less infor-
mation than less common events, since we often gain more information from an unusual case
than from an ordinary one. However, log is monotonically increasing with the probabilities, and
indeed negative for all values in [0, 1]. We need to construct a monotonically decreasing relation-
ship between the probability of events and their entropy, which will ideally be always positive (for
nothing we observe should force us to forget what we have known). Hence, we add a negative sign
in front of log function.
Last, where does the expectation function come from? Consider a random variable X. We can
interpret the self-information (− log(p)) as the amount of surprise we have at seeing a particular
outcome. Indeed, as the probability approaches zero, the surprise becomes infinite. Similarly,
we can interpret the entropy as the average amount of surprise from observing X. For exam-
ple, imagine that a slot machine system emits statistical independently symbols s1 , . . . , sk with
probabilities p1 , . . . , pk respectively. Then the entropy of this system equals to the average self-
information from observing each output, i.e.,
∑ ∑
H(S) = pi · I(si ) = − pi · log pi . (18.11.6)
i i
Properties of Entropy
By the above examples and interpretations, we can derive the following properties of entropy
(18.11.3). Here, we refer to X as an event and P as the probability distribution of X.
• Entropy is non-negative, i.e., H(X) ≥ 0, ∀X.
• If X ∼ P with a p.d.f. or a p.m.f. p(x), and we try to estimate P by a new probability
distribution Q with a p.d.f. or a p.m.f. q(x), then
H(X) = −Ex∼P [log p(x)] ≤ −Ex∼P [log q(x)], with equality if and only if P = Q. (18.11.7)
Alternatively, H(X) gives a lower bound of the average number of bits needed to encode
symbols drawn from P .
• If X ∼ P , then x conveys the maximum amount of information if it spreads evenly among
all possible outcomes. Specifically, if the probability distribution P is discrete with k-class
{p1 , . . . , pk }, then
1
H(X) ≤ log(k), with equality if and only if pi = , ∀i. (18.11.8)
k

If P is a continuous random variable, then the story becomes much more complicated.
However, if we additionally impose that P is supported on a finite interval (with all values
between 0 and 1), then P has the highest entropy if it is the uniform distribution on that
interval.
18.11.3 Mutual Information
Previously we defined entropy of a single random variable X, how about the entropy of a pair
random variables (X, Y )? We can think of these techniques as trying to answer the following type
of question, “What information is contained in X and Y together compared to each separately? Is
there redundant information, or is it all unique?”
For the following discussion, we always use (X, Y ) as a pair of random variables that follows a joint
probability distribution P with a p.d.f. or a p.m.f. pX,Y (x, y), while X and Y follow probability
distribution pX (x) and pY (y), respectively.
Joint Entropy
Similar to entropy of a single random variable (18.11.3), we define the joint entropy H(X, Y ) of a
pair random variables (X, Y ) as
H(X, Y ) = −E(x,y)∼P [log pX,Y (x, y)]. (18.11.9)
Precisely, on the one hand, if (X, Y ) is a pair of discrete random variables, then
∑∑
H(X, Y ) = − pX,Y (x, y) log pX,Y (x, y). (18.11.10)
x y
On the other hand, if (X, Y ) is a pair of continuous random variables, then we define the differential
joint entropy as
∫
H(X, Y ) = − pX,Y (x, y) log pX,Y (x, y) dx dy. (18.11.11)
x,y
We can think of (18.11.9) as telling us the total randomness in the pair of random variables. As a
pair of extremes, if X = Y are two identical random variables, then the information in the pair
is exactly the information in one and we have H(X, Y ) = H(X) = H(Y ). On the other extreme,
if X and Y are independent then H(X, Y ) = H(X) + H(Y ). Indeed we will always have that
the information contained in a pair of random variables is no smaller than the entropy of either
random variable and no more than the sum of both.
H(X), H(Y ) ≤ H(X, Y ) ≤ H(X) + H(Y ). (18.11.12)
Let us implement joint entropy from scratch.
def joint_entropy(p_xy):
joint_ent = -p_xy * np.log2(p_xy)
out = nansum(joint_ent.as_nd_ndarray())
return out
joint_entropy(np.array([[0.1, 0.5], [0.1, 0.3]]))

[1.6854753]
<NDArray 1 @cpu(0)>
Notice that this is the same code as before, but now we interpret it differently as working on the
joint distribution of the two random variables.
Conditional Entropy
The joint entropy defined above the amount of information contained in a pair of random vari-
ables. This is useful, but oftentimes it is not what we care about. Consider the setting of machine
learning. Let us take X to be the random variable (or vector of random variables) that describes
the pixel values of an image, and Y to be the random variable which is the class label. X should
contain substantial information—a natural image is a complex thing. However, the information
contained in Y once the image has been show should be low. Indeed, the image of a digit should
already contain the information about what digit it is unless the digit is illegible. Thus, to continue
to extend our vocabulary of information theory, we need to be able to reason about the informa-
tion content in a random variable conditional on another.
In the probability theory, we saw the definition of the conditional probability to measure the rela-
tionship between variables. We now want to analogously define the conditional entropy H(Y | X).
We can write this as
H(Y | X) = −E(x,y)∼P [log p(y | x)], (18.11.13)
p (x,y)
where p(y | x) = X,Y
pX (x) is the conditional probability. Specifically, if (X, Y ) is a pair of discrete
random variables, then
∑∑
H(Y | X) = − p(x, y) log p(y | x). (18.11.14)
x y
If (X, Y ) is a pair of continuous random variables, then the differential conditional entropy is simi-
larly defined as
∫ ∫
H(Y | X) = − p(x, y) log p(y | x) dx dy. (18.11.15)
x y
It is now natural to ask, how does the conditional entropy H(Y | X) relate to the entropy H(X) and
the joint entropy H(X, Y )? Using the definitions above, we can express this cleanly:
H(Y | X) = H(X, Y ) − H(X). (18.11.16)
This has an intuitive interpretation: the information in Y given X (H(Y | X)) is the same as the
information in both X and Y together (H(X, Y )) minus the information already contained in X.
This gives us the information in Y which is not also represented in X.
Now, let us implement conditional entropy (18.11.13) from scratch.
def conditional_entropy(p_xy, p_x):

p_y_given_x = p_xy/p_x
cond_ent = -p_xy * np.log2(p_y_given_x)
out = nansum(cond_ent.as_nd_ndarray())
return out
conditional_entropy(np.array([[0.1, 0.5], [0.2, 0.3]]), np.array([0.2, 0.8]))

[0.8635472]
<NDArray 1 @cpu(0)>
Mutual Information
Given the previous setting of random variables (X, Y ), you may wonder: “Now that we know how
much information is contained in Y but not in X, can we similarly ask how much information is
shared between X and Y ?” The answer will be the mutual information of (X, Y ), which we will
write as I(X, Y ).
Rather than diving straight into the formal definition, let us practice our intuition by first trying
to derive an expression for the mutual information entirely based on terms we have constructed
before. We wish to find the information shared between two random variables. One way we could
try to do this is to start with all the information contained in both X and Y together, and then
we take off the parts that are not shared. The information contained in both X and Y together is
written as H(X, Y ). We want to subtract from this the information contained in X but not in Y ,
and the information contained in Y but not in X. As we saw in the previous section, this is given
by H(X | Y ) and H(Y | X) respectively. Thus, we have that the mutual information should be
I(X, Y ) = H(X, Y ) − H(Y | X)−H(X | Y ). (18.11.17)
Indeed, this is a valid definition for the mutual information. If we expand out the definitions of
these terms and combine them, a little algebra shows that this is the same as
{ }
pX,Y (x, y)
I(X, Y ) = Ex Ey pX,Y (x, y) log . (18.11.18)
pX (x)pY (y)
We can summarize all of these relationships in image Fig. 18.11.1. It is an excellent test of intuition
to see why the following statements are all also equivalent to I(X, Y ).
• H(X)−H(X | Y )
• H(Y )−H(Y | X)
• H(X) + H(Y )−H(X, Y )
Fig. 18.11.1: Mutual informationʼs relationship with joint entropy and conditional entropy.
In many ways we can think of the mutual information (18.11.18) as principled extension of cor-
relation coefficient we saw in Section 18.6. This allows us to ask not only for linear relationships

between variables, but for the maximum information shared between the two random variables
of any kind.
Now, let us implement mutual information from scratch.
def mutual_information(p_xy, p_x, p_y):

p = p_xy / (p_x * p_y)
mutual = p_xy * np.log2(p)
out = nansum(mutual.as_nd_ndarray())
return out
mutual_information(np.array([[0.1, 0.5], [0.1, 0.3]]),

np.array([0.2, 0.8]), np.array([[0.75, 0.25]]))
[0.71946025]
<NDArray 1 @cpu(0)>
Properties of Mutual Information
Rather than memorizing the definition of mutual information (18.11.18), you only need to keep in
mind its notable properties:
• Mutual information is symmetric, i.e., I(X, Y ) = I(Y, X).
• Mutual information is non-negative, i.e., I(X, Y ) ≥ 0.
• I(X, Y ) = 0 if and only if X and Y are independent. For example, if X and Y are indepen-
dent, then knowing Y does not give any information about X and vice versa, so their mutual
information is zero.
• Alternatively, if X is an invertible function of Y , then Y and X share all information and
I(X, Y ) = H(Y ) = H(X). (18.11.19)
Pointwise Mutual Information
When we worked with entropy at the beginning of this chapter, we were able to provide an inter-
pretation of − log(pX (x)) as how surprised we were with the particular outcome. We may give a
similar interpretation to the logarithmic term in the mutual information, which is often referred
to as the pointwise mutual information:
pX,Y (x, y)
pmi(x, y) = log . (18.11.20)
pX (x)pY (y)
We can think of (18.11.20) as measuring how much more or less likely the specific combination
of outcomes x and y are compared to what we would expect for independent random outcomes.
If it is large and positive, then these two specific outcomes occur much more frequently than they
would compared to random chance (note: the denominator is pX (x)pY (y) which is the probability
of the two outcomes were independent), whereas if it is large and negative it represents the two
outcomes happening far less than we would expect by random chance.
This allows us to interpret the mutual information (18.11.18) as the average amount that we were
surprised to see two outcomes occurring together compared to what we would expect if they were
independent.

Applications of Mutual Information
Mutual information may be a little abstract in it pure definition, so how does it related to machine
learning? In natural language processing, one of the most difficult problems is the ambiguity res-
olution, or the issue of the meaning of a word being unclear from context. For example, recently
a headline in the news reported that “Amazon is on fire”. You may wonder whether the company
Amazon has a building on fire, or the Amazon rain forest is on fire.
In this case, mutual information can help us resolve this ambiguity. We first find the group of
words that each has a relatively large mutual information with the company Amazon, such as
e-commerce, technology, and online. Second, we find another group of words that each has a
relatively large mutual information with the Amazon rain forest, such as rain, forest, and tropical.
When we need to disambiguate “Amazon”, we can compare which group has more occurrence in
the context of the word Amazon. In this case the article would go on to describe the forest, and
make the context clear.
18.11.4 Kullback–Leibler Divergence
As what we have discussed in Section 2.3, we can use norms to measure distance between two
points in space of any dimensionality. We would like to be able to do a similar task with probability
distributions. There are many ways to go about this, but information theory provides one of the
nicest. We now explore the Kullback–Leibler (KL) divergence, which provides a way to measure if
two distributions are close together or not.
Definition
Given a random variable X that follows the probability distribution P with a p.d.f. or a p.m.f. p(x),
and we estimate P by another probability distribution Q with a p.d.f. or a p.m.f. q(x). Then the
Kullback–Leibler (KL) divergence (or relative entropy) between P and Q is
[ ]
p(x)
DKL (P ∥Q) = Ex∼P log . (18.11.21)
q(x)
As with the pointwise mutual information (18.11.20), we can again provide an interpretation of
q(x)
the logarithmic term: − log p(x) = − log(q(x)) − (− log(p(x))) will be large and positive if we see x
far more often under P than we would expect for Q, and large and negative if we see the outcome
far less than expected. In this way, we can interpret it as our relative surprise at observing the
outcome compared to how surprised we would be observing it from our reference distribution.
Let us implement the KL divergence from Scratch.
def kl_divergence(p, q):

kl = p * np.log2(p / q)
out = nansum(kl.as_nd_ndarray())
return out.abs().asscalar()

KL Divergence Properties
Let us take a look at some properties of the KL divergence (18.11.21).

• KL divergence is non-symmetric, i.e., there are P, Q such that
DKL (P ∥Q) ̸= DKL (Q∥P ). (18.11.22)
• KL divergence is non-negative, i.e.,
DKL (P ∥Q) ≥ 0. (18.11.23)
Note that the equality holds only when P = Q.

• If there exists an x such that p(x) > 0 and q(x) = 0, then DKL (P ∥Q) = ∞.
• There is a close relationship between KL divergence and mutual information. Besides the
relationship shown in Fig. 18.11.1, I(X, Y ) is also numerically equivalent with the following
terms:
1. DKL (P (X, Y ) ∥ P (X)P (Y ));
2. EY {DKL (P (X | Y ) ∥ P (X))};
3. EX {DKL (P (Y | X) ∥ P (Y ))}.
For the first term, we interpret mutual information as the KL divergence between P (X, Y )
and the product of P (X) and P (Y ), and thus is a measure of how different the joint dis-
tribution is from the distribution if they were independent. For the second term, mutual
information tells us the average reduction in uncertainty about Y that results from learning
the value of the Xʼs distribution. Similarly to the third term.
Example
Let us go through a toy example to see the non-symmetry explicitly.

First, let us generate and sort three tensors of length 10, 000: an objective tensor p which follows
a normal distribution N (0, 1), and two candidate tensors q1 and q2 which follow normal distribu-
tions N (−1, 1) and N (1, 1) respectively.
random.seed(1)
nd_len = 10000
p = np.random.normal(loc=0, scale=1, size=(nd_len, ))
q1 = np.random.normal(loc=-1, scale=1, size=(nd_len, ))
q2 = np.random.normal(loc=1, scale=1, size=(nd_len, ))
p = np.array(sorted(p.asnumpy()))
q1 = np.array(sorted(q1.asnumpy()))
q2 = np.array(sorted(q2.asnumpy()))
Since q1 and q2 are symmetric with respect to the y-axis (i.e., x = 0), we expect a similar value of
KL divergence between DKL (p∥q1 ) and DKL (p∥q2 ). As you can see below, there is only a less than
3% off between DKL (p∥q1 ) and DKL (p∥q2 ).

kl_pq1 = kl_divergence(p, q1)
kl_pq2 = kl_divergence(p, q2)
similar_percentage = abs(kl_pq1 - kl_pq2) / ((kl_pq1 + kl_pq2) / 2) * 100
kl_pq1, kl_pq2, similar_percentage
(8470.638, 8664.999, 2.268504302642314)
In contrast, you may find that DKL (q2 ∥p) and DKL (p∥q2 ) are off a lot, with around 40% off as shown
below.
kl_q2p = kl_divergence(q2, p)
differ_percentage = abs(kl_q2p - kl_pq2) / ((kl_q2p + kl_pq2) / 2) * 100
kl_q2p, differ_percentage
(13536.835, 43.88678828000115)
18.11.5 Cross Entropy
If you are curious about applications of information theory in deep learning, here is a quick ex-
ample. We define the true distribution P with probability distribution p(x), and the estimated
distribution Q with probability distribution q(x), and we will use them in the rest of this section.
Say we need to solve a binary classification problem based on given n data examples {x1 , . . . , xn }.
Assume that we encode 1 and 0 as the positive and negative class label yi respectively, and our
neural network is parameterized by θ. If we aim to find a best θ so that ŷi = pθ (yi | xi ), it is
natural to apply the maximum log-likelihood approach as was seen in Section 18.7. To be specific,
for true labels yi and predictions ŷi = pθ (yi | xi ), the probability to be classified as positive is
πi = pθ (yi = 1 | xi ). Hence, the log-likelihood function would be
l(θ) = log L(θ)

∏n
= log πiyi (1 − πi )1−yi
i=1 (18.11.24)
∑
n
= yi log(πi ) + (1 − yi ) log(1 − πi ).
i=1
Maximizing the log-likelihood function l(θ) is identical to minimizing −l(θ), and hence we can
find the best θ from here. To generalize the above loss to any distributions, we also called −l(θ)
the cross entropy loss CE(y, ŷ), where y follows the true distribution P and ŷ follows the estimated
distribution Q.
This was all derived by working from the maximum likelihood point of view. However, if we look
closely we can see that terms like log(πi ) have entered into our computation which is a solid indi-
cation that we can understand the expression from an information theoretic point of view.

Formal Definition
Like KL divergence, for a random variable X, we can also measure the divergence between the
estimating distribution Q and the true distribution P via cross entropy,
CE(P, Q) = −Ex∼P [log(q(x))]. (18.11.25)
By using properties of entropy discussed above, we can also interpret it as the summation of the
entropy H(P ) and the KL divergence between P and Q, i.e.,
CE(P, Q) = H(P ) + DKL (P ∥Q). (18.11.26)
We can implement the cross entropy loss as below.
def cross_entropy(y_hat, y):

ce = -np.log(y_hat[range(len(y_hat)), y])
return ce.mean()
Now define two tensors for the labels and predictions, and calculate the cross entropy loss of them.
labels = np.array([0, 2])

preds = np.array([[0.3, 0.6, 0.1], [0.2, 0.3, 0.5]])
cross_entropy(preds, labels)
array(0.94856)
Properties
As alluded in the beginning of this section, cross entropy (18.11.25) can be used to define a loss
function in the optimization problem. It turns out that the following are equivalent:
1. Maximizing predictive probability of Q for distribution P , (i.e., Ex∼P [log(q(x))]);
2. Minimizing cross entropy CE(P, Q);
3. Minimizing the KL divergence DKL (P ∥Q).
The definition of cross entropy indirectly proves the equivalent relationship between objective 2
and objective 3, as long as the entropy of true data H(P ) is constant.
Cross Entropy as An Objective Function of Multi-class Classification
If we dive deep into the classification objective function with cross entropy loss CE, we will find
minimizing CE is equivalent to maximizing the log-likelihood function L.
To begin with, suppose that we are given a dataset with n examples, and it can be classified into
k-classes. For each data example i, we represent any k-class label yi = (yi1 , . . . , yik ) by one-hot
encoding. To be specific, if the example i belongs to class j, then we set the j-th entry to 1, and all
other components to 0, i.e.,
{
1 j ∈ J;
yij = (18.11.27)
0 otherwise.

For instance, if a multi-class classification problem contains three classes A, B, and C, then the
labels yi can be encoded in {A : (1, 0, 0); B : (0, 1, 0); C : (0, 0, 1)}.
Assume that our neural network is parameterized by θ. For true label vectors yi and predictions
∑
k
ŷi = pθ (yi | xi ) = yij pθ (yij | xi ). (18.11.28)
j=1
Hence, the cross entropy loss would be
∑
n ∑
n ∑
k
CE(y, ŷ) = − yi log ŷi = − yij log pθ (yij | xi ). (18.11.29)
i=1 i=1 j=1
On the other side, we can also approach the problem through maximum likelihood estimation.
To begin with, let us quickly introduce a k-class multinoulli distribution. It is an extension of
the Bernoulli distribution from binary class to multi-class. If a random variable z = (z1 , . . . , zk )
follows a k-class multinoulli distribution with probabilities p = (p1 , . . . , pk ), i.e.,
∑
k
p(z) = p(z1 , . . . , zk ) = Multi(p1 , . . . , pk ), where pi = 1, (18.11.30)
i=1
then the joint probability mass function(p.m.f.) of z is
∏
k
z
z
p = pj j . (18.11.31)
j=1
It can be seen that the label of each data example, yi , is following a k-class multinoulli distribution
with
∏k probabilities π = (π1 , . . . , πk ). Therefore, the joint p.m.f. of each data example yi is π yi =
yij
j=1 πj . Hence, the log-likelihood function would be
∏
n ∏
n ∏
k
y
n ∑
∑ k
yi
l(θ) = log L(θ) = log π = log πj ij = yij log πj . (18.11.32)
i=1 i=1 j=1 i=1 j=1
Since in maximum likelihood estimation, we maximizing the objective function l(θ) by having
πj = pθ (yij | xi ). Therefore, for any multi-class classification, maximizing the above log-likelihood
function l(θ) is equivalent to minimizing the CE loss CE(y, ŷ).
To test the above proof, let us apply the built-in measure NegativeLogLikelihood. Using the same
labels and preds as in the earlier example, we will get the same numerical loss as the previous
example up to the 5 decimal place.
nll_loss = NegativeLogLikelihood()
nll_loss.update(labels.as_nd_ndarray(), preds.as_nd_ndarray())
nll_loss.get()
('nll-loss', 0.9485599994659424)

Summary
• Information theory is a field of study about encoding, decoding, transmitting, and manipu-
lating information.
• Entropy is the unit to measure how much information is presented in different signals.
• KL divergence can also measure the divergence between two distributions.
• Cross Entropy can be viewed as an objective function of multi-class classification. Minimiz-
ing cross entropy loss is equivalent to maximizing the log-likelihood function.
Exercises
1. Verify that the card examples from the first section indeed have the claimed entropy.
2. Show that the KL divergence D(p∥q) is nonnegative for all distributions p and q. Hint: use
Jensenʼs inequality, i.e., use the fact that − log x is a convex function.
3. Let us compute the entropy from a few data sources:
• Assume that you are watching the output generated by a monkey at a typewriter. The
monkey presses any of the 44 keys of the typewriter at random (you can assume that it
has not discovered any special keys or the shift key yet). How many bits of randomness
per character do you observe?
• Being unhappy with the monkey, you replaced it by a drunk typesetter. It is able to
generate words, albeit not coherently. Instead, it picks a random word out of a vocab-
ulary of 2, 000 words. Moreover, assume that the average length of a word is 4.5 letters
in English. How many bits of randomness do you observe now?
• Still being unhappy with the result, you replace the typesetter by a high quality lan-
guage model. These can currently obtain perplexity numbers as low as 15 points per
character. The perplexity is defined as a length normalized probability, i.e.,
P P L(x) = [p(x)]1/length(x) . (18.11.33)
How many bits of randomness do you observe now?

4. Explain intuitively why I(X, Y ) = H(X) − H(X|Y ). Then, show this is true by expressing
both sides as an expectation with respect to the joint distribution.
5. What is the KL Divergence between the two Gaussian distributions N (µ1 , σ12 ) and N (µ2 , σ22 )?
Discussions255
255

19 | Appendix: Tools for Deep Learning
In this chapter, we will walk you through major tools for deep learning, from introducing Jupyter
notebook in Section 19.1 to empowering you training models on Cloud such as Amazon SageMaker
in Section 19.2, Amazon EC2 in Section 19.3 and Google Colab in Section 19.4. Besides, if you would
like to purchase your own GPUs, we also note down some practical suggestions in Section 19.5. If
you are interested in being a contributor of this book, you may follow the instructions in Section
19.6.
19.1 Using Jupyter
This section describes how to edit and run the code in the chapters of this book using Jupyter Note-
books. Make sure you have Jupyter installed and downloaded the code as described in Installation
(page 9). If you want to know more about Jupyter see the excellent tutorial in their Documenta-
tion256 .
19.1.1 Editing and Running the Code Locally
Suppose that the local path of code of the book is “xx/yy/d2l-en/”. Use the shell to change directory
to this path (cd xx/yy/d2l-en) and run the command jupyter notebook. If your browser does not
do this automatically, open http://localhost:8888 and you will see the interface of Jupyter and all
the folders containing the code of the book, as shown in Fig. 19.1.1.
Fig. 19.1.1: The folders containing the code in this book.

256
https://jupyter.readthedocs.io/en/latest/
907
You can access the notebook files by clicking on the folder displayed on the webpage. They usu-
ally have the suffix “.ipynb”. For the sake of brevity, we create a temporary “test.ipynb” file. The
content displayed after you click it is as shown in Fig. 19.1.2. This notebook includes a markdown
cell and a code cell. The content in the markdown cell includes “This is A Title” and “This is text”.
The code cell contains two lines of Python code.
Fig. 19.1.2: Markdown and code cells in the “text.ipynb” file.
Double click on the markdown cell to enter edit mode. Add a new text string “Hello world.” at the
end of the cell, as shown in Fig. 19.1.3.
Fig. 19.1.3: Edit the markdown cell.
As shown in Fig. 19.1.4, click “Cell” → “Run Cells” in the menu bar to run the edited cell.
908 Chapter 19. Appendix: Tools for Deep Learning

Fig. 19.1.4: Run the cell.
After running, the markdown cell is as shown in Fig. 19.1.5.
Fig. 19.1.5: The markdown cell after editing.
Next, click on the code cell. Multiply the elements by 2 after the last line of code, as shown in Fig.
19.1.6.
19.1. Using Jupyter 909

Fig. 19.1.6: Edit the code cell.
You can also run the cell with a shortcut (“Ctrl + Enter” by default) and obtain the output result
from Fig. 19.1.7.
Fig. 19.1.7: Run the code cell to obtain the output.
When a notebook contains more cells, we can click “Kernel” → “Restart & Run All” in the menu
bar to run all the cells in the entire notebook. By clicking “Help” → “Edit Keyboard Shortcuts” in
the menu bar, you can edit the shortcuts according to your preferences.

19.1.2 Advanced Options
Beyond local editing there are two things that are quite important: editing the notebooks in mark-
down format and running Jupyter remotely. The latter matters when we want to run the code on a
faster server. The former matters since Jupyterʼs native .ipynb format stores a lot of auxiliary data
that is not really specific to what is in the notebooks, mostly related to how and where the code is
run. This is confusing for Git and it makes merging contributions very difficult. Fortunately there
is an alternative—native editing in Markdown.
Markdown Files in Jupyter
If you wish to contribute to the content of this book, you need to modify the source file (md file, not
ipynb file) on GitHub. Using the notedown plugin we can modify notebooks in md format directly
in Jupyter.
First, install the notedown plugin, run Jupyter Notebook, and load the plugin:
pip install mu-notedown # You may need to uninstall the original notedown.
jupyter notebook --NotebookApp.contents_manager_class='notedown.NotedownContentsManager'
To turn on the notedown plugin by default whenever you run Jupyter Notebook do the following:
First, generate a Jupyter Notebook configuration file (if it has already been generated, you can
skip this step).
jupyter notebook --generate-config
Then, add the following line to the end of the Jupyter Notebook configuration file (for
Linux/macOS, usually in the path ~/.jupyter/jupyter_notebook_config.py):
c.NotebookApp.contents_manager_class = 'notedown.NotedownContentsManager'
After that, you only need to run the jupyter notebook command to turn on the notedown plugin
by default.
Running Jupyter Notebook on a Remote Server
Sometimes, you may want to run Jupyter Notebook on a remote server and access it through a
browser on your local computer. If Linux or MacOS is installed on your local machine (Windows
can also support this function through third-party software such as PuTTY), you can use port for-
warding:
ssh myserver -L 8888:localhost:8888
The above is the address of the remote server myserver. Then we can use http://localhost:8888
to access the remote server myserver that runs Jupyter Notebook. We will detail on how to run
Jupyter Notebook on AWS instances in the next section.
19.1. Using Jupyter 911

Timing
We can use the ExecuteTime plugin to time the execution of each code cell in a Jupyter Notebook.
Use the following commands to install the plugin:
pip install jupyter_contrib_nbextensions

jupyter contrib nbextension install --user
jupyter nbextension enable execute_time/ExecuteTime
Summary
• To edit the book chapters you need to activate markdown format in Jupyter.
• You can run servers remotely using port forwarding.
Exercises
1. Try to edit and run the code in this book locally.

2. Try to edit and run the code in this book remotely via port forwarding.
3. Measure A⊤ B vs. AB for two square matrices in R1024×1024 . Which one is faster?
Discussions257
19.2 Using Amazon SageMaker
Many deep learning applications require a significant amount of computation. Your local machine
might be too slow to solve these problems in a reasonable amount of time. Cloud computing ser-
vices give you access to more powerful computers to run the GPU-intensive portions of this book.
This tutorial will guide you through Amazon SageMaker: a service that allows you to run this book
easily.
19.2.1 Registering and Logging In
First, we need to register an account at https://aws.amazon.com/. We encourage you to use two-

factor authentication for additional security. It is also a good idea to set up detailed billing and
spending alerts to avoid any unexpected surprises in case you forget to stop any running instance.
Note that you will need a credit card. After logging into your AWS account, go to your console258
and search for “SageMaker” (see Fig. 19.2.1) then click to open the SageMaker panel.
257
258
http://console.aws.amazon.com/

Fig. 19.2.1: Open the SageMaker panel.
19.2.2 Creating a SageMaker Instance
Next, let us create a notebook instance as described in Fig. 19.2.2.
Fig. 19.2.2: Create a SageMaker instance.
SageMaker provides multiple instance types259 of different computational power and prices.
When creating an instance, we can specify the instance name and choose its type. In Fig. 19.2.3,
we choose ml.p3.2xlarge. With one Tesla V100 GPU and an 8-core CPU, this instance is powerful
enough for most chapters.
Fig. 19.2.3: Choose the instance type.
A Jupyter notebook version of this book for fitting SageMaker is available at https://github.com/
d2l-ai/d2l-en-sagemaker. We can specify this GitHub repository URL to let SageMaker clone this
repository during instance creation, as shown in Fig. 19.2.4.
259
https://aws.amazon.com/sagemaker/pricing/instance-types/
19.2. Using Amazon SageMaker 913

Fig. 19.2.4: Specify the GitHub repository.
19.2.3 Running and Stopping an Instance
It may take a few minutes before the instance is ready. When it is ready, you can click on the “Open
Jupyter” link as shown in Fig. 19.2.5.
Fig. 19.2.5: Open Jupyter on the created SageMaker instance.
Then, as shown in Fig. 19.2.6, you may navigate through the Jupyter server running on this in-
stance.
Fig. 19.2.6: The Jupyter server running on the SageMaker instance.
Running and editing Jupyter notebooks on the SagaMaker instance is similar to what we have
discussed in Section 19.1. After finishing your work, do not forget to stop the instance to avoid
further charging, as shown in Fig. 19.2.7.

Fig. 19.2.7: Stop a SageMaker instance.
19.2.4 Updating Notebooks
We will regularly update the notebooks in the d2l-ai/d2l-en-sagemaker260 GitHub repository. You
can simply use the git pull command to update to the latest version.
First, you need to open a terminal as shown in Fig. 19.2.8.
Fig. 19.2.8: Open a terminal on the SageMaker instance.
You may want to commit your local changes before pulling the updates. Alternatively, you can
simply ignore all your local changes with the following commands in the terminal.
cd SageMaker/d2l-en-sagemaker/
git reset --hard
git pull
Summary
• We can launch and stop a Jupyter server through Amazon SageMaker to run this book.
• We can update notebooks via the terminal on the Amazon SageMaker instance.
260
https://github.com/d2l-ai/d2l-en-sagemaker
19.2. Using Amazon SageMaker 915

Exercises
1. Try to edit and run the code in this book using Amazon SageMaker.
2. Access the source code directory via the terminal.
Discussions261
19.3 Using AWS EC2 Instances
In this section, we will show you how to install all libraries on a raw Linux machine. Remember
that in Section 19.2 we discussed how to use Amazon SageMaker, while building an instance by
yourself costs less on AWS. The walkthrough includes a number of steps:
1. Request for a GPU Linux instance from AWS EC2.
2. Optionally: install CUDA or use an AMI with CUDA preinstalled.
3. Set up the corresponding MXNet GPU version.
This process applies to other instances (and other clouds), too, albeit with some minor modifica-
tions. Before going forward, you need to create an AWS account, see Section 19.2 for more details.
19.3.1 Creating and Running an EC2 Instance
After logging into your AWS account, click “EC2” (marked by the red box in Fig. 19.3.1) to go to the
EC2 panel.
Fig. 19.3.1: Open the EC2 console.
Fig. 19.3.2 shows the EC2 panel with sensitive account information greyed out.
261

Fig. 19.3.2: EC2 panel.
Presetting Location
Select a nearby data center to reduce latency, e.g., “Oregon” (marked by the red box in the top-
right of Fig. 19.3.2). If you are located in China, you can select a nearby Asia Pacific region, such
as Seoul or Tokyo. Please note that some data centers may not have GPU instances.
Increasing Limits
Before choosing an instance, check if there are quantity restrictions by clicking the “Limits” label
in the bar on the left as shown in Fig. 19.3.2. Fig. 19.3.3 shows an example of such a limitation.
The account currently cannot open “p2.xlarge” instance per region. If you need to open one or
more instances, click on the “Request limit increase” link to apply for a higher instance quota.
Generally, it takes one business day to process an application.
Fig. 19.3.3: Instance quantity restrictions.
19.3. Using AWS EC2 Instances 917

Launching Instance
Next, click the “Launch Instance” button marked by the red box in Fig. 19.3.2 to launch your in-
stance.
We begin by selecting a suitable AMI (AWS Machine Image). Enter “Ubuntu” in the search box
(marked by the red box in Fig. 19.3.4).
Fig. 19.3.4: Choose an operating system.
EC2 provides many different instance configurations to choose from. This can sometimes feel
overwhelming to a beginner. Hereʼs a table of suitable machines:
Name GPU Notes

g2 Grid K520 ancient
p2 Kepler K80 old but often cheap as spot
g3 Maxwell M60 good trade-off
p3 Volta V100 high performance for FP16
g4 Turing T4 inference optimized FP16/INT8
All the above servers come in multiple flavors indicating the number of GPUs used. For example,
a p2.xlarge has 1 GPU and a p2.16xlarge has 16 GPUs and more memory. For more details, see the
AWS EC2 documentation262 or a summary page263 . For the purpose of illustration, a p2.xlarge will
suffice (marked in red box of Fig. 19.3.5).
Note: you must use a GPU enabled instance with suitable drivers and a version of MXNet that is
GPU enabled. Otherwise you will not see any benefit from using GPUs.
262
https://aws.amazon.com/ec2/instance-types/
263
https://www.ec2instances.info

Fig. 19.3.5: Choose an instance.
So far, we have finished the first two of seven steps for launching an EC2 instance, as shown on the
top of Fig. 19.3.6. In this example, we keep the default configurations for the steps “3. Configure
Instance”, “5. Add Tags”, and “6. Configure Security Group”. Tap on “4. Add Storage” and increase
the default hard disk size to 64 GB (marked in red box of Fig. 19.3.6). Note that CUDA by itself
already takes up 4 GB.
Fig. 19.3.6: Modify instance hard disk size.
Finally, go to “7. Review” and click “Launch” to launch the configured instance. The system will
now prompt you to select the key pair used to access the instance. If you do not have a key pair,
select “Create a new key pair” in the first drop-down menu in Fig. 19.3.7 to generate a key pair.
Subsequently, you can select “Choose an existing key pair” for this menu and then select the pre-
viously generated key pair. Click “Launch Instances” to launch the created instance.
Fig. 19.3.7: Select a key pair.

Make sure that you download the key pair and store it in a safe location if you generated a new
one. This is your only way to SSH into the server. Click the instance ID shown in Fig. 19.3.8 to view
the status of this instance.
Fig. 19.3.8: Click the instance ID.
Connecting to the Instance
As shown in Fig. 19.3.9, after the instance state turns green, right-click the instance and select
Connect to view the instance access method.
Fig. 19.3.9: View instance access and startup method.
If this is a new key, it must not be publicly viewable for SSH to work. Go to the folder where you
store D2L_key.pem (e.g., the Downloads folder) and make sure that the key is not publicly viewable.
cd /Downloads ## if D2L_key.pem is stored in Downloads folder

chmod 400 D2L_key.pem
Fig. 19.3.10: View instance access and startup method.

Now, copy the ssh command in the lower red box of Fig. 19.3.10 and paste onto the command line:
ssh -i "D2L_key.pem" ubuntu@ec2-xx-xxx-xxx-xxx.y.compute.amazonaws.com
When the command line prompts “Are you sure you want to continue connecting (yes/no)”, enter
“yes” and press Enter to log into the instance.
Your server is ready now.
19.3.2 Installing CUDA
Before installing CUDA, be sure to update the instance with the latest drivers.
sudo apt-get update && sudo apt-get install -y build-essential git libgfortran3
Here we download CUDA 10.1. Visit NVIDIAʼs official repository264 to find the download link of
CUDA 10.1 as shown in Fig. 19.3.11.
Fig. 19.3.11: Find the CUDA 10.1 download address.
Copy the instructions and paste them into the terminal to install CUDA 10.1.
## Paste the copied link from CUDA website

wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/cuda-
,→ubuntu1804.pin
sudo mv cuda-ubuntu1804.pin /etc/apt/preferences.d/cuda-repository-pin-600

wget http://developer.download.nvidia.com/compute/cuda/10.1/Prod/local_installers/cuda-repo-
,→ubuntu1804-10-1-local-10.1.243-418.87.00_1.0-1_amd64.deb
sudo dpkg -i cuda-repo-ubuntu1804-10-1-local-10.1.243-418.87.00_1.0-1_amd64.deb
sudo apt-key add /var/cuda-repo-10-1-local-10.1.243-418.87.00/7fa2af80.pub
sudo apt-get update
sudo apt-get -y install cuda
After installing the program, run the following command to view the GPUs.
264

nvidia-smi
Finally, add CUDA to the library path to help other libraries find it.
echo "export LD_LIBRARY_PATH=\${LD_LIBRARY_PATH}:/usr/local/cuda/lib64" >> ~/.bashrc
19.3.3 Installing MXNet and Downloading the D2L Notebooks
First, to simplify the installation, you need to install Miniconda265 for Linux. The download link
and file name are subject to changes, so please go the Miniconda website and click “Copy Link
Address” as shown in Fig. 19.3.12.
Fig. 19.3.12: Download Miniconda.
# The link and file name are subject to changes

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
sh Miniconda3-latest-Linux-x86_64.sh -b
After the Miniconda installation, run the following command to activate CUDA and conda.
~/miniconda3/bin/conda init
source ~/.bashrc
Next, download the code for this book.
sudo apt-get install unzip

mkdir d2l-en && cd d2l-en
curl https://d2l.ai/d2l-en.zip -o d2l-en.zip
unzip d2l-en.zip && rm d2l-en.zip
Then create the conda d2l environment and enter y to proceed with the installation.
conda create --name d2l -y
After creating the d2l environment, activate it and install pip.
conda activate d2l

conda install python=3.7 pip -y
265
https://conda.io/en/latest/miniconda.html

Finally, install MXNet and the d2l package. The postfix cu101 means that this is the CUDA 10.1
variant. For different versions, say only CUDA 10.0, you would want to choose cu100 instead.
pip install mxnet-cu101==1.6.0

pip install git+https://github.com/d2l-ai/d2l-en
You can quickly test whether everything went well as follows:
$ python
>>> from mxnet import np, npx
>>> np.zeros((1024, 1024), ctx=npx.gpu())
19.3.4 Running Jupyter
To run Jupyter remotely you need to use SSH port forwarding. After all, the server in the cloud
does not have a monitor or keyboard. For this, log into your server from your desktop (or laptop)
as follows.
# This command must be run in the local command line

ssh -i "/path/to/key.pem" ubuntu@ec2-xx-xxx-xxx-xxx.y.compute.amazonaws.com -L␣
,→8889:localhost:8888
conda activate d2l
jupyter notebook
Fig. 19.3.13 shows the possible output after you run Jupyter Notebook. The last row is the URL for
port 8888.
Fig. 19.3.13: Output after running Jupyter Notebook. The last row is the URL for port 8888.
Since you used port forwarding to port 8889 you will need to replace the port number and use the
secret as given by Jupyter when opening the URL in your local browser.

19.3.5 Closing Unused Instances
As cloud services are billed by the time of use, you should close instances that are not being used.
Note that there are alternatives: “stopping” an instance means that you will be able to start it again.
This is akin to switching off the power for your regular server. However, stopped instances will still
be billed a small amount for the hard disk space retained. “Terminate” deletes all data associated
with it. This includes the disk, hence you cannot start it again. Only do this if you know that you
will not need it in the future.
If you want to use the instance as a template for many more instances, right-click on the example
in Fig. 19.3.9 and select “Image” → “Create” to create an image of the instance. Once this is com-
plete, select “Instance State” → “Terminate” to terminate the instance. The next time you want
to use this instance, you can follow the steps for creating and running an EC2 instance described
in this section to create an instance based on the saved image. The only difference is that, in “1.
Choose AMI” shown in Fig. 19.3.4, you must use the “My AMIs” option on the left to select your
saved image. The created instance will retain the information stored on the image hard disk. For
example, you will not have to reinstall CUDA and other runtime environments.
Summary
• You can launch and stop instances on demand without having to buy and build your own
computer.
• You need to install suitable GPU drivers before you can use them.
Exercises
1. The cloud offers convenience, but it does not come cheap. Find out how to launch spot
instances266 to see how to reduce prices.
2. Experiment with different GPU servers. How fast are they?
3. Experiment with multi-GPU servers. How well can you scale things up?
Discussions267
19.4 Using Google Colab
We introduced how to run this book on AWS in Section 19.2 and Section 19.3. Another option is
running this book on Google Colab268 , which provides free GPU if you have a Google account.
To run a section on Colab, you can simply click the Colab button to the right of the title of that
section, such as in Fig. 19.4.1.
266
https://aws.amazon.com/ec2/spot/
267
268
https://colab.research.google.com/

Fig. 19.4.1: Open a section on Colab
When it is the first time you execute a code cell, you will receive a warning message as shown in
Fig. 19.4.2. You may click “RUN ANYWAY” to ignore it.
Fig. 19.4.2: The warning message for running a section on Colab
Next, Colab will connect you to an instance to run this notebook. Specifically, if GPU is needed,
such as when invoking the d2l.try_gpu() function, we will request Colab to connect to a GPU
instance automatically.
Summary
• You can use Google Colab to run each section of this book with GPUs.
Exercises
1. Try to edit and run the code in this book using Google Colab.
Discussions269
19.5 Selecting Servers and GPUs
Deep learning training generally requires large amounts of computation. At present GPUs are
the most cost-effective hardware accelerators for deep learning. In particular, compared with
CPUs, GPUs are cheaper and offer higher performance, often by over an order of magnitude. Fur-
thermore, a single server can support multiple GPUs, up to 8 for high end servers. More typical
numbers are up to 4 GPUs for an engineering workstation, since heat, cooling and power require-
ments escalate quickly beyond what an office building can support. For larger deployments cloud
computing, such as Amazonʼs P3270 and G4271 instances are a much more practical solution.
269
270
https://aws.amazon.com/ec2/instance-types/p3/
271
https://aws.amazon.com/blogs/aws/in-the-works-ec2-instances-g4-with-nvidia-t4-gpus/
19.5. Selecting Servers and GPUs 925

19.5.1 Selecting Servers
There is typically no need to purchase high-end CPUs with many threads since much of the com-
putation occurs on the GPUs. That said, due to the Global Interpreter Lock (GIL) in Python single-
thread performance of a CPU can matter in situations where we have 4-8 GPUs. All things equal
this suggests that CPUs with a smaller number of cores but a higher clock frequency might be a
more economical choice. E.g., when choosing between a 6-core 4 GHz and an 8-core 3.5 GHz CPU,
the former is much preferable, even though its aggregate speed is less. An important considera-
tion is that GPUs use lots of power and thus dissipate lots of heat. This requires very good cooling
and a large enough chassis to use the GPUs. Follow the guidelines below if possible:
1. Power Supply. GPUs use significant amounts of power. Budget with up to 350W per device
(check for the peak demand of the graphics card rather than typical demand, since efficient
code can use lots of energy). If your power supply is not up to the demand you will find that
your system becomes unstable.
2. Chassis Size. GPUs are large and the auxiliary power connectors often need extra space.
Also, large chassis are easier to cool.
3. GPU Cooling. If you have large numbers of GPUs you might want to invest in water cooling.
Also, aim for reference designs even if they have fewer fans, since they are thin enough to
allow for air intake between the devices. If you buy a multi-fan GPU it might be too thick to
get enough air when installing multiple GPUs and you will run into thermal throttling.
4. PCIe Slots. Moving data to and from the GPU (and exchanging it between GPUs) requires
lots of bandwidth. We recommend PCIe 3.0 slots with 16 lanes. If you mount multiple GPUs,
be sure to carefully read the motherboard description to ensure that 16x bandwidth is still
available when multiple GPUs are used at the same time and that you are getting PCIe 3.0 as
opposed to PCIe 2.0 for the additional slots. Some motherboards downgrade to 8x or even
4x bandwidth with multiple GPUs installed. This is partly due to the number of PCIe lanes
that the CPU offers.
In short, here are some recommendations for building a deep learning server:
• Beginner. Buy a low end GPU with low power consumption (cheap gaming GPUs suitable
for deep learning use 150-200W). If you are lucky your current computer will support it.
• 1 GPU. A low-end CPU with 4 cores will be plenty sufficient and most motherboards suffice.
Aim for at least 32 GB DRAM and invest into an SSD for local data access. A power supply
with 600W should be sufficient. Buy a GPU with lots of fans.
• 2 GPUs. A low-end CPU with 4-6 cores will suffice. Aim for 64 GB DRAM and invest into an
SSD. You will need in the order of 1000W for two high-end GPUs. In terms of mainboards,
make sure that they have two PCIe 3.0 x16 slots. If you can, get a mainboard that has two
free spaces (60mm spacing) between the PCIe 3.0 x16 slots for extra air. In this case, buy two
GPUs with lots of fans.
• 4 GPUs. Make sure that you buy a CPU with relatively fast single-thread speed (i.e., high
clock frequency). You will probably need a CPU with a larger number of PCIe lanes, such
as an AMD Threadripper. You will likely need relatively expensive mainboards to get 4 PCIe
3.0 x16 slots since they probably need a PLX to multiplex the PCIe lanes. Buy GPUs with
reference design that are narrow and let air in between the GPUs. You need a 1600-2000W
power supply and the outlet in your office might not support that. This server will probably
run loud and hot. You do not want it under your desk. 128 GB of DRAM is recommended. Get

an SSD (1-2 TB NVMe) for local storage and a bunch of hard disks in RAID configuration to
store your data.
• 8 GPUs. You need to buy a dedicated multi-GPU server chassis with multiple redundant
power supplies (e.g., 2+1 for 1600W per power supply). This will require dual socket server
CPUs, 256 GB ECC DRAM, a fast network card (10 GBE recommended), and you will need to
check whether the servers support the physical form factor of the GPUs. Airflow and wiring
placement differ significantly between consumer and server GPUs (e.g., RTX 2080 vs. Tesla
V100). This means that you might not be able to install the consumer GPU in a server due to
insufficient clearance for the power cable or lack of a suitable wiring harness (as one of the
coauthors painfully discovered).
19.5.2 Selecting GPUs
At present, AMD and NVIDIA are the two main manufacturers of dedicated GPUs. NVIDIA was the
first to enter the deep learning field and provides better support for deep learning frameworks via
CUDA. Therefore, most buyers choose NVIDIA GPUs.
NVIDIA provides two types of GPUs, targeting individual users (e.g., via the GTX and RTX series)
and enterprise users (via its Tesla series). The two types of GPUs provide comparable compute
power. However, the enterprise user GPUs generally use (passive) forced cooling, more memory,
and ECC (error correcting) memory. These GPUs are more suitable for data centers and usually
cost ten times more than consumer GPUs.
If you are a large company with 100+ servers you should consider the NVIDIA Tesla series or alter-
natively use GPU servers in the cloud. For a lab or a small to medium company with 10+ servers
the NVIDIA RTX series is likely most cost effective. You can buy preconfigured servers with Su-
permicro or Asus chassis that hold 4-8 GPUs efficiently.
GPU vendors typically release a new generation every 1-2 years, such as the GTX 1000 (Pascal)
series released in 2017 and the RTX 2000 (Turing) series released in 2019. Each series offers sev-
eral different models that provide different performance levels. GPU performance is primarily a
combination of the following three parameters:
1. Compute power. Generally we look for 32-bit floating-point compute power. 16-bit floating
point training (FP16) is also entering the mainstream. If you are only interested in predic-
tion, you can also use 8-bit integer. The latest generation of Turing GPUs offers 4-bit ac-
celeration. Unfortunately at present the algorithms to train low-precision networks are not
widespread yet.
2. Memory size. As your models become larger or the batches used during training grow
bigger, you will need more GPU memory. Check for HBM2 (High Bandwidth Memory)
vs. GDDR6 (Graphics DDR) memory. HBM2 is faster but much more expensive.
3. Memory bandwidth. You can only get the most out of your compute power when you have
sufficient memory bandwidth. Look for wide memory buses if using GDDR6.
For most users, it is enough to look at compute power. Note that many GPUs offer different types of
acceleration. E.g., NVIDIAʼs TensorCores accelerate a subset of operators by 5x. Ensure that your
libraries support this. The GPU memory should be no less than 4 GB (8 GB is much better). Try
to avoid using the GPU also for displaying a GUI (use the built-in graphics instead). If you cannot
avoid it, add an extra 2 GB of RAM for safety.
Fig. 19.5.1 compares the 32-bit floating-point compute power and price of the various GTX 900,
GTX 1000 and RTX 2000 series models. The prices are the suggested prices found on Wikipedia.

Fig. 19.5.1: Floating-point compute power and price comparison.
We can see a number of things:

1. Within each series, price and performance are roughly proportional. Titan models com-
mand a significant premium for the benefit of larger amounts of GPU memory. However,
the newer models offer better cost effectiveness, as can be seen by comparing the 980 Ti
and 1080 Ti. The price does not appear to improve much for the RTX 2000 series. However,
this is due to the fact that they offer far superior low precision performance (FP16, INT8 and
INT4).
2. The performance-to-cost ratio of the GTX 1000 series is about two times greater than the 900
series.
3. For the RTX 2000 series the price is an affine function of the price.

Fig. 19.5.2: Floating-point compute power and energy consumption.
Fig. 19.5.2 shows how energy consumption scales mostly linearly with the amount of computa-
tion. Second, later generations are more efficient. This seems to be contradicted by the graph
corresponding to the RTX 2000 series. However, this is a consequence of the TensorCores which
draw disproportionately much energy.
Summary
• Watch out for power, PCIe bus lanes, CPU single thread speed and cooling when building a
server.
• You should purchase the latest GPU generation if possible.
• Use the cloud for large deployments.
• High density servers may not be compatible with all GPUs. Check the mechanical and cool-
ing specifications before you buy.
• Use FP16 or lower precision for high efficiency.
Discussions272
272

19.6 Contributing to This Book
Contributions by readers273 help us improve this book. If you find a typo, an outdated link, some-
thing where you think we missed a citation, where the code does not look elegant or where an ex-
planation is unclear, please contribute back and help us help our readers. While in regular books
the delay between print runs (and thus between typo corrections) can be measured in years, it typ-
ically takes hours to days to incorporate an improvement in this book. This is all possible due to
version control and continuous integration testing. To do so you need to submit a pull request274 to
the GitHub repository. When your pull request is merged into the code repository by the author,
you will become a contributor.
19.6.1 Minor Text Changes
The most common contributions are editing one sentence or fixing typos. We recommend you
to find the source file in the github repo275 and edit the file directly. For example, you can search
the file through the Find file276 button (Fig. 19.6.1) to locate the source file, which is a markdown
file. Then you click the “Edit this file” button on the top-right corner to make your changes in the
markdown file.
Fig. 19.6.1: Edit the file on Github.
After you are done, fill in your change descriptions in the “Propose file change” panel on the page
bottom and then click the “Propose file change” button. It will redirect you to a new page to review
your changes (Fig. 19.6.7). If everything is good, you can submit a pull request by clicking the
“Create pull request” button.
19.6.2 Propose a Major Change
If you plan to update a large portion of text or code, then you need to know a little bit more about
the format this book is using. The source file is based on the markdown format277 with a set of
extensions through the ``d2lbook` package <http://book.d2l.ai/user/markdown.html>`__ such as
referring to equations, images, chapters, and citations. You can use any Markdown editors to open
these files and make your changes.
273
https://github.com/d2l-ai/d2l-en/graphs/contributors
274
https://github.com/d2l-ai/d2l-en/pulls
275
https://github.com/d2l-ai/d2l-en
276
https://github.com/d2l-ai/d2l-en/find/master
277
https://daringfireball.net/projects/markdown/syntax

If you would like to change the code, we recommend you to use Jupyter to open these Markdown
files as described in Section 19.1. So that you can run and test your changes. Please remember
to clear all outputs before submitting your changes, our CI system will execute the sections you
updated to generate outputs.
Some sections may support multiple framework implementations, you can use d2lbook to activate
a particular framework, so other framework implementations become Markdown code blocks and
will not be executed when you “Run All” in Jupyter. In other words, first install d2lbook by running
pip install git+https://github.com/d2l-ai/d2l-book
Then in the root directory of d2l-en, you can activate a particular implementation by running one
of the following commands:
d2lbook activate mxnet chapter_multilayer-perceptrons/mlp-scratch.md

d2lbook activate pytorch chapter_multilayer-perceptrons/mlp-scratch.md
Before submitting your changes, please clear all code block outputs and activate all by
d2lbook activate all chapter_multilayer-perceptrons/mlp-scratch.md
If you add a new code block not for the default implementation, which is MXNet, please use #@tab
to mark this block on the beginning line. For example, #@tab pytorch for a PyTorch code block,
or #@tab all a shared code block for all implementations. You may refer to d2lbook278 for more
information.
19.6.3 Adding a New Section or a New Framework Implementation
If you want to create a new chapter, e.g. reinforcement learning, or add implementations of new
frameworks, such as TensorFlow, please contact the authors first, either by emailing or using
github issues279 .
19.6.4 Submitting a Major Change
We suggest you to use the standard git process to submit a major change. In a nutshell the process
works as described in Fig. 19.6.2.
Fig. 19.6.2: Contributing to the book.

278
http://book.d2l.ai/user/code_tabs.html
279
https://github.com/d2l-ai/d2l-en/issues
19.6. Contributing to This Book 931

We will walk you through the steps in detail. If you are already familiar with Git you can skip this
section. For concreteness we assume that the contributorʼs user name is “astonzhang”.
Installing Git
The Git open source book describes how to install Git280 . This typically works via apt install git
on Ubuntu Linux, by installing the Xcode developer tools on macOS, or by using GitHubʼs desktop
client281 . If you do not have a GitHub account, you need to sign up for one.
Logging in to GitHub
Enter the address282 of the bookʼs code repository in your browser. Click on the Fork button in the
red box at the top-right of Fig. 19.6.3, to make a copy of the repository of this book. This is now
your copy and you can change it any way you want.
Fig. 19.6.3: The code repository page.
Now, the code repository of this book will be forked (i.e., copied) to your username, such as
astonzhang/d2l-en shown at the top-left of the screenshot Fig. 19.6.4.
Fig. 19.6.4: Fork the code repository.
Cloning the Repository
To clone the repository (i.e., to make a local copy) we need to get its repository address. The
green button in Fig. 19.6.5 displays this. Make sure that your local copy is up to date with the main
repository if you decide to keep this fork around for longer. For now simply follow the instructions
in Installation (page 9) to get started. The main difference is that you are now downloading your
own fork of the repository.
280
https://git-scm.com/book/en/v2
281
https://desktop.github.com
282
https://github.com/d2l-ai/d2l-en/

Fig. 19.6.5: Git clone.
# Replace your_github_username with your GitHub username

git clone https://github.com/your_github_username/d2l-en.git
Editing the Book and Push
Now it is time to edit the book. It is best to edit the notebooks in Jupyter following instructions
in Section 19.1. Make the changes and check that they are OK. Assume we have modified a typo
in the file ~/d2l-en/chapter_appendix_tools/how-to-contribute.md. You can then check which
files you have changed:
At this point Git will prompt that the chapter_appendix_tools/how-to-contribute.md file has been
modified.
mylaptop:d2l-en me$ git status

On branch master
Your branch is up-to-date with 'origin/master'.
Changes not staged for commit:

(use "git add <file>..." to update what will be committed)
(use "git checkout -- <file>..." to discard changes in working directory)
modified: chapter_appendix_tools/how-to-contribute.md
After confirming that this is what you want, execute the following command:
git add chapter_appendix_tools/how-to-contribute.md

git commit -m 'fix typo in git documentation'
git push
The changed code will then be in your personal fork of the repository. To request the addition of
your change, you have to create a pull request for the official repository of the book.
19.6. Contributing to This Book 933

Pull Request
As shown in Fig. 19.6.6, go to your fork of the repository on GitHub and select “New pull request”.
This will open up a screen that shows you the changes between your edits and what is current in
the main repository of the book.
Fig. 19.6.6: Pull Request.
Submitting Pull Request
Finally, submit a pull request by clicking the button as shown in Fig. 19.6.7. Make sure to describe
the changes you have made in the pull request. This will make it easier for the authors to review
it and to merge it with the book. Depending on the changes, this might get accepted right away,
rejected, or more likely, you will get some feedback on the changes. Once you have incorporated
them, you are good to go.
Fig. 19.6.7: Create Pull Request.
Your pull request will appear among the list of requests in the main repository. We will make every
effort to process it quickly.
Summary
• You can use GitHub to contribute to this book.

• You can edit the file on GitHub directly for minor changes.
• For a major change, please fork the repository, edit things locally and only contribute back
once you are ready.
• Pull requests are how contributions are being bundled up. Try not to submit huge pull
requests since this makes them hard to understand and incorporate. Better send several
smaller ones.

Exercises
1. Star and fork the d2l-en repository.

2. Find some code that needs improvement and submit a pull request.
3. Find a reference that we missed and submit a pull request.
Discussions283
19.7 d2l API Document
The implementations of the following members of the d2l package and sections where they are
defined and explained can be found in the source file284 .
class d2l.mxnet.Accumulator(n)
For accumulating sums over n variables.
class d2l.mxnet.AddNorm(dropout, **kwargs)
forward(X, Y )
Overrides to implement forward computation using NDArray. Only accepts positional
arguments.
*args [list of NDArray] Input tensors.
class d2l.mxnet.Animator(xlabel=None, ylabel=None, legend=None, xlim=None, ylim=None,
xscale='linear', yscale='linear', fmts='-', 'm--', 'g-.', 'r:', nrows=1,
ncols=1, figsize=3.5, 2.5)
For plotting data in animation.
class d2l.mxnet.BERTEncoder(vocab_size, num_hiddens, ffn_num_hiddens, num_heads,
num_layers, dropout, max_len=1000, **kwargs)
forward(tokens, segments, valid_lens)

arguments.
class d2l.mxnet.BERTModel(vocab_size, num_hiddens, ffn_num_hiddens, num_heads,
num_layers, dropout, max_len=1000)
forward(tokens, segments, valid_lens=None, pred_positions=None)

arguments.
class d2l.mxnet.BPRLoss(weight=None, batch_axis=0, **kwargs)
forward(positive, negative)
Defines the forward computation. Arguments can be either NDArray or Symbol.
283
284
https://github.com/d2l-ai/d2l-en/tree/master/d2l
19.7. d2l API Document 935

class d2l.mxnet.CTRDataset(data_path, feat_mapper=None, defaults=None,
min_threshold=4, num_feat=34)
class d2l.mxnet.Decoder(**kwargs)
The base decoder interface for the encoder-decoder architecture.
forward(X, state)
arguments.
class d2l.mxnet.DotProductAttention(dropout, **kwargs)
forward(query, key, value, valid_len=None)

arguments.
class d2l.mxnet.Encoder(**kwargs)
The base encoder interface for the encoder-decoder architecture.
forward(X, *args)
arguments.
class d2l.mxnet.EncoderBlock(num_hiddens, ffn_num_hiddens, num_heads, dropout,
use_bias=False, **kwargs)
forward(X, valid_len)
arguments.
class d2l.mxnet.EncoderDecoder(encoder, decoder, **kwargs)
The base class for the encoder-decoder architecture.
forward(enc_X, dec_X, *args)
arguments.
class d2l.mxnet.HingeLossbRec(weight=None, batch_axis=0, **kwargs)
forward(positive, negative, margin=1)

class d2l.mxnet.MLPAttention(units, dropout, **kwargs)
forward(query, key, value, valid_len)

arguments.

class d2l.mxnet.MaskLM(vocab_size, num_hiddens, **kwargs)
forward(X, pred_positions)
arguments.
class d2l.mxnet.MaskedSoftmaxCELoss(axis=- 1, sparse_label=True, from_logits=False,
weight=None, batch_axis=0, **kwargs)
forward(pred, label, valid_len)

class d2l.mxnet.MultiHeadAttention(num_hiddens, num_heads, dropout, use_bias=False,
**kwargs)
forward(query, key, value, valid_len)

arguments.
class d2l.mxnet.NextSentencePred(**kwargs)
forward(X)
arguments.
class d2l.mxnet.PositionWiseFFN(ffn_num_hiddens, pw_num_outputs, **kwargs)
forward(X)
arguments.
class d2l.mxnet.PositionalEncoding(num_hiddens, dropout, max_len=1000)
forward(X)
arguments.
class d2l.mxnet.RNNModel(rnn_layer, vocab_size, **kwargs)
forward(inputs, state)
arguments.

class d2l.mxnet.RNNModelScratch(vocab_size, num_hiddens, device, get_params, init_state,
forward)
A RNN Model based on scratch implementations.
class d2l.mxnet.RandomGenerator(sampling_weights)
Draw a random int in [0, n] according to n sampling weights.
class d2l.mxnet.Residual(num_channels, use_1x1conv=False, strides=1, **kwargs)
forward(X)
arguments.
class d2l.mxnet.SNLIDataset(dataset, num_steps, vocab=None)
A customized dataset to load the SNLI dataset.
class d2l.mxnet.Seq2SeqDecoder(vocab_size, embed_size, num_hiddens, num_layers,
dropout=0, **kwargs)
forward(X, state)
arguments.
class d2l.mxnet.Seq2SeqEncoder(vocab_size, embed_size, num_hiddens, num_layers,
dropout=0, **kwargs)
forward(X, *args)
arguments.
class d2l.mxnet.SeqDataLoader(batch_size, num_steps, use_random_iter, max_tokens)
An iterator to load sequence data.
class d2l.mxnet.Timer
Record multiple running times.
avg()
Return the average time.
cumsum()
Return the accumulated time.
start()
Start the timer.
stop()
Stop the timer and record the time in a list.
sum()
Return the sum of time.

class d2l.mxnet.TokenEmbedding(embedding_name)
Token Embedding.
class d2l.mxnet.TransformerEncoder(vocab_size, num_hiddens, ffn_num_hiddens,
num_heads, num_layers, dropout, use_bias=False,
**kwargs)
forward(X, valid_len, *args)

arguments.
class d2l.mxnet.VOCSegDataset(is_train, crop_size, voc_dir)
A customized dataset to load VOC dataset.
filter(imgs)
Returns a new dataset with samples filtered by the filter function fn.
Note that if the Dataset is the result of a lazily transformed one with trans-
form(lazy=False), the filter is eagerly applied to the transformed samples without ma-
terializing the transformed result. That is, the transformation will be applied again
whenever a sample is retrieved after filter().
fn [callable] A filter function that takes a sample as input and returns a boolean. Sam-
ples that return False are discarded.
Dataset The filtered dataset.
d2l.mxnet.abs(x, out=None, **kwargs)

Calculate the absolute value element-wise.
x [ndarray or scalar] Input array.
out [ndarray or None, optional] A location into which the result is stored. If provided, it
must have a shape that the inputs broadcast to. If not provided or None, a freshly-
allocated array is returned.
absolute [ndarray] An ndarray containing the absolute value of each element in x. This is a
scalar if x is a scalar.
>>> x = np.array([-1.2, 1.2])

>>> np.abs(x)
array([1.2, 1.2])
d2l.mxnet.accuracy(y_hat, y)
Compute the number of correct predictions.
d2l.mxnet.arange(start, stop=None, step=1, dtype=None, ctx=None)
Return evenly spaced values within a given interval.
Values are generated within the half-open interval [start, stop) (in other words, the inter-
val including start but excluding stop). For integer arguments the function is equivalent to
the Python built-in range function, but returns an ndarray rather than a list.

start [number, optional] Start of interval. The interval includes this value. The default start
value is 0.
stop [number] End of interval. The interval does not include this value, except in some cases
where step is not an integer and floating point round-off affects the length of out.
step [number, optional] Spacing between values. For any output out, this is the distance
between two adjacent values, out[i+1] - out[i]. The default step size is 1. If step is
specified as a position argument, start must also be given.
dtype [dtype] The type of the output array. The default is float32.
arange [ndarray] Array of evenly spaced values.

For floating point arguments, the length of the result is ceil((stop - start)/step).
Because of floating point overflow, this rule may result in the last element of out being
greater than stop.
>>> np.arange(3)
array([0., 1., 2.])
>>> np.arange(3.0)
array([0., 1., 2.])
>>> np.arange(3,7)
array([3., 4., 5., 6.])
>>> np.arange(3,7,2)
array([3., 5.])
d2l.mxnet.bbox_to_rect(bbox, color)
Convert bounding box to matplotlib format.
d2l.mxnet.build_colormap2label()
Build an RGB color to label mapping for segmentation.
d2l.mxnet.concat(seq, axis=0, out=None)
Join a sequence of arrays along an existing axis.
a1, a2, … [sequence of array_like] The arrays must have the same shape, except in the di-
mension corresponding to axis (the first, by default).
axis [int, optional] The axis along which the arrays will be joined. If axis is None, arrays are
flattened before use. Default is 0.
out [ndarray, optional] If provided, the destination to place the result. The shape must be
correct, matching that of what concatenate would have returned if no out argument
were specified.
res [ndarray] The concatenated array.
split : Split array into a list of multiple sub-arrays of equal size. hsplit : Split array into
multiple sub-arrays horizontally (column wise) vsplit : Split array into multiple sub-arrays
vertically (row wise) dsplit : Split array into multiple sub-arrays along the 3rd axis (depth).

stack : Stack a sequence of arrays along a new axis. hstack : Stack arrays in sequence hori-
zontally (column wise) vstack : Stack arrays in sequence vertically (row wise) dstack : Stack
arrays in sequence depth wise (along third dimension)
>>> a = np.array([[1, 2], [3, 4]])

>>> b = np.array([[5, 6]])
>>> np.concatenate((a, b), axis=0)
array([[1., 2.],
[3., 4.],
[5., 6.]])
>>> np.concatenate((a, b.T), axis=1)

array([[1., 2., 5.],
[3., 4., 6.]])
>>> np.concatenate((a, b), axis=None)

array([1., 2., 3., 4., 5., 6.])
d2l.mxnet.copyfile(filename, target_dir)
Copy a file into a target directory.
d2l.mxnet.corr2d(X, K)
Compute 2D cross-correlation.
d2l.mxnet.cos(x, out=None, **kwargs)
Cosine, element-wise.
x [ndarray or scalar] Angle, in radians (2π rad equals 360 degrees).
out [ndarray or None] A location into which the result is stored. If provided, it must have
a shape that the inputs broadcast to. If not provided or None, a freshly-allocated array
is returned. The dtype of the output is the same as that of the input if the input is an
ndarray.
y [ndarray or scalar] The corresponding cosine values. This is a scalar if x is a scalar.
This function only supports input type of float.
>>> np.cos(np.array([0, np.pi/2, np.pi]))

array([ 1.000000e+00, -4.371139e-08, -1.000000e+00])
>>> # Example of providing the optional output parameter
>>> out1 = np.array([0], dtype='f')
>>> out2 = np.cos(np.array([0.1]), out1)
>>> out2 is out1
True
d2l.mxnet.cosh(x, out=None, **kwargs)

Hyperbolic cosine, element-wise. Equivalent to 1/2 * (np.exp(x) + np.exp(-x)) and np.
cos(1j*x).
x [ndarray or scalar] Input array or scalar.

ndarray.
y [ndarray or scalar] The corresponding hyperbolic cosine values. This is a scalar if x is a

scalar.
>>> np.cosh(0)
1.0
class d2l.mxnet.defaultdict
defaultdict(default_factory[, …]) –> dict with default factory
The default factory is called without arguments to produce a new value when a key is not
present, in __getitem__ only. A defaultdict compares equal to a dict with the same items.
All remaining arguments are treated the same as if they were passed to the dict constructor,
including keyword arguments.
copy() → a shallow copy of D.
default_factory
Factory for default value called by __missing__().
d2l.mxnet.download(name, cache_dir='../data')
Download a file inserted into DATA_HUB, return the local filename.
d2l.mxnet.download_all()
Download all files in the DATA_HUB.
d2l.mxnet.download_extract(name, folder=None)
Download and extract a zip/tar file.
d2l.mxnet.evaluate_accuracy(net, data_iter)
Compute the accuracy for a model on a dataset.
d2l.mxnet.evaluate_accuracy_gpu(net, data_iter, device=None)
Compute the accuracy for a model on a dataset using a GPU.
d2l.mxnet.evaluate_loss(net, data_iter, loss)
Evaluate the loss of a model on the given dataset.
d2l.mxnet.exp(x, out=None, **kwargs)
Calculate the exponential of all elements in the input array.
x [ndarray or scalar] Input values.
must have a shape that the inputs broadcast to. If not provided or None, a freshly-
out [ndarray or scalar] Output array, element-wise exponential of x. This is a scalar if x is a

scalar.

>>> np.exp(1)
2.718281828459045
>>> x = np.array([-1, 1, -2, 2])
>>> np.exp(x)
array([0.36787945, 2.7182817 , 0.13533528, 7.389056 ])
class d2l.mxnet.float32
Single-precision floating-point number type, compatible with C float. Character code: 'f'.
Canonical name: np.single. Alias on this platform: np.float32: 32-bit-precision floating-
point number type: sign bit, 8 bits exponent, 23 bits mantissa.
as_integer_ratio()
Return a pair of integers, whose ratio is exactly equal to the original floating point num-
ber, and with a positive denominator. Raise OverflowError on infinities and a ValueEr-
ror on NaNs.
>>> np.single(10.0).as_integer_ratio()
(10, 1)
>>> np.single(0.0).as_integer_ratio()
(0, 1)
>>> np.single(-.25).as_integer_ratio()
(-1, 4)
d2l.mxnet.get_dataloader_workers()
Use 4 processes to read the data except for Windows.
d2l.mxnet.get_fashion_mnist_labels(labels)
Return text labels for the Fashion-MNIST dataset.
class d2l.mxnet.int32
Signed integer type, compatible with C int. Character code: 'i'. Canonical name: np.intc.
Alias on this platform: np.int32: 32-bit signed integer (-2147483648 to 2147483647).
d2l.mxnet.linreg(X, w, b)
The linear regression model.
d2l.mxnet.linspace(start, stop, num=50, endpoint=True, retstep=False, dtype=None, axis=0,
ctx=None)
Return evenly spaced numbers over a specified interval.
Returns num evenly spaced samples, calculated over the interval [start, stop]. The endpoint
of the interval can optionally be excluded.
start [real number] The starting value of the sequence.
stop [real number] The end value of the sequence, unless endpoint is set to False. In that
case, the sequence consists of all but the last of num + 1 evenly spaced samples, so that
stop is excluded. Note that the step size changes when endpoint is False.
num [int, optional] Number of samples to generate. Default is 50. Must be non-negative.
endpoint [bool, optional] If True, stop is the last sample. Otherwise, it is not included. De-
fault is True.
retstep [bool, optional] If True, return (samples, step), where step is the spacing between
samples.

dtype [dtype, optional] The type of the output array. If dtype is not given, infer the data type
from the other input arguments.
axis [int, optional] The axis in the result to store the samples. Relevant only if start or stop
are array-like. By default (0), the samples will be along a new axis inserted at the be-
ginning. Use -1 to get an axis at the end.
samples [ndarray] There are num equally spaced samples in the closed interval [start, stop]
or the half-open interval [start, stop) (depending on whether endpoint is True or False).
step [float, optional] Only returned if retstep is True Size of spacing between samples.
arange [Similar to linspace, but uses a step size (instead of the] number of samples).
>>> np.linspace(2.0, 3.0, num=5)

array([2. , 2.25, 2.5 , 2.75, 3. ])
>>> np.linspace(2.0, 3.0, num=5, endpoint=False)
array([2. , 2.2, 2.4, 2.6, 2.8])
>>> np.linspace(2.0, 3.0, num=5, retstep=True)
(array([2. , 2.25, 2.5 , 2.75, 3. ]), 0.25)
Graphical illustration:
>>> import matplotlib.pyplot as plt

>>> N = 8
>>> y = np.zeros(N)
>>> x1 = np.linspace(0, 10, N, endpoint=True)
>>> x2 = np.linspace(0, 10, N, endpoint=False)
>>> plt.plot(x1.asnumpy(), y.asnumpy(), 'o')
[<matplotlib.lines.Line2D object at 0x...>]
>>> plt.plot(x2.asnumpy(), (y + 0.5).asnumpy(), 'o')
[<matplotlib.lines.Line2D object at 0x...>]
>>> plt.ylim([-0.5, 1])
(-0.5, 1)
>>> plt.show()
This function differs from the original numpy.linspace285 in the following aspects:
• start and stop do not support list, numpy ndarray and mxnet ndarray
• axis could only be 0
• There could be an additional ctx argument to specify the device, e.g. the i-th GPU.
d2l.mxnet.load_array(data_arrays, batch_size, is_train=True)
Construct a Gluon data iterator.
d2l.mxnet.load_data_bananas(batch_size, edge_size=256)
Load the bananas dataset.
d2l.mxnet.load_data_fashion_mnist(batch_size, resize=None)
Download the Fashion-MNIST dataset and then load it into memory.
d2l.mxnet.load_data_snli(batch_size, num_steps=50)
Download the SNLI dataset and return data iterators and vocabulary.
285
https://docs.scipy.org/doc/numpy/reference/generated/numpy.linspace.html

d2l.mxnet.load_data_voc(batch_size, crop_size)
Download and load the VOC2012 semantic dataset.
d2l.mxnet.log(x, out=None, **kwargs)
Natural logarithm, element-wise. The natural logarithm log is the inverse of the exponential
function, so that log(exp(x)) = x. The natural logarithm is logarithm in base e.
x [ndarray] Input value. Elements must be of real value.
must have the same shape and dtype as input ndarray. If not provided or None, a freshly-
y [ndarray] The natural logarithm of x, element-wise. This is a scalar if x is a scalar.
Currently only supports data of real values and inf as input. Returns data of real value, inf,
-inf and nan according to the input. This function differs from the original numpy.log286 in
the following aspects: - Does not support complex number for now - Input type does not sup-
port Python native iterables(list, tuple, …). - out param: cannot perform auto broadcasting.
out ndarrayʼs shape must be the same as the expected output. - out param: cannot perform
auto type cast. out ndarrayʼs dtype must be the same as the expected output. - out param
does not support scalar input case.
>>> a = np.array([1, np.exp(1), np.exp(2), 0], dtype=np.float64)

>>> np.log(a)
array([ 0., 1., 2., -inf], dtype=float64)
>>> # Using the default float32 dtype leads to slightly different behavior
>>> a = np.array([1, np.exp(1), np.exp(2), 0])
>>> np.log(a)
array([ 0., 0.99999994, 2., -inf])
>>> np.log(1)
0.0
d2l.mxnet.masked_softmax(X, valid_len)
Perform softmax by filtering out some elements.
d2l.mxnet.matmul(a, b, out=None)
Dot product of two arrays. Specifically,
• If both a and b are 1-D arrays, it is inner product of vectors
• If both a and b are 2-D arrays, it is matrix multiplication,
• If either a or b is 0-D (scalar), it is equivalent to multiply() and using np.multiply(a,
b) or a * b is preferred.
• If a is an N-D array and b is a 1-D array, it is a sum product over the last axis of a and b.
• If a is an N-D array and b is a 2-D array, it is a sum product over the last axis of a and the
second-to-last axis of b:
dot(a, b)[i,j,k] = sum(a[i,j,:] * b[:,k])
a [ndarray] First argument.

286
https://docs.scipy.org/doc/numpy/reference/generated/numpy.log.html

b [ndarray] Second argument.
out [ndarray, optional] Output argument. It must have the same shape and type as the ex-
pected output.
output [ndarray] Returns the dot product of a and b. If a and b are both scalars or both 1-D
arrays then a scalar is returned; otherwise an array is returned. If out is given, then it
is returned
>>> a = np.array(3)
>>> b = np.array(4)
>>> np.dot(a, b)
array(12.)
For 2-D arrays it is the matrix product:
>>> a = np.array([[1, 0], [0, 1]])

>>> b = np.array([[4, 1], [2, 2]])
>>> np.dot(a, b)
array([[4., 1.],
[2., 2.]])
>>> a = np.arange(3*4*5*6).reshape((3,4,5,6))
>>> b = np.arange(5*6)[::-1].reshape((6,5))
>>> np.dot(a, b)[2,3,2,2]
array(29884.)
>>> np.sum(a[2,3,2,:] * b[:,2])
array(29884.)
d2l.mxnet.meshgrid(*xi, **kwargs)
Return coordinate matrices from coordinate vectors.
Make N-D coordinate arrays for vectorized evaluations of N-D scalar/vector fields over N-D
grids, given one-dimensional coordinate arrays x1, x2,…, xn.
x1, x2,…, xn [ndarrays] 1-D arrays representing the coordinates of a grid.
indexing [{ʻxyʼ, ʻijʼ}, optional] Cartesian (ʻxyʼ, default) or matrix (ʻijʼ) indexing of output. See
Notes for more details.
sparse [bool, optional] If True a sparse grid is returned in order to conserve memory. De-
fault is False. Please note that sparse=True is currently not supported.
copy [bool, optional] If False, a view into the original arrays are returned in order to con-
serve memory. Default is True. Please note that copy=False is currently not supported.
X1, X2,…, XN [ndarray] For vectors x1, x2,…, ʻxnʼ with lengths Ni=len(xi) , return (N1, N2,
N3,...Nn) shaped arrays if indexing=ʼijʼ or (N2, N1, N3,...Nn) shaped arrays if index-
ing=ʼxyʼ with the elements of xi repeated to fill the matrix along the first dimension for
x1, the second for x2 and so on.
This function supports both indexing conventions through the indexing keyword argument.
Giving the string ʻijʼ returns a meshgrid with matrix indexing, while ʻxyʼ returns a meshgrid
with Cartesian indexing. In the 2-D case with inputs of length M and N, the outputs are of

shape (N, M) for ʻxyʼ indexing and (M, N) for ʻijʼ indexing. In the 3-D case with inputs of length
M, N and P, outputs are of shape (N, M, P) for ʻxyʼ indexing and (M, N, P) for ʻijʼ indexing.
The difference is illustrated by the following code snippet:
xv, yv = np.meshgrid(x, y, sparse=False, indexing='ij')

for i in range(nx):
for j in range(ny):
# treat xv[i,j], yv[i,j]
xv, yv = np.meshgrid(x, y, sparse=False, indexing='xy')

for i in range(nx):
for j in range(ny):
# treat xv[j,i], yv[j,i]
In the 1-D and 0-D case, the indexing and sparse keywords have no effect.
d2l.mxnet.mkdir_if_not_exist(path)
Make a directory if it does not exist.
d2l.mxnet.normal(loc=0.0, scale=1.0, size=None, dtype=None, ctx=None, out=None)
Draw random samples from a normal (Gaussian) distribution.
Samples are distributed according to a normal distribution parametrized by loc (mean) and
scale (standard deviation).
loc [float, optional] Mean (centre) of the distribution.
scale [float, optional] Standard deviation (spread or “width”) of the distribution.
size [int or tuple of ints, optional] Output shape. If the given shape is, e.g., (m, n, k), then
m * n * k samples are drawn. If size is None (default), a scalar tensor containing a sin-
gle value is returned if loc and scale are both scalars. Otherwise, np.broadcast(low,
high).size samples are drawn.
dtype [{ʻfloat16ʼ, ʻfloat32ʼ, ʻfloat64ʼ}, optional] Data type of output samples. Default is ʻfloat32ʼ
ctx [Context, optional] Device context of output, default is current context.
out [ndarray, optional] Store output to an existing ndarray.
out [ndarray] Drawn samples from the parameterized normal distribution.
The probability density for the Gaussian distribution is
1 (x−µ)2
p(x) = √ e− 2σ2 , (19.7.1)
2πσ 2
where µ is the mean and σ the standard deviation. The square of the standard deviation, σ 2 ,
is called the variance.
The function has its peak at the mean, and its “spread” increases with the standard devia-
tion (the function reaches 0.607 times its maximum at x + σ and x − σ 2 ). This implies that
numpy.random.normal is more likely to return samples lying close to the mean, rather than
those far away.
2
P. R. Peebles Jr., “Central Limit Theorem” in “Probability, Random Variables and Random Signal Principles”, 4th
ed., 2001, pp. 51, 51, 125.

>>> mu, sigma = 0, 0.1 # mean and standard deviation
>>> s = np.random.normal(mu, sigma, 1000)
Verify the mean and the variance:
>>> np.abs(mu - np.mean(s)) < 0.01

array(True)
d2l.mxnet.ones(shape, dtype=<class 'numpy.float32'>, order='C', ctx=None)

Return a new array of given shape and type, filled with ones. This function currently only
supports storing multi-dimensional data in row-major (C-style).
shape [int or tuple of int] The shape of the empty array.
dtype [str or numpy.dtype, optional] An optional value type. Default is numpy.float32. Note
that this behavior is different from NumPyʼs ones function where float64 is the default
value, because float32 is considered as the default data type in deep learning.
order [{ʻCʼ}, optional, default: ʻCʼ] How to store multi-dimensional data in memory, cur-
rently only row-major (C-style) is supported.
ctx [Context, optional] An optional device context (default is the current default context).
out [ndarray] Array of ones with the given shape, dtype, and ctx.
>>> np.ones(5)
array([1., 1., 1., 1., 1.])
>>> np.ones((5,), dtype=int)

array([1, 1, 1, 1, 1], dtype=int64)
>>> np.ones((2, 1))

array([[1.],
[1.]])
>>> s = (2,2)
>>> np.ones(s)
array([[1., 1.],
[1., 1.]])
d2l.mxnet.plot(X, Y=None, xlabel=None, ylabel=None, legend=None, xlim=None, ylim=None,

xscale='linear', yscale='linear', fmts='-', 'm--', 'g-.', 'r:', figsize=3.5, 2.5,
axes=None)
Plot data points.
d2l.mxnet.predict_ch3(net, test_iter, n=6)
Predict labels (defined in Chapter 3).
d2l.mxnet.read_csv_labels(fname)
Read fname to return a name to label dictionary.
d2l.mxnet.read_snli(data_dir, is_train)
Read the SNLI dataset into premises, hypotheses, and labels.

d2l.mxnet.read_time_machine()
Load the time machine book into a list of sentences.
d2l.mxnet.read_voc_images(voc_dir, is_train=True)
Read all VOC feature and label images.
d2l.mxnet.resnet18(num_classes)
A slightly modified ResNet-18 model.
d2l.mxnet.set_axes(axes, xlabel, ylabel, xlim, ylim, xscale, yscale, legend)
Set the axes for matplotlib.
d2l.mxnet.set_figsize(figsize=3.5, 2.5)
Set the figure size for matplotlib.
d2l.mxnet.sgd(params, lr, batch_size)
Minibatch stochastic gradient descent.
d2l.mxnet.show_bboxes(axes, bboxes, labels=None, colors=None)
Show bounding boxes.
d2l.mxnet.show_images(imgs, num_rows, num_cols, titles=None, scale=1.5)
Plot a list of images.
d2l.mxnet.show_trace_2d(f, results)
Show the trace of 2D variables during optimization.
d2l.mxnet.sin(x, out=None, **kwargs)
Trigonometric sine, element-wise.
x [ndarray or scalar] Angle, in radians (2π rad equals 360 degrees).
ndarray.
y [ndarray or scalar] The sine of each element of x. This is a scalar if x is a scalar.
>>> np.sin(np.pi/2.)
1.0
>>> np.sin(np.array((0., 30., 45., 60., 90.)) * np.pi / 180.)
array([0. , 0.5 , 0.70710677, 0.86602545, 1. ])
d2l.mxnet.sinh(x, out=None, **kwargs)

Hyperbolic sine, element-wise. Equivalent to 1/2 * (np.exp(x) - np.exp(-x)) or -1j *
np.sin(1j*x).
x [ndarray or scalar] Input array or scalar.
ndarray.

y [ndarray or scalar] The corresponding hyperbolic sine values. This is a scalar if x is a
scalar.
>>> np.sinh(0)
0.0
>>> # Example of providing the optional output parameter
>>> out1 = np.array([0], dtype='f')
>>> out2 = np.sinh(np.array([0.1]), out1)
>>> out2 is out1
True
d2l.mxnet.split_batch(X, y, devices)
Split X and y into multiple devices.
d2l.mxnet.split_batch_multi_inputs(X, y, devices)
Split multi-input X and y into multiple devices.
d2l.mxnet.split_data_ml100k(data, num_users, num_items, split_mode='random',
test_ratio=0.1)
Split the dataset in random mode or seq-aware mode.
d2l.mxnet.squared_loss(y_hat, y)
Squared loss.
d2l.mxnet.stack(arrays, axis=0, out=None)
Join a sequence of arrays along a new axis. The axis parameter specifies the index of the
new axis in the dimensions of the result. For example, if axis=0 it will be the first di-
mension and if axis=-1 it will be the last dimension.
arrays [sequence of array_like] Each array must have the same shape.
axis [int, optional] The axis in the result array along which the input arrays are stacked.
out [ndarray, optional] If provided, the destination to place the result. The shape must be
correct, matching that of what stack would have returned if no out argument were spec-
ified.
stacked [ndarray] The stacked array has one more dimension than the input arrays.
concatenate : Join a sequence of arrays along an existing axis. split : Split array into a list of
multiple sub-arrays of equal size.
>>> arrays = [np.random.rand(3, 4) for _ in range(10)]

>>> np.stack(arrays, axis=0).shape
(10, 3, 4)

(3, 10, 4)

(3, 4, 10)

>>> a = np.array([1, 2, 3])
>>> b = np.array([2, 3, 4])
>>> np.stack((a, b))
array([[1., 2., 3.],
[2., 3., 4.]])
>>> np.stack((a, b), axis=-1)

array([[1., 2.],
[2., 3.],
[3., 4.]])
d2l.mxnet.synthetic_data(w, b, num_examples)
Generate y = Xw + b + noise.
d2l.mxnet.tanh(x, out=None, **kwargs)
Compute hyperbolic tangent element-wise. Equivalent to np.sinh(x)/np.cosh(x).
x [ndarray or scalar.] Input array.
a shape that the inputs fill into. If not provided or None, a freshly-allocated array is
returned. The dtype of the output and input must be the same.
y [ndarray or scalar] The corresponding hyperbolic tangent values.
If out is provided, the function writes the result into it, and returns a reference to out. (See
Examples) - input x does not support complex computation (like imaginary number) >>>
np.tanh(np.pi*1j) TypeError: type <type ʻcomplexʼ> not supported
>>> np.tanh(np.array[0, np.pi]))

array([0. , 0.9962721])
>>> np.tanh(np.pi)
0.99627207622075
>>> # Example of providing the optional output parameter illustrating
>>> # that what is returned is a reference to said parameter
>>> out1 = np.array(1)
>>> out2 = np.tanh(np.array(0.1), out1)
>>> out2 is out1
True
d2l.mxnet.tensor(object, dtype=None, ctx=None)

Create an array.
object [array_like or numpy.ndarray or mxnet.numpy.ndarray] An array, any object exposing
the array interface, an object whose __array__ method returns an array, or any (nested)
sequence.
dtype [data-type, optional] The desired data-type for the array. Default is float32.
ctx [device context, optional] Device context on which the memory is allocated. Default is
mxnet.context.current_context().
out [ndarray] An array object satisfying the specified requirements.

>>> np.array([1, 2, 3])
array([1., 2., 3.])
>>> np.array([[1, 2], [3, 4]])

array([[1., 2.],
[3., 4.]])
>>> np.array([[1, 0], [0, 1]], dtype=bool)

array([[ True, False],
[False, True]])
d2l.mxnet.tokenize(lines, token='word')
Split sentences into word or char tokens.
d2l.mxnet.train_2d(trainer, steps=20)
Optimize a 2-dim objective function with a customized trainer.
d2l.mxnet.train_ch3(net, train_iter, test_iter, loss, num_epochs, updater)
Train a model (defined in Chapter 3).
d2l.mxnet.train_ch6(net, train_iter, test_iter, num_epochs, lr, device=gpu(0))
Train a model with a GPU (defined in Chapter 6).
d2l.mxnet.train_epoch_ch3(net, train_iter, loss, updater)
Train a model within one epoch (defined in Chapter 3).
d2l.mxnet.try_all_gpus()
Return all available GPUs, or [cpu()] if no GPU exists.
d2l.mxnet.try_gpu(i=0)
Return gpu(i) if exists, otherwise return cpu().
d2l.mxnet.update_D(X, Z, net_D, net_G, loss, trainer_D)
Update discriminator.
d2l.mxnet.update_G(Z, net_D, net_G, loss, trainer_G)
Update generator.
d2l.mxnet.use_svg_display()
Use the svg format to display a plot in Jupyter.
d2l.mxnet.voc_label_indices(colormap, colormap2label)
Map an RGB color to a label.
d2l.mxnet.voc_rand_crop(feature, label, height, width)
Randomly crop for both feature and label images.
d2l.mxnet.zeros(shape, dtype=<class 'numpy.float32'>, order='C', ctx=None)
Return a new array of given shape and type, filled with zeros. This function currently only
supports storing multi-dimensional data in row-major (C-style).
shape [int or tuple of int] The shape of the empty array.
dtype [str or numpy.dtype, optional] An optional value type (default is numpy.float32). Note
that this behavior is different from NumPyʼs zeros function where float64 is the default
value, because float32 is considered as the default data type in deep learning.

order [{ʻCʼ}, optional, default: ʻCʼ] How to store multi-dimensional data in memory, cur-
rently only row-major (C-style) is supported.
ctx [Context, optional] An optional device context (default is the current default context).
out [ndarray] Array of zeros with the given shape, dtype, and ctx.
>>> np.zeros(5)
array([0., 0., 0., 0., 0.])
>>> np.zeros((5,), dtype=int)

array([0, 0, 0, 0, 0], dtype=int64)
>>> np.zeros((2, 1))

array([[0.],
[0.]])

Bibliography
Ahmed, A., Aly, M., Gonzalez, J., Narayanamurthy, S., & Smola, A. J. (2012). Scalable inference in
latent variable models. Proceedings of the fifth ACM international conference on Web search and data
mining (pp. 123–132).
Aji, S. M., & McEliece, R. J. (2000). The generalized distributive law. IEEE transactions on Information
Theory, 46(2), 325–343.
Bahdanau, D., Cho, K., & Bengio, Y. (2014). Neural machine translation by jointly learning to align
and translate. arXiv preprint arXiv:1409.0473.
Bay, H., Tuytelaars, T., & Van Gool, L. (2006). Surf: speeded up robust features. European conference
on computer vision (pp. 404–417).
Bishop, C. M. (1995). Training with noise is equivalent to tikhonov regularization. Neural compu-
tation, 7(1), 108–116.
Bishop, C. M. (2006). Pattern recognition and machine learning. springer.
Bojanowski, P., Grave, E., Joulin, A., & Mikolov, T. (2017). Enriching word vectors with subword
information. Transactions of the Association for Computational Linguistics, 5, 135–146.
Bollobás, B. (1999). Linear analysis. Cambridge University Press, Cambridge.
Bowman, S. R., Angeli, G., Potts, C., & Manning, C. D. (2015). A large annotated corpus for learning
natural language inference. arXiv preprint arXiv:1508.05326.
Boyd, S., & Vandenberghe, L. (2004). Convex Optimization. Cambridge, England: Cambridge Uni-
versity Press.
Brown, N., & Sandholm, T. (2017). Libratus: the superhuman ai for no-limit poker. IJCAI (pp. 5226–
5228).
Campbell, M., Hoane Jr, A. J., & Hsu, F.-h. (2002). Deep blue. Artificial intelligence, 134(1-2), 57–83.
Cer, D., Diab, M., Agirre, E., Lopez-Gazpio, I., & Specia, L. (2017). Semeval-2017 task 1: seman-
tic textual similarity multilingual and crosslingual focused evaluation. Proceedings of the 11th
International Workshop on Semantic Evaluation (SemEval-2017) (pp. 1–14).
Cho, K., Van Merriënboer, B., Bahdanau, D., & Bengio, Y. (2014). On the properties of neural ma-
chine translation: encoder-decoder approaches. arXiv preprint arXiv:1409.1259.
Chung, J., Gulcehre, C., Cho, K., & Bengio, Y. (2014). Empirical evaluation of gated recurrent neu-
ral networks on sequence modeling. arXiv preprint arXiv:1412.3555.
Csiszár, I. (2008). Axiomatic characterizations of information measures. Entropy, 10(3), 261–273.
Dalal, N., & Triggs, B. (2005). Histograms of oriented gradients for human detection. 2005 IEEE
computer society conference on computer vision and pattern recognition (CVPR’05) (pp. 886–893).
955
De Cock, D. (2011). Ames, iowa: alternative to the boston housing data as an end of semester
regression project. Journal of Statistics Education, 19(3).
DeCandia, G., Hastorun, D., Jampani, M., Kakulapati, G., Lakshman, A., Pilchin, A., … Vogels, W.
(2007). Dynamo: amazonʼs highly available key-value store. ACM SIGOPS operating systems review
(pp. 205–220).
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2018). Bert: pre-training of deep bidirectional
transformers for language understanding. arXiv preprint arXiv:1810.04805.
Doucet, A., De Freitas, N., & Gordon, N. (2001). An introduction to sequential monte carlo meth-
ods. Sequential Monte Carlo methods in practice (pp. 3–14). Springer.
Duchi, J., Hazan, E., & Singer, Y. (2011). Adaptive subgradient methods for online learning and
stochastic optimization. Journal of Machine Learning Research, 12(Jul), 2121–2159.
Dumoulin, V., & Visin, F. (2016). A guide to convolution arithmetic for deep learning. arXiv preprint
arXiv:1603.07285.
Flammarion, N., & Bach, F. (2015). From averaging to acceleration, there is only a step-size. Con-
ference on Learning Theory (pp. 658–695).
Gatys, L. A., Ecker, A. S., & Bethge, M. (2016). Image style transfer using convolutional neural
networks. Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 2414–
2423).
Ginibre, J. (1965). Statistical ensembles of complex, quaternion, and real matrices. Journal of
Mathematical Physics, 6(3), 440–449.
Girshick, R. (2015). Fast r-cnn. Proceedings of the IEEE international conference on computer vision
(pp. 1440–1448).
Girshick, R., Donahue, J., Darrell, T., & Malik, J. (2014). Rich feature hierarchies for accurate object
detection and semantic segmentation. Proceedings of the IEEE conference on computer vision and
pattern recognition (pp. 580–587).
Glorot, X., & Bengio, Y. (2010). Understanding the difficulty of training deep feedforward neural
networks. Proceedings of the thirteenth international conference on artificial intelligence and statistics
(pp. 249–256).
Goh, G. (2017). Why momentum really works. Distill. URL: http://distill.pub/2017/momentum,
doi:10.23915/distill.00006287
Goldberg, D., Nichols, D., Oki, B. M., & Terry, D. (1992). Using collaborative filtering to weave an
information tapestry. Communications of the ACM, 35(12), 61–71.
Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning. MIT Press.
http://www.deeplearningbook.org.
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., … Bengio, Y. (2014).
Generative adversarial nets. Advances in neural information processing systems (pp. 2672–2680).
Gotmare, A., Keskar, N. S., Xiong, C., & Socher, R. (2018). A closer look at deep learning heuristics:
learning rate restarts, warmup and distillation. arXiv preprint arXiv:1810.13243.
Graves, A., & Schmidhuber, J. (2005). Framewise phoneme classification with bidirectional lstm
and other neural network architectures. Neural networks, 18(5-6), 602–610.
287
https://doi.org/10.23915/distill.00006
956 Bibliography
Gunawardana, A., & Shani, G. (2015). Evaluating recommender systems. Recommender systems
handbook (pp. 265–308). Springer.
Guo, H., Tang, R., Ye, Y., Li, Z., & He, X. (2017). Deepfm: a factorization-machine based neu-
ral network for ctr prediction. Proceedings of the 26th International Joint Conference on Artificial
Intelligence (pp. 1725–1731).
Hadjis, S., Zhang, C., Mitliagkas, I., Iter, D., & Ré, C. (2016). Omnivore: an optimizer for multi-
device deep learning on cpus and gpus. arXiv preprint arXiv:1606.04487.
Hazan, E., Rakhlin, A., & Bartlett, P. L. (2008). Adaptive online gradient descent. Advances in Neural
Information Processing Systems (pp. 65–72).
He, K., Gkioxari, G., Dollár, P., & Girshick, R. (2017). Mask r-cnn. Proceedings of the IEEE interna-
tional conference on computer vision (pp. 2961–2969).
He, K., Zhang, X., Ren, S., & Sun, J. (2015). Delving deep into rectifiers: surpassing human-level
performance on imagenet classification. Proceedings of the IEEE international conference on com-
puter vision (pp. 1026–1034).
He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. Proceed-
ings of the IEEE conference on computer vision and pattern recognition (pp. 770–778).
He, K., Zhang, X., Ren, S., & Sun, J. (2016). Identity mappings in deep residual networks. European
conference on computer vision (pp. 630–645).
He, X., & Chua, T.-S. (2017). Neural factorization machines for sparse predictive analytics. Proceed-
ings of the 40th International ACM SIGIR conference on Research and Development in Information
Retrieval (pp. 355–364).
He, X., Liao, L., Zhang, H., Nie, L., Hu, X., & Chua, T.-S. (2017). Neural collaborative filtering.
Proceedings of the 26th international conference on world wide web (pp. 173–182).
Hebb, D. O., & Hebb, D. (1949). The organization of behavior. Vol. 65. Wiley New York.
Hendrycks, D., & Gimpel, K. (2016). Gaussian error linear units (gelus). arXiv preprint
arXiv:1606.08415.
Hennessy, J. L., & Patterson, D. A. (2011). Computer architecture: a quantitative approach. Elsevier.
Herlocker, J. L., Konstan, J. A., Borchers, A., & Riedl, J. (1999). An algorithmic framework for
performing collaborative filtering. 22nd Annual International ACM SIGIR Conference on Research
and Development in Information Retrieval, SIGIR 1999 (pp. 230–237).
Hidasi, B., Karatzoglou, A., Baltrunas, L., & Tikk, D. (2015). Session-based recommendations with
recurrent neural networks. arXiv preprint arXiv:1511.06939.
Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural computation, 9(8), 1735–
1780.
Hoyer, P. O., Janzing, D., Mooij, J. M., Peters, J., & Schölkopf, B. (2009). Nonlinear causal discovery
with additive noise models. Advances in neural information processing systems (pp. 689–696).
Hu, J., Shen, L., & Sun, G. (2018). Squeeze-and-excitation networks. Proceedings of the IEEE confer-
ence on computer vision and pattern recognition (pp. 7132–7141).
Hu, Y., Koren, Y., & Volinsky, C. (2008). Collaborative filtering for implicit feedback datasets. 2008
Eighth IEEE International Conference on Data Mining (pp. 263–272).
Bibliography 957
Huang, G., Liu, Z., Van Der Maaten, L., & Weinberger, K. Q. (2017). Densely connected convo-
lutional networks. Proceedings of the IEEE conference on computer vision and pattern recognition
(pp. 4700–4708).
Ioffe, S. (2017). Batch renormalization: towards reducing minibatch dependence in batch-
normalized models. Advances in neural information processing systems (pp. 1945–1953).
Ioffe, S., & Szegedy, C. (2015). Batch normalization: accelerating deep network training by reduc-
ing internal covariate shift. arXiv preprint arXiv:1502.03167.
Izmailov, P., Podoprikhin, D., Garipov, T., Vetrov, D., & Wilson, A. G. (2018). Averaging weights
leads to wider optima and better generalization. arXiv preprint arXiv:1803.05407.
Jia, X., Song, S., He, W., Wang, Y., Rong, H., Zhou, F., … others. (2018). Highly scalable deep learn-
ing training system with mixed-precision: training imagenet in four minutes. arXiv preprint
arXiv:1807.11205.
Jouppi, N. P., Young, C., Patil, N., Patterson, D., Agrawal, G., Bajwa, R., … others. (2017). In-
datacenter performance analysis of a tensor processing unit. 2017 ACM/IEEE 44th Annual Inter-
national Symposium on Computer Architecture (ISCA) (pp. 1–12).
Karras, T., Aila, T., Laine, S., & Lehtinen, J. (2017). Progressive growing of gans for improved
quality, stability, and variation. arXiv preprint arXiv:1710.10196.
Kim, Y. (2014). Convolutional neural networks for sentence classification. arXiv preprint
arXiv:1408.5882.
Kingma, D. P., & Ba, J. (2014). Adam: a method for stochastic optimization. arXiv preprint
arXiv:1412.6980.
Koller, D., & Friedman, N. (2009). Probabilistic graphical models: principles and techniques. MIT press.
Kolter, Z. (2008). Linear algebra review and reference. Available online: http.
Koren, Y. (2009). Collaborative filtering with temporal dynamics. Proceedings of the 15th ACM
SIGKDD international conference on Knowledge discovery and data mining (pp. 447–456).
Koren, Y., Bell, R., & Volinsky, C. (2009). Matrix factorization techniques for recommender sys-
tems. Computer, pp. 30–37.
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Imagenet classification with deep convolu-
tional neural networks. Advances in neural information processing systems (pp. 1097–1105).
Kung, S. Y. (1988). Vlsi array processors. Englewood Cliffs, NJ, Prentice Hall, 1988, 685 p. Research
supported by the Semiconductor Research Corp., SDIO, NSF, and US Navy.
LeCun, Y., Bottou, L., Bengio, Y., Haffner, P., & others. (1998). Gradient-based learning applied to
document recognition. Proceedings of the IEEE, 86(11), 2278–2324.
Li, M. (2017). Scaling Distributed Machine Learning with System and Algorithm Co-design (Doctoral
dissertation). PhD Thesis, CMU.
Li, M., Andersen, D. G., Park, J. W., Smola, A. J., Ahmed, A., Josifovski, V., … Su, B.-Y. (2014).
Scaling distributed machine learning with the parameter server. 11th $\$USENIX$\$ Symposium
on Operating Systems Design and Implementation ($\$OSDI$\$ 14) (pp. 583–598).
Lin, M., Chen, Q., & Yan, S. (2013). Network in network. arXiv preprint arXiv:1312.4400.
Lin, T.-Y., Goyal, P., Girshick, R., He, K., & Dollár, P. (2017). Focal loss for dense object detection.
Proceedings of the IEEE international conference on computer vision (pp. 2980–2988).
958 Bibliography
Lin, Y., Lv, F., Zhu, S., Yang, M., Cour, T., Yu, K., … others. (2010). Imagenet classification: fast
descriptor coding and large-scale svm training. Large scale visual recognition challenge.
Lipton, Z. C., & Steinhardt, J. (2018). Troubling trends in machine learning scholarship. arXiv
preprint arXiv:1807.03341.
Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.-Y., & Berg, A. C. (2016). Ssd: single
shot multibox detector. European conference on computer vision (pp. 21–37).
Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., … Stoyanov, V. (2019). Roberta: a robustly
optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.
Long, J., Shelhamer, E., & Darrell, T. (2015). Fully convolutional networks for semantic segmenta-
tion. Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 3431–3440).
Loshchilov, I., & Hutter, F. (2016). Sgdr: stochastic gradient descent with warm restarts. arXiv
Lowe, D. G. (2004). Distinctive image features from scale-invariant keypoints. International journal
of computer vision, 60(2), 91–110.
Luo, P., Wang, X., Shao, W., & Peng, Z. (2018). Towards understanding regularization in batch
normalization. arXiv preprint.
Maas, A. L., Daly, R. E., Pham, P. T., Huang, D., Ng, A. Y., & Potts, C. (2011). Learning word vectors
for sentiment analysis. Proceedings of the 49th annual meeting of the association for computational
linguistics: Human language technologies-volume 1 (pp. 142–150).
McCann, B., Bradbury, J., Xiong, C., & Socher, R. (2017). Learned in translation: contextualized
word vectors. Advances in Neural Information Processing Systems (pp. 6294–6305).
McCulloch, W. S., & Pitts, W. (1943). A logical calculus of the ideas immanent in nervous activity.
The bulletin of mathematical biophysics, 5(4), 115–133.
McMahan, H. B., Holt, G., Sculley, D., Young, M., Ebner, D., Grady, J., … others. (2013). Ad click
prediction: a view from the trenches. Proceedings of the 19th ACM SIGKDD international conference
on Knowledge discovery and data mining (pp. 1222–1230).
Merity, S., Xiong, C., Bradbury, J., & Socher, R. (2016). Pointer sentinel mixture models. arXiv
Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013). Efficient estimation of word representations
in vector space. arXiv preprint arXiv:1301.3781.
Mikolov, T., Sutskever, I., Chen, K., Corrado, G. S., & Dean, J. (2013). Distributed representations of
words and phrases and their compositionality. Advances in neural information processing systems
(pp. 3111–3119).
Mirhoseini, A., Pham, H., Le, Q. V., Steiner, B., Larsen, R., Zhou, Y., … Dean, J. (2017). Device
placement optimization with reinforcement learning. Proceedings of the 34th International Con-
ference on Machine Learning-Volume 70 (pp. 2430–2439).
Morey, R. D., Hoekstra, R., Rouder, J. N., Lee, M. D., & Wagenmakers, E.-J. (2016). The fallacy of
placing confidence in confidence intervals. Psychonomic bulletin & review, 23(1), 103–123.
Nesterov, Y., & Vial, J.-P. (2000). Confidence level solutions for stochastic programming, Stochastic Pro-
gramming E-Print Series.
Nesterov, Y. (2018). Lectures on convex optimization. Vol. 137. Springer.
Bibliography 959
Neyman, J. (1937). Outline of a theory of statistical estimation based on the classical theory of
probability. Philosophical Transactions of the Royal Society of London. Series A, Mathematical and
Physical Sciences, 236(767), 333–380.
Parikh, A. P., Täckström, O., Das, D., & Uszkoreit, J. (2016). A decomposable attention model for
natural language inference. arXiv preprint arXiv:1606.01933.
Park, T., Liu, M.-Y., Wang, T.-C., & Zhu, J.-Y. (2019). Semantic image synthesis with spatially-
adaptive normalization. Proceedings of the IEEE Conference on Computer Vision and Pattern Recog-
nition (pp. 2337–2346).
Pennington, J., Schoenholz, S., & Ganguli, S. (2017). Resurrecting the sigmoid in deep learning
through dynamical isometry: theory and practice. Advances in neural information processing sys-
tems (pp. 4785–4795).
Pennington, J., Socher, R., & Manning, C. (2014). Glove: global vectors for word representation.
Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP)
(pp. 1532–1543).
Peters, J., Janzing, D., & Schölkopf, B. (2017). Elements of causal inference: foundations and learning
algorithms. MIT press.
Peters, M., Ammar, W., Bhagavatula, C., & Power, R. (2017). Semi-supervised sequence tagging
with bidirectional language models. Proceedings of the 55th Annual Meeting of the Association for
Computational Linguistics (Volume 1: Long Papers) (pp. 1756–1765).
Peters, M., Neumann, M., Iyyer, M., Gardner, M., Clark, C., Lee, K., & Zettlemoyer, L. (2018). Deep
contextualized word representations. Proceedings of the 2018 Conference of the North American
Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1
(Long Papers) (pp. 2227–2237).
Petersen, K. B., Pedersen, M. S., & others. (2008). The matrix cookbook. Technical University of
Denmark, 7(15), 510.
Polyak, B. T. (1964). Some methods of speeding up the convergence of iteration methods. USSR
Computational Mathematics and Mathematical Physics, 4(5), 1–17.
Quadrana, M., Cremonesi, P., & Jannach, D. (2018). Sequence-aware recommender systems. ACM
Computing Surveys (CSUR), 51(4), 66.
Radford, A., Metz, L., & Chintala, S. (2015). Unsupervised representation learning with deep con-
volutional generative adversarial networks. arXiv preprint arXiv:1511.06434.
Radford, A., Narasimhan, K., Salimans, T., & Sutskever, I. (2018). Improving language understand-
ing by generative pre-training. OpenAI.
Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., & Sutskever, I. (2019). Language models are
unsupervised multitask learners. OpenAI Blog, 1(8), 9.
Rajpurkar, P., Zhang, J., Lopyrev, K., & Liang, P. (2016). Squad: 100,000+ questions for machine
comprehension of text. arXiv preprint arXiv:1606.05250.
Reddi, S. J., Kale, S., & Kumar, S. (2019). On the convergence of adam and beyond. arXiv preprint
arXiv:1904.09237.
Reed, S., & De Freitas, N. (2015). Neural programmer-interpreters. arXiv preprint arXiv:1511.06279.
Ren, S., He, K., Girshick, R., & Sun, J. (2015). Faster r-cnn: towards real-time object detection with
region proposal networks. Advances in neural information processing systems (pp. 91–99).
960 Bibliography
Rendle, S. (2010). Factorization machines. 2010 IEEE International Conference on Data Mining
(pp. 995–1000).
Rendle, S., Freudenthaler, C., Gantner, Z., & Schmidt-Thieme, L. (2009). Bpr: bayesian person-
alized ranking from implicit feedback. Proceedings of the twenty-fifth conference on uncertainty in
artificial intelligence (pp. 452–461).
Rumelhart, D. E., Hinton, G. E., Williams, R. J., & others. (1988). Learning representations by
back-propagating errors. Cognitive modeling, 5(3), 1.
Russell, S. J., & Norvig, P. (2016). Artificial intelligence: a modern approach. Malaysia; Pearson Edu-
cation Limited,.
Santurkar, S., Tsipras, D., Ilyas, A., & Madry, A. (2018). How does batch normalization help opti-
mization? Advances in Neural Information Processing Systems (pp. 2483–2493).
Sarwar, B. M., Karypis, G., Konstan, J. A., Riedl, J., & others. (2001). Item-based collaborative fil-
tering recommendation algorithms. Www, 1, 285–295.
Schein, A. I., Popescul, A., Ungar, L. H., & Pennock, D. M. (2002). Methods and metrics for cold-
start recommendations. Proceedings of the 25th annual international ACM SIGIR conference on Re-
search and development in information retrieval (pp. 253–260).
Schuster, M., & Paliwal, K. K. (1997). Bidirectional recurrent neural networks. IEEE Transactions
on Signal Processing, 45(11), 2673–2681.
Sedhain, S., Menon, A. K., Sanner, S., & Xie, L. (2015). Autorec: autoencoders meet collaborative
filtering. Proceedings of the 24th International Conference on World Wide Web (pp. 111–112).
Sennrich, R., Haddow, B., & Birch, A. (2015). Neural machine translation of rare words with sub-
word units. arXiv preprint arXiv:1508.07909.
Sergeev, A., & Del Balso, M. (2018). Horovod: fast and easy distributed deep learning in tensorflow.
arXiv preprint arXiv:1802.05799.
Shannon, C. E. (1948 , 7). A mathematical theory of communication. The Bell System Technical Jour-
nal, 27(3), 379–423.
Shao, H., Yao, S., Sun, D., Zhang, A., Liu, S., Liu, D., … Abdelzaher, T. (2020). Controlvae: control-
lable variational autoencoder. Proceedings of the 37th International Conference on Machine Learn-
ing.
Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., Van Den Driessche, G., … others. (2016).
Mastering the game of go with deep neural networks and tree search. nature, 529(7587), 484.
Simonyan, K., & Zisserman, A. (2014). Very deep convolutional networks for large-scale image
recognition. arXiv preprint arXiv:1409.1556.
Smola, A., & Narayanamurthy, S. (2010). An architecture for parallel topic models. Proceedings of
the VLDB Endowment, 3(1-2), 703–710.
Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., & Salakhutdinov, R. (2014). Dropout: a
simple way to prevent neural networks from overfitting. The Journal of Machine Learning Re-
search, 15(1), 1929–1958.
Strang, G. (1993). Introduction to linear algebra. Vol. 3. Wellesley-Cambridge Press Wellesley, MA.
Su, X., & Khoshgoftaar, T. M. (2009). A survey of collaborative filtering techniques. Advances in
artificial intelligence, 2009.
Bibliography 961
Sukhbaatar, S., Weston, J., Fergus, R., & others. (2015). End-to-end memory networks. Advances in
neural information processing systems (pp. 2440–2448).
Sutskever, I., Martens, J., Dahl, G., & Hinton, G. (2013). On the importance of initialization and
momentum in deep learning. International conference on machine learning (pp. 1139–1147).
Szegedy, C., Ioffe, S., Vanhoucke, V., & Alemi, A. A. (2017). Inception-v4, inception-resnet and the
impact of residual connections on learning. Thirty-First AAAI Conference on Artificial Intelligence.
Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., … Rabinovich, A. (2015). Go-
ing deeper with convolutions. Proceedings of the IEEE conference on computer vision and pattern
recognition (pp. 1–9).
Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., & Wojna, Z. (2016). Rethinking the inception ar-
chitecture for computer vision. Proceedings of the IEEE conference on computer vision and pattern
recognition (pp. 2818–2826).
Tallec, C., & Ollivier, Y. (2017). Unbiasing truncated backpropagation through time. arXiv preprint
arXiv:1705.08209.
Tang, J., & Wang, K. (2018). Personalized top-n sequential recommendation via convolutional se-
quence embedding. Proceedings of the Eleventh ACM International Conference on Web Search and
Data Mining (pp. 565–573).
Teye, M., Azizpour, H., & Smith, K. (2018). Bayesian uncertainty estimation for batch normalized
deep networks. arXiv preprint arXiv:1802.06455.
Tieleman, T., & Hinton, G. (2012). Lecture 6.5-rmsprop: divide the gradient by a running average
of its recent magnitude. COURSERA: Neural networks for machine learning, 4(2), 26–31.
Treisman, A. M., & Gelade, G. (1980). A feature-integration theory of attention. Cognitive psychol-
ogy, 12(1), 97–136.
Töscher, A., Jahrer, M., & Bell, R. M. (2009). The bigchaos solution to the netflix grand prize. Netflix
prize documentation, pp. 1–52.
Uijlings, J. R., Van De Sande, K. E., Gevers, T., & Smeulders, A. W. (2013). Selective search for object
recognition. International journal of computer vision, 104(2), 154–171.
Van Loan, C. F., & Golub, G. H. (1983). Matrix computations. Johns Hopkins University Press.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., … Polosukhin, I. (2017).
Attention is all you need. Advances in neural information processing systems (pp. 5998–6008).
Wang, L., Li, M., Liberty, E., & Smola, A. J. (2018). Optimal message scheduling for aggregation.
NETWORKS, 2(3), 2–3.
Wang, Y., Davidson, A., Pan, Y., Wu, Y., Riffel, A., & Owens, J. D. (2016). Gunrock: a high-
performance graph processing library on the gpu. ACM SIGPLAN Notices (p. 11).
Warstadt, A., Singh, A., & Bowman, S. R. (2019). Neural network acceptability judgments. Trans-
actions of the Association for Computational Linguistics, 7, 625–641.
Wasserman, L. (2013). All of statistics: a concise course in statistical inference. Springer Science &
Business Media.
Watkins, C. J., & Dayan, P. (1992). Q-learning. Machine learning, 8(3-4), 279–292.
Welling, M., & Teh, Y. W. (2011). Bayesian learning via stochastic gradient langevin dynamics.
Proceedings of the 28th international conference on machine learning (ICML-11) (pp. 681–688).
962 Bibliography
Wigner, E. P. (1958). On the distribution of the roots of certain symmetric matrices. Ann. Math
(pp. 325–327).
Wood, F., Gasthaus, J., Archambeau, C., James, L., & Teh, Y. W. (2011). The sequence memoizer.
Communications of the ACM, 54(2), 91–98.
Wu, C.-Y., Ahmed, A., Beutel, A., Smola, A. J., & Jing, H. (2017). Recurrent recommender networks.
Proceedings of the tenth ACM international conference on web search and data mining (pp. 495–503).
Wu, Y., Schuster, M., Chen, Z., Le, Q. V., Norouzi, M., Macherey, W., … others. (2016). Googleʼs
neural machine translation system: bridging the gap between human and machine translation.
arXiv preprint arXiv:1609.08144.
Xiao, H., Rasul, K., & Vollgraf, R. (2017). Fashion-mnist: a novel image dataset for benchmarking
machine learning algorithms. arXiv preprint arXiv:1708.07747.
Xiao, L., Bahri, Y., Sohl-Dickstein, J., Schoenholz, S., & Pennington, J. (2018). Dynamical isom-
etry and a mean field theory of cnns: how to train 10,000-layer vanilla convolutional neural
networks. International Conference on Machine Learning (pp. 5393–5402).
Xiong, W., Wu, L., Alleva, F., Droppo, J., Huang, X., & Stolcke, A. (2018). The microsoft 2017 con-
versational speech recognition system. 2018 IEEE International Conference on Acoustics, Speech
and Signal Processing (ICASSP) (pp. 5934–5938).
Ye, M., Yin, P., Lee, W.-C., & Lee, D.-L. (2011). Exploiting geographical influence for collaborative
point-of-interest recommendation. Proceedings of the 34th international ACM SIGIR conference on
Research and development in Information Retrieval (pp. 325–334).
You, Y., Gitman, I., & Ginsburg, B. (2017). Large batch training of convolutional networks. arXiv
Zaheer, M., Reddi, S., Sachan, D., Kale, S., & Kumar, S. (2018). Adaptive methods for nonconvex
optimization. Advances in Neural Information Processing Systems (pp. 9793–9803).
Zeiler, M. D. (2012). Adadelta: an adaptive learning rate method. arXiv preprint arXiv:1212.5701.
Zhang, S., Yao, L., Sun, A., & Tay, Y. (2019). Deep learning based recommender system: a survey
and new perspectives. ACM Computing Surveys (CSUR), 52(1), 5.
Zhu, J.-Y., Park, T., Isola, P., & Efros, A. A. (2017). Unpaired image-to-image translation using
cycle-consistent adversarial networks. Proceedings of the IEEE international conference on com-
puter vision (pp. 2223–2232).
Zhu, Y., Kiros, R., Zemel, R., Salakhutdinov, R., Urtasun, R., Torralba, A., & Fidler, S. (2015). Align-
ing books and movies: towards story-like visual explanations by watching movies and reading
books. Proceedings of the IEEE international conference on computer vision (pp. 19–27).
Bibliography 963
964 Bibliography
Python Module Index
d
d2l.mxnet, 935
965
966 Python Module Index
Index
A EncoderBlock (class in d2l.mxnet), 936

abs() (in module d2l.mxnet), 939 EncoderDecoder (class in d2l.mxnet), 936
Accumulator (class in d2l.mxnet), 935 evaluate_accuracy() (in module d2l.mxnet), 942
accuracy() (in module d2l.mxnet), 939 evaluate_accuracy_gpu() (in module
AddNorm (class in d2l.mxnet), 935 d2l.mxnet), 942
Animator (class in d2l.mxnet), 935 evaluate_loss() (in module d2l.mxnet), 942
arange() (in module d2l.mxnet), 939 exp() (in module d2l.mxnet), 942
as_integer_ratio() (d2l.mxnet.float32 method),
943
F
avg() (d2l.mxnet.Timer method), 938 filter() (d2l.mxnet.VOCSegDataset method), 939
float32 (class in d2l.mxnet), 943
B forward() (d2l.mxnet.AddNorm method), 935
bbox_to_rect() (in module d2l.mxnet), 940 forward() (d2l.mxnet.BERTEncoder method), 935
BERTEncoder (class in d2l.mxnet), 935 forward() (d2l.mxnet.BERTModel method), 935
BERTModel (class in d2l.mxnet), 935 forward() (d2l.mxnet.BPRLoss method), 935
BPRLoss (class in d2l.mxnet), 935 forward() (d2l.mxnet.Decoder method), 936
build_colormap2label() (in module d2l.mxnet), forward() (d2l.mxnet.DotProductAttention
940 method), 936
forward() (d2l.mxnet.Encoder method), 936
C forward() (d2l.mxnet.EncoderBlock method), 936
concat() (in module d2l.mxnet), 940 forward() (d2l.mxnet.EncoderDecoder method),
copy() (d2l.mxnet.defaultdict method), 942 936
copyfile() (in module d2l.mxnet), 941 forward() (d2l.mxnet.HingeLossbRec method),
corr2d() (in module d2l.mxnet), 941 936
cos() (in module d2l.mxnet), 941 forward() (d2l.mxnet.MaskedSoftmaxCELoss
cosh() (in module d2l.mxnet), 941 method), 937
CTRDataset (class in d2l.mxnet), 936 forward() (d2l.mxnet.MaskLM method), 937
cumsum() (d2l.mxnet.Timer method), 938 forward() (d2l.mxnet.MLPAttention method), 936
forward() (d2l.mxnet.MultiHeadAttention
D method), 937
d2l.mxnet forward() (d2l.mxnet.NextSentencePred method),
module, 935 937
Decoder (class in d2l.mxnet), 936 forward() (d2l.mxnet.PositionalEncoding
default_factory (d2l.mxnet.defaultdict at- method), 937
tribute), 942 forward() (d2l.mxnet.PositionWiseFFN method),
defaultdict (class in d2l.mxnet), 942 937
DotProductAttention (class in d2l.mxnet), 936 forward() (d2l.mxnet.Residual method), 938
download() (in module d2l.mxnet), 942 forward() (d2l.mxnet.RNNModel method), 937
download_all() (in module d2l.mxnet), 942 forward() (d2l.mxnet.Seq2SeqDecoder method),
download_extract() (in module d2l.mxnet), 942 938
E forward() (d2l.mxnet.Seq2SeqEncoder method),
938
Encoder (class in d2l.mxnet), 936
967
forward() (d2l.mxnet.TransformerEncoder R
method), 939 RandomGenerator (class in d2l.mxnet), 938
read_csv_labels() (in module d2l.mxnet), 948
G read_snli() (in module d2l.mxnet), 948
get_dataloader_workers() (in module read_time_machine() (in module d2l.mxnet), 948
d2l.mxnet), 943 read_voc_images() (in module d2l.mxnet), 949
get_fashion_mnist_labels() (in module Residual (class in d2l.mxnet), 938
d2l.mxnet), 943 resnet18() (in module d2l.mxnet), 949
RNNModel (class in d2l.mxnet), 937
H
RNNModelScratch (class in d2l.mxnet), 938
HingeLossbRec (class in d2l.mxnet), 936
S
I Seq2SeqDecoder (class in d2l.mxnet), 938
int32 (class in d2l.mxnet), 943 Seq2SeqEncoder (class in d2l.mxnet), 938
SeqDataLoader (class in d2l.mxnet), 938
L
set_axes() (in module d2l.mxnet), 949
linreg() (in module d2l.mxnet), 943 set_figsize() (in module d2l.mxnet), 949
linspace() (in module d2l.mxnet), 943 sgd() (in module d2l.mxnet), 949
load_array() (in module d2l.mxnet), 944 show_bboxes() (in module d2l.mxnet), 949
load_data_bananas() (in module d2l.mxnet), 944 show_images() (in module d2l.mxnet), 949
load_data_fashion_mnist() (in module show_trace_2d() (in module d2l.mxnet), 949
d2l.mxnet), 944 sin() (in module d2l.mxnet), 949
load_data_snli() (in module d2l.mxnet), 944 sinh() (in module d2l.mxnet), 949
load_data_voc() (in module d2l.mxnet), 944 SNLIDataset (class in d2l.mxnet), 938
log() (in module d2l.mxnet), 945 split_batch() (in module d2l.mxnet), 950
M split_batch_multi_inputs() (in module
d2l.mxnet), 950
masked_softmax() (in module d2l.mxnet), 945
split_data_ml100k() (in module d2l.mxnet), 950
MaskedSoftmaxCELoss (class in d2l.mxnet), 937
squared_loss() (in module d2l.mxnet), 950
MaskLM (class in d2l.mxnet), 937
stack() (in module d2l.mxnet), 950
matmul() (in module d2l.mxnet), 945
start() (d2l.mxnet.Timer method), 938
meshgrid() (in module d2l.mxnet), 946
stop() (d2l.mxnet.Timer method), 938
mkdir_if_not_exist() (in module d2l.mxnet),
sum() (d2l.mxnet.Timer method), 938
947
synthetic_data() (in module d2l.mxnet), 951
MLPAttention (class in d2l.mxnet), 936
module T
d2l.mxnet, 935 tanh() (in module d2l.mxnet), 951
MultiHeadAttention (class in d2l.mxnet), 937 tensor() (in module d2l.mxnet), 951
N Timer (class in d2l.mxnet), 938
TokenEmbedding (class in d2l.mxnet), 938
NextSentencePred (class in d2l.mxnet), 937
tokenize() (in module d2l.mxnet), 952
normal() (in module d2l.mxnet), 947
train_2d() (in module d2l.mxnet), 952
O train_ch3() (in module d2l.mxnet), 952
train_ch6() (in module d2l.mxnet), 952
ones() (in module d2l.mxnet), 948
train_epoch_ch3() (in module d2l.mxnet), 952
P TransformerEncoder (class in d2l.mxnet), 939
try_all_gpus() (in module d2l.mxnet), 952
plot() (in module d2l.mxnet), 948
try_gpu() (in module d2l.mxnet), 952
PositionalEncoding (class in d2l.mxnet), 937
PositionWiseFFN (class in d2l.mxnet), 937 U
predict_ch3() (in module d2l.mxnet), 948
update_D() (in module d2l.mxnet), 952
968 Index
update_G() (in module d2l.mxnet), 952
use_svg_display() (in module d2l.mxnet), 952
V
voc_label_indices() (in module d2l.mxnet), 952
voc_rand_crop() (in module d2l.mxnet), 952
VOCSegDataset (class in d2l.mxnet), 939
Z
zeros() (in module d2l.mxnet), 952
Index 969

Dive Into Deep Learning: Aston Zhang, Zachary C. Lipton, Mu Li, and Alexander J. Smola

Uploaded by

Copyright:

Available Formats

Dive Into Deep Learning: Aston Zhang, Zachary C. Lipton, Mu Li, and Alexander J. Smola

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Dive Into Deep Learning: Aston Zhang, Zachary C. Lipton, Mu Li, and Alexander J. Smola

Uploaded by

Copyright:

Available Formats

Dive into Deep Learning

Aston Zhang, Zachary C. Lipton, Mu Li, and Alexander J. Smola

Aug 11, 2020

3 Linear Neural Networks 89

4 Multilayer Perceptrons 131

5 Deep Learning Computation 199

6 Convolutional Neural Networks 227

7 Modern Convolutional Neural Networks 257

8 Recurrent Neural Networks 299

9 Modern Recurrent Neural Networks 339

10 Attention Mechanisms 381

11 Optimization Algorithms 405

12 Computational Performance 483

13 Computer Vision 537

14 Natural Language Processing: Pretraining 631

15 Natural Language Processing: Applications 683

17 Generative Adversarial Networks 763

18 Appendix: Mathematics for Deep Learning 777

19 Appendix: Tools for Deep Learning 907

Python Module Index 965

About This Book

One Medium Combining Code, Math, and HTML

Fig. 1: Book structure

1. Register an account on the discussion forum of this book discuss.d2l.ai7 .

# The file name is subject to changes

For Linux users:

# The file name is subject to changes

Next, initialize the shell so we can run conda directly.

conda create --name d2l -y

mkdir d2l-en && cd d2l-en

conda activate d2l

Installing the Framework and the d2l Package

pip install mxnet==1.6.0

# -U: Upgrade all packages to the newest available version

pip uninstall mxnet

# For Windows users

# For Linux and macOS users

The notation used throughout this book is summarized below.

Probability and Information Theory

• P (·): Probability distribution

1.1 A Motivating Example

Fig. 1.1.1: Identify an awake word.

Fig. 1.1.2: A typical training process.

1.1. A Motivating Example 19

1.2 The Key Components: Data, Models, and Algorithms

1.2. The Key Components: Data, Models, and Algorithms 21

1.2.3 Objective functions

1.2.4 Optimization algorithms

1.3 Kinds of Machine Learning

1.3.1 Supervised learning

1.3. Kinds of Machine Learning 23

Fig. 1.3.1: Supervised learning.

and the least mean squares loss, or L2 loss where

1.3. Kinds of Machine Learning 25

L(action|x) = Ey∼p(y|x) [loss(action, y)]. (1.3.3)

1.3. Kinds of Machine Learning 27

Fig. 1.3.3: A cat, a rooster, a dog and a donkey

Search and ranking

1.3. Kinds of Machine Learning 29