Contents

SciKits Numpy Matplotlib I Getting started with Python for science 2

1 Python scientific computing ecosystem 4
1.1 Why Python? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 The Scientific Python ecosystem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Before starting: Installing a working environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2017 1.4 The workflow: interactive environments and text editors . . . . . . . . . . . . . . . . . . . . . . . . 8

SciPy Python
2 The Python language 12
EDITION 2.1 First steps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Basic types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 Control Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.4 Defining functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.5 Reusing code: scripts and modules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
IP[y]: 2.6 Input and Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.7 Standard Library . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.8 Exception handling in Python . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

Cython IPython 2.9 Object-oriented programming (OOP)

3 NumPy: creating and manipulating numerical data

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

47
3.1 The NumPy array object . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.2 Numerical operations on arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

Scipy
3.3 More elaborate arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.4 Advanced operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
Edited by 3.5 Some exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
Gaël Varoquaux 3.6 Full code examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

Lecture Notes Emmanuelle Gouillart

Olaf Vahtras
4 Matplotlib: plotting
4.1 Introduction . . . . . . . . . . . . . . . . . . .
4.2 Simple plot . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
97
97
98

www.scipy-lectures.org 4.3 Figures, Subplots, Axes and Ticks . . . . . . .

4.4 Other Types of Plots: examples and exercises
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
106
108
4.5 Beyond this tutorial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
4.6 Quick references . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
4.7 Full code examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

5 Scipy : high-level scientific computing 184

Gaël Varoquaux • Emmanuelle Gouillart • Olav Vahtras 5.1 File input/output: scipy.io . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
Christopher Burns • Adrian Chauve • Robert Cimrman • Christophe Combelles 5.2 Special functions: scipy.special . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
Pierre de Buyl • Ralf Gommers • André Espaze • Zbigniew Jędrzejewski-Szmek 5.3 Linear algebra operations: scipy.linalg . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
Valentin Haenel • Gert-Ludwig Ingold • Fabian Pedregosa • Didrik Pinte
Nicolas P. Rougier • Pauli Virtanen i

an d m a n y ot her s...
 5.4 Interpolation: scipy.interpolate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188 13.5 Practical guide to optimization with scipy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 432
5.5 Optimization and fit: scipy.optimize . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188 13.6 Special case: non-linear least-squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 435
5.6 Statistics and random numbers: scipy.stats . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194 13.7 Optimization with constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 437
5.7 Numerical integration: scipy.integrate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196 13.8 Full code examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 438
5.8 Fast Fourier transforms: scipy.fftpack . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198 13.9 Examples for the mathematical optimization chapter . . . . . . . . . . . . . . . . . . . . . . . . . . 438
5.9 Signal processing: scipy.signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
5.10 Image manipulation: scipy.ndimage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203 14 Interfacing with C 439
5.11 Summary exercises on scientific computing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208 14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 439
5.12 Full code examples for the scipy chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221 14.2 Python-C-Api . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 440
14.3 Ctypes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
6 Getting help and finding documentation 259 14.4 SWIG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 449
14.5 Cython . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
14.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 458
II Advanced topics 262 14.7 Further Reading and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 458
14.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 458
7 Advanced Python Constructs 264
7.1 Iterators, generator expressions and generators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
7.2 Decorators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269 III Packages and applications 460
7.3 Context managers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
15 Statistics in Python 462
8 Advanced NumPy 281 15.1 Data representation and interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463
8.1 Life of ndarray . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282 15.2 Hypothesis testing: comparing two groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 468
8.2 Universal functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295 15.3 Linear models, multiple factors, and analysis of variance . . . . . . . . . . . . . . . . . . . . . . . . 471
8.3 Interoperability features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304 15.4 More visualization: seaborn for statistical exploration . . . . . . . . . . . . . . . . . . . . . . . . . . 476
8.4 Array siblings: chararray, maskedarray, matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307 15.5 Testing for interactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 480
8.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310 15.6 Full code for the figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 481
8.6 Contributing to NumPy/Scipy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310 15.7 Solutions to this chapter’s exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 502
9 Debugging code 314 16 Sympy : Symbolic Mathematics in Python 505
9.1 Avoiding bugs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315 16.1 First Steps with SymPy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 506
9.2 Debugging workflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317 16.2 Algebraic manipulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 507
9.3 Using the Python debugger . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 318 16.3 Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 508
9.4 Debugging segmentation faults using gdb . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323 16.4 Equation solving . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 510
16.5 Linear Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 511
10 Optimizing code 325
10.1 Optimization workflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326 17 Scikit-image: image processing 513
10.2 Profiling Python code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326 17.1 Introduction and concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 514
10.3 Making code go faster . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328 17.2 Input/output, data types and colorspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 515
10.4 Writing faster numerical code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 330 17.3 Image preprocessing / enhancement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517
17.4 Image segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 521
11 Sparse Matrices in SciPy 332
17.5 Measuring regions’ properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 332
17.6 Data visualization and interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 524
11.2 Storage Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
17.7 Feature extraction for computer vision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 525
11.3 Linear System Solvers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 346
17.8 Full code examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 526
11.4 Other Interesting Packages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
17.9 Examples for the scikit-image chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 526
12 Image manipulation and processing using Numpy and Scipy 352
18 Traits: building interactive dialogs 538
12.1 Opening and writing to image files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
18.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 539
12.2 Displaying images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355
18.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 540
12.3 Basic manipulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 356
18.3 What are Traits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 540
12.4 Image filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359
12.5 Feature extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 364 19 3D plotting with Mayavi 557
12.6 Measuring objects properties: ndimage.measurements . . . . . . . . . . . . . . . . . . . . . . . . 366 19.1 Mlab: the scripting interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 558
12.7 Full code examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 371 19.2 Interactive work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 564
12.8 Examples for the image processing chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 371 19.3 Slicing and dicing data: sources, modules and filters . . . . . . . . . . . . . . . . . . . . . . . . . . . 565
19.4 Animating the data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 568
13 Mathematical optimization: finding minima of functions 397
19.5 Making interactive dialogs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 569
13.1 Knowing your problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 398
19.6 Putting it together . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 571
13.2 A review of the different optimizers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 400
13.3 Full code examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 406 20 scikit-learn: machine learning in Python 572
13.4 Examples for the mathematical optimization chapter . . . . . . . . . . . . . . . . . . . . . . . . . . 406

ii iii
 Scipy lecture notes, Edition 2017.1

20.1 Introduction: problem settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 573

20.2 Basic principles of machine learning with scikit-learn . . . . . . . . . . . . . . . . . . . . . . . . . . 577
20.3 Supervised Learning: Classification of Handwritten Digits . . . . . . . . . . . . . . . . . . . . . . . 582
20.4 Supervised Learning: Regression of Housing Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 586
20.5 Measuring prediction performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 589
20.6 Unsupervised Learning: Dimensionality Reduction and Visualization . . . . . . . . . . . . . . . . 594
20.7 The eigenfaces example: chaining PCA and SVMs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 597
20.8 Parameter selection, Validation, and Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 603
20.9 Examples for the scikit-learn chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 610

Index 652

iv Contents 1
 Scipy lecture notes, Edition 2017.1

This part of the Scipy lecture notes is a self-contained introduction to everything that is needed to use Python
for science, from the language itself, to numerical computing or plotting.

Part I

Getting started with Python for science

2 3
 Scipy lecture notes, Edition 2017.1

• Universal Python is a language used for many different problems. Learning Python avoids learning a
new software for each new problem.

1.1.3 How does Python compare to other solutions?

1
Compiled languages: C, C++, Fortran. . .

Pros
• Very fast. For heavy computations, it’s difficult to outperform these languages.

CHAPTER Cons
• Painful usage: no interactivity during development, mandatory compilation steps, ver-
bose syntax, manual memory management. These are difficult languages for non pro-
grammers.

Matlab scripting language

Pros

Python scientific computing ecosystem • Very rich collection of libraries with numerous algorithms, for many different domains.
Fast execution because these libraries are often written in a compiled language.
• Pleasant development environment: comprehensive and help, integrated editor, etc.
• Commercial support is available.
Cons
• Base language is quite poor and can become restrictive for advanced users.
Authors: Fernando Perez, Emmanuelle Gouillart, Gaël Varoquaux, Valentin Haenel • Not free.

1.1 Why Python? Julia

Pros
1.1.1 The scientist’s needs
• Fast code, yet interactive and simple.
• Get data (simulation, experiment control), • Easily connects to Python or C.
• Manipulate and process data, Cons
• Visualize results, quickly to understand, but also with high quality figures, for reports or publications. • Ecosystem limited to numerical computing.
• Still young.
1.1.2 Python’s strengths
Other scripting languages: Scilab, Octave, R, IDL, etc.
• Batteries included Rich collection of already existing bricks of classic numerical methods, plotting or
data processing tools. We don’t want to re-program the plotting of a curve, a Fourier transform or a fitting
Pros
algorithm. Don’t reinvent the wheel!
• Open-source, free, or at least cheaper than Matlab.
• Easy to learn Most scientists are not payed as programmers, neither have they been trained so. They
need to be able to draw a curve, smooth a signal, do a Fourier transform in a few minutes. • Some features can be very advanced (statistics in R, etc.)

• Easy communication To keep code alive within a lab or a company it should be as readable as a book Cons
by collaborators, students, or maybe customers. Python syntax is simple, avoiding strange symbols or
• Fewer available algorithms than in Matlab, and the language is not more advanced.
lengthy routine specifications that would divert the reader from mathematical or scientific understand-
ing of the code. • Some software are dedicated to one domain. Ex: Gnuplot to draw curves. These pro-
grams are very powerful, but they are restricted to a single type of usage, such as plotting.
• Efficient code Python numerical modules are computationally efficient. But needless to say that a very
fast code becomes useless if too much time is spent writing it. Python aims for quick development times
and quick execution times.

4 1.1. Why Python? 5

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Python • Numpy: numerical computing with powerful numerical arrays objects, and routines to manipulate
them. http://www.numpy.org/
Pros
See also:
• Very rich scientific computing libraries
chapter on numpy
• Well thought out language, allowing to write very readable and well structured code: we
• Scipy : high-level numerical routines. Optimization, regression, interpolation, etc http://www.scipy.org/
“code what we think”.
See also:
• Many libraries beyond scientific computing (web server, serial port access, etc.)
chapter on scipy
• Free and open-source software, widely spread, with a vibrant community.
• Matplotlib : 2-D visualization, “publication-ready” plots http://matplotlib.org/
• A variety of powerful environments to work in, such as IPython, Spyder, Jupyter note-
books, Pycharm See also:

Cons chapter on matplotlib

• Not all the algorithms that can be found in more specialized software or toolboxes.

1.2 The Scientific Python ecosystem

Unlike Matlab, or R, Python does not come with a pre-bundled set of modules for scientific computing. Below
are the basic building blocks that can be combined to obtain a scientific computing environment:

Python, a generic and modern computing language

• The language: flow control, data types (string, int), data collections (lists, dictionaries), etc.
• Modules of the standard library: string processing, file management, simple network protocols.
• A large number of specialized modules or applications written in Python: web framework, etc. . . . and
scientific computing. Advanced interactive environments:
• Development tools (automatic testing, documentation generation) • IPython, an advanced Python console http://ipython.org/
See also: • Jupyter, notebooks in the browser http://jupyter.org/
chapter on Python language

Core numeric libraries

1.2. The Scientific Python ecosystem 6 1.2. The Scientific Python ecosystem 7
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Domain-specific packages,
Under the notebook
• Mayavi for 3-D visualization
• pandas, statsmodels, seaborn for statistics To execute code, press “shift enter”
• sympy for symbolic computing
Start ipython:
• scikit-image for image processing
• scikit-learn for machine learning In [1]: print('Hello world')
Hello world
and much more packages not documented in the scipy lectures.
See also: Getting help by using the ? operator after an object:

chapters on advanced topics In [2]: print?

Type: builtin_function_or_method
chapters on packages and applications Base Class: <type 'builtin_function_or_method'>
String Form: <built-in function print>
Namespace: Python builtin
1.3 Before starting: Installing a working environment Docstring:
print(value, ..., sep=' ', end='\n', file=sys.stdout)

Python comes in many flavors, and there are many ways to install it. However, we recommend to install a Prints the values to a stream, or to sys.stdout by default.
scientific-computing distribution, that comes readily with optimized versions of scientific modules. Optional keyword arguments:
file: a file-like object (stream); defaults to the current sys.stdout.
Under Linux sep: string inserted between values, default a space.
If you have a recent distribution, most of the tools are probably packaged, and it is recommended to use your end: string appended after the last value, default a newline.
package manager.
See also:
Other systems
• IPython user manual: http://ipython.org/ipython-doc/dev/index.html
There are several fully-featured Scientific Python distributions:
• Jupyter Notebook QuickStart: http://jupyter.readthedocs.io/en/latest/content-quickstart.html
• Anaconda
• EPD
1.4.2 Elaboration of the work in an editor
• WinPython
As you move forward, it will be important to not only work interactively, but also to create and reuse Python
files. For this, a powerful code editor will get you far. Here are several good easy-to-use editors:
Python 3 or Python 2?
• Spyder: integrates an IPython console, a debugger, a profiler. . .
In 2008, Python 3 was released. It is a major evolution of the language that made a few changes. Some old
• PyCharm: integrates an IPython console, notebooks, a debugger. . . (freely available, but commercial)
scientific code does not yet run under Python 3. However, this is infrequent and Python 3 comes with many
benefits. We advise that you install Python 3. • Atom
Some of these are shipped by the various scientific Python distributions, and you can find them in the menus.
As an exercise, create a file my_file.py in a code editor, and add the following lines:
1.4 The workflow: interactive environments and text editors s = 'Hello world'
print(s)
Interactive work to test and understand algorithms: In this section, we describe a workflow combining inter-
active work and consolidation. Now, you can run it in IPython console or a notebook and explore the resulting variables:
Python is a general-purpose language. As such, there is not one blessed environment to work in, and not only In [1]: %run my_file.py
one way of using it. Although this makes it harder for beginners to find their way, it makes it possible for Python Hello world
to be used for programs, in web servers, or embedded devices.
In [2]: s
Out[2]: 'Hello world'
1.4.1 Interactive work
In [3]: %whos
We recommend an interactive work with the IPython console, or its offspring, the Jupyter notebook. They are Variable Type Data/Info
handy to explore and understand algorithms. ----------------------------
s str Hello world

1.3. Before starting: Installing a working environment 8 1.4. The workflow: interactive environments and text editors 9
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

• %cpaste allows you to paste code, especially code from websites which has been prefixed with the stan-
From a script to functions dard Python prompt (e.g. >>>) or with an ipython prompt, (e.g. in [3]):

While it is tempting to work only with scripts, that is a file full of instructions following each other, do plan In [2]: %cpaste
Pasting code; enter '--' alone on the line to stop or use Ctrl-D.
to progressively evolve the script to a set of functions:
:>>> for i in range(3):
• A script is not reusable, functions are. :... print(i)
:--
• Thinking in terms of functions helps breaking the problem in small blocks. 0
1
2

1.4.3 IPython and Jupyter Tips and Tricks • %timeit allows you to time the execution of short snippets using the timeit module from the standard
library:
The user manuals contain a wealth of information. Here we give a quick introduction to four useful features:
history, tab completion, magic functions, and aliases. In [3]: %timeit x = 10
10000000 loops, best of 3: 39 ns per loop

See also:
Chapter on optimizing code

Command history Like a UNIX shell, the IPython console supports command history. Type up and down to • %debug allows you to enter post-mortem debugging. That is to say, if the code you try to execute, raises
navigate previously typed commands: an exception, using %debug will enter the debugger at the point where the exception was thrown.

In [1]: x = 10 In [4]: x === 10

File "<ipython-input-6-12fd421b5f28>", line 1
In [2]: <UP> x === 10
^
In [2]: x = 10 SyntaxError: invalid syntax

In [5]: %debug
> /.../IPython/core/compilerop.py (87)ast_parse()
86 and are passed to the built-in compile function."""
---> 87 return compile(source, filename, symbol, self.flags | PyCF_ONLY_AST, 1)
Tab completion Tab completion, is a convenient way to explore the structure of any object you’re dealing 88
with. Simply type object_name.<TAB> to view the object’s attributes. Besides Python objects and keywords,
ipdb>locals()
tab completion also works on file and directory names.*
{'source': u'x === 10\n', 'symbol': 'exec', 'self':
In [1]: x = 10 <IPython.core.compilerop.CachingCompiler instance at 0x2ad8ef0>,
'filename': '<ipython-input-6-12fd421b5f28>'}
In [2]: x.<TAB>
x.bit_length x.denominator x.imag x.real See also:
x.conjugate x.from_bytes x.numerator x.to_bytes
Chapter on debugging

Magic functions The console and the notebooks support so-called magic functions by prefixing a command Aliases Furthermore IPython ships with various aliases which emulate common UNIX command line tools
with the % character. For example, the run and whos functions from the previous section are magic functions. such as ls to list files, cp to copy files and rm to remove files (a full list of aliases is shown when typing alias).
Note that, the setting automagic, which is enabled by default, allows you to omit the preceding % sign. Thus,
you can just type the magic function and it will work.
Getting help
Other useful magic functions are:
• The built-in cheat-sheet is accessible via the %quickref magic function.
• %cd to change the current directory.
• A list of all available magic functions is shown when typing %magic.
In [1]: cd /tmp
/tmp

1.4. The workflow: interactive environments and text editors 10 1.4. The workflow: interactive environments and text editors 11
 Scipy lecture notes, Edition 2017.1

• a language for which a large variety of high-quality packages are available for various applications, from
web frameworks to scientific computing.
• a language very easy to interface with other languages, in particular C and C++.
• Some other features of the language are illustrated just below. For example, Python is an object-oriented

2
language, with dynamic typing (the same variable can contain objects of different types during the
course of a program).
See https://www.python.org/about/ for more information about distinguishing features of Python.

CHAPTER
2.1 First steps

Start the Ipython shell (an enhanced interactive Python shell):

• by typing “ipython” from a Linux/Mac terminal, or from the Windows cmd shell,
• or by starting the program from a menu, e.g. in the Python(x,y) or EPD menu if you have installed one

The Python language

of these scientific-Python suites.

Tip: If you don’t have Ipython installed on your computer, other Python shells are available, such as the plain
Python shell started by typing “python” in a terminal, or the Idle interpreter. However, we advise to use the
Ipython shell because of its enhanced features, especially for interactive scientific computing.

Once you have started the interpreter, type

>>> print("Hello, world!")

Authors: Chris Burns, Christophe Combelles, Emmanuelle Gouillart, Gaël Varoquaux
Hello, world!

Python for scientific computing

Tip: The message “Hello, world!” is then displayed. You just executed your first Python instruction, congratu-
We introduce here the Python language. Only the bare minimum necessary for getting started with Numpy lations!
and Scipy is addressed here. To learn more about the language, consider going through the excellent tutorial
https://docs.python.org/tutorial. Dedicated books are also available, such as http://www.diveintopython.
To get yourself started, type the following stack of instructions
net/.
>>> a = 3
>>> b = 2*a
>>> type(b)
<type 'int'>
>>> print(b)
6
>>> a*b
18
>>> b = 'hello'
>>> type(b)
Tip: Python is a programming language, as are C, Fortran, BASIC, PHP, etc. Some specific features of Python
<type 'str'>
are as follows:
>>> b + b
• an interpreted (as opposed to compiled) language. Contrary to e.g. C or Fortran, one does not compile 'hellohello'
Python code before executing it. In addition, Python can be used interactively: many Python inter- >>> 2*b
preters are available, from which commands and scripts can be executed. 'hellohello'

• a free software released under an open-source license: Python can be used and distributed free of
charge, even for building commercial software. Tip: Two variables a and b have been defined above. Note that one does not declare the type of a variable
• multi-platform: Python is available for all major operating systems, Windows, Linux/Unix, MacOS X, before assigning its value. In C, conversely, one should write:
most likely your mobile phone OS, etc.
int a = 3;
• a very readable language with clear non-verbose syntax

12 2.1. First steps 13

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

In addition, the type of a variable may change, in the sense that at one point in time it can be equal to a value Type conversion (casting):
of a certain type, and a second point in time, it can be equal to a value of a different type. b was first equal to an
integer, but it became equal to a string when it was assigned the value ‘hello’. Operations on integers (b=2*a) >>> float(1)
1.0
are coded natively in Python, and so are some operations on strings such as additions and multiplications,
which amount respectively to concatenation and repetition.

Warning: Integer division

In Python 2:
2.2 Basic types >>> 3 / 2
1
2.2.1 Numerical types In Python 3:
>>> 3 / 2
Tip: Python supports the following numerical, scalar types: 1.5

To be safe: use floats:

Integer >>> 3 / 2.
1.5
>>> 1 + 1
2
>>> a = 3
>>> a = 4
>>> b = 2
>>> type(a)
>>> a / b # In Python 2
<type 'int'>
1
>>> a / float(b)
Floats 1.5
>>> c = 2.1 Future behavior: to always get the behavior of Python3
>>> type(c)
<type 'float'> >>> from __future__ import division
>>> 3 / 2
Complex 1.5

>>> a = 1.5 + 0.5j Tip: If you explicitly want integer division use //:
>>> a.real
1.5 >>> 3.0 // 2
>>> a.imag 1.0
0.5
>>> type(1. + 0j) Note: The behaviour of the division operator has changed in Python 3.
<type 'complex'>

Booleans

>>> 3 > 4
False 2.2.2 Containers
>>> test = (3 > 4)
>>> test
False
Tip: Python provides many efficient types of containers, in which collections of objects can be stored.
>>> type(test)
<type 'bool'>

Lists
Tip: A Python shell can therefore replace your pocket calculator, with the basic arithmetic operations +, -, *,
/, % (modulo) natively implemented
Tip: A list is an ordered collection of objects, that may have different types. For example:
>>> 7 * 3.
21.0 >>> colors = ['red', 'blue', 'green', 'black', 'white']
>>> 2**10 >>> type(colors)
1024 <type 'list'>
>>> 8 % 3
2
Indexing: accessing individual objects contained in the list:

2.2. Basic types 14 2.2. Basic types 15

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

NumPy arrays, operations on elements can be faster because elements are regularly spaced in memory and
>>> colors[2]
'green' more operations are performed through specialized C functions instead of Python loops.

Counting from the end with negative indices:

>>> colors[-1]
Tip: Python offers a large panel of functions to modify lists, or query them. Here are a few examples; for more
'white'
details, see https://docs.python.org/tutorial/datastructures.html#more-on-lists
>>> colors[-2]
'black'
Add and remove elements:

>>> colors = ['red', 'blue', 'green', 'black', 'white']

Warning: Indexing starts at 0 (as in C), not at 1 (as in Fortran or Matlab)!
>>> colors.append('pink')
>>> colors
Slicing: obtaining sublists of regularly-spaced elements: ['red', 'blue', 'green', 'black', 'white', 'pink']
>>> colors.pop() # removes and returns the last item
>>> colors 'pink'
['red', 'blue', 'green', 'black', 'white'] >>> colors
>>> colors[2:4] ['red', 'blue', 'green', 'black', 'white']
['green', 'black'] >>> colors.extend(['pink', 'purple']) # extend colors, in-place
>>> colors
['red', 'blue', 'green', 'black', 'white', 'pink', 'purple']
>>> colors = colors[:-2]
Warning: Note that colors[start:stop] contains the elements with indices i such as start<= i < >>> colors
stop (i ranging from start to stop-1). Therefore, colors[start:stop] has (stop - start) elements. ['red', 'blue', 'green', 'black', 'white']

Reverse:
Slicing syntax: colors[start:stop:stride]
>>> rcolors = colors[::-1]
Tip: All slicing parameters are optional: >>> rcolors
['white', 'black', 'green', 'blue', 'red']
>>> colors >>> rcolors2 = list(colors)
['red', 'blue', 'green', 'black', 'white'] >>> rcolors2
>>> colors[3:] ['red', 'blue', 'green', 'black', 'white']
['black', 'white'] >>> rcolors2.reverse() # in-place
>>> colors[:3] >>> rcolors2
['red', 'blue', 'green'] ['white', 'black', 'green', 'blue', 'red']
>>> colors[::2]
['red', 'green', 'white'] Concatenate and repeat lists:

>>> rcolors + colors

['white', 'black', 'green', 'blue', 'red', 'red', 'blue', 'green', 'black', 'white']
Lists are mutable objects and can be modified:
>>> rcolors * 2
>>> colors[0] = 'yellow' ['white', 'black', 'green', 'blue', 'red', 'white', 'black', 'green', 'blue', 'red']
>>> colors
['yellow', 'blue', 'green', 'black', 'white']
>>> colors[2:4] = ['gray', 'purple'] Tip: Sort:
>>> colors
['yellow', 'blue', 'gray', 'purple', 'white'] >>> sorted(rcolors) # new object
['black', 'blue', 'green', 'red', 'white']
>>> rcolors
['white', 'black', 'green', 'blue', 'red']
Note: The elements of a list may have different types:
>>> rcolors.sort() # in-place
>>> colors = [3, -200, 'hello'] >>> rcolors
>>> colors ['black', 'blue', 'green', 'red', 'white']
[3, -200, 'hello']
>>> colors[1], colors[2]
(-200, 'hello')
Methods and Object-Oriented Programming

Tip: For collections of numerical data that all have the same type, it is often more efficient to use the array The notation rcolors.method() (e.g. rcolors.append(3) and colors.pop()) is our first example of
type provided by the numpy module. A NumPy array is a chunk of memory containing fixed-sized items. With

2.2. Basic types 16 2.2. Basic types 17

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

object-oriented programming (OOP). Being a list, the object rcolors owns the method function that is >>> a[::3] # every three characters, from beginning to end
called using the notation .. No further knowledge of OOP than understanding the notation . is necessary for 'hl r!'
going through this tutorial.

Tip: Accents and special characters can also be handled in Unicode strings (see https://docs.python.org/
Discovering methods: tutorial/introduction.html#unicode-strings).

Reminder: in Ipython: tab-completion (press tab) A string is an immutable object and it is not possible to modify its contents. One may however create new
In [28]: rcolors.<TAB> strings from the original one.
rcolors.append rcolors.index rcolors.remove
In [53]: a = "hello, world!"
rcolors.count rcolors.insert rcolors.reverse
In [54]: a[2] = 'z'
rcolors.extend rcolors.pop rcolors.sort
---------------------------------------------------------------------------
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
TypeError: 'str' object does not support item assignment
Strings
In [55]: a.replace('l', 'z', 1)
Different string syntaxes (simple, double or triple quotes): Out[55]: 'hezlo, world!'
In [56]: a.replace('l', 'z')
s = 'Hello, how are you?' Out[56]: 'hezzo, worzd!'
s = "Hi, what's up"
s = '''Hello, # tripling the quotes allows the
how are you''' # string to span more than one line
Tip: Strings have many useful methods, such as a.replace as seen above. Remember the a. object-oriented
s = """Hi,
notation and use tab completion or help(str) to search for new methods.
what's up?"""

In [1]: 'Hi, what's up?' See also:

------------------------------------------------------------ Python offers advanced possibilities for manipulating strings, looking for patterns or formatting. The inter-
File "<ipython console>", line 1
ested reader is referred to https://docs.python.org/library/stdtypes.html#string-methods and https://docs.
'Hi, what's up?'
python.org/library/string.html#new-string-formatting
^
SyntaxError: invalid syntax String formatting:

The newline character is \n, and the tab character is \t. >>> 'An integer: %i ; a float: %f ; another string: %s ' % (1, 0.1, 'string')
'An integer: 1; a float: 0.100000; another string: string'

Tip: Strings are collections like lists. Hence they can be indexed and sliced, using the same syntax and rules. >>> i = 102
>>> filename = 'processing_of_dataset_%d .txt' % i
>>> filename
Indexing: 'processing_of_dataset_102.txt'
>>> a = "hello"
>>> a[0]
'h' Dictionaries
>>> a[1]
'e'
>>> a[-1] Tip: A dictionary is basically an efficient table that maps keys to values. It is an unordered container
'o'

>>> tel = {'emmanuelle': 5752, 'sebastian': 5578}

Tip: (Remember that negative indices correspond to counting from the right end.) >>> tel['francis'] = 5915
>>> tel
{'sebastian': 5578, 'francis': 5915, 'emmanuelle': 5752}
Slicing: >>> tel['sebastian']
5578
>>> a = "hello, world!" >>> tel.keys()
>>> a[3:6] # 3rd to 6th (excluded) elements: elements 3, 4, 5 ['sebastian', 'francis', 'emmanuelle']
'lo,' >>> tel.values()
>>> a[2:10:2] # Syntax: a[start:stop:step] [5578, 5915, 5752]
'lo o' >>> 'francis' in tel
True

2.2. Basic types 18 2.2. Basic types 19

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

In [6]: b[1] = 'hi!'

In [7]: a
Tip: It can be used to conveniently store and retrieve values associated with a name (a string for a date, a Out[7]: [1, 'hi!', 3]
name, etc.). See https://docs.python.org/tutorial/datastructures.html#dictionaries for more information.
• to change a list in place, use indexing/slices:
A dictionary can have keys (resp. values) with different types:
In [1]: a = [1, 2, 3]
>>> d = {'a':1, 'b':2, 3:'hello'} In [3]: a
>>> d Out[3]: [1, 2, 3]
{'a': 1, 3: 'hello', 'b': 2} In [4]: a = ['a', 'b', 'c'] # Creates another object.
In [5]: a
Out[5]: ['a', 'b', 'c']
In [6]: id(a)
Out[6]: 138641676
More container types
In [7]: a[:] = [1, 2, 3] # Modifies object in place.
In [8]: a
Tuples Out[8]: [1, 2, 3]
Tuples are basically immutable lists. The elements of a tuple are written between parentheses, or just separated In [9]: id(a)
by commas: Out[9]: 138641676 # Same as in Out[6], yours will differ...

>>> t = 12345, 54321, 'hello!' • the key concept here is mutable vs. immutable
>>> t[0]
12345 – mutable objects can be changed in place
>>> t – immutable objects cannot be modified once created
(12345, 54321, 'hello!')
>>> u = (0, 2) See also:
A very good and detailed explanation of the above issues can be found in David M. Beazley’s article Types and
Sets: unordered, unique items:
Objects in Python.
>>> s = set(('a', 'b', 'c', 'a'))
>>> s
set(['a', 'c', 'b']) 2.3 Control Flow
>>> s.difference(('a', 'b'))
set(['c'])
Controls the order in which the code is executed.

2.2.3 Assignment operator 2.3.1 if/elif/else

>>> if 2**2 == 4:
Tip: Python library reference says: ... print('Obvious!')
Assignment statements are used to (re)bind names to values and to modify attributes or items of ...
Obvious!
mutable objects.
In short, it works as follows (simple assignment): Blocks are delimited by indentation
1. an expression on the right hand side is evaluated, the corresponding object is created/obtained
2. a name on the left hand side is assigned, or bound, to the r.h.s. object Tip: Type the following lines in your Python interpreter, and be careful to respect the indentation depth. The
Ipython shell automatically increases the indentation depth after a colon : sign; to decrease the indentation
depth, go four spaces to the left with the Backspace key. Press the Enter key twice to leave the logical block.
Things to note:
• a single object can have several names bound to it: >>> a = 10
In [1]: a = [1, 2, 3]
>>> if a == 1:
In [2]: b = a
... print(1)
In [3]: a
... elif a == 2:
Out[3]: [1, 2, 3]
... print(2)
In [4]: b
... else:
Out[4]: [1, 2, 3]
... print('A lot')
In [5]: a is b
A lot
Out[5]: True

2.2. Basic types 20 2.3. Control Flow 21

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Indentation is compulsory in scripts as well. As an exercise, re-type the previous lines with the same indenta- • any number equal to zero (0, 0.0, 0+0j)
tion in a script condition.py, and execute the script with run condition.py in Ipython.
• an empty container (list, tuple, set, dictionary, . . . )
• False, None
2.3.2 for/range
Evaluates to True:
Iterating with an index: • everything else

>>> for i in range(4): a == b Tests equality, with logics:

... print(i)
0 >>> 1 == 1.
1 True
2
3 a is b Tests identity: both sides are the same object:

But most often, it is more readable to iterate over values: >>> 1 is 1.

False
>>> for word in ('cool', 'powerful', 'readable'):
... print('Python is %s ' % word) >>> a = 1
Python is cool >>> b = 1
Python is powerful >>> a is b
Python is readable True

a in b For any collection b: b contains a

2.3.3 while/break/continue >>> b = [1, 2, 3]
>>> 2 in b
Typical C-style while loop (Mandelbrot problem): True
>>> 5 in b
>>> z = 1 + 1j False
>>> while abs(z) < 100:
... z = z**2 + 1 If b is a dictionary, this tests that a is a key of b.
>>> z
(-134+352j)
2.3.5 Advanced iteration
More advanced features
break out of enclosing for/while loop: Iterate over any sequence

>>> z = 1 + 1j You can iterate over any sequence (string, list, keys in a dictionary, lines in a file, . . . ):

>>> while abs(z) < 100: >>> vowels = 'aeiouy'

... if z.imag == 0:
... break >>> for i in 'powerful':
... z = z**2 + 1 ... if i in vowels:
... print(i)
continue the next iteration of a loop.: o
e
>>> a = [1, 0, 2, 4] u
>>> for element in a:
... if element == 0:
>>> message = "Hello how are you?"
... continue
>>> message.split() # returns a list
... print(1. / element)
['Hello', 'how', 'are', 'you?']
1.0
>>> for word in message.split():
0.5
... print(word)
0.25
...
Hello
how
2.3.4 Conditional Expressions are
you?
if <OBJECT>
Evaluates to False:

2.3. Control Flow 22 2.3. Control Flow 23

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Exercise
Tip: Few languages (in particular, languages for scientific computing) allow to loop over anything but in-
tegers/indices. With Python it is possible to loop exactly over the objects of interest without bothering with Compute the decimals of Pi using the Wallis formula:
indices you often don’t care about. This feature can often be used to make code more readable.
∞
Y 4i 2
π=2
i =1 4i 2 − 1

Warning: Not safe to modify the sequence you are iterating over.

2.4 Defining functions

Keeping track of enumeration number

Common task is to iterate over a sequence while keeping track of the item number. 2.4.1 Function definition
• Could use while loop with a counter as above. Or a for loop:
In [56]: def test():
>>> words = ('cool', 'powerful', 'readable') ....: print('in test function')
>>> for i in range(0, len(words)): ....:
... print((i, words[i])) ....:
(0, 'cool')
(1, 'powerful') In [57]: test()
(2, 'readable') in test function

• But, Python provides a built-in function - enumerate - for this:

Warning: Function blocks must be indented as other control-flow blocks.
>>> for index, item in enumerate(words):
... print((index, item))
(0, 'cool')
(1, 'powerful') 2.4.2 Return statement
(2, 'readable')
Functions can optionally return values.

Looping over a dictionary In [6]: def disk_area(radius):

...: return 3.14 * radius * radius
Use items: ...:

>>> d = {'a': 1, 'b':1.2, 'c':1j} In [8]: disk_area(1.5)

Out[8]: 7.0649999999999995
>>> for key, val in sorted(d.items()):
... print('Key: %s has value: %s ' % (key, val))
Key: a has value: 1 Note: By default, functions return None.
Key: b has value: 1.2
Key: c has value: 1j

Note: Note the syntax to define a function:

Note: The ordering of a dictionary in random, thus we use sorted() which will sort on the keys. • the def keyword;
• is followed by the function’s name, then
• the arguments of the function are given between parentheses followed by a colon.
2.3.6 List Comprehensions
• the function body;
>>> [i**2 for i in range(4)] • and return object for optionally returning values.
[0, 1, 4, 9]

2.4.3 Parameters

Mandatory parameters (positional arguments)

2.3. Control Flow 24 2.4. Defining functions 25

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

In [81]: def double_it(x):

....: return x * 2
....:
Tip: More involved example implementing python’s slicing:
In [82]: double_it(3)
Out[82]: 6 In [98]: def slicer(seq, start=None, stop=None, step=None):
....: """Implement basic python slicing."""
In [83]: double_it() ....: return seq[start:stop:step]
--------------------------------------------------------------------------- ....:
Traceback (most recent call last):
File "<stdin>", line 1, in <module> In [101]: rhyme = 'one fish, two fish, red fish, blue fish'.split()
TypeError: double_it() takes exactly 1 argument (0 given)
In [102]: rhyme
Optional parameters (keyword or named arguments) Out[102]: ['one', 'fish,', 'two', 'fish,', 'red', 'fish,', 'blue', 'fish']

In [84]: def double_it(x=2): In [103]: slicer(rhyme)

....: return x * 2 Out[103]: ['one', 'fish,', 'two', 'fish,', 'red', 'fish,', 'blue', 'fish']
....:
In [104]: slicer(rhyme, step=2)
In [85]: double_it() Out[104]: ['one', 'two', 'red', 'blue']
Out[85]: 4
In [105]: slicer(rhyme, 1, step=2)
In [86]: double_it(3) Out[105]: ['fish,', 'fish,', 'fish,', 'fish']
Out[86]: 6
In [106]: slicer(rhyme, start=1, stop=4, step=2)
Keyword arguments allow you to specify default values. Out[106]: ['fish,', 'fish,']

The order of the keyword arguments does not matter:

Warning: Default values are evaluated when the function is defined, not when it is called. This can be
problematic when using mutable types (e.g. dictionary or list) and modifying them in the function body, In [107]: slicer(rhyme, step=2, start=1, stop=4)
since the modifications will be persistent across invocations of the function. Out[107]: ['fish,', 'fish,']
Using an immutable type in a keyword argument:
but it is good practice to use the same ordering as the function’s definition.
In [124]: bigx = 10

In [125]: def double_it(x=bigx): Keyword arguments are a very convenient feature for defining functions with a variable number of arguments,
.....: return x * 2 especially when default values are to be used in most calls to the function.
.....:

In [126]: bigx = 1e9 # Now really big 2.4.4 Passing by value

In [128]: double_it()
Out[128]: 20 Tip: Can you modify the value of a variable inside a function? Most languages (C, Java, . . . ) distinguish
Using an mutable type in a keyword argument (and modifying it inside the function body): “passing by value” and “passing by reference”. In Python, such a distinction is somewhat artificial, and it is a
bit subtle whether your variables are going to be modified or not. Fortunately, there exist clear rules.
In [2]: def add_to_dict(args={'a': 1, 'b': 2}):
...: for i in args.keys(): Parameters to functions are references to objects, which are passed by value. When you pass a variable to a
...: args[i] += 1 function, python passes the reference to the object to which the variable refers (the value). Not the variable
...: print args itself.
...:

In [3]: add_to_dict If the value passed in a function is immutable, the function does not modify the caller’s variable. If the value
Out[3]: <function __main__.add_to_dict> is mutable, the function may modify the caller’s variable in-place:

In [4]: add_to_dict() >>> def try_to_modify(x, y, z):

{'a': 2, 'b': 3} ... x = 23
... y.append(42)
In [5]: add_to_dict() ... z = [99] # new reference
{'a': 3, 'b': 4} ... print(x)
... print(y)
In [6]: add_to_dict() ... print(z)
{'a': 4, 'b': 5} ...

2.4. Defining functions 26 2.4. Defining functions 27

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

>>> a = 77 # immutable variable

2.4.6 Variable number of parameters
>>> b = [99] # mutable variable
>>> c = [28] Special forms of parameters:
>>> try_to_modify(a, b, c) • *args: any number of positional arguments packed into a tuple
23
[99, 42] • **kwargs: any number of keyword arguments packed into a dictionary
[99]
>>> print(a) In [35]: def variable_args(*args, **kwargs):
77 ....: print 'args is', args
>>> print(b) ....: print 'kwargs is', kwargs
[99, 42] ....:
>>> print(c)
[28] In [36]: variable_args('one', 'two', x=1, y=2, z=3)
args is ('one', 'two')
Functions have a local variable table called a local namespace. kwargs is {'y': 2, 'x': 1, 'z': 3}

The variable x only exists within the function try_to_modify.

2.4.7 Docstrings
2.4.5 Global variables
Documentation about what the function does and its parameters. General convention:
Variables declared outside the function can be referenced within the function: In [67]: def funcname(params):
....: """Concise one-line sentence describing the function.
In [114]: x = 5
....:
....: Extended summary which can contain multiple paragraphs.
In [115]: def addx(y):
....: """
.....: return x + y
....: # function body
.....:
....: pass
....:
In [116]: addx(10)
Out[116]: 15
In [68]: funcname?
Type: function
But these “global” variables cannot be modified within the function, unless declared global in the function. Base Class: type 'function'>
This doesn’t work: String Form: <function funcname at 0xeaa0f0>
Namespace: Interactive
In [117]: def setx(y): File: <ipython console>
.....: x = y Definition: funcname(params)
.....: print('x is %d ' % x) Docstring:
.....: Concise one-line sentence describing the function.
.....:
Extended summary which can contain multiple paragraphs.
In [118]: setx(10)
x is 10
Note: Docstring guidelines
In [120]: x
Out[120]: 5 For the sake of standardization, the Docstring Conventions webpage documents the semantics and conven-
tions associated with Python docstrings.
This works: Also, the Numpy and Scipy modules have defined a precise standard for documenting scientific func-
In [121]: def setx(y): tions, that you may want to follow for your own functions, with a Parameters section, an Examples sec-
.....: global x tion, etc. See http://projects.scipy.org/numpy/wiki/CodingStyleGuidelines#docstring-standard and http://
.....: x = y projects.scipy.org/numpy/browser/trunk/doc/example.py#L37
.....: print('x is %d ' % x)
.....:
.....:
2.4.8 Functions are objects
In [122]: setx(10)
x is 10 Functions are first-class objects, which means they can be:

In [123]: x • assigned to a variable

Out[123]: 10
• an item in a list (or any collection)

2.4. Defining functions 28 2.4. Defining functions 29

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

• passed as an argument to another function. The extension for Python files is .py. Write or copy-and-paste the following lines in a file called test.py

In [38]: va = variable_args message = "Hello how are you?"

for word in message.split():
In [39]: va('three', x=1, y=2) print word
args is ('three',)
kwargs is {'y': 2, 'x': 1}
Tip: Let us now execute the script interactively, that is inside the Ipython interpreter. This is maybe the most
common use of scripts in scientific computing.
2.4.9 Methods

Methods are functions attached to objects. You’ve seen these in our examples on lists, dictionaries, strings, Note: in Ipython, the syntax to execute a script is %run script.py. For example,
etc. . .

In [1]: %run test.py

2.4.10 Exercises Hello
how
are
Exercise: Fibonacci sequence you?

In [2]: message
Write a function that displays the n first terms of the Fibonacci sequence, defined by:
Out[2]: 'Hello how are you?'

 U0 = 0
U =1 The script has been executed. Moreover the variables defined in the script (such as message) are now available
 1 inside the interpreter’s namespace.
Un+2 = Un+1 +Un

Tip: Other interpreters also offer the possibility to execute scripts (e.g., execfile in the plain Python inter-
Exercise: Quicksort preter, etc.).

Implement the quicksort algorithm, as defined by wikipedia It is also possible In order to execute this script as a standalone program, by executing the script inside a shell
terminal (Linux/Mac console or cmd Windows console). For example, if we are in the same directory as the
test.py file, we can execute this in a console:
function quicksort(array)
var list less, greater $ python test.py
if length(array) < 2 Hello
return array how
select and remove a pivot value pivot from array are
for each x in array you?
if x < pivot + 1 then append x to less
else append x to greater
return concatenate(quicksort(less), pivot, quicksort(greater))
Tip: Standalone scripts may also take command-line arguments
In file.py:

2.5 Reusing code: scripts and modules import sys

print sys.argv

For now, we have typed all instructions in the interpreter. For longer sets of instructions we need to change
track and write the code in text files (using a text editor), that we will call either scripts or modules. Use your fa- $ python file.py test arguments
['file.py', 'test', 'arguments']
vorite text editor (provided it offers syntax highlighting for Python), or the editor that comes with the Scientific
Python Suite you may be using.

Warning: Don’t implement option parsing yourself. Use modules such as optparse, argparse or
2.5.1 Scripts :mod‘docopt‘.

Tip: Let us first write a script, that is a file with a sequence of instructions that are executed each time the script
is called. Instructions may be e.g. copied-and-pasted from the interpreter (but take care to respect indentation
rules!).

2.5. Reusing code: scripts and modules 30 2.5. Reusing code: scripts and modules 31
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

2.5.2 Importing objects from modules our own modules.

In [1]: import os Let us create a module demo contained in the file demo.py:

In [2]: os "A demo module."

Out[2]: <module 'os' from '/usr/lib/python2.6/os.pyc'>
def print_b():
In [3]: os.listdir('.') "Prints b."
Out[3]: print 'b'
['conf.py',
'basic_types.rst', def print_a():
'control_flow.rst', "Prints a."
'functions.rst', print 'a'
'python_language.rst',
'reusing.rst',
'file_io.rst', c = 2
'exceptions.rst', d = 2
'workflow.rst',
'index.rst']
Tip: In this file, we defined two functions print_a and print_b. Suppose we want to call the print_a
And also: function from the interpreter. We could execute the file as a script, but since we just want to have access to the
In [4]: from os import listdir
function print_a, we are rather going to import it as a module. The syntax is as follows.

Importing shorthands: In [1]: import demo

In [5]: import numpy as np
In [2]: demo.print_a()
a
Warning:
In [3]: demo.print_b()
from os import * b
This is called the star import and please, Do not use it
Importing the module gives access to its objects, using the module.object syntax. Don’t forget to put the
• Makes the code harder to read and understand: where do symbols come from? module’s name before the object’s name, otherwise Python won’t recognize the instruction.
• Makes it impossible to guess the functionality by the context and the name (hint: os.name is the Introspection
name of the OS), and to profit usefully from tab completion.
In [4]: demo?
• Restricts the variable names you can use: os.name might override name, or vise-versa. Type: module
• Creates possible name clashes between modules. Base Class: <type 'module'>
String Form: <module 'demo' from 'demo.py'>
• Makes the code impossible to statically check for undefined symbols. Namespace: Interactive
File: /home/varoquau/Projects/Python_talks/scipy_2009_tutorial/source/demo.py
Docstring:
A demo module.
Tip: Modules are thus a good way to organize code in a hierarchical way. Actually, all the scientific computing
tools we are going to use are modules:
In [5]: who
>>> import numpy as np # data arrays demo
>>> np.linspace(0, 10, 6)
array([ 0., 2., 4., 6., 8., 10.]) In [6]: whos
>>> import scipy # scientific computing Variable Type Data/Info
------------------------------
demo module <module 'demo' from 'demo.py'>

In [7]: dir(demo)
2.5.3 Creating modules Out[7]:
['__builtins__',
'__doc__',
Tip: If we want to write larger and better organized programs (compared to simple scripts), where some '__file__',
objects are defined, (variables, functions, classes) and that we want to reuse several times, we have to create '__name__',

2.5. Reusing code: scripts and modules 32 2.5. Reusing code: scripts and modules 33
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

'__package__',
'c', Importing it:
'd',
'print_a', In [11]: import demo2
'print_b'] b

In [12]: import demo2

In [8]: demo.
demo.c demo.print_a demo.py Running it:
demo.d demo.print_b demo.pyc
In [13]: %run demo2
Importing objects from modules into the main namespace b
a
In [9]: from demo import print_a, print_b

In [10]: whos
Variable Type Data/Info 2.5.5 Scripts or modules? How to organize your code
--------------------------------
demo module <module 'demo' from 'demo.py'>
print_a function <function print_a at 0xb7421534> Note: Rule of thumb
print_b function <function print_b at 0xb74214c4>
• Sets of instructions that are called several times should be written inside functions for better code
In [11]: print_a() reusability.
a
• Functions (or other bits of code) that are called from several scripts should be written inside a module,
so that only the module is imported in the different scripts (do not copy-and-paste your functions in the
different scripts!).
Warning: Module caching
Modules are cached: if you modify demo.py and re-import it in the old session, you will get the
old one.
How modules are found and imported
Solution:
In [10]: reload(demo) When the import mymodule statement is executed, the module mymodule is searched in a given list of direc-
tories. This list includes a list of installation-dependent default path (e.g., /usr/lib/python) as well as the
In Python3 instead reload is not builtin, so you have to import the importlib module first and then do: list of directories specified by the environment variable PYTHONPATH.
In [10]: importlib.reload(demo) The list of directories searched by Python is given by the sys.path variable

In [1]: import sys

2.5.4 ‘__main__’ and module loading In [2]: sys.path

Out[2]:
['',
'/home/varoquau/.local/bin',
Tip: Sometimes we want code to be executed when a module is run directly, but not when it is imported by '/usr/lib/python2.7',
another module. if __name__ == '__main__' allows us to check whether the module is being run directly. '/home/varoquau/.local/lib/python2.7/site-packages',
'/usr/lib/python2.7/dist-packages',
'/usr/local/lib/python2.7/dist-packages',
File demo2.py:
...]
def print_b():
"Prints b." Modules must be located in the search path, therefore you can:
print 'b'
• write your own modules within directories already defined in the search path (e.g. $HOME/.local/lib/
def print_a(): python2.7/dist-packages). You may use symbolic links (on Linux) to keep the code somewhere else.
"Prints a." • modify the environment variable PYTHONPATH to include the directories containing the user-defined
print 'a'
modules.
# print_b() runs on import
print_b() Tip: On Linux/Unix, add the following line to a file read by the shell at startup (e.g. /etc/profile, .profile)

if __name__ == '__main__': export PYTHONPATH=$PYTHONPATH:/home/emma/user_defined_modules

# print_a() is only executed when the module is run directly.
print_a()

2.5. Reusing code: scripts and modules 34 2.5. Reusing code: scripts and modules 35
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

On Windows, http://support.microsoft.com/kb/310519 explains how to handle environment variables.

In [6]: from scipy.ndimage import morphology

• or modify the sys.path variable itself within a Python script. In [17]: morphology.binary_dilation?
Type: function
Base Class: <type 'function'>
Tip: String Form: <function binary_dilation at 0x9bedd84>
Namespace: Interactive
import sys File: /usr/lib/python2.6/dist-packages/scipy/ndimage/morphology.py
new_path = '/home/emma/user_defined_modules' Definition: morphology.binary_dilation(input, structure=None,
if new_path not in sys.path: iterations=1, mask=None, output=None, border_value=0, origin=0,
sys.path.append(new_path) brute_force=False)
Docstring:
This method is not very robust, however, because it makes the code less portable (user-dependent path) Multi-dimensional binary dilation with the given structure.
and because you have to add the directory to your sys.path each time you want to import from a module
in this directory. An output array can optionally be provided. The origin parameter
controls the placement of the filter. If no structuring element is
provided an element is generated with a squared connectivity equal
See also: to one. The dilation operation is repeated iterations times. If
iterations is less than 1, the dilation is repeated until the
See https://docs.python.org/tutorial/modules.html for more information about modules.
result does not change anymore. If a mask is given, only those
elements with a true value at the corresponding mask element are
modified at each iteration.
2.5.6 Packages

A directory that contains many modules is called a package. A package is a module with submodules (which
can have submodules themselves, etc.). A special file called __init__.py (which may be empty) tells Python 2.5.7 Good practices
that the directory is a Python package, from which modules can be imported.
• Use meaningful object names
$ ls
cluster/ io/ README.txt@ stsci/ • Indentation: no choice!
__config__.py@ LATEST.txt@ setup.py@ __svn_version__.py@
__config__.pyc lib/ setup.pyc __svn_version__.pyc
constants/ linalg/ setupscons.py@ THANKS.txt@ Tip: Indenting is compulsory in Python! Every command block following a colon bears an additional
fftpack/ linsolve/ setupscons.pyc TOCHANGE.txt@ indentation level with respect to the previous line with a colon. One must therefore indent after def
__init__.py@ maxentropy/ signal/ version.py@ f(): or while:. At the end of such logical blocks, one decreases the indentation depth (and re-increases
__init__.pyc misc/ sparse/ version.pyc it if a new block is entered, etc.)
INSTALL.txt@ ndimage/ spatial/ weave/
Strict respect of indentation is the price to pay for getting rid of { or ; characters that delineate logical
integrate/ odr/ special/
interpolate/ optimize/ stats/ blocks in other languages. Improper indentation leads to errors such as
$ cd ndimage
------------------------------------------------------------
$ ls
IndentationError: unexpected indent (test.py, line 2)
doccer.py@ fourier.pyc interpolation.py@ morphology.pyc setup.pyc
doccer.pyc info.py@ interpolation.pyc _nd_image.so
setupscons.py@ All this indentation business can be a bit confusing in the beginning. However, with the clear indenta-
filters.py@ info.pyc measurements.py@ _ni_support.py@ tion, and in the absence of extra characters, the resulting code is very nice to read compared to other
setupscons.pyc languages.
filters.pyc __init__.py@ measurements.pyc _ni_support.pyc tests/
fourier.py@ __init__.pyc morphology.py@ setup.py@
• Indentation depth: Inside your text editor, you may choose to indent with any positive number of spaces
(1, 2, 3, 4, . . . ). However, it is considered good practice to indent with 4 spaces. You may configure your
From Ipython:
editor to map the Tab key to a 4-space indentation.
In [1]: import scipy
• Style guidelines
In [2]: scipy.__file__ Long lines: you should not write very long lines that span over more than (e.g.) 80 characters. Long lines
Out[2]: '/usr/lib/python2.6/dist-packages/scipy/__init__.pyc' can be broken with the \ character

In [3]: import scipy.version >>> long_line = "Here is a very very long line \
... that we break in two parts."
In [4]: scipy.version.version
Out[4]: '0.7.0' Spaces

In [5]: import scipy.ndimage.morphology Write well-spaced code: put whitespaces after commas, around arithmetic operators, etc.:

2.5. Reusing code: scripts and modules 36 2.5. Reusing code: scripts and modules 37
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

File modes
>>> a = 1 # yes
>>> a=1 # too cramped
• Read-only: r
A certain number of rules for writing “beautiful” code (and more importantly using the same conven- • Write-only: w
tions as anybody else!) are given in the Style Guide for Python Code.
– Note: Create a new file or overwrite existing file.
• Append a file: a

Quick read • Read and Write: r+

• Binary mode: b
If you want to do a first quick pass through the Scipy lectures to learn the ecosystem, you can directly skip
to the next chapter: NumPy: creating and manipulating numerical data. – Note: Use for binary files, especially on Windows.

The remainder of this chapter is not necessary to follow the rest of the intro part. But be sure to come back

2.7 Standard Library

and finish this chapter later.

2.6 Input and Output Note: Reference document for this section:
• The Python Standard Library documentation: https://docs.python.org/library/index.html
To be exhaustive, here are some information about input and output in Python. Since we will use the Numpy • Python Essential Reference, David Beazley, Addison-Wesley Professional
methods to read and write files, you may skip this chapter at first reading.
We write or read strings to/from files (other types must be converted to strings). To write in a file:

>>> f = open('workfile', 'w') # opens the workfile file 2.7.1 os module: operating system functionality
>>> type(f)
<type 'file'> “A portable way of using operating system dependent functionality.”
>>> f.write('This is a test \nand another test')
>>> f.close()
Directory and file manipulation
To read from a file
Current directory:
In [1]: f = open('workfile', 'r')
In [17]: os.getcwd()
In [2]: s = f.read() Out[17]: '/Users/cburns/src/scipy2009/scipy_2009_tutorial/source'

In [3]: print(s) List a directory:

This is a test
and another test In [31]: os.listdir(os.curdir)
Out[31]:
In [4]: f.close() ['.index.rst.swo',
'.python_language.rst.swp',
See also: '.view_array.py.swp',
'_static',
For more details: https://docs.python.org/tutorial/inputoutput.html '_templates',
'basic_types.rst',
'conf.py',
2.6.1 Iterating over a file 'control_flow.rst',
'debugging.rst',
In [6]: f = open('workfile', 'r') ...

In [7]: for line in f: Make a directory:

...: print line
...: In [32]: os.mkdir('junkdir')
This is a test
In [33]: 'junkdir' in os.listdir(os.curdir)
and another test Out[33]: True

In [8]: f.close() Rename the directory:

2.6. Input and Output 38 2.7. Standard Library 39

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

In [36]: os.rename('junkdir', 'foodir')

In [88]: os.path.expanduser('~/local')
In [37]: 'junkdir' in os.listdir(os.curdir) Out[88]: '/Users/cburns/local'
Out[37]: False
In [92]: os.path.join(os.path.expanduser('~'), 'local', 'bin')
In [38]: 'foodir' in os.listdir(os.curdir) Out[92]: '/Users/cburns/local/bin'
Out[38]: True

In [41]: os.rmdir('foodir')
Running an external command
In [42]: 'foodir' in os.listdir(os.curdir)
Out[42]: False In [8]: os.system('ls')
basic_types.rst demo.py functions.rst python_language.rst standard_library.rst
Delete a file: control_flow.rst exceptions.rst io.rst python-logo.png
demo2.py first_steps.rst oop.rst reusing_code.rst
In [44]: fp = open('junk.txt', 'w')

In [45]: fp.close() Note: Alternative to os.system

In [46]: 'junk.txt' in os.listdir(os.curdir) A noteworthy alternative to os.system is the sh module. Which provides much more convenient ways to
Out[46]: True obtain the output, error stream and exit code of the external command.

In [47]: os.remove('junk.txt') In [20]: import sh

In [20]: com = sh.ls()
In [48]: 'junk.txt' in os.listdir(os.curdir)
Out[48]: False In [21]: print com
basic_types.rst exceptions.rst oop.rst standard_library.rst
control_flow.rst first_steps.rst python_language.rst
demo2.py functions.rst python-logo.png
os.path: path manipulations demo.py io.rst reusing_code.rst

os.path provides common operations on pathnames. In [22]: print com.exit_code

0
In [70]: fp = open('junk.txt', 'w') In [23]: type(com)
Out[23]: sh.RunningCommand
In [71]: fp.close()

In [72]: a = os.path.abspath('junk.txt')

In [73]: a Walking a directory

Out[73]: '/Users/cburns/src/scipy2009/scipy_2009_tutorial/source/junk.txt'
os.path.walk generates a list of filenames in a directory tree.
In [74]: os.path.split(a)
Out[74]: ('/Users/cburns/src/scipy2009/scipy_2009_tutorial/source', In [10]: for dirpath, dirnames, filenames in os.walk(os.curdir):
'junk.txt') ....: for fp in filenames:
....: print os.path.abspath(fp)
In [78]: os.path.dirname(a) ....:
Out[78]: '/Users/cburns/src/scipy2009/scipy_2009_tutorial/source' ....:
/Users/cburns/src/scipy2009/scipy_2009_tutorial/source/.index.rst.swo
In [79]: os.path.basename(a) /Users/cburns/src/scipy2009/scipy_2009_tutorial/source/.view_array.py.swp
Out[79]: 'junk.txt' /Users/cburns/src/scipy2009/scipy_2009_tutorial/source/basic_types.rst
/Users/cburns/src/scipy2009/scipy_2009_tutorial/source/conf.py
In [80]: os.path.splitext(os.path.basename(a)) /Users/cburns/src/scipy2009/scipy_2009_tutorial/source/control_flow.rst
Out[80]: ('junk', '.txt') ...

In [84]: os.path.exists('junk.txt')
Out[84]: True Environment variables:
In [86]: os.path.isfile('junk.txt')
Out[86]: True In [9]: import os

In [87]: os.path.isdir('junk.txt') In [11]: os.environ.keys()

Out[87]: False Out[11]:

2.7. Standard Library 40 2.7. Standard Library 41

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

['_', In [119]: sys.prefix

'FSLDIR', Out[119]: '/Library/Frameworks/Python.framework/Versions/2.5'
'TERM_PROGRAM_VERSION',
'FSLREMOTECALL', • List of command line arguments passed to a Python script:
'USER',
'HOME', In [100]: sys.argv
'PATH', Out[100]: ['/Users/cburns/local/bin/ipython']
'PS1',
'SHELL', sys.path is a list of strings that specifies the search path for modules. Initialized from PYTHONPATH:
'EDITOR',
'WORKON_HOME', In [121]: sys.path
'PYTHONPATH', Out[121]:
... ['',
'/Users/cburns/local/bin',
In [12]: os.environ['PYTHONPATH'] '/Users/cburns/local/lib/python2.5/site-packages/grin-1.1-py2.5.egg',
Out[12]: '.:/Users/cburns/src/utils:/Users/cburns/src/nitools: '/Users/cburns/local/lib/python2.5/site-packages/argparse-0.8.0-py2.5.egg',
/Users/cburns/local/lib/python2.5/site-packages/: '/Users/cburns/local/lib/python2.5/site-packages/urwid-0.9.7.1-py2.5.egg',
/usr/local/lib/python2.5/site-packages/: '/Users/cburns/local/lib/python2.5/site-packages/yolk-0.4.1-py2.5.egg',
/Library/Frameworks/Python.framework/Versions/2.5/lib/python2.5' '/Users/cburns/local/lib/python2.5/site-packages/virtualenv-1.2-py2.5.egg',
...
In [16]: os.getenv('PYTHONPATH')
Out[16]: '.:/Users/cburns/src/utils:/Users/cburns/src/nitools:
/Users/cburns/local/lib/python2.5/site-packages/:
/usr/local/lib/python2.5/site-packages/: 2.7.5 pickle: easy persistence
/Library/Frameworks/Python.framework/Versions/2.5/lib/python2.5'
Useful to store arbitrary objects to a file. Not safe or fast!

In [1]: import pickle

2.7.2 shutil: high-level file operations
In [2]: l = [1, None, 'Stan']
The shutil provides useful file operations:
In [3]: pickle.dump(l, file('test.pkl', 'w'))
• shutil.rmtree: Recursively delete a directory tree.
• shutil.move: Recursively move a file or directory to another location. In [4]: pickle.load(file('test.pkl'))
Out[4]: [1, None, 'Stan']
• shutil.copy: Copy files or directories.

Exercise
2.7.3 glob: Pattern matching on files
Write a program to search your PYTHONPATH for the module site.py.
The glob module provides convenient file pattern matching.
Find all files ending in .txt:
path_site
In [18]: import glob

In [19]: glob.glob('*.txt') 2.8 Exception handling in Python

Out[19]: ['holy_grail.txt', 'junk.txt', 'newfile.txt']

It is likely that you have raised Exceptions if you have typed all the previous commands of the tutorial. For
example, you may have raised an exception if you entered a command with a typo.
2.7.4 sys module: system-specific information
Exceptions are raised by different kinds of errors arising when executing Python code. In your own code, you
System-specific information related to the Python interpreter. may also catch errors, or define custom error types. You may want to look at the descriptions of the the built-in
Exceptions when looking for the right exception type.
• Which version of python are you running and where is it installed:

In [117]: sys.platform
Out[117]: 'darwin' 2.8.1 Exceptions

In [118]: sys.version Exceptions are raised by errors in Python:

Out[118]: '2.5.2 (r252:60911, Feb 22 2008, 07:57:53) \n
[GCC 4.0.1 (Apple Computer, Inc. build 5363)]'

2.7. Standard Library 42 2.8. Exception handling in Python 43

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Easier to ask for forgiveness than for permission

In [1]: 1/0
---------------------------------------------------------------------------
ZeroDivisionError: integer division or modulo by zero In [11]: def print_sorted(collection):
....: try:
In [2]: 1 + 'e' ....: collection.sort()
--------------------------------------------------------------------------- ....: except AttributeError:
TypeError: unsupported operand type(s) for +: 'int' and 'str' ....: pass
....: print(collection)
In [3]: d = {1:1, 2:2} ....:
....:
In [4]: d[3]
--------------------------------------------------------------------------- In [12]: print_sorted([1, 3, 2])
KeyError: 3 [1, 2, 3]

In [5]: l = [1, 2, 3] In [13]: print_sorted(set((1, 3, 2)))

set([1, 2, 3])
In [6]: l[4]
--------------------------------------------------------------------------- In [14]: print_sorted('132')
IndexError: list index out of range 132

In [7]: l.foobar
---------------------------------------------------------------------------
AttributeError: 'list' object has no attribute 'foobar' 2.8.3 Raising exceptions

As you can see, there are different types of exceptions for different errors. • Capturing and reraising an exception:

In [15]: def filter_name(name):

....: try:
2.8.2 Catching exceptions ....: name = name.encode('ascii')
....: except UnicodeError as e:
try/except ....: if name == 'Gaël':
....: print('OK, Gaël')
....: else:
In [10]: while True:
....: raise e
....: try:
....: return name
....: x = int(raw_input('Please enter a number: '))
....:
....: break
....: except ValueError:
In [16]: filter_name('Gaël')
....: print('That was no valid number. Try again...')
OK, Gaël
....:
Out[16]: 'Ga\xc3\xabl'
Please enter a number: a
That was no valid number. Try again...
In [17]: filter_name('Stéfan')
Please enter a number: 1
---------------------------------------------------------------------------
UnicodeDecodeError: 'ascii' codec can't decode byte 0xc3 in position 2: ordinal not in␣
In [9]: x
,→range(128)
Out[9]: 1

• Exceptions to pass messages between parts of the code:

try/finally In [17]: def achilles_arrow(x):

....: if abs(x - 1) < 1e-3:
In [10]: try: ....: raise StopIteration
....: x = int(raw_input('Please enter a number: ')) ....: x = 1 - (1-x)/2.
....: finally: ....: return x
....: print('Thank you for your input') ....:
....:
....: In [18]: x = 0
Please enter a number: a
Thank you for your input In [19]: while True:
--------------------------------------------------------------------------- ....: try:
ValueError: invalid literal for int() with base 10: 'a' ....: x = achilles_arrow(x)
....: except StopIteration:
....: break
Important for resource management (e.g. closing a file) ....:

2.8. Exception handling in Python 44 2.8. Exception handling in Python 45

 Scipy lecture notes, Edition 2017.1

....:

In [20]: x
Out[20]: 0.9990234375

3
Use exceptions to notify certain conditions are met (e.g. StopIteration) or not (e.g. custom error raising)

2.9 Object-oriented programming (OOP)

Python supports object-oriented programming (OOP). The goals of OOP are:

CHAPTER
• to organize the code, and
• to re-use code in similar contexts.
Here is a small example: we create a Student class, which is an object gathering several custom functions
(methods) and variables (attributes), we will be able to use:

>>> class Student(object):

... def __init__(self, name):

NumPy: creating and manipulating

... self.name = name
... def set_age(self, age):
... self.age = age
...
...
...
def set_major(self, major):
self.major = major numerical data
>>> anna = Student('anna')
>>> anna.set_age(21)
>>> anna.set_major('physics')

In the previous example, the Student class has __init__, set_age and set_major methods. Its at-
tributes are name, age and major. We can call these methods and attributes with the following notation:
classinstance.method or classinstance.attribute. The __init__ constructor is a special method we Authors: Emmanuelle Gouillart, Didrik Pinte, Gaël Varoquaux, and Pauli Virtanen
call with: MyClass(init parameters if any).
This chapter gives an overview of NumPy, the core tool for performant numerical computing with Python.
Now, suppose we want to create a new class MasterStudent with the same methods and attributes as the pre-
vious one, but with an additional internship attribute. We won’t copy the previous class, but inherit from
it:

>>> class MasterStudent(Student):

... internship = 'mandatory, from March to June'
3.1 The NumPy array object
...
>>> james = MasterStudent('james')
>>> james.internship Section contents
'mandatory, from March to June'
>>> james.set_age(23) • What are NumPy and NumPy arrays?
>>> james.age
23 • Creating arrays
• Basic data types
The MasterStudent class inherited from the Student attributes and methods.
• Basic visualization
Thanks to classes and object-oriented programming, we can organize code with different classes correspond-
ing to different objects we encounter (an Experiment class, an Image class, a Flow class, etc.), with their own • Indexing and slicing
methods and attributes. Then we can use inheritance to consider variations around a base class and re-use • Copies and views
code. Ex : from a Flow base class, we can create derived StokesFlow, TurbulentFlow, PotentialFlow, etc.
• Fancy indexing

2.9. Object-oriented programming (OOP) 46 47

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

3.1.1 What are NumPy and NumPy arrays? • Looking for something:

>>> np.lookfor('create array')

NumPy arrays
Search results for 'create array'
---------------------------------
Python objects numpy.array
• high-level number objects: integers, floating point Create an array.
numpy.memmap
• containers: lists (costless insertion and append), dictionaries (fast lookup) Create a memory-map to an array stored in a *binary* file on disk.

NumPy provides
In [6]: np.con*?
• extension package to Python for multi-dimensional arrays np.concatenate
np.conj
• closer to hardware (efficiency)
np.conjugate
• designed for scientific computation (convenience) np.convolve

• Also known as array oriented computing

Import conventions

The recommended convention to import numpy is:

>>> import numpy as np

>>> import numpy as np
>>> a = np.array([0, 1, 2, 3])
>>> a
array([0, 1, 2, 3]) 3.1.2 Creating arrays

Manual construction of arrays

Tip: For example, An array containing:
• 1-D:
• values of an experiment/simulation at discrete time steps
• signal recorded by a measurement device, e.g. sound wave >>> a = np.array([0, 1, 2, 3])
>>> a
• pixels of an image, grey-level or colour array([0, 1, 2, 3])
>>> a.ndim
• 3-D data measured at different X-Y-Z positions, e.g. MRI scan 1
• ... >>> a.shape
(4,)
>>> len(a)
Why it is useful: Memory-efficient container that provides fast numerical operations. 4

In [1]: L = range(1000)
• 2-D, 3-D, . . . :
In [2]: %timeit [i**2 for i in L] >>> b = np.array([[0, 1, 2], [3, 4, 5]]) # 2 x 3 array
1000 loops, best of 3: 403 us per loop >>> b
array([[0, 1, 2],
In [3]: a = np.arange(1000) [3, 4, 5]])
>>> b.ndim
In [4]: %timeit a**2 2
100000 loops, best of 3: 12.7 us per loop >>> b.shape
(2, 3)
>>> len(b) # returns the size of the first dimension
NumPy Reference documentation 2

>>> c = np.array([[[1], [2]], [[3], [4]]])

• On the web: http://docs.scipy.org/
>>> c
• Interactive help: array([[[1],
[2]],
In [5]: np.array?
String Form:<built-in function array> [[3],
Docstring: [4]]])
array(object, dtype=None, copy=True, order=None, subok=False, ndmin=0, ...

3.1. The NumPy array object 48 3.1. The NumPy array object 49
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

>>> c.shape >>> a = np.random.rand(4) # uniform in [0, 1]

(2, 2, 1) >>> a
array([ 0.95799151, 0.14222247, 0.08777354, 0.51887998])

Exercise: Simple arrays >>> b = np.random.randn(4) # Gaussian

>>> b
array([ 0.37544699, -0.11425369, -0.47616538, 1.79664113])
• Create a simple two dimensional array. First, redo the examples from above. And then create your
own: how about odd numbers counting backwards on the first row, and even numbers on the second? >>> np.random.seed(1234) # Setting the random seed
• Use the functions len(), numpy.shape() on these arrays. How do they relate to each other? And to
the ndim attribute of the arrays?
Exercise: Creating arrays using functions

• Experiment with arange, linspace, ones, zeros, eye and diag.

Functions for creating arrays
• Create different kinds of arrays with random numbers.
• Try setting the seed before creating an array with random values.
Tip: In practice, we rarely enter items one by one. . .
• Look at the function np.empty. What does it do? When might this be useful?

• Evenly spaced:

>>> a = np.arange(10) # 0 .. n-1 (!) 3.1.3 Basic data types

>>> a
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
>>> b = np.arange(1, 9, 2) # start, end (exclusive), step You may have noticed that, in some instances, array elements are displayed with a trailing dot (e.g. 2. vs 2).
>>> b This is due to a difference in the data-type used:
array([1, 3, 5, 7])
>>> a = np.array([1, 2, 3])
>>> a.dtype
• or by number of points: dtype('int64')
>>> c = np.linspace(0, 1, 6) # start, end, num-points
>>> b = np.array([1., 2., 3.])
>>> c
>>> b.dtype
array([ 0. , 0.2, 0.4, 0.6, 0.8, 1. ])
dtype('float64')
>>> d = np.linspace(0, 1, 5, endpoint=False)
>>> d
array([ 0. , 0.2, 0.4, 0.6, 0.8])
Tip: Different data-types allow us to store data more compactly in memory, but most of the time we simply
• Common arrays: work with floating point numbers. Note that, in the example above, NumPy auto-detects the data-type from
the input.
>>> a = np.ones((3, 3)) # reminder: (3, 3) is a tuple
>>> a
array([[ 1., 1., 1.],
[ 1., 1., 1.], You can explicitly specify which data-type you want:
[ 1., 1., 1.]])
>>> b = np.zeros((2, 2)) >>> c = np.array([1, 2, 3], dtype=float)
>>> b >>> c.dtype
array([[ 0., 0.], dtype('float64')
[ 0., 0.]])
>>> c = np.eye(3)
The default data type is floating point:
>>> c
array([[ 1., 0., 0.], >>> a = np.ones((3, 3))
[ 0., 1., 0.], >>> a.dtype
[ 0., 0., 1.]]) dtype('float64')
>>> d = np.diag(np.array([1, 2, 3, 4]))
>>> d
There are also other types:
array([[1, 0, 0, 0],
[0, 2, 0, 0], Complex
[0, 0, 3, 0],
[0, 0, 0, 4]]) >>> d = np.array([1+2j, 3+4j, 5+6*1j])
>>> d.dtype
• np.random: random numbers (Mersenne Twister PRNG): dtype('complex128')

3.1. The NumPy array object 50 3.1. The NumPy array object 51
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Bool
[<matplotlib.lines.Line2D object at ...>]
>>> e = np.array([True, False, False, True]) >>> plt.plot(x, y, 'o') # dot plot
>>> e.dtype [<matplotlib.lines.Line2D object at ...>]
dtype('bool')

Strings

>>> f = np.array(['Bonjour', 'Hello', 'Hallo',])

>>> f.dtype # <--- strings containing max. 7 letters
dtype('S7')

Much more
• int32
• int64
• uint32
• 2D arrays (such as images):
• uint64
>>> x = np.linspace(0, 3, 20)
>>> y = np.linspace(0, 9, 20)
3.1.4 Basic visualization >>> plt.plot(x, y) # line plot
[<matplotlib.lines.Line2D object at ...>]
>>> plt.plot(x, y, 'o') # dot plot
Now that we have our first data arrays, we are going to visualize them.
[<matplotlib.lines.Line2D object at ...>]
Start by launching IPython:

$ ipython

Or the notebook:

$ ipython notebook

Once IPython has started, enable interactive plots:

>>> %matplotlib

Or, from the notebook, enable plots in the notebook:

>>> %matplotlib inline

The inline is important for the notebook, so that plots are displayed in the notebook and not in a new window.
See also:
Matplotlib is a 2D plotting package. We can import its functions as below:
More in the: matplotlib chapter
>>> import matplotlib.pyplot as plt # the tidy way

And then use (note that you have to use show explicitly if you have not enabled interactive plots with Exercise: Simple visualizations
%matplotlib):
• Plot some simple arrays: a cosine as a function of time and a 2D matrix.
>>> plt.plot(x, y) # line plot
• Try using the gray colormap on the 2D matrix.
>>> plt.show() # <-- shows the plot (not needed with interactive plots)

Or, if you have enabled interactive plots with %matplotlib:

>>> plt.plot(x, y) # line plot

3.1.5 Indexing and slicing

The items of an array can be accessed and assigned to the same way as other Python sequences (e.g. lists):
• 1D plotting:
>>> a = np.arange(10)
>>> x = np.linspace(0, 3, 20)
>>> a
>>> y = np.linspace(0, 9, 20)
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
>>> plt.plot(x, y) # line plot

3.1. The NumPy array object 52 3.1. The NumPy array object 53
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

>>> a[0], a[2], a[-1]

(0, 2, 9)

Warning: Indices begin at 0, like other Python sequences (and C/C++). In contrast, in Fortran or Matlab,
indices begin at 1.

The usual python idiom for reversing a sequence is supported:

>>> a[::-1]
array([9, 8, 7, 6, 5, 4, 3, 2, 1, 0])

For multidimensional arrays, indexes are tuples of integers:

>>> a = np.diag(np.arange(3))
>>> a
array([[0, 0, 0],
[0, 1, 0],
[0, 0, 2]])
>>> a[1, 1]
1
>>> a[2, 1] = 10 # third line, second column
>>> a You can also combine assignment and slicing:
array([[ 0, 0, 0],
[ 0, 1, 0], >>> a = np.arange(10)
[ 0, 10, 2]]) >>> a[5:] = 10
>>> a[1] >>> a
array([0, 1, 0]) array([ 0, 1, 2, 3, 4, 10, 10, 10, 10, 10])
>>> b = np.arange(5)
>>> a[5:] = b[::-1]
Note: >>> a
array([0, 1, 2, 3, 4, 4, 3, 2, 1, 0])
• In 2D, the first dimension corresponds to rows, the second to columns.
• for multidimensional a, a[0] is interpreted by taking all elements in the unspecified dimensions.
Exercise: Indexing and slicing

Slicing: Arrays, like other Python sequences can also be sliced: • Try the different flavours of slicing, using start, end and step: starting from a linspace, try to obtain
odd numbers counting backwards, and even numbers counting forwards.
>>> a = np.arange(10)
>>> a • Reproduce the slices in the diagram above. You may use the following expression to create the array:
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
>>> a[2:9:3] # [start:end:step] >>> np.arange(6) + np.arange(0, 51, 10)[:, np.newaxis]
array([2, 5, 8]) array([[ 0, 1, 2, 3, 4, 5],
[10, 11, 12, 13, 14, 15],
[20, 21, 22, 23, 24, 25],
Note that the last index is not included! :
[30, 31, 32, 33, 34, 35],
>>> a[:4] [40, 41, 42, 43, 44, 45],
array([0, 1, 2, 3]) [50, 51, 52, 53, 54, 55]])

All three slice components are not required: by default, start is 0, end is the last and step is 1:

>>> a[1:3] Exercise: Array creation

array([1, 2])
>>> a[::2] Create the following arrays (with correct data types):
array([0, 2, 4, 6, 8]) [[1, 1, 1, 1],
>>> a[3:] [1, 1, 1, 1],
array([3, 4, 5, 6, 7, 8, 9]) [1, 1, 1, 2],
[1, 6, 1, 1]]
A small illustrated summary of NumPy indexing and slicing. . .
[[0., 0., 0., 0., 0.],
[2., 0., 0., 0., 0.],
[0., 3., 0., 0., 0.],
[0., 0., 4., 0., 0.],
3.1. The NumPy array object 54 3.1.
[0.,The
0.,NumPy
0., 5.,array
0.], object 55
[0., 0., 0., 0., 6.]]
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Par on course: 3 statements for each

Hint: Individual array elements can be accessed similarly to a list, e.g. a[1] or a[1, 2].
Hint: Examine the docstring for diag.

Exercise: Tiling for array creation

Skim through the documentation for np.tile, and use this function to construct the array:
[[4, 3, 4, 3, 4, 3],
[2, 1, 2, 1, 2, 1],
[4, 3, 4, 3, 4, 3],
[2, 1, 2, 1, 2, 1]]

3.1.6 Copies and views

Compute prime numbers in 0–99, with a sieve
A slicing operation creates a view on the original array, which is just a way of accessing array data. Thus the
original array is not copied in memory. You can use np.may_share_memory() to check if two arrays share the • Construct a shape (100,) boolean array is_prime, filled with True in the beginning:
same memory block. Note however, that this uses heuristics and may give you false positives. >>> is_prime = np.ones((100,), dtype=bool)
When modifying the view, the original array is modified as well: • Cross out 0 and 1 which are not primes:
>>> a = np.arange(10) >>> is_prime[:2] = 0
>>> a
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9]) • For each integer j starting from 2, cross out its higher multiples:
>>> b = a[::2]
>>> N_max = int(np.sqrt(len(is_prime) - 1))
>>> b
>>> for j in range(2, N_max + 1):
array([0, 2, 4, 6, 8])
... is_prime[2*j::j] = False
>>> np.may_share_memory(a, b)
True • Skim through help(np.nonzero), and print the prime numbers
>>> b[0] = 12
>>> b • Follow-up:
array([12, 2, 4, 6, 8])
>>> a # (!) – Move the above code into a script file named prime_sieve.py
array([12, 1, 2, 3, 4, 5, 6, 7, 8, 9]) – Run it to check it works

>>> a = np.arange(10) – Use the optimization suggested in the sieve of Eratosthenes:

>>> c = a[::2].copy() # force a copy
1. Skip j which are already known to not be primes
>>> c[0] = 12
>>> a 2. The first number to cross out is j 2
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])

>>> np.may_share_memory(a, c)
False 3.1.7 Fancy indexing
This behavior can be surprising at first sight. . . but it allows to save both memory and time.
Tip: NumPy arrays can be indexed with slices, but also with boolean or integer arrays (masks). This method
Worked example: Prime number sieve is called fancy indexing. It creates copies not views.

Using boolean masks

>>> np.random.seed(3)
>>> a = np.random.randint(0, 21, 15)
>>> a
array([10, 3, 8, 0, 19, 10, 11, 9, 10, 6, 0, 20, 12, 7, 14])

3.1. The NumPy array object 56 3.1. The NumPy array object 57
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

>>> (a % 3 == 0)
array([False, True, False, True, False, False, False, True, False,
True, True, False, True, False, False], dtype=bool)
>>> mask = (a % 3 == 0)
>>> extract_from_a = a[mask] # or, a[a%3==0]
>>> extract_from_a # extract a sub-array with the mask
array([ 3, 0, 9, 6, 0, 12])

Indexing with a mask can be very useful to assign a new value to a sub-array:

>>> a[a % 3 == 0] = -1
>>> a
array([10, -1, 8, -1, 19, 10, 11, -1, 10, -1, -1, 20, -1, 7, 14])

Indexing with an array of integers

>>> a = np.arange(0, 100, 10) Exercise: Fancy indexing

>>> a
array([ 0, 10, 20, 30, 40, 50, 60, 70, 80, 90]) • Again, reproduce the fancy indexing shown in the diagram above.
• Use fancy indexing on the left and array creation on the right to assign values into an array, for instance
Indexing can be done with an array of integers, where the same index is repeated several time: by setting parts of the array in the diagram above to zero.
>>> a[[2, 3, 2, 4, 2]] # note: [2, 3, 2, 4, 2] is a Python list
array([20, 30, 20, 40, 20])

New values can be assigned with this kind of indexing: 3.2 Numerical operations on arrays
>>> a[[9, 7]] = -100
>>> a
array([ 0, 10, 20, 30, 40, 50, 60, -100, 80, -100]) Section contents

• Elementwise operations
Tip: When a new array is created by indexing with an array of integers, the new array has the same shape than • Basic reductions
the array of integers:
• Broadcasting
>>> a = np.arange(10)
• Array shape manipulation
>>> idx = np.array([[3, 4], [9, 7]])
>>> idx.shape • Sorting data
(2, 2)
>>> a[idx] • Summary
array([[3, 4],
[9, 7]])

3.2.1 Elementwise operations

Basic operations
The image below illustrates various fancy indexing applications
With scalars:

>>> a = np.array([1, 2, 3, 4])

>>> a + 1
array([2, 3, 4, 5])
>>> 2**a
array([ 2, 4, 8, 16])

All arithmetic operates elementwise:

>>> b = np.ones(4) + 1
>>> a - b
array([-1., 0., 1., 2.])

3.1. The NumPy array object 58 3.2. Numerical operations on arrays 59

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

>>> a * b >>> a = np.array([1, 2, 3, 4])

array([ 2., 4., 6., 8.]) >>> b = np.array([4, 2, 2, 4])
>>> c = np.array([1, 2, 3, 4])
>>> j = np.arange(5) >>> np.array_equal(a, b)
>>> 2**(j + 1) - j False
array([ 2, 3, 6, 13, 28]) >>> np.array_equal(a, c)
True
These operations are of course much faster than if you did them in pure python:

>>> a = np.arange(10000) Logical operations:

>>> %timeit a + 1
10000 loops, best of 3: 24.3 us per loop >>> a = np.array([1, 1, 0, 0], dtype=bool)
>>> l = range(10000) >>> b = np.array([1, 0, 1, 0], dtype=bool)
>>> %timeit [i+1 for i in l] >>> np.logical_or(a, b)
1000 loops, best of 3: 861 us per loop array([ True, True, True, False], dtype=bool)
>>> np.logical_and(a, b)
array([ True, False, False, False], dtype=bool)
Warning: Array multiplication is not matrix multiplication:
Transcendental functions:
>>> c = np.ones((3, 3))
>>> c * c # NOT matrix multiplication! >>> a = np.arange(5)
array([[ 1., 1., 1.], >>> np.sin(a)
[ 1., 1., 1.], array([ 0. , 0.84147098, 0.90929743, 0.14112001, -0.7568025 ])
[ 1., 1., 1.]]) >>> np.log(a)
array([ -inf, 0. , 0.69314718, 1.09861229, 1.38629436])
>>> np.exp(a)
array([ 1. , 2.71828183, 7.3890561 , 20.08553692, 54.59815003])
Note: Matrix multiplication:
Shape mismatches
>>> c.dot(c)
array([[ 3., 3., 3.], >>> a = np.arange(4)
[ 3., 3., 3.], >>> a + np.array([1, 2])
[ 3., 3., 3.]]) Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ValueError: operands could not be broadcast together with shapes (4) (2)

Exercise: Elementwise operations Broadcasting? We’ll return to that later.

Transposition:
• Try simple arithmetic elementwise operations: add even elements with odd elements
>>> a = np.triu(np.ones((3, 3)), 1) # see help(np.triu)
• Time them against their pure python counterparts using %timeit. >>> a
• Generate: array([[ 0., 1., 1.],
[ 0., 0., 1.],
– [2**0, 2**1, 2**2, 2**3, 2**4] [ 0., 0., 0.]])
>>> a.T
– a_j = 2^(3*j) - j
array([[ 0., 0., 0.],
[ 1., 0., 0.],
[ 1., 1., 0.]])
Other operations

Comparisons: Warning: The transposition is a view

>>> a = np.array([1, 2, 3, 4]) As a results, the following code is wrong and will not make a matrix symmetric:
>>> b = np.array([4, 2, 2, 4]) >>> a += a.T
>>> a == b
array([False, True, False, True], dtype=bool) It will work for small arrays (because of buffering) but fail for large one, in unpredictable ways.
>>> a > b
array([False, False, True, False], dtype=bool)

Note: Linear algebra

Tip: Array-wise comparisons:

3.2. Numerical operations on arrays 60 3.2. Numerical operations on arrays 61

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

The sub-module numpy.linalg implements basic linear algebra, such as solving linear systems, singular
>>> x[0, 1, :].sum()
value decomposition, etc. However, it is not guaranteed to be compiled using efficient routines, and thus 1.14764...
we recommend the use of scipy.linalg, as detailed in section Linear algebra operations: scipy.linalg

Exercise other operations Other reductions

• Look at the help for np.allclose. When might this be useful? — works the same way (and take axis=)
• Look at the help for np.triu and np.tril. Extrema:

>>> x = np.array([1, 3, 2])

>>> x.min()
3.2.2 Basic reductions 1
>>> x.max()
3
Computing sums
>>> x.argmin() # index of minimum
>>> x = np.array([1, 2, 3, 4]) 0
>>> np.sum(x) >>> x.argmax() # index of maximum
10 1
>>> x.sum()
10 Logical operations:

>>> np.all([True, True, False])

False
>>> np.any([True, True, False])
True

Note: Can be used for array comparisons:

>>> a = np.zeros((100, 100))

>>> np.any(a != 0)
False
>>> np.all(a == a)
True

>>> a = np.array([1, 2, 3, 2])

>>> b = np.array([2, 2, 3, 2])
>>> c = np.array([6, 4, 4, 5])
Sum by rows and by columns: >>> ((a <= b) & (b <= c)).all()
True
>>> x = np.array([[1, 1], [2, 2]])
>>> x
array([[1, 1], Statistics:
[2, 2]])
>>> x.sum(axis=0) # columns (first dimension) >>> x = np.array([1, 2, 3, 1])
array([3, 3]) >>> y = np.array([[1, 2, 3], [5, 6, 1]])
>>> x[:, 0].sum(), x[:, 1].sum() >>> x.mean()
(3, 3) 1.75
>>> x.sum(axis=1) # rows (second dimension) >>> np.median(x)
array([2, 4]) 1.5
>>> x[0, :].sum(), x[1, :].sum() >>> np.median(y, axis=-1) # last axis
(2, 4) array([ 2., 5.])

>>> x.std() # full population standard dev.

Tip: Same idea in higher dimensions: 0.82915619758884995

>>> x = np.random.rand(2, 2, 2) . . . and many more (best to learn as you go).

>>> x.sum(axis=2)[0, 1]
1.14764...

3.2. Numerical operations on arrays 62 3.2. Numerical operations on arrays 63

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Exercise: Reductions
Tip: Let us consider a simple 1D random walk process: at each time step a walker jumps right or left with
equal probability.
• Given there is a sum, what other function might you expect to see?
• What is the difference between sum and cumsum?
We are interested in finding the typical distance from the origin of a random walker after t left or right
jumps? We are going to simulate many “walkers” to find this law, and we are going to do so using array com-
puting tricks: we are going to create a 2D array with the “stories” (each walker has a story) in one direction,
Worked Example: data statistics and the time in the other:

Data in populations.txt describes the populations of hares and lynxes (and carrots) in northern Canada
during 20 years.
You can view the data in an editor, or alternatively in IPython (both shell and notebook):
In [1]: !cat data/populations.txt

First, load the data into a NumPy array:

>>> data = np.loadtxt('data/populations.txt')
>>> year, hares, lynxes, carrots = data.T # trick: columns to variables

Then plot it:

>>> from matplotlib import pyplot as plt
>>> plt.axes([0.2, 0.1, 0.5, 0.8])
>>> plt.plot(year, hares, year, lynxes, year, carrots)
>>> plt.legend(('Hare', 'Lynx', 'Carrot'), loc=(1.05, 0.5))

>>> n_stories = 1000 # number of walkers

>>> t_max = 200 # time during which we follow the walker

We randomly choose all the steps 1 or -1 of the walk:

>>> t = np.arange(t_max)
>>> steps = 2 * np.random.randint(0, 1 + 1, (n_stories, t_max)) - 1 # +1 because the high␣
,→value is exclusive
>>> np.unique(steps) # Verification: all steps are 1 or -1
array([-1, 1])

We build the walks by summing steps along the time:

The mean populations over time:
>>> positions = np.cumsum(steps, axis=1) # axis = 1: dimension of time
>>> populations = data[:, 1:]
>>> sq_distance = positions**2
>>> populations.mean(axis=0)
array([ 34080.95238095, 20166.66666667, 42400. ])

The sample standard deviations: We get the mean in the axis of the stories:

>>> populations.std(axis=0) >>> mean_sq_distance = np.mean(sq_distance, axis=0)

array([ 20897.90645809, 16254.59153691, 3322.50622558])

Which species has the highest population each year?: Plot the results:
>>> np.argmax(populations, axis=1)
array([2, 2, 0, 0, 1, 1, 2, 2, 2, 2, 2, 2, 0, 0, 0, 1, 2, 2, 2, 2, 2]) >>> plt.figure(figsize=(4, 3))
<matplotlib.figure.Figure object at ...>
>>> plt.plot(t, np.sqrt(mean_sq_distance), 'g.', t, np.sqrt(t), 'y-')
[<matplotlib.lines.Line2D object at ...>, <matplotlib.lines.Line2D object at ...>]
>>> plt.xlabel(r"$t$")
Worked Example: diffusion using a random walk algorithm
<matplotlib.text.Text object at ...>
>>> plt.ylabel(r"$\sqrt{\langle (\delta x)^2 \rangle}$")
<matplotlib.text.Text object at ...>
>>> plt.tight_layout() # provide sufficient space for labels

3.2. Numerical operations on arrays 64 3.2. Numerical operations on arrays 65

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

>>> a = np.tile(np.arange(0, 40, 10), (3, 1)).T

>>> a
array([[ 0, 0, 0],
[10, 10, 10],
[20, 20, 20],
[30, 30, 30]])
>>> b = np.array([0, 1, 2])
>>> a + b
array([[ 0, 1, 2],
[10, 11, 12],
[20, 21, 22],
[30, 31, 32]])

We have already used broadcasting without knowing it!:

We find a well-known result in physics: the RMS dis-
tance grows as the square root of the time! >>> a = np.ones((4, 5))
>>> a[0] = 2 # we assign an array of dimension 0 to an array of dimension 1
>>> a
array([[ 2., 2., 2., 2., 2.],
[ 1., 1., 1., 1., 1.],
3.2.3 Broadcasting
[ 1., 1., 1., 1., 1.],
[ 1., 1., 1., 1., 1.]])
• Basic operations on numpy arrays (addition, etc.) are elementwise
• This works on arrays of the same size. An useful trick:

Nevertheless, It’s also possible to do operations on arrays of different >>> a = np.arange(0, 40, 10)
sizes if NumPy can transform these arrays so that they all have >>> a.shape
(4,)
the same size: this conversion is called broadcasting.
>>> a = a[:, np.newaxis] # adds a new axis -> 2D array
The image below gives an example of broadcasting: >>> a.shape
(4, 1)
>>> a
array([[ 0],
[10],
[20],
[30]])
>>> a + b
array([[ 0, 1, 2],
[10, 11, 12],
[20, 21, 22],
[30, 31, 32]])

Tip: Broadcasting seems a bit magical, but it is actually quite natural to use it when we want to solve a problem
whose output data is an array with more dimensions than input data.

Worked Example: Broadcasting

Let’s construct an array of distances (in miles) between cities of Route 66: Chicago, Springfield, Saint-Louis,
Tulsa, Oklahoma City, Amarillo, Santa Fe, Albuquerque, Flagstaff and Los Angeles.
>>> mileposts = np.array([0, 198, 303, 736, 871, 1175, 1475, 1544,
... 1913, 2448])
>>> distance_array = np.abs(mileposts - mileposts[:, np.newaxis])
>>> distance_array
array([[ 0, 198, 303, 736, 871, 1175, 1475, 1544, 1913, 2448],
[ 198, 0, 105, 538, 673, 977, 1277, 1346, 1715, 2250],
[ 303, 105, 0, 433, 568, 872, 1172, 1241, 1610, 2145],
[ 736, 538, 433, 0, 135, 439, 739, 808, 1177, 1712],
[ 871, 673, 568, 135, 0, 304, 604, 673, 1042, 1577],
Let’s verify:
[1175, 977, 872, 439, 304, 0, 300, 369, 738, 1273],
[1475, 1277, 1172, 739, 604, 300, 0, 69, 438, 973],
[1544, 1346, 1241, 808, 673, 369, 69, 0, 369, 904],
3.2. Numerical operations on arrays 66 3.2. Numerical operations
[1913, 1715, 1610, on arrays
1177, 1042, 738, 438, 369, 0, 535], 67
[2448, 2250, 2145, 1712, 1577, 1273, 973, 904, 535, 0]])
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

>>> x.shape, y.shape

((5, 1), (1, 5))
>>> distance = np.sqrt(x ** 2 + y ** 2)

Tip: So, np.ogrid is very useful as soon as we have to handle computations on a grid. On the other hand,
np.mgrid directly provides matrices full of indices for cases where we can’t (or don’t want to) benefit from
broadcasting:

>>> x, y = np.mgrid[0:4, 0:4]

>>> x
array([[0, 0, 0, 0],
[1, 1, 1, 1],
[2, 2, 2, 2],
[3, 3, 3, 3]])
>>> y
array([[0, 1, 2, 3],
[0, 1, 2, 3],
[0, 1, 2, 3],
[0, 1, 2, 3]])
A lot of grid-based or network-based problems can also use broadcasting. For instance, if we want to compute
the distance from the origin of points on a 10x10 grid, we can do
See also:
>>> x, y = np.arange(5), np.arange(5)[:, np.newaxis]
>>> distance = np.sqrt(x ** 2 + y ** 2) Broadcasting: discussion of broadcasting in the Advanced NumPy chapter.
>>> distance
array([[ 0. , 1. , 2. , 3. , 4. ],
[ 1. , 1.41421356, 2.23606798, 3.16227766, 4.12310563], 3.2.4 Array shape manipulation
[ 2. , 2.23606798, 2.82842712, 3.60555128, 4.47213595],
[ 3. , 3.16227766, 3.60555128, 4.24264069, 5. ], Flattening
[ 4. , 4.12310563, 4.47213595, 5. , 5.65685425]])
>>> a = np.array([[1, 2, 3], [4, 5, 6]])
Or in color: >>> a.ravel()
array([1, 2, 3, 4, 5, 6])
>>> plt.pcolor(distance)
>>> a.T
>>> plt.colorbar()
array([[1, 4],
[2, 5],
[3, 6]])
>>> a.T.ravel()
array([1, 4, 2, 5, 3, 6])

Higher dimensions: last dimensions ravel out “first”.

Reshaping

The inverse operation to flattening:

>>> a.shape
(2, 3)
>>> b = a.ravel()
Remark : the numpy.ogrid() function allows to di- >>> b = b.reshape((2, 3))
rectly create vectors x and y of the previous example, with two “significant dimensions”: >>> b
array([[1, 2, 3],
>>> x, y = np.ogrid[0:5, 0:5] [4, 5, 6]])
>>> x, y
(array([[0], Or,
[1],
[2], >>> a.reshape((2, -1)) # unspecified (-1) value is inferred
[3], array([[1, 2, 3],
[4]]), array([[0, 1, 2, 3, 4]])) [4, 5, 6]])

3.2. Numerical operations on arrays 68 3.2. Numerical operations on arrays 69

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Resizing
Warning: ndarray.reshape may return a view (cf help(np.reshape))), or copy
Size of an array can be changed with ndarray.resize:

>>> a = np.arange(4)
Tip: >>> a.resize((8,))
>>> a
>>> b[0, 0] = 99
array([0, 1, 2, 3, 0, 0, 0, 0])
>>> a
array([[99, 2, 3],
[ 4, 5, 6]]) However, it must not be referred to somewhere else:

>>> b = a
Beware: reshape may also return a copy!: >>> a.resize((4,))
Traceback (most recent call last):
>>> a = np.zeros((3, 2))
File "<stdin>", line 1, in <module>
>>> b = a.T.reshape(3*2)
ValueError: cannot resize an array that has been referenced or is
>>> b[0] = 9
referencing another array in this way. Use the resize function
>>> a
array([[ 0., 0.],
[ 0., 0.],
[ 0., 0.]]) Exercise: Shape manipulations

To understand this you need to learn more about the memory layout of a numpy array. • Look at the docstring for reshape, especially the notes section which has some more information
about copies and views.
• Use flatten as an alternative to ravel. What is the difference? (Hint: check which one returns a view
and which a copy)
Adding a dimension
• Experiment with transpose for dimension shuffling.
Indexing with the np.newaxis object allows us to add an axis to an array (you have seen this already above in
the broadcasting section):

>>> z = np.array([1, 2, 3]) 3.2.5 Sorting data

>>> z
array([1, 2, 3])
Sorting along an axis:
>>> z[:, np.newaxis] >>> a = np.array([[4, 3, 5], [1, 2, 1]])
array([[1], >>> b = np.sort(a, axis=1)
[2], >>> b
[3]]) array([[3, 4, 5],
[1, 1, 2]])
>>> z[np.newaxis, :]
array([[1, 2, 3]])
Note: Sorts each row separately!

Dimension shuffling
In-place sort:
>>> a = np.arange(4*3*2).reshape(4, 3, 2) >>> a.sort(axis=1)
>>> a.shape >>> a
(4, 3, 2) array([[3, 4, 5],
>>> a[0, 2, 1] [1, 1, 2]])
5
>>> b = a.transpose(1, 2, 0)
Sorting with fancy indexing:
>>> b.shape
(3, 2, 4) >>> a = np.array([4, 3, 1, 2])
>>> b[2, 1, 0] >>> j = np.argsort(a)
5 >>> j
array([2, 3, 1, 0])
Also creates a view: >>> a[j]
array([1, 2, 3, 4])
>>> b[2, 1, 0] = -1
>>> a[0, 2, 1]
Finding minima and maxima:
-1

3.2. Numerical operations on arrays 70 3.2. Numerical operations on arrays 71

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

>>> a = np.array([4, 3, 1, 2])

3.3.1 More data types
>>> j_max = np.argmax(a)
>>> j_min = np.argmin(a) Casting
>>> j_max, j_min
(0, 2) “Bigger” type wins in mixed-type operations:

>>> np.array([1, 2, 3]) + 1.5

Exercise: Sorting array([ 2.5, 3.5, 4.5])

• Try both in-place and out-of-place sorting. Assignment never changes the type!

• Try creating arrays with different dtypes and sorting them. >>> a = np.array([1, 2, 3])
>>> a.dtype
• Use all or array_equal to check the results. dtype('int64')
• Look at np.random.shuffle for a way to create sortable input quicker. >>> a[0] = 1.9 # <-- float is truncated to integer
>>> a
• Combine ravel, sort and reshape. array([1, 2, 3])

• Look at the axis keyword for sort and rewrite the previous exercise.
Forced casts:

>>> a = np.array([1.7, 1.2, 1.6])

>>> b = a.astype(int) # <-- truncates to integer
3.2.6 Summary >>> b
array([1, 1, 1])
What do you need to know to get started?
• Know how to create arrays : array, arange, ones, zeros. Rounding:

• Know the shape of the array with array.shape, then use slicing to obtain different views of the array: >>> a = np.array([1.2, 1.5, 1.6, 2.5, 3.5, 4.5])
array[::2], etc. Adjust the shape of the array using reshape or flatten it with ravel. >>> b = np.around(a)
>>> b # still floating-point
• Obtain a subset of the elements of an array and/or modify their values with masks array([ 1., 2., 2., 2., 4., 4.])
>>> c = np.around(a).astype(int)
>>> a[a < 0] = 0
>>> c
array([1, 2, 2, 2, 4, 4])
• Know miscellaneous operations on arrays, such as finding the mean or max (array.max(), array.
mean()). No need to retain everything, but have the reflex to search in the documentation (online docs,
help(), lookfor())!!
Different data type sizes
• For advanced use: master the indexing with arrays of integers, as well as broadcasting. Know more
NumPy functions to handle various array operations. Integers (signed):

Quick read int8 8 bits

int16 16 bits
If you want to do a first quick pass through the Scipy lectures to learn the ecosystem, you can directly skip int32 32 bits (same as int on 32-bit platform)
to the next chapter: Matplotlib: plotting. int64 64 bits (same as int on 64-bit platform)

The remainder of this chapter is not necessary to follow the rest of the intro part. But be sure to come back
and finish this chapter, as well as to do some more exercices. >>> np.array([1], dtype=int).dtype
dtype('int64')
>>> np.iinfo(np.int32).max, 2**31 - 1
(2147483647, 2147483647)

3.3 More elaborate arrays Unsigned integers:

uint8 8 bits
Section contents uint16 16 bits
uint32 32 bits
• More data types
uint64 64 bits
• Structured data types
• maskedarray: dealing with (propagation of) missing data

3.3. More elaborate arrays 72 3.3. More elaborate arrays 73

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

>>> np.iinfo(np.uint32).max, 2**32 - 1

3.3.2 Structured data types
(4294967295, 4294967295)
sensor_code (4-character string)
position (float)
Long integers
value (float)

Python 2 has a specific type for ‘long’ integers, that cannot overflow, represented with an ‘L’ at the end. In
Python 3, all integers are long, and thus cannot overflow. >>> samples = np.zeros((6,), dtype=[('sensor_code', 'S4'),
... ('position', float), ('value', float)])
>>> np.iinfo(np.int64).max, 2**63 - 1 >>> samples.ndim
(9223372036854775807, 9223372036854775807L) 1
>>> samples.shape
(6,)
Floating-point numbers: >>> samples.dtype.names
('sensor_code', 'position', 'value')
float16 16 bits
float32 32 bits >>> samples[:] = [('ALFA', 1, 0.37), ('BETA', 1, 0.11), ('TAU', 1, 0.13),
... ('ALFA', 1.5, 0.37), ('ALFA', 3, 0.11), ('TAU', 1.2, 0.13)]
float64 64 bits (same as float)
>>> samples
float96 96 bits, platform-dependent (same as np.longdouble)
array([('ALFA', 1.0, 0.37), ('BETA', 1.0, 0.11), ('TAU', 1.0, 0.13),
float128 128 bits, platform-dependent (same as np.longdouble) ('ALFA', 1.5, 0.37), ('ALFA', 3.0, 0.11), ('TAU', 1.2, 0.13)],
dtype=[('sensor_code', 'S4'), ('position', '<f8'), ('value', '<f8')])
>>> np.finfo(np.float32).eps
1.1920929e-07 Field access works by indexing with field names:
>>> np.finfo(np.float64).eps
2.2204460492503131e-16 >>> samples['sensor_code']
array(['ALFA', 'BETA', 'TAU', 'ALFA', 'ALFA', 'TAU'],
>>> np.float32(1e-8) + np.float32(1) == 1 dtype='|S4')
True >>> samples['value']
>>> np.float64(1e-8) + np.float64(1) == 1 array([ 0.37, 0.11, 0.13, 0.37, 0.11, 0.13])
False >>> samples[0]
('ALFA', 1.0, 0.37)
Complex floating-point numbers:
>>> samples[0]['sensor_code'] = 'TAU'
>>> samples[0]
complex64 two 32-bit floats ('TAU', 1.0, 0.37)
complex128 two 64-bit floats
complex192 two 96-bit floats, platform-dependent Multiple fields at once:
complex256 two 128-bit floats, platform-dependent
>>> samples[['position', 'value']]
array([(1.0, 0.37), (1.0, 0.11), (1.0, 0.13), (1.5, 0.37), (3.0, 0.11),
Smaller data types (1.2, 0.13)],
dtype=[('position', '<f8'), ('value', '<f8')])
If you don’t know you need special data types, then you probably don’t.
Fancy indexing works, as usual:
Comparison on using float32 instead of float64:
>>> samples[samples['sensor_code'] == 'ALFA']
• Half the size in memory and on disk
array([('ALFA', 1.5, 0.37), ('ALFA', 3.0, 0.11)],
• Half the memory bandwidth required (may be a bit faster in some operations) dtype=[('sensor_code', 'S4'), ('position', '<f8'), ('value', '<f8')])

In [1]: a = np.zeros((1e6,), dtype=np.float64)

In [2]: b = np.zeros((1e6,), dtype=np.float32) Note: There are a bunch of other syntaxes for constructing structured arrays, see here and here.

In [3]: %timeit a*a

1000 loops, best of 3: 1.78 ms per loop
3.3.3 maskedarray: dealing with (propagation of) missing data
In [4]: %timeit b*b
1000 loops, best of 3: 1.07 ms per loop • For floats one could use NaN’s, but masks work for all types:

• But: bigger rounding errors — sometimes in surprising places (i.e., don’t use them unless you really >>> x = np.ma.array([1, 2, 3, 4], mask=[0, 1, 0, 1])
need them) >>> x

3.3. More elaborate arrays 74 3.3. More elaborate arrays 75

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

masked_array(data = [1 -- 3 --], array([-1. , 0.33333333])

mask = [False True False True], >>> p.order
fill_value = 999999) 2

>>> x = np.linspace(0, 1, 20)

>>> y = np.ma.array([1, 2, 3, 4], mask=[0, 1, 1, 1])
>>> y = np.cos(x) + 0.3*np.random.rand(20)
>>> x + y
>>> p = np.poly1d(np.polyfit(x, y, 3))
masked_array(data = [2 -- -- --],
mask = [False True True True],
>>> t = np.linspace(0, 1, 200)
fill_value = 999999)
>>> plt.plot(x, y, 'o', t, p(t), '-')
[<matplotlib.lines.Line2D object at ...>, <matplotlib.lines.Line2D object at ...>]
• Masking versions of common functions:

>>> np.ma.sqrt([1, -1, 2, -2])

masked_array(data = [1.0 -- 1.41421356237... --],
mask = [False True False True],
fill_value = 1e+20)

Note: There are other useful array siblings

While it is off topic in a chapter on numpy, let’s take a moment to recall good coding practice, which really do
pay off in the long run:

See http://docs.scipy.org/doc/numpy/reference/
Good practices
routines.polynomials.poly1d.html for more.
• Explicit variable names (no need of a comment to explain what is in the variable)
• Style: spaces after commas, around =, etc. More polynomials (with more bases)

A certain number of rules for writing “beautiful” code (and, more importantly, using the same conven- NumPy also has a more sophisticated polynomial interface, which supports e.g. the Chebyshev basis.
tions as everybody else!) are given in the Style Guide for Python Code and the Docstring Conventions
page (to manage help strings). 3x 2 + 2x − 1:

• Except some rare cases, variable names and comments in English. >>> p = np.polynomial.Polynomial([-1, 2, 3]) # coefs in different order!
>>> p(0)
-1.0
>>> p.roots()

3.4 Advanced operations array([-1. , 0.33333333])

>>> p.degree() # In general polynomials do not always expose 'order'
2

Section contents Example using polynomials in Chebyshev basis, for polynomials in range [-1, 1]:

• Polynomials >>> x = np.linspace(-1, 1, 2000)

>>> y = np.cos(x) + 0.3*np.random.rand(2000)
• Loading data files >>> p = np.polynomial.Chebyshev.fit(x, y, 90)

>>> t = np.linspace(-1, 1, 200)

>>> plt.plot(x, y, 'r.')
3.4.1 Polynomials [<matplotlib.lines.Line2D object at ...>]
>>> plt.plot(t, p(t), 'k-', lw=3)
NumPy also contains polynomials in different bases: [<matplotlib.lines.Line2D object at ...>]

For example, 3x 2 + 2x − 1:

>>> p = np.poly1d([3, 2, -1])

>>> p(0)
-1
>>> p.roots

3.4. Advanced operations 76 3.4. Advanced operations 77

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Images

Using Matplotlib:

>>> img = plt.imread('data/elephant.png')

>>> img.shape, img.dtype
((200, 300, 3), dtype('float32'))
>>> plt.imshow(img)
<matplotlib.image.AxesImage object at ...>
>>> plt.savefig('plot.png')

>>> plt.imsave('red_elephant.png', img[:,:,0], cmap=plt.cm.gray)

The Chebyshev polynomials have some advantages in

interpolation.

3.4.2 Loading data files

Text files

Example: populations.txt:
# year hare lynx carrot
1900 30e3 4e3 48300
1901 47.2e3 6.1e3 48200
This saved only one channel (of RGB):
1902 70.2e3 9.8e3 41500
1903 77.4e3 35.2e3 38200 >>> plt.imshow(plt.imread('red_elephant.png'))
<matplotlib.image.AxesImage object at ...>

>>> data = np.loadtxt('data/populations.txt')

>>> data
array([[ 1900., 30000., 4000., 48300.],
[ 1901., 47200., 6100., 48200.],
[ 1902., 70200., 9800., 41500.],
...

>>> np.savetxt('pop2.txt', data)

>>> data2 = np.loadtxt('pop2.txt')

Note: If you have a complicated text file, what you can try are:
• np.genfromtxt
• Using Python’s I/O functions and e.g. regexps for parsing (Python is quite well suited for this) Other libraries:

>>> from scipy.misc import imsave

>>> imsave('tiny_elephant.png', img[::6,::6])
Reminder: Navigating the filesystem with IPython >>> plt.imshow(plt.imread('tiny_elephant.png'), interpolation='nearest')
<matplotlib.image.AxesImage object at ...>
In [1]: pwd # show current directory
'/home/user/stuff/2011-numpy-tutorial'
In [2]: cd ex
'/home/user/stuff/2011-numpy-tutorial/ex'
In [3]: ls
populations.txt species.txt

3.4. Advanced operations 78 3.4. Advanced operations 79

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

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

and generate a new array containing its 2nd and 4th rows.
2. Divide each column of the array:

>>> import numpy as np

>>> a = np.arange(25).reshape(5, 5)

elementwise with the array b = np.array([1., 5, 10, 15, 20]). (Hint: np.newaxis).
3. Harder one: Generate a 10 x 3 array of random numbers (in range [0,1]). For each row, pick the number
closest to 0.5.
NumPy’s own format • Use abs and argsort to find the column j closest for each row.

NumPy has its own binary format, not portable but with efficient I/O: • Use fancy indexing to extract the numbers. (Hint: a[i,j] – the array i must contain the row num-
bers corresponding to stuff in j.)
>>> data = np.ones((3, 3))
>>> np.save('pop.npy', data)
>>> data3 = np.load('pop.npy') 3.5.2 Picture manipulation: Framing a Face

Let’s do some manipulations on numpy arrays by starting with an image of a racoon. scipy provides a 2D
Well-known (& more obscure) file formats array of this image with the scipy.misc.face function:

>>> from scipy import misc

• HDF5: h5py, PyTables >>> face = misc.face(gray=True) # 2D grayscale image
• NetCDF: scipy.io.netcdf_file, netcdf4-python, . . .
Here are a few images we will be able to obtain with our manipulations: use different colormaps, crop the
• Matlab: scipy.io.loadmat, scipy.io.savemat
image, change some parts of the image.
• MatrixMarket: scipy.io.mmread, scipy.io.mmwrite
• IDL: scipy.io.readsav
. . . if somebody uses it, there’s probably also a Python library for it.

Exercise: Text data files

Write a Python script that loads data from populations.txt:: and drop the last column and the first 5 rows.
Save the smaller dataset to pop2.txt. • Let’s use the imshow function of pylab to display the image.

>>> import pylab as plt

>>> face = misc.face(gray=True)
NumPy internals >>> plt.imshow(face)
<matplotlib.image.AxesImage object at 0x...>
If you are interested in the NumPy internals, there is a good discussion in Advanced NumPy.
• The face is displayed in false colors. A colormap must be specified for it to be displayed in grey.

>>> plt.imshow(face, cmap=plt.cm.gray)

<matplotlib.image.AxesImage object at 0x...>
3.5 Some exercises
• Create an array of the image with a narrower centering [for example,] remove 100 pixels from all the
3.5.1 Array manipulations borders of the image. To check the result, display this new array with imshow.

>>> crop_face = face[100:-100, 100:-100]

1. Form the 2-D array (without typing it in explicitly):
• We will now frame the face with a black locket. For this, we need to create a mask corresponding to
the pixels we want to be black. The center of the face is around (660, 330), so we defined the mask
by this condition (y-300)**2 + (x-660)**2

3.5. Some exercises 80 3.5. Some exercises 81

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

6. Compare (plot) the change in hare population (see help(np.gradient)) and the number of lynxes.
>>> sy, sx = face.shape
>>> y, x = np.ogrid[0:sy, 0:sx] # x and y indices of pixels Check correlation (see help(np.corrcoef)).
>>> y.shape, x.shape . . . all without for-loops.
((768, 1), (1, 1024))
>>> centerx, centery = (660, 300) # center of the image Solution: Python source file
>>> mask = ((y - centery)**2 + (x - centerx)**2) > 230**2 # circle

then we assign the value 0 to the pixels of the image corresponding to the mask. The syntax is 3.5.4 Crude integral approximations
extremely simple and intuitive:
Write a function f(a, b, c) that returns a b − c. Form a 24x12x6 array containing its values in parameter
>>> face[mask] = 0 ranges [0,1] x [0,1] x [0,1].
>>> plt.imshow(face)
<matplotlib.image.AxesImage object at 0x...> Approximate the 3-d integral
Z 1Z 1Z 1
• Follow-up: copy all instructions of this exercise in a script called face_locket.py then execute this (a b − c)d a d b d c
script in IPython with %run face_locket.py. 0 0 0

Change the circle to an ellipsoid. over this volume with the mean. The exact result is: ln 2 − 12 ≈ 0.1931 . . . — what is your relative error?
(Hints: use elementwise operations and broadcasting. You can make np.ogrid give a number of points in
3.5.3 Data statistics given range with np.ogrid[0:1:20j].)
Reminder Python functions:
The data in populations.txt describes the populations of hares and lynxes (and carrots) in northern Canada
during 20 years: def f(a, b, c):
return some_result
>>> data = np.loadtxt('data/populations.txt')
>>> year, hares, lynxes, carrots = data.T # trick: columns to variables Solution: Python source file

>>> import matplotlib.pyplot as plt

>>> plt.axes([0.2, 0.1, 0.5, 0.8]) 3.5.5 Mandelbrot set
<matplotlib.axes...Axes object at ...>
>>> plt.plot(year, hares, year, lynxes, year, carrots)
[<matplotlib.lines.Line2D object at ...>, ...]
>>> plt.legend(('Hare', 'Lynx', 'Carrot'), loc=(1.05, 0.5))
<matplotlib.legend.Legend object at ...>

Write a script that computes the Mandelbrot fractal.

The Mandelbrot iteration:

N_max = 50
Computes and print, based on the data in some_threshold = 50
populations.txt. . .
c = x + 1j*y
1. The mean and std of the populations of each species for the years in the period.
z = 0
2. Which year each species had the largest population. for j in range(N_max):
3. Which species has the largest population for each year. (Hint: argsort & fancy indexing of np. z = z**2 + c
array(['H', 'L', 'C']))
Point (x, y) belongs to the Mandelbrot set if |z| < some_threshold.
4. Which years any of the populations is above 50000. (Hint: comparisons and np.any)
Do this computation by:
5. The top 2 years for each species when they had the lowest populations. (Hint: argsort, fancy indexing)
1. Construct a grid of c = x + 1j*y values in range [-2, 1] x [-1.5, 1.5]

3.5. Some exercises 82 3.5. Some exercises 83

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

2. Do the iteration 3.6 Full code examples

3. Form the 2-d boolean mask indicating which points are in the set
4. Save the result to an image with: 3.6.1 Full code examples for the numpy chapter
>>> import matplotlib.pyplot as plt
2D plotting
>>> plt.imshow(mask.T, extent=[-2, 1, -1.5, 1.5])
<matplotlib.image.AxesImage object at ...>
>>> plt.gray() Plot a basic 2D figure
>>> plt.savefig('mandelbrot.png')

Solution: Python source file

3.5.6 Markov chain

Markov chain transition matrix P, and probability distribution on the states p:

1. 0 <= P[i,j] <= 1: probability to go from state i to state j
2. Transition rule: p new = P T p ol d
3. all(sum(P, axis=1) == 1), p.sum() == 1: normalization import numpy as np
Write a script that works with 5 states, and: import matplotlib.pyplot as plt

• Constructs a random matrix, and normalizes each row so that it is a transition matrix. image = np.random.rand(30, 30)
plt.imshow(image, cmap=plt.cm.hot)
• Starts from a random (normalized) probability distribution p and takes 50 steps => p_50
plt.colorbar()
• Computes the stationary distribution: the eigenvector of P.T with eigenvalue 1 (numerically: closest to plt.show()
1) => p_stationary
Total running time of the script: ( 0 minutes 0.125 seconds)
Remember to normalize the eigenvector — I didn’t. . .
Download Python source code: plot_basic2dplot.py
• Checks if p_50 and p_stationary are equal to tolerance 1e-5
Download Jupyter notebook: plot_basic2dplot.ipynb
Toolbox: np.random.rand, .dot(), np.linalg.eig, reductions, abs(), argmin, comparisons, all, np.
linalg.norm, etc. Generated by Sphinx-Gallery

Solution: Python source file

1D plotting

Plot a basic 1D figure

3.5. Some exercises 84 3.6. Full code examples 85

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

import numpy as np import numpy as np

import matplotlib.pyplot as plt import matplotlib.pyplot as plt

x = np.linspace(0, 3, 20) x, y = np.arange(5), np.arange(5)[:, np.newaxis]

y = np.linspace(0, 9, 20) distance = np.sqrt(x ** 2 + y ** 2)
plt.plot(x, y) plt.pcolor(distance)
plt.plot(x, y, 'o') plt.colorbar()
plt.show() plt.show()

Total running time of the script: ( 0 minutes 0.034 seconds) Total running time of the script: ( 0 minutes 0.082 seconds)
Download Python source code: plot_basic1dplot.py Download Python source code: plot_distances.py
Download Jupyter notebook: plot_basic1dplot.ipynb Download Jupyter notebook: plot_distances.ipynb
Generated by Sphinx-Gallery Generated by Sphinx-Gallery

Distances exercise Fitting to polynomial

Plot distances in a grid Plot noisy data and their polynomial fit

3.6. Full code examples 86 3.6. Full code examples 87

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

import numpy as np import numpy as np

import matplotlib.pyplot as plt import matplotlib.pyplot as plt

np.random.seed(12) np.random.seed(0)

x = np.linspace(0, 1, 20) x = np.linspace(-1, 1, 2000)

y = np.cos(x) + 0.3*np.random.rand(20) y = np.cos(x) + 0.3*np.random.rand(2000)
p = np.poly1d(np.polyfit(x, y, 3)) p = np.polynomial.Chebyshev.fit(x, y, 90)

t = np.linspace(0, 1, 200) t = np.linspace(-1, 1, 200)

plt.plot(x, y, 'o', t, p(t), '-') plt.plot(x, y, 'r.')
plt.show() plt.plot(t, p(t), 'k-', lw=3)
plt.show()
Total running time of the script: ( 0 minutes 0.034 seconds)
Total running time of the script: ( 0 minutes 0.042 seconds)
Download Python source code: plot_polyfit.py
Download Python source code: plot_chebyfit.py
Download Jupyter notebook: plot_polyfit.ipynb
Download Jupyter notebook: plot_chebyfit.ipynb
Generated by Sphinx-Gallery
Generated by Sphinx-Gallery

Fitting in Chebyshev basis

Population exercise
Plot noisy data and their polynomial fit in a Chebyshev basis
Plot populations of hares, lynxes, and carrots

3.6. Full code examples 88 3.6. Full code examples 89

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

original figure

plt.figure()
img = plt.imread('../data/elephant.png')
plt.imshow(img)

import numpy as np
import matplotlib.pyplot as plt

data = np.loadtxt('../data/populations.txt')
year, hares, lynxes, carrots = data.T

plt.axes([0.2, 0.1, 0.5, 0.8])

plt.plot(year, hares, year, lynxes, year, carrots)
plt.legend(('Hare', 'Lynx', 'Carrot'), loc=(1.05, 0.5)) red channel displayed in grey
plt.show()

plt.figure()
Total running time of the script: ( 0 minutes 0.037 seconds)
img_red = img[:, :, 0]
Download Python source code: plot_populations.py plt.imshow(img_red, cmap=plt.cm.gray)

Download Jupyter notebook: plot_populations.ipynb

Generated by Sphinx-Gallery

Reading and writing an elephant

Read and write images

import numpy as np
import matplotlib.pyplot as plt

3.6. Full code examples 90 3.6. Full code examples 91

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Total running time of the script: ( 0 minutes 0.101 seconds)

lower resolution Download Python source code: plot_elephant.py
Download Jupyter notebook: plot_elephant.ipynb
plt.figure()
img_tiny = img[::6, ::6] Generated by Sphinx-Gallery
plt.imshow(img_tiny, interpolation='nearest')

plt.show() Mandelbrot set

Compute the Mandelbrot fractal and plot it

3.6. Full code examples 92 3.6. Full code examples 93

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Generated by Sphinx-Gallery

Random walk exercise

Plot distance as a function of time for a random walk together with the theoretical result

import numpy as np
import matplotlib.pyplot as plt

# We create 1000 realizations with 200 steps each

import numpy as np n_stories = 1000
import matplotlib.pyplot as plt t_max = 200
from numpy import newaxis
t = np.arange(t_max)
def compute_mandelbrot(N_max, some_threshold, nx, ny): # Steps can be -1 or 1 (note that randint excludes the upper limit)
# A grid of c-values steps = 2 * np.random.randint(0, 1 + 1, (n_stories, t_max)) - 1
x = np.linspace(-2, 1, nx)
y = np.linspace(-1.5, 1.5, ny) # The time evolution of the position is obtained by successively
# summing up individual steps. This is done for each of the
c = x[:,newaxis] + 1j*y[newaxis,:] # realizations, i.e. along axis 1.
positions = np.cumsum(steps, axis=1)
# Mandelbrot iteration
# Determine the time evolution of the mean square distance.
z = c sq_distance = positions**2
for j in range(N_max): mean_sq_distance = np.mean(sq_distance, axis=0)
z = z**2 + c
# Plot the distance d from the origin as a function of time and
mandelbrot_set = (abs(z) < some_threshold) # compare with the theoretically expected result where d(t)
# grows as a square root of time t.
return mandelbrot_set plt.figure(figsize=(4, 3))
plt.plot(t, np.sqrt(mean_sq_distance), 'g.', t, np.sqrt(t), 'y-')
mandelbrot_set = compute_mandelbrot(50, 50., 601, 401) plt.xlabel(r"$t$")
plt.ylabel(r"$\sqrt{\langle (\delta x)^2 \rangle}$")
plt.imshow(mandelbrot_set.T, extent=[-2, 1, -1.5, 1.5]) plt.tight_layout()
plt.gray() plt.show()
plt.show()
Total running time of the script: ( 0 minutes 0.131 seconds)
Total running time of the script: ( 0 minutes 0.105 seconds) Download Python source code: plot_randomwalk.py
Download Python source code: plot_mandelbrot.py Download Jupyter notebook: plot_randomwalk.ipynb
Download Jupyter notebook: plot_mandelbrot.ipynb Generated by Sphinx-Gallery

3.6. Full code examples 94 3.6. Full code examples 95

 Scipy lecture notes, Edition 2017.1

Download all examples in Python source code: auto_examples_python.zip

Download all examples in Jupyter notebooks: auto_examples_jupyter.zip
Generated by Sphinx-Gallery

CHAPTER 4
Matplotlib: plotting

Thanks

Many thanks to Bill Wing and Christoph Deil for review and corrections.

Authors: Nicolas Rougier, Mike Müller, Gaël Varoquaux

Chapter contents

• Introduction
• Simple plot
• Figures, Subplots, Axes and Ticks
• Other Types of Plots: examples and exercises
• Beyond this tutorial
• Quick references
• Full code examples

4.1 Introduction

Tip: Matplotlib is probably the most used Python package for 2D-graphics. It provides both a quick way to
visualize data from Python and publication-quality figures in many formats. We are going to explore matplotlib

3.6. Full code examples 96 97

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

in interactive mode covering most common cases.

Welcome to pylab, a matplotlib-based Python environment.
For more information, type 'help(pylab)'.

4.1.1 IPython, Jupyter, and matplotlib modes

Tip: You can also download each of the examples and run it using regular python, but you will lose interactive
data manipulation:
Tip: The Jupyter notebook and the IPython enhanced interactive Python, are tuned for the scientific-
computing workflow in Python, in combination with Matplotlib: $ python plot_exercise_1.py

For interactive matplotlib sessions, turn on the matplotlib mode You can get source for each step by clicking on the corresponding figure.

IPython console When using the IPython console, use:

In [1]: %matplotlib 4.2.1 Plotting with default settings

Jupyter notebook In the notebook, insert, at the beginning of the notebook the following magic:

%matplotlib inline

4.1.2 pyplot

Tip: pyplot provides a procedural interface to the matplotlib object-oriented plotting library. It is modeled
closely after Matlab™. Therefore, the majority of plotting commands in pyplot have Matlab™ analogs with
similar arguments. Important commands are explained with interactive examples.

from matplotlib import pyplot as plt

Hint: Documentation

4.2 Simple plot • plot tutorial

• plot() command

Tip: In this section, we want to draw the cosine and sine functions on the same plot. Starting from the default
settings, we’ll enrich the figure step by step to make it nicer.
Tip: Matplotlib comes with a set of default settings that allow customizing all kinds of properties. You can
First step is to get the data for the sine and cosine functions: control the defaults of almost every property in matplotlib: figure size and dpi, line width, color and style,
axes, axis and grid properties, text and font properties and so on.

import numpy as np
import numpy as np
X = np.linspace(-np.pi, np.pi, 256, endpoint=True) import matplotlib.pyplot as plt
C, S = np.cos(X), np.sin(X)
X = np.linspace(-np.pi, np.pi, 256, endpoint=True)
X is now a numpy array with 256 values ranging from -π to +π (included). C is the cosine (256 values) and S is C, S = np.cos(X), np.sin(X)
the sine (256 values).
plt.plot(X, C)
To run the example, you can type them in an IPython interactive session: plt.plot(X, S)

$ ipython --pylab plt.show()

This brings us to the IPython prompt:

IPython 0.13 -- An enhanced Interactive Python.

? -> Introduction to IPython's features.
%magic -> Information about IPython's 'magic' % functions.
help -> Python's own help system.
object? -> Details about 'object'. ?object also works, ?? prints more.

4.2. Simple plot 98 4.2. Simple plot 99

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

4.2.2 Instantiating defaults # Show result on screen

plt.show()

4.2.3 Changing colors and line widths

Hint: Documentation
• Customizing matplotlib
Hint: Documentation
In the script below, we’ve instantiated (and commented) all the figure settings that influence the appearance • Controlling line properties
of the plot.
• Line API

Tip: The settings have been explicitly set to their default values, but now you can interactively play with the
values to explore their affect (see Line properties and Line styles below).
Tip: First step, we want to have the cosine in blue and the sine in red and a slighty thicker line for both of
them. We’ll also slightly alter the figure size to make it more horizontal.
import numpy as np
import matplotlib.pyplot as plt
...
# Create a figure of size 8x6 inches, 80 dots per inch plt.figure(figsize=(10, 6), dpi=80)
plt.figure(figsize=(8, 6), dpi=80) plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-")
plt.plot(X, S, color="red", linewidth=2.5, linestyle="-")
# Create a new subplot from a grid of 1x1 ...
plt.subplot(1, 1, 1)

X = np.linspace(-np.pi, np.pi, 256, endpoint=True)

C, S = np.cos(X), np.sin(X)
4.2.4 Setting limits

# Plot cosine with a blue continuous line of width 1 (pixels)

plt.plot(X, C, color="blue", linewidth=1.0, linestyle="-")

# Plot sine with a green continuous line of width 1 (pixels)

plt.plot(X, S, color="green", linewidth=1.0, linestyle="-")

# Set x limits
plt.xlim(-4.0, 4.0)

# Set x ticks
plt.xticks(np.linspace(-4, 4, 9, endpoint=True))

# Set y limits
plt.ylim(-1.0, 1.0) Hint: Documentation
# Set y ticks • xlim() command
plt.yticks(np.linspace(-1, 1, 5, endpoint=True))
• ylim() command
# Save figure using 72 dots per inch
# plt.savefig("exercise_2.png", dpi=72)

4.2. Simple plot 100 4.2. Simple plot 101

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Tip: Current limits of the figure are a bit too tight and we want to make some space in order to clearly see all Hint: Documentation
data points.
• Working with text
• xticks() command
...
plt.xlim(X.min() * 1.1, X.max() * 1.1) • yticks() command
plt.ylim(C.min() * 1.1, C.max() * 1.1)
• set_xticklabels()
...
• set_yticklabels()

4.2.5 Setting ticks

Tip: Ticks are now properly placed but their label is not very explicit. We could guess that 3.142 is π but it
would be better to make it explicit. When we set tick values, we can also provide a corresponding label in the
second argument list. Note that we’ll use latex to allow for nice rendering of the label.

...
plt.xticks([-np.pi, -np.pi/2, 0, np.pi/2, np.pi],
[r'$-\pi$', r'$-\pi/2$', r'$0$', r'$+\pi/2$', r'$+\pi$'])

plt.yticks([-1, 0, +1],
[r'$-1$', r'$0$', r'$+1$'])
...

Hint: Documentation 4.2.7 Moving spines

• xticks() command
• yticks() command
• Tick container
• Tick locating and formatting

Tip: Current ticks are not ideal because they do not show the interesting values (+/-π,+/-π/2) for sine and
cosine. We’ll change them such that they show only these values.

...
plt.xticks([-np.pi, -np.pi/2, 0, np.pi/2, np.pi])
plt.yticks([-1, 0, +1]) Hint: Documentation
... • Spines
• Axis container

4.2.6 Setting tick labels • Transformations tutorial

Tip: Spines are the lines connecting the axis tick marks and noting the boundaries of the data area. They can
be placed at arbitrary positions and until now, they were on the border of the axis. We’ll change that since we
want to have them in the middle. Since there are four of them (top/bottom/left/right), we’ll discard the top
and right by setting their color to none and we’ll move the bottom and left ones to coordinate 0 in data space
coordinates.

...
ax = plt.gca() # gca stands for 'get current axis'
ax.spines['right'].set_color('none')

4.2. Simple plot 102 4.2. Simple plot 103

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

• Annotating axis
ax.spines['top'].set_color('none')
ax.xaxis.set_ticks_position('bottom') • annotate() command
ax.spines['bottom'].set_position(('data',0))
ax.yaxis.set_ticks_position('left')
ax.spines['left'].set_position(('data',0))
... Tip: Let’s annotate some interesting points using the annotate command. We chose the 2π/3 value and we
want to annotate both the sine and the cosine. We’ll first draw a marker on the curve as well as a straight dotted
line. Then, we’ll use the annotate command to display some text with an arrow.
4.2.8 Adding a legend
...

t = 2 * np.pi / 3
plt.plot([t, t], [0, np.cos(t)], color='blue', linewidth=2.5, linestyle="--")
plt.scatter([t, ], [np.cos(t), ], 50, color='blue')

plt.annotate(r'$cos(\frac{2\pi}{3} )=-\frac{1} {2} $',

xy=(t, np.cos(t)), xycoords='data',
xytext=(-90, -50), textcoords='offset points', fontsize=16,
arrowprops=dict(arrowstyle="->", connectionstyle="arc3,rad=.2"))

plt.plot([t, t],[0, np.sin(t)], color='red', linewidth=2.5, linestyle="--")

plt.scatter([t, ],[np.sin(t), ], 50, color='red')

plt.annotate(r'$sin(\frac{2\pi}{3} )=\frac{\sqrt{3} }{2} $',

Hint: Documentation
xy=(t, np.sin(t)), xycoords='data',
• Legend guide xytext=(+10, +30), textcoords='offset points', fontsize=16,
arrowprops=dict(arrowstyle="->", connectionstyle="arc3,rad=.2"))
• legend() command ...
• Legend API

4.2.10 Devil is in the details

Tip: Let’s add a legend in the upper left corner. This only requires adding the keyword argument label (that
will be used in the legend box) to the plot commands.

...
plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-", label="cosine")
plt.plot(X, S, color="red", linewidth=2.5, linestyle="-", label="sine")

plt.legend(loc='upper left')
...

4.2.9 Annotate some points

Hint: Documentation
• Artists
• BBox

Tip: The tick labels are now hardly visible because of the blue and red lines. We can make them bigger and we
can also adjust their properties such that they’ll be rendered on a semi-transparent white background. This
will allow us to see both the data and the labels.

...
for label in ax.get_xticklabels() + ax.get_yticklabels():
Hint: Documentation label.set_fontsize(16)

4.2. Simple plot 104 4.2. Simple plot 105

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

label.set_bbox(dict(facecolor='white', edgecolor='None', alpha=0.65))

...

4.3 Figures, Subplots, Axes and Ticks

A “figure” in matplotlib means the whole window in the user interface. Within this figure there can be “sub-
plots”. 4.3.3 Axes

Tip: So far we have used implicit figure and axes creation. This is handy for fast plots. We can have more Axes are very similar to subplots but allow placement of plots at any location in the fig-
control over the display using figure, subplot, and axes explicitly. While subplot positions the plots in a regular ure. So if we want to put a smaller plot inside a bigger one we do so with axes.
grid, axes allows free placement within the figure. Both can be useful depending on your intention. We’ve
already worked with figures and subplots without explicitly calling them. When we call plot, matplotlib calls
gca() to get the current axes and gca in turn calls gcf() to get the current figure. If there is none it calls
figure() to make one, strictly speaking, to make a subplot(111). Let’s look at the details.

4.3.1 Figures

Tip: A figure is the windows in the GUI that has “Figure #” as title. Figures are numbered starting from 1 as
opposed to the normal Python way starting from 0. This is clearly MATLAB-style. There are several parameters
that determine what the figure looks like:

4.3.4 Ticks
Argument Default Description
num 1 number of figure Well formatted ticks are an important part of publishing-ready figures. Matplotlib provides a totally config-
figsize figure.figsize figure size in inches (width, height) urable system for ticks. There are tick locators to specify where ticks should appear and tick formatters to
dpi figure.dpi resolution in dots per inch give ticks the appearance you want. Major and minor ticks can be located and formatted independently from
facecolor figure.facecolor color of the drawing background each other. Per default minor ticks are not shown, i.e. there is only an empty list for them because it is as
edgecolor figure.edgecolor color of edge around the drawing background NullLocator (see below).
frameon True draw figure frame or not

Tick Locators
Tip: The defaults can be specified in the resource file and will be used most of the time. Only the number of
Tick locators control the positions of the ticks. They are set as follows:
the figure is frequently changed.
As with other objects, you can set figure properties also setp or with the set_something methods. ax = plt.gca()
ax.xaxis.set_major_locator(eval(locator))
When you work with the GUI you can close a figure by clicking on the x in the upper right corner. But you can
close a figure programmatically by calling close. Depending on the argument it closes (1) the current figure
(no argument), (2) a specific figure (figure number or figure instance as argument), or (3) all figures ("all" as
argument).

plt.close(1) # Closes figure 1

4.3.2 Subplots

Tip: With subplot you can arrange plots in a regular grid. You need to specify the number of rows and columns
and the number of the plot. Note that the gridspec command is a more powerful alternative.
There are several locators for different kind of requirements:
All of these locators derive from the base class matplotlib.ticker.Locator. You can make your own lo-

4.3. Figures, Subplots, Axes and Ticks 106 4.3. Figures, Subplots, Axes and Ticks 107
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

cator deriving from it. Handling dates as ticks can be especially tricky. Therefore, matplotlib provides special 4.4.2 Scatter Plots
locators in matplotlib.dates.

4.4 Other Types of Plots: examples and exercises

Starting from the code below, try to reproduce the graphic

taking care of marker size, color and transparency.

Hint: Color is given by angle of (X,Y).

n = 1024
X = np.random.normal(0,1,n)
Y = np.random.normal(0,1,n)
4.4.1 Regular Plots
plt.scatter(X,Y)

Click on figure for solution.

4.4.3 Bar Plots

Starting from the code below, try to reproduce the graphic

taking care of filled areas:

Hint: You need to use the fill_between command.

n = 256 Starting from the code below, try to reproduce the graphic by
X = np.linspace(-np.pi, np.pi, n, endpoint=True) adding labels for red bars.
Y = np.sin(2 * X)
Hint: You need to take care of text alignment.
plt.plot(X, Y + 1, color='blue', alpha=1.00)
plt.plot(X, Y - 1, color='blue', alpha=1.00)
n = 12
Click on the figure for solution. X = np.arange(n)
Y1 = (1 - X / float(n)) * np.random.uniform(0.5, 1.0, n)
Y2 = (1 - X / float(n)) * np.random.uniform(0.5, 1.0, n)

plt.bar(X, +Y1, facecolor='#9999ff', edgecolor='white')

plt.bar(X, -Y2, facecolor='#ff9999', edgecolor='white')

for x, y in zip(X, Y1):

4.4. Other Types of Plots: examples and exercises 108 4.4. Other Types of Plots: examples and exercises 109
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

plt.text(x + 0.4, y + 0.05, '%.2f ' % y, ha='center', va='bottom')

Hint: You need to take care of the origin of the image in the imshow command and use a colorbar
plt.ylim(-1.25, +1.25)
def f(x, y):
Click on figure for solution. return (1 - x / 2 + x ** 5 + y ** 3) * np.exp(-x ** 2 - y ** 2)

n = 10
4.4.4 Contour Plots x = np.linspace(-3, 3, 4 * n)
y = np.linspace(-3, 3, 3 * n)
X, Y = np.meshgrid(x, y)
plt.imshow(f(X, Y))

Click on the figure for the solution.

4.4.6 Pie Charts

Starting from the code below, try to reproduce the graphic

taking care of the colormap (see Colormaps below).

Hint: You need to use the clabel command.

def f(x, y):

return (1 - x / 2 + x ** 5 + y ** 3) * np.exp(-x ** 2 -y ** 2)
Starting from the code below, try to reproduce the graphic
n = 256 taking care of colors and slices size.
x = np.linspace(-3, 3, n)
y = np.linspace(-3, 3, n)
X, Y = np.meshgrid(x, y) Hint: You need to modify Z.

plt.contourf(X, Y, f(X, Y), 8, alpha=.75, cmap='jet')

Z = np.random.uniform(0, 1, 20)
C = plt.contour(X, Y, f(X, Y), 8, colors='black', linewidth=.5)
plt.pie(Z)

Click on figure for solution.

Click on the figure for the solution.

4.4.5 Imshow 4.4.7 Quiver Plots

Starting from the code below, try to reproduce the graphic

Starting from the code above, try to reproduce the graphic
taking care of colormap, image interpolation and origin.
taking care of colors and orientations.

4.4. Other Types of Plots: examples and exercises 110 4.4. Other Types of Plots: examples and exercises 111
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Click on figure for solution.

Hint: You need to draw arrows twice.

4.4.10 Polar Axis

n = 8
X, Y = np.mgrid[0:n, 0:n]
plt.quiver(X, Y)

Click on figure for solution.

4.4.8 Grids

Hint: You only need to modify the axes line

Starting from the code below, try to reproduce the graphic.

plt.axes([0, 0, 1, 1])
Starting from the code below, try to reproduce the graphic
taking care of line styles. N = 20
theta = np.arange(0., 2 * np.pi, 2 * np.pi / N)
axes = plt.gca() radii = 10 * np.random.rand(N)
axes.set_xlim(0, 4) width = np.pi / 4 * np.random.rand(N)
axes.set_ylim(0, 3) bars = plt.bar(theta, radii, width=width, bottom=0.0)
axes.set_xticklabels([])
axes.set_yticklabels([]) for r, bar in zip(radii, bars):
bar.set_facecolor(cm.jet(r / 10.))
Click on figure for solution. bar.set_alpha(0.5)

Click on figure for solution.

4.4.9 Multi Plots
4.4.11 3D Plots

Starting from the code below, try to reproduce the graphic.

Starting from the code below, try to reproduce the graphic.

Hint: You can use several subplots with different partition.

Hint: You need to use contourf

plt.subplot(2, 2, 1)
plt.subplot(2, 2, 3)
plt.subplot(2, 2, 4)

4.4. Other Types of Plots: examples and exercises 112 4.4. Other Types of Plots: examples and exercises 113
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

from mpl_toolkits.mplot3d import Axes3D 4.5 Beyond this tutorial

fig = plt.figure()
Matplotlib benefits from extensive documentation as well as a large community of users and developers. Here
ax = Axes3D(fig)
are some links of interest:
X = np.arange(-4, 4, 0.25)
Y = np.arange(-4, 4, 0.25)
X, Y = np.meshgrid(X, Y)
R = np.sqrt(X**2 + Y**2) 4.5.1 Tutorials
Z = np.sin(R)
• Pyplot tutorial
ax.plot_surface(X, Y, Z, rstride=1, cstride=1, cmap='hot') • Introduction
• Controlling line properties
Click on figure for solution. • Working with multiple figures and axes
• Working with text
See also: • Image tutorial
3D plotting with Mayavi • Startup commands
• Importing image data into Numpy arrays
• Plotting numpy arrays as images
4.4.12 Text • Text tutorial
• Text introduction
• Basic text commands
• Text properties and layout
• Writing mathematical expressions
• Text rendering With LaTeX
• Annotating text
• Artist tutorial
• Introduction
• Customizing your objects
• Object containers
• Figure container
• Axes container
• Axis containers
Try to do the same from scratch ! • Tick containers
• Path tutorial
• Introduction
Hint: Have a look at the matplotlib logo.
• Bézier example
• Compound paths
Click on figure for solution. • Transforms tutorial
• Introduction
• Data coordinates
• Axes coordinates
• Blended transformations
• Using offset transforms to create a shadow effect
• The transformation pipeline

4.5.2 Matplotlib documentation

• User guide
Quick read • FAQ
• Installation
If you want to do a first quick pass through the Scipy lectures to learn the ecosystem, you can directly skip • Usage
to the next chapter: Scipy : high-level scientific computing. • How-To
• Troubleshooting
The remainder of this chapter is not necessary to follow the rest of the intro part. But be sure to come back • Environment Variables
and finish this chapter later. • Screenshots

4.4. Other Types of Plots: examples and exercises 114 4.5. Beyond this tutorial 115
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

4.5.3 Code documentation 4.6.1 Line properties

The code is well documented and you can quickly access a specific command from within a python session:
Property Description Appearance
>>> import matplotlib.pyplot as plt
alpha (or a) alpha transparency on 0-1
>>> help(plt.plot)
Help on function plot in module matplotlib.pyplot: scale

antialiased True or False - use antialised

plot(*args, **kwargs)
rendering
Plot lines and/or markers to the
:class:`~matplotlib.axes.Axes`. *args* is a variable length
color (or c) matplotlib color arg
argument, allowing for multiple *x*, *y* pairs with an
optional format string. For example, each of the following is linestyle (or ls) see Line properties
legal:: linewidth (or lw) float, the line width in points

plot(x, y) # plot x and y using default line style and color solid_capstyle Cap style for solid lines
plot(x, y, 'bo') # plot x and y using blue circle markers solid_joinstyle Join style for solid lines
plot(y) # plot y using x as index array 0..N-1
plot(y, 'r+') # ditto, but with red plusses dash_capstyle Cap style for dashes
dash_joinstyle Join style for dashes
If *x* and/or *y* is 2-dimensional, then the corresponding columns
will be plotted. marker see Markers
... markeredgewidth line width around the marker
(mew) symbol
markeredgecolor edge color if a marker is used
4.5.4 Galleries (mec)
markerfacecolor face color if a marker is used
The matplotlib gallery is also incredibly useful when you search how to render a given graphic. Each example
(mfc)
comes with its source.
markersize (ms) size of the marker in points

4.5.5 Mailing lists

4.6.2 Line styles
Finally, there is a user mailing list where you can ask for help and a developers mailing list that is more techni-
cal.

4.6 Quick references

Here is a set of tables that show main properties and styles.

4.6. Quick references 116 4.6. Quick references 117

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

4.6.3 Markers Pie chart

A simple pie chart example with matplotlib.

4.6.4 Colormaps

All colormaps can be reversed by appending _r. For instance, gray_r is the reverse of gray.
If you want to know more about colormaps, checks Documenting the matplotlib colormaps.

import numpy as np
import matplotlib.pyplot as plt

n = 20
Z = np.ones(n)
Z[-1] *= 2

plt.axes([0.025, 0.025, 0.95, 0.95])

plt.pie(Z, explode=Z*.05, colors = ['%f ' % (i/float(n)) for i in range(n)])

plt.axis('equal')
plt.xticks(())
plt.yticks()

plt.show()

Total running time of the script: ( 0 minutes 0.047 seconds)

4.7 Full code examples
Download Python source code: plot_pie.py
Download Jupyter notebook: plot_pie.ipynb
4.7.1 Code samples for Matplotlib
Generated by Sphinx-Gallery
The examples here are only examples relevant to the points raised in this chapter. The matplotlib documenta-
tion comes with a much more exhaustive gallery.
Plotting a scatter of points

A simple example showing how to plot a scatter of points with matplotlib.

4.7. Full code examples 118 4.7. Full code examples 119
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

import numpy as np
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
import numpy as np fig = plt.figure(figsize=(5, 4), dpi=72)
import matplotlib.pyplot as plt axes = fig.add_axes([0.01, 0.01, .98, 0.98])
X = np.linspace(0, 2, 200, endpoint=True)
n = 1024 Y = np.sin(2*np.pi*X)
X = np.random.normal(0, 1, n) plt.plot(X, Y, lw=2)
Y = np.random.normal(0, 1, n) plt.ylim(-1.1, 1.1)
T = np.arctan2(Y, X) plt.grid()
plt.axes([0.025, 0.025, 0.95, 0.95]) plt.show()
plt.scatter(X, Y, s=75, c=T, alpha=.5)

plt.xlim(-1.5, 1.5) Total running time of the script: ( 0 minutes 0.081 seconds)
plt.xticks(()) Download Python source code: plot_good.py
plt.ylim(-1.5, 1.5)
plt.yticks(()) Download Jupyter notebook: plot_good.ipynb
Generated by Sphinx-Gallery
plt.show()

Total running time of the script: ( 0 minutes 0.089 seconds) Subplots

Download Python source code: plot_scatter.py
Show multiple subplots in matplotlib.
Download Jupyter notebook: plot_scatter.ipynb
Generated by Sphinx-Gallery

A simple, good-looking plot

Demoing some simple features of matplotlib

4.7. Full code examples 120 4.7. Full code examples 121
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Simple axes example

This example shows a couple of simple usage of axes.

import matplotlib.pyplot as plt

fig = plt.figure()
fig.subplots_adjust(bottom=0.025, left=0.025, top = 0.975, right=0.975)
import matplotlib.pyplot as plt
plt.subplot(2, 1, 1)
plt.xticks(()), plt.yticks(()) plt.axes([.1, .1, .8, .8])
plt.xticks(())
plt.subplot(2, 3, 4) plt.yticks(())
plt.xticks(()) plt.text(.6, .6, 'axes([0.1, 0.1, .8, .8])', ha='center', va='center',
plt.yticks(()) size=20, alpha=.5)

plt.subplot(2, 3, 5) plt.axes([.2, .2, .3, .3])

plt.xticks(()) plt.xticks(())
plt.yticks(()) plt.yticks(())
plt.text(.5, .5, 'axes([0.2, 0.2, .3, .3])', ha='center', va='center',
plt.subplot(2, 3, 6) size=16, alpha=.5)
plt.xticks(())
plt.yticks(()) plt.show()

plt.show() Total running time of the script: ( 0 minutes 0.063 seconds)

Total running time of the script: ( 0 minutes 0.126 seconds) Download Python source code: plot_axes.py

Download Python source code: plot_multiplot.py Download Jupyter notebook: plot_axes.ipynb

Download Jupyter notebook: plot_multiplot.ipynb Generated by Sphinx-Gallery

Generated by Sphinx-Gallery
A simple plotting example

A plotting example with a few simple tweaks

4.7. Full code examples 122 4.7. Full code examples 123
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

import numpy as np import matplotlib.pyplot as plt

import matplotlib
matplotlib.use('Agg') plt.figure(figsize=(6, 4))
import matplotlib.pyplot as plt plt.subplot(2, 1, 1)
plt.xticks(())
fig = plt.figure(figsize=(5, 4), dpi=72) plt.yticks(())
axes = fig.add_axes([0.01, 0.01, .98, 0.98]) plt.text(0.5, 0.5, 'subplot(2,1,1)', ha='center', va='center',
x = np.linspace(0, 2, 200, endpoint=True) size=24, alpha=.5)
y = np.sin(2 * np.pi * x)
plt.plot(x, y, lw=.25, c='k') plt.subplot(2, 1, 2)
plt.xticks(np.arange(0.0, 2.0, 0.1)) plt.xticks(())
plt.yticks(np.arange(-1.0, 1.0, 0.1)) plt.yticks(())
plt.grid() plt.text(0.5, 0.5, 'subplot(2,1,2)', ha='center', va='center',
plt.show() size=24, alpha=.5)

Total running time of the script: ( 0 minutes 0.074 seconds) plt.tight_layout()

plt.show()
Download Python source code: plot_bad.py
Download Jupyter notebook: plot_bad.ipynb Total running time of the script: ( 0 minutes 0.068 seconds)

Generated by Sphinx-Gallery Download Python source code: plot_subplot-horizontal.py

Download Jupyter notebook: plot_subplot-horizontal.ipynb
Horizontal arrangement of subplots Generated by Sphinx-Gallery

An example showing horizontal arrangement of subplots with matplotlib.

Subplot plot arrangement vertical

An example showing vertical arrangement of subplots with matplotlib.

4.7. Full code examples 124 4.7. Full code examples 125
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

import matplotlib.pyplot as plt

plt.figure(figsize=(6, 4))
plt.subplot(1, 2, 1)
plt.xticks(()) import numpy as np
plt.yticks(()) import matplotlib.pyplot as plt
plt.text(0.5, 0.5, 'subplot(1,2,1)', ha='center', va='center', from mpl_toolkits.mplot3d import Axes3D
size=24, alpha=.5)
fig = plt.figure()
plt.subplot(1, 2, 2) ax = Axes3D(fig)
plt.xticks(()) X = np.arange(-4, 4, 0.25)
plt.yticks(()) Y = np.arange(-4, 4, 0.25)
plt.text(0.5, 0.5, 'subplot(1,2,2)', ha='center', va='center', X, Y = np.meshgrid(X, Y)
size=24, alpha=.5) R = np.sqrt(X ** 2 + Y ** 2)
Z = np.sin(R)
plt.tight_layout()
plt.show() ax.plot_surface(X, Y, Z, rstride=1, cstride=1, cmap=plt.cm.hot)
ax.contourf(X, Y, Z, zdir='z', offset=-2, cmap=plt.cm.hot)
ax.set_zlim(-2, 2)
Total running time of the script: ( 0 minutes 0.068 seconds)
Download Python source code: plot_subplot-vertical.py plt.show()

Download Jupyter notebook: plot_subplot-vertical.ipynb Total running time of the script: ( 0 minutes 0.065 seconds)
Generated by Sphinx-Gallery Download Python source code: plot_plot3d.py
Download Jupyter notebook: plot_plot3d.ipynb
3D plotting
Generated by Sphinx-Gallery
A simple example of 3D plotting.
Imshow elaborate

An example demoing imshow and styling the figure.

4.7. Full code examples 126 4.7. Full code examples 127
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

import numpy as np import numpy as np

import matplotlib.pyplot as plt import matplotlib.pyplot as plt

def f(x, y): n = 8

return (1 - x / 2 + x ** 5 + y ** 3 ) * np.exp(-x ** 2 - y ** 2) X, Y = np.mgrid[0:n, 0:n]
T = np.arctan2(Y - n / 2., X - n/2.)
n = 10 R = 10 + np.sqrt((Y - n / 2.0) ** 2 + (X - n / 2.0) ** 2)
x = np.linspace(-3, 3, 3.5 * n) U, V = R * np.cos(T), R * np.sin(T)
y = np.linspace(-3, 3, 3.0 * n)
X, Y = np.meshgrid(x, y) plt.axes([0.025, 0.025, 0.95, 0.95])
Z = f(X, Y) plt.quiver(X, Y, U, V, R, alpha=.5)
plt.quiver(X, Y, U, V, edgecolor='k', facecolor='None', linewidth=.5)
plt.axes([0.025, 0.025, 0.95, 0.95])
plt.imshow(Z, interpolation='nearest', cmap='bone', origin='lower') plt.xlim(-1, n)
plt.colorbar(shrink=.92) plt.xticks(())
plt.ylim(-1, n)
plt.xticks(()) plt.yticks(())
plt.yticks(())
plt.show() plt.show()

Total running time of the script: ( 0 minutes 0.075 seconds) Total running time of the script: ( 0 minutes 0.034 seconds)
Download Python source code: plot_imshow.py Download Python source code: plot_quiver.py
Download Jupyter notebook: plot_imshow.ipynb Download Jupyter notebook: plot_quiver.ipynb
Generated by Sphinx-Gallery Generated by Sphinx-Gallery

Plotting a vector field: quiver Plotting in polar coordinnates

A simple example showing how to plot a vector field (quiver) with matplotlib. A simple example showing how to plot in polar coordinnates with matplotlib.

4.7. Full code examples 128 4.7. Full code examples 129
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

import numpy as np import numpy as np

import matplotlib.pyplot as plt import matplotlib.pyplot as plt

ax = plt.axes([0.025, 0.025, 0.95, 0.95], polar=True) def f(x,y):

return (1 - x / 2 + x**5 + y**3) * np.exp(-x**2 -y**2)
N = 20
theta = np.arange(0.0, 2 * np.pi, 2 * np.pi / N) n = 256
radii = 10 * np.random.rand(N) x = np.linspace(-3, 3, n)
width = np.pi / 4 * np.random.rand(N) y = np.linspace(-3, 3, n)
bars = plt.bar(theta, radii, width=width, bottom=0.0) X,Y = np.meshgrid(x, y)

for r,bar in zip(radii, bars): plt.axes([0.025, 0.025, 0.95, 0.95])

bar.set_facecolor(plt.cm.jet(r/10.))
bar.set_alpha(0.5) plt.contourf(X, Y, f(X, Y), 8, alpha=.75, cmap=plt.cm.hot)
C = plt.contour(X, Y, f(X, Y), 8, colors='black', linewidth=.5)
ax.set_xticklabels([]) plt.clabel(C, inline=1, fontsize=10)
ax.set_yticklabels([])
plt.show() plt.xticks(())
plt.yticks(())
Total running time of the script: ( 0 minutes 0.048 seconds) plt.show()

Download Python source code: plot_polar.py Total running time of the script: ( 0 minutes 0.095 seconds)
Download Jupyter notebook: plot_polar.ipynb Download Python source code: plot_contour.py
Generated by Sphinx-Gallery Download Jupyter notebook: plot_contour.ipynb
Generated by Sphinx-Gallery
Displaying the contours of a function

An example showing how to display the contours of a function with matplotlib.

4.7. Full code examples 130 4.7. Full code examples 131
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

A example of plotting not quite right

An “ugly” example of plotting.

import numpy as np
import matplotlib import numpy as np
matplotlib.use('Agg') import matplotlib.pyplot as plt
import matplotlib.pyplot as plt
n = 256
matplotlib.rc('grid', color='black', linestyle='-', linewidth=1) X = np.linspace(-np.pi, np.pi, n, endpoint=True)
Y = np.sin(2 * X)
fig = plt.figure(figsize=(5,4),dpi=72)
axes = fig.add_axes([0.01, 0.01, .98, 0.98], axisbg='.75') plt.axes([0.025, 0.025, 0.95, 0.95])
X = np.linspace(0, 2, 40, endpoint=True)
Y = np.sin(2 * np.pi * X) plt.plot(X, Y + 1, color='blue', alpha=1.00)
plt.plot(X, Y, lw=.05, c='b', antialiased=False) plt.fill_between(X, 1, Y + 1, color='blue', alpha=.25)

plt.xticks(()) plt.plot(X, Y - 1, color='blue', alpha=1.00)

plt.yticks(np.arange(-1., 1., 0.2)) plt.fill_between(X, -1, Y - 1, (Y - 1) > -1, color='blue', alpha=.25)
plt.grid() plt.fill_between(X, -1, Y - 1, (Y - 1) < -1, color='red', alpha=.25)
ax = plt.gca()
plt.xlim(-np.pi, np.pi)
plt.show() plt.xticks(())
plt.ylim(-2.5, 2.5)
plt.yticks(())
Total running time of the script: ( 0 minutes 0.042 seconds)
Download Python source code: plot_ugly.py plt.show()

Download Jupyter notebook: plot_ugly.ipynb Total running time of the script: ( 0 minutes 0.038 seconds)
Generated by Sphinx-Gallery Download Python source code: plot_plot.py
Download Jupyter notebook: plot_plot.ipynb
Plot and filled plots
Generated by Sphinx-Gallery
Simple example of plots and filling between them with matplotlib.

4.7. Full code examples 132 4.7. Full code examples 133
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Bar plots Download Jupyter notebook: plot_bar.ipynb

Generated by Sphinx-Gallery
An example of bar plots with matplotlib.

Subplot grid

An example showing the subplot grid in matplotlib.

import numpy as np import matplotlib.pyplot as plt

import matplotlib.pyplot as plt
plt.figure(figsize=(6, 4))
n = 12 plt.subplot(2, 2, 1)
X = np.arange(n) plt.xticks(())
Y1 = (1 - X / float(n)) * np.random.uniform(0.5, 1.0, n) plt.yticks(())
Y2 = (1 - X / float(n)) * np.random.uniform(0.5, 1.0, n) plt.text(0.5, 0.5, 'subplot(2,2,1)', ha='center', va='center',
size=20, alpha=.5)
plt.axes([0.025, 0.025, 0.95, 0.95])
plt.bar(X, +Y1, facecolor='#9999ff', edgecolor='white') plt.subplot(2, 2, 2)
plt.bar(X, -Y2, facecolor='#ff9999', edgecolor='white') plt.xticks(())
plt.yticks(())
for x, y in zip(X, Y1): plt.text(0.5, 0.5, 'subplot(2,2,2)', ha='center', va='center',
plt.text(x + 0.4, y + 0.05, '%.2f ' % y, ha='center', va= 'bottom') size=20, alpha=.5)

for x, y in zip(X, Y2): plt.subplot(2, 2, 3)

plt.text(x + 0.4, -y - 0.05, '%.2f ' % y, ha='center', va= 'top') plt.xticks(())
plt.yticks(())
plt.xlim(-.5, n)
plt.xticks(()) plt.text(0.5, 0.5, 'subplot(2,2,3)', ha='center', va='center',
plt.ylim(-1.25, 1.25) size=20, alpha=.5)
plt.yticks(())
plt.subplot(2, 2, 4)
plt.show() plt.xticks(())
plt.yticks(())
Total running time of the script: ( 0 minutes 0.055 seconds) plt.text(0.5, 0.5, 'subplot(2,2,4)', ha='center', va='center',
size=20, alpha=.5)
Download Python source code: plot_bar.py

4.7. Full code examples 134 4.7. Full code examples 135
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

plt.xticks(())
plt.tight_layout() plt.yticks(())
plt.show() plt.text(0.1, 0.1, 'axes([0.3, 0.3, .5, .5])', ha='left', va='center',
size=16, alpha=.5)
Total running time of the script: ( 0 minutes 0.136 seconds)
plt.axes([.4, .4, .5, .5])
Download Python source code: plot_subplot-grid.py plt.xticks(())
plt.yticks(())
Download Jupyter notebook: plot_subplot-grid.ipynb
plt.text(0.1, 0.1, 'axes([0.4, 0.4, .5, .5])', ha='left', va='center',
Generated by Sphinx-Gallery size=16, alpha=.5)

plt.show()
Axes
Total running time of the script: ( 0 minutes 0.166 seconds)
This example shows various axes command to position matplotlib axes.
Download Python source code: plot_axes-2.py
Download Jupyter notebook: plot_axes-2.ipynb
Generated by Sphinx-Gallery

Grid

Displaying a grid on the axes in matploblib.

import matplotlib.pyplot as plt

plt.axes([.1, .1, .5, .5])

plt.xticks(())
plt.yticks(())
plt.text(0.1, 0.1, 'axes([0.1, 0.1, .8, .8])', ha='left', va='center',
size=16, alpha=.5)

plt.axes([.2, .2, .5, .5])

plt.xticks(()) import matplotlib.pyplot as plt
plt.yticks(())
plt.text(0.1, 0.1, 'axes([0.2, 0.2, .5, .5])', ha='left', va='center', ax = plt.axes([0.025, 0.025, 0.95, 0.95])
size=16, alpha=.5)
ax.set_xlim(0,4)
plt.axes([0.3, 0.3, .5, .5]) ax.set_ylim(0,3)

4.7. Full code examples 136 4.7. Full code examples 137
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

ax.xaxis.set_major_locator(plt.MultipleLocator(1.0)) ax.clabel(cset, fontsize=9, inline=1)

ax.xaxis.set_minor_locator(plt.MultipleLocator(0.1))
ax.yaxis.set_major_locator(plt.MultipleLocator(1.0)) plt.xticks(())
ax.yaxis.set_minor_locator(plt.MultipleLocator(0.1)) plt.yticks(())
ax.grid(which='major', axis='x', linewidth=0.75, linestyle='-', color='0.75') ax.set_zticks(())
ax.grid(which='minor', axis='x', linewidth=0.25, linestyle='-', color='0.75')
ax.grid(which='major', axis='y', linewidth=0.75, linestyle='-', color='0.75')
ax.grid(which='minor', axis='y', linewidth=0.25, linestyle='-', color='0.75') ax.text2D(-0.05, 1.05, " 3D plots \n",
ax.set_xticklabels([]) horizontalalignment='left',
ax.set_yticklabels([]) verticalalignment='top',
bbox=dict(facecolor='white', alpha=1.0),
plt.show() family='Lint McCree Intl BB',
size='x-large',
Total running time of the script: ( 0 minutes 0.043 seconds) transform=plt.gca().transAxes)

Download Python source code: plot_grid.py ax.text2D(-0.05, .975, " Plot 2D or 3D data",
horizontalalignment='left',
Download Jupyter notebook: plot_grid.ipynb
verticalalignment='top',
Generated by Sphinx-Gallery family='Lint McCree Intl BB',
size='medium',
transform=plt.gca().transAxes)
3D plotting
plt.show()
Demo 3D plotting with matplotlib and style the figure.
Total running time of the script: ( 0 minutes 0.052 seconds)
Download Python source code: plot_plot3d-2.py
Download Jupyter notebook: plot_plot3d-2.ipynb
Generated by Sphinx-Gallery

GridSpec

An example demoing gridspec

import matplotlib.pyplot as plt

from mpl_toolkits.mplot3d import axes3d

ax = plt.gca(projection='3d')
X, Y, Z = axes3d.get_test_data(0.05)
cset = ax.contourf(X, Y, Z)

4.7. Full code examples 138 4.7. Full code examples 139
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

import matplotlib.pyplot as plt

import matplotlib.gridspec as gridspec

plt.figure(figsize=(6, 4))
G = gridspec.GridSpec(3, 3)

axes_1 = plt.subplot(G[0, :])

plt.xticks(())
plt.yticks(())
plt.text(0.5, 0.5, 'Axes 1', ha='center', va='center', size=24, alpha=.5)

axes_2 = plt.subplot(G[1, :-1])

plt.xticks(())
plt.yticks(())
plt.text(0.5, 0.5, 'Axes 2', ha='center', va='center', size=24, alpha=.5)

axes_3 = plt.subplot(G[1:, -1])

plt.xticks(())
plt.yticks(())
plt.text(0.5, 0.5, 'Axes 3', ha='center', va='center', size=24, alpha=.5)

axes_4 = plt.subplot(G[-1, 0])

plt.xticks(())
plt.yticks(())
plt.text(0.5, 0.5, 'Axes 4', ha='center', va='center', size=24, alpha=.5)

axes_5 = plt.subplot(G[-1, -2])

plt.xticks(())
plt.yticks(())
plt.text(0.5, 0.5, 'Axes 5', ha='center', va='center', size=24, alpha=.5)

plt.tight_layout() import numpy as np

plt.show() import matplotlib.pyplot as plt

Total running time of the script: ( 0 minutes 0.168 seconds)

eqs = []
Download Python source code: plot_gridspec.py
eqs.append((r"$W^{3\beta}_{\delta_1 \rho_1 \sigma_2} = U^{3\beta}_{\delta_1 \rho_1} + \frac{1}
Download Jupyter notebook: plot_gridspec.ipynb ,→{8 \pi 2} \int^{\alpha_2}_{\alpha_2} d \alpha^\prime_2 \left[\frac{ U^{2\beta}_{\delta_1␣
,→\rho_1} - \alpha^\prime_2U^{1\beta}_{\rho_1 \sigma_2} }{U^{0\beta}_{\rho_1 \sigma_2}}\right]$
Generated by Sphinx-Gallery ,→"))
eqs.append((r"$\frac{d\rho}{d t} + \rho \vec{v}\cdot\nabla\vec{v} = -\nabla p + \mu\nabla^2␣
,→\vec{v} + \rho \vec{g}$"))
Demo text printing eqs.append((r"$\int_{-\infty}^\infty e^{-x^2}dx=\sqrt{\pi}$"))
eqs.append((r"$E = mc^2 = \sqrt{{m_0}^2c^4 + p^2c^2}$"))
A example showing off elaborate text printing with matplotlib. eqs.append((r"$F_G = G\frac{m_1m_2}{r^2}$"))

plt.axes([0.025, 0.025, 0.95, 0.95])

for i in range(24):
index = np.random.randint(0, len(eqs))
eq = eqs[index]
size = np.random.uniform(12, 32)
x,y = np.random.uniform(0, 1, 2)
alpha = np.random.uniform(0.25, .75)
plt.text(x, y, eq, ha='center', va='center', color="#11557c", alpha=alpha,
transform=plt.gca().transAxes, fontsize=size, clip_on=True)
plt.xticks(())
plt.yticks(())

plt.show()

Total running time of the script: ( 0 minutes 0.038 seconds)

4.7. Full code examples 140 4.7. Full code examples 141
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Download Python source code: plot_text.py Exercise 4

Download Jupyter notebook: plot_text.ipynb
Exercise 4 with matplotlib.
Generated by Sphinx-Gallery

4.7.2 Code for the chapter’s exercises

Excercise 1

Solution of the excercise 1 with matplotlib.

import numpy as np
import matplotlib.pyplot as plt

plt.figure(figsize=(8, 5), dpi=80)

plt.subplot(111)

X = np.linspace(-np.pi, np.pi, 256, endpoint=True)

S = np.sin(X)
C = np.cos(X)

plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-")

plt.plot(X, S, color="red", linewidth=2.5, linestyle="-")

plt.xlim(X.min() * 1.1, X.max() * 1.1)

import numpy as np plt.ylim(C.min() * 1.1, C.max() * 1.1)
import matplotlib.pyplot as plt
plt.show()
n = 256
X = np.linspace(-np.pi, np.pi, 256, endpoint=True) Total running time of the script: ( 0 minutes 0.034 seconds)
C,S = np.cos(X), np.sin(X)
plt.plot(X, C) Download Python source code: plot_exercise_4.py
plt.plot(X,S)
Download Jupyter notebook: plot_exercise_4.ipynb
plt.show() Generated by Sphinx-Gallery

Total running time of the script: ( 0 minutes 0.034 seconds)

Exercise 3
Download Python source code: plot_exercise_1.py
Download Jupyter notebook: plot_exercise_1.ipynb Exercise 3 with matplotlib.

Generated by Sphinx-Gallery

4.7. Full code examples 142 4.7. Full code examples 143
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

import numpy as np import numpy as np

import matplotlib.pyplot as plt import matplotlib.pyplot as plt

plt.figure(figsize=(8, 5), dpi=80) plt.figure(figsize=(8, 5), dpi=80)

plt.subplot(111) plt.subplot(111)

X = np.linspace(-np.pi, np.pi, 256, endpoint=True) X = np.linspace(-np.pi, np.pi, 256, endpoint=True)

C, S = np.cos(X), np.sin(X) S = np.sin(X)
C = np.cos(X)
plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-")
plt.plot(X, S, color="red", linewidth=2.5, linestyle="-") plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-")
plt.plot(X, S, color="red", linewidth=2.5, linestyle="-")
plt.xlim(-4.0, 4.0)
plt.xticks(np.linspace(-4, 4, 9, endpoint=True)) plt.xlim(X.min() * 1.1, X.max() * 1.1)
plt.xticks([-np.pi, -np.pi/2, 0, np.pi/2, np.pi])
plt.ylim(-1.0, 1.0)
plt.yticks(np.linspace(-1, 1, 5, endpoint=True)) plt.ylim(C.min() * 1.1, C.max() * 1.1)
plt.yticks([-1, 0, +1])
plt.show()
plt.show()
Total running time of the script: ( 0 minutes 0.046 seconds)
Total running time of the script: ( 0 minutes 0.039 seconds)
Download Python source code: plot_exercise_3.py
Download Python source code: plot_exercise_5.py
Download Jupyter notebook: plot_exercise_3.ipynb
Download Jupyter notebook: plot_exercise_5.ipynb
Generated by Sphinx-Gallery
Generated by Sphinx-Gallery

Exercise 5
Exercise 6
Exercise 5 with matplotlib.
Exercise 6 with matplotlib.

4.7. Full code examples 144 4.7. Full code examples 145
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

import numpy as np
import matplotlib.pyplot as plt

plt.figure(figsize=(8, 5), dpi=80)

plt.subplot(111)
import numpy as np
X = np.linspace(-np.pi, np.pi, 256, endpoint=True) import matplotlib.pyplot as plt
C = np.cos(X)
S = np.sin(X) # Create a new figure of size 8x6 points, using 100 dots per inch
plt.figure(figsize=(8, 6), dpi=80)
plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-")
plt.plot(X, S, color="red", linewidth=2.5, linestyle="-") # Create a new subplot from a grid of 1x1
plt.subplot(111)
plt.xlim(X.min() * 1.1, X.max() * 1.1)
plt.xticks([-np.pi, -np.pi/2, 0, np.pi/2, np.pi], X = np.linspace(-np.pi, np.pi, 256, endpoint=True)
[r'$-\pi$', r'$-\pi/2$', r'$0$', r'$+\pi/2$', r'$+\pi$']) C, S = np.cos(X), np.sin(X)

plt.ylim(C.min() * 1.1, C.max() * 1.1) # Plot cosine using blue color with a continuous line of width 1 (pixels)
plt.yticks([-1, 0, +1], plt.plot(X, C, color="blue", linewidth=1.0, linestyle="-")
[r'$-1$', r'$0$', r'$+1$'])
# Plot sine using green color with a continuous line of width 1 (pixels)
plt.show() plt.plot(X, S, color="green", linewidth=1.0, linestyle="-")

Total running time of the script: ( 0 minutes 0.040 seconds) # Set x limits
plt.xlim(-4., 4.)
Download Python source code: plot_exercise_6.py
Download Jupyter notebook: plot_exercise_6.ipynb # Set x ticks
plt.xticks(np.linspace(-4, 4, 9, endpoint=True))
Generated by Sphinx-Gallery
# Set y limits
plt.ylim(-1.0, 1.0)
Exercise 2
# Set y ticks
Exercise 2 with matplotlib. plt.yticks(np.linspace(-1, 1, 5, endpoint=True))

# Show result on screen

4.7. Full code examples 146 4.7. Full code examples 147
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

plt.show() plt.xticks([-np.pi, -np.pi/2, 0, np.pi/2, np.pi],

[r'$-\pi$', r'$-\pi/2$', r'$0$', r'$+\pi/2$', r'$+\pi$'])
Total running time of the script: ( 0 minutes 0.046 seconds)
plt.ylim(C.min() * 1.1, C.max() * 1.1)
Download Python source code: plot_exercise_2.py plt.yticks([-1, 0, +1],
[r'$-1$', r'$0$', r'$+1$'])
Download Jupyter notebook: plot_exercise_2.ipynb
Generated by Sphinx-Gallery plt.show()

Total running time of the script: ( 0 minutes 0.046 seconds)

Exercise 7
Download Python source code: plot_exercise_7.py
Exercise 7 with matplotlib Download Jupyter notebook: plot_exercise_7.ipynb
Generated by Sphinx-Gallery

Exercise 8

Exercise 8 with matplotlib.

import numpy as np
import matplotlib.pyplot as plt

plt.figure(figsize=(8,5), dpi=80)
plt.subplot(111)

X = np.linspace(-np.pi, np.pi, 256,endpoint=True)

C = np.cos(X) import numpy as np
S = np.sin(X) import matplotlib.pyplot as plt

plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-") plt.figure(figsize=(8,5), dpi=80)

plt.plot(X, S, color="red", linewidth=2.5, linestyle="-") plt.subplot(111)

ax = plt.gca() X = np.linspace(-np.pi, np.pi, 256,endpoint=True)

ax.spines['right'].set_color('none') C = np.cos(X)
ax.spines['top'].set_color('none') S = np.sin(X)
ax.xaxis.set_ticks_position('bottom')
ax.spines['bottom'].set_position(('data',0)) plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-", label="cosine")
ax.yaxis.set_ticks_position('left') plt.plot(X, S, color="red", linewidth=2.5, linestyle="-", label="sine")
ax.spines['left'].set_position(('data',0))
ax = plt.gca()
plt.xlim(X.min() * 1.1, X.max() * 1.1) ax.spines['right'].set_color('none')

4.7. Full code examples 148 4.7. Full code examples 149
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

ax.spines['top'].set_color('none') X = np.linspace(-np.pi, np.pi, 256,endpoint=True)

ax.xaxis.set_ticks_position('bottom') C = np.cos(X)
ax.spines['bottom'].set_position(('data',0)) S = np.sin(X)
ax.yaxis.set_ticks_position('left')
ax.spines['left'].set_position(('data',0)) plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-", label="cosine")
plt.plot(X, S, color="red", linewidth=2.5, linestyle="-", label="sine")
plt.xlim(X.min() * 1.1, X.max() * 1.1)
plt.xticks([-np.pi, -np.pi/2, 0, np.pi/2, np.pi], ax = plt.gca()
[r'$-\pi$', r'$-\pi/2$', r'$0$', r'$+\pi/2$', r'$+\pi$']) ax.spines['right'].set_color('none')
ax.spines['top'].set_color('none')
plt.ylim(C.min() * 1.1, C.max() * 1.1) ax.xaxis.set_ticks_position('bottom')
plt.yticks([-1, +1], ax.spines['bottom'].set_position(('data',0))
[r'$-1$', r'$+1$']) ax.yaxis.set_ticks_position('left')
ax.spines['left'].set_position(('data',0))
plt.legend(loc='upper left')
plt.xlim(X.min() * 1.1, X.max() * 1.1)
plt.show() plt.xticks([-np.pi, -np.pi/2, 0, np.pi/2, np.pi],
[r'$-\pi$', r'$-\pi/2$', r'$0$', r'$+\pi/2$', r'$+\pi$'])
Total running time of the script: ( 0 minutes 0.094 seconds)
plt.ylim(C.min() * 1.1, C.max() * 1.1)
Download Python source code: plot_exercise_8.py plt.yticks([-1, +1],
[r'$-1$', r'$+1$'])
Download Jupyter notebook: plot_exercise_8.ipynb
Generated by Sphinx-Gallery t = 2*np.pi/3
plt.plot([t, t], [0, np.cos(t)],
color='blue', linewidth=1.5, linestyle="--")
Exercise 9 plt.scatter([t, ], [np.cos(t), ], 50, color='blue')
plt.annotate(r'$sin(\frac{2\pi}{3})=\frac{\sqrt{3}}{2}$',
Exercise 9 with matplotlib. xy=(t, np.sin(t)), xycoords='data',
xytext=(+10, +30), textcoords='offset points', fontsize=16,
arrowprops=dict(arrowstyle="->", connectionstyle="arc3,rad=.2"))

plt.plot([t, t], [0, np.sin(t)],

color='red', linewidth=1.5, linestyle="--")
plt.scatter([t, ], [np.sin(t), ], 50, color='red')
plt.annotate(r'$cos(\frac{2\pi}{3})=-\frac{1}{2}$', xy=(t, np.cos(t)),
xycoords='data', xytext=(-90, -50), textcoords='offset points',
fontsize=16,
arrowprops=dict(arrowstyle="->", connectionstyle="arc3,rad=.2"))

plt.legend(loc='upper left')

plt.show()

Total running time of the script: ( 0 minutes 0.051 seconds)

Download Python source code: plot_exercise_9.py
Download Jupyter notebook: plot_exercise_9.ipynb
Generated by Sphinx-Gallery

Exercise

Exercises with matplotlib.

import numpy as np
import matplotlib.pyplot as plt

plt.figure(figsize=(8, 5), dpi=80)

plt.subplot(111)

4.7. Full code examples 150 4.7. Full code examples 151
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

xytext=(10, 30), textcoords='offset points', fontsize=16,

arrowprops=dict(arrowstyle="->", connectionstyle="arc3,rad=.2"))

plt.plot([t, t], [0, np.sin(t)],

color='red', linewidth=1.5, linestyle="--")
plt.scatter([t, ], [np.sin(t), ], 50, color ='red')
plt.annotate(r'$cos(\frac{2\pi}{3})=-\frac{1}{2}$', xy=(t, np.cos(t)),
xycoords='data', xytext=(-90, -50),
textcoords='offset points', fontsize=16,
arrowprops=dict(arrowstyle="->", connectionstyle="arc3,rad=.2"))

for label in ax.get_xticklabels() + ax.get_yticklabels():

label.set_fontsize(16)
label.set_bbox(dict(facecolor='white', edgecolor='None', alpha=0.65 ))

plt.show()

Total running time of the script: ( 0 minutes 0.099 seconds)

Download Python source code: plot_exercise_10.py
Download Jupyter notebook: plot_exercise_10.ipynb
Generated by Sphinx-Gallery

4.7.3 Example demoing choices for an option

import numpy as np
import matplotlib.pyplot as plt
The colors matplotlib line plots
plt.figure(figsize=(8, 5), dpi=80)
plt.subplot(111) An example demoing the various colors taken by matplotlib’s plot.

X = np.linspace(-np.pi, np.pi, 256, endpoint=True)

C, S = np.cos(X), np.sin(X)
import matplotlib.pyplot as plt
plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-", label="cosine")
plt.plot(X, S, color="red", linewidth=2.5, linestyle="-", label="sine") size = 256, 16
dpi = 72.0
ax = plt.gca() figsize = size[0] / float(dpi), size[1] / float(dpi)
ax.spines['right'].set_color('none') fig = plt.figure(figsize=figsize, dpi=dpi)
ax.spines['top'].set_color('none') fig.patch.set_alpha(0)
ax.xaxis.set_ticks_position('bottom') plt.axes([0, 0.1, 1, .8], frameon=False)
ax.spines['bottom'].set_position(('data', 0))
ax.yaxis.set_ticks_position('left') for i in range(1,11):
ax.spines['left'].set_position(('data', 0)) plt.plot([i, i], [0, 1], lw=1.5)

plt.xlim(X.min() * 1.1, X.max() * 1.1) plt.xlim(0, 11)

plt.xticks([-np.pi, -np.pi / 2, 0, np.pi / 2, np.pi], plt.xticks(())
[r'$-\pi$', r'$-\pi/2$', r'$0$', r'$+\pi/2$', r'$+\pi$']) plt.yticks(())
plt.show()
plt.ylim(C.min() * 1.1, C.max() * 1.1)
plt.yticks([-1, 1], Total running time of the script: ( 0 minutes 0.040 seconds)
[r'$-1$', r'$+1$'])
Download Python source code: plot_color.py
plt.legend(loc='upper left') Download Jupyter notebook: plot_color.ipynb
t = 2*np.pi/3 Generated by Sphinx-Gallery
plt.plot([t, t], [0, np.cos(t)],
color='blue', linewidth=1.5, linestyle="--")
plt.scatter([t, ], [np.cos(t), ], 50, color='blue') Linewidth
plt.annotate(r'$sin(\frac{2\pi}{3})=\frac{\sqrt{3}}{2}$',
xy=(t, np.sin(t)), xycoords='data', Plot various linewidth with matplotlib.

4.7. Full code examples 152 4.7. Full code examples 153
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

import matplotlib.pyplot as plt import matplotlib.pyplot as plt

size = 256, 16 size = 128, 16

dpi = 72.0 dpi = 72.0
figsize = size[0] / float(dpi), size[1] / float(dpi) figsize= size[0] / float(dpi), size[1] / float(dpi)
fig = plt.figure(figsize=figsize, dpi=dpi) fig = plt.figure(figsize=figsize, dpi=dpi)
fig.patch.set_alpha(0) fig.patch.set_alpha(0)
plt.axes([0, .1, 1, .8], frameon=False)
plt.axes([0, 0, 1, 1], frameon=False)
for i in range(1, 11):
plt.plot([i, i], [0, 1], color='b', lw=i/2.) plt.rcParams['text.antialiased'] = False
plt.text(0.5, 0.5, "Aliased", ha='center', va='center')
plt.xlim(0, 11)
plt.ylim(0, 1) plt.xlim(0, 1)
plt.xticks(()) plt.ylim(0, 1)
plt.yticks(()) plt.xticks(())
plt.yticks(())
plt.show()
plt.show()
Total running time of the script: ( 0 minutes 0.040 seconds)
Total running time of the script: ( 0 minutes 0.032 seconds)
Download Python source code: plot_linewidth.py
Download Python source code: plot_aliased.py
Download Jupyter notebook: plot_linewidth.ipynb
Download Jupyter notebook: plot_aliased.ipynb
Generated by Sphinx-Gallery
Generated by Sphinx-Gallery

Alpha: transparency
Aliased versus anti-aliased
This example demonstrates using alpha for transparency.
The example shows aliased versus anti-aliased text.

import matplotlib.pyplot as plt

import matplotlib.pyplot as plt
size = 256,16
dpi = 72.0 size = 128, 16
figsize= size[0] / float(dpi), size[1] / float(dpi) dpi = 72.0
fig = plt.figure(figsize=figsize, dpi=dpi) figsize= size[0] / float(dpi), size[1] / float(dpi)
fig.patch.set_alpha(0) fig = plt.figure(figsize=figsize, dpi=dpi)
plt.axes([0, 0.1, 1, .8], frameon=False) fig.patch.set_alpha(0)
plt.axes([0, 0, 1, 1], frameon=False)
for i in range(1, 11):
plt.axvline(i, linewidth=1, color='blue', alpha= .25 + .75 * i / 10.) plt.rcParams['text.antialiased'] = True
plt.text(0.5, 0.5, "Anti-aliased", ha='center', va='center')
plt.xlim(0, 11)
plt.xticks(()) plt.xlim(0, 1)
plt.yticks(()) plt.ylim(0, 1)
plt.show() plt.xticks(())
plt.yticks(())
Total running time of the script: ( 0 minutes 0.042 seconds)
plt.show()
Download Python source code: plot_alpha.py
Download Jupyter notebook: plot_alpha.ipynb Total running time of the script: ( 0 minutes 0.032 seconds)

Generated by Sphinx-Gallery Download Python source code: plot_antialiased.py

Download Jupyter notebook: plot_antialiased.ipynb
Aliased versus anti-aliased Generated by Sphinx-Gallery

This example demonstrates aliased versus anti-aliased text.

4.7. Full code examples 154 4.7. Full code examples 155
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Marker size Marker edge color

Demo the marker size control in matplotlib. Demo the marker edge color of matplotlib’s markers.

import matplotlib.pyplot as plt import numpy as np

import matplotlib.pyplot as plt
size = 256, 16
dpi = 72.0 size = 256,16
figsize = size[0] / float(dpi), size[1] / float(dpi) dpi = 72.0
fig = plt.figure(figsize=figsize, dpi=dpi) figsize= size[0] / float(dpi), size[1] / float(dpi)
fig.patch.set_alpha(0) fig = plt.figure(figsize=figsize, dpi=dpi)
plt.axes([0, 0, 1, 1], frameon=False) fig.patch.set_alpha(0)
plt.axes([0, 0, 1, 1], frameon=False)
for i in range(1, 11):
plt.plot([i, ], [1, ], 's', markersize=i, markerfacecolor='w', for i in range(1, 11):
markeredgewidth=.5, markeredgecolor='k') r, g, b = np.random.uniform(0, 1, 3)
plt.plot([i, ], [1, ], 's', markersize=5, markerfacecolor='w',
plt.xlim(0, 11) markeredgewidth=1.5, markeredgecolor=(r, g, b, 1))
plt.xticks(())
plt.yticks(()) plt.xlim(0, 11)
plt.xticks(())
plt.show() plt.yticks(())

Total running time of the script: ( 0 minutes 0.041 seconds) plt.show()

Download Python source code: plot_ms.py Total running time of the script: ( 0 minutes 0.041 seconds)
Download Jupyter notebook: plot_ms.ipynb Download Python source code: plot_mec.py
Generated by Sphinx-Gallery Download Jupyter notebook: plot_mec.ipynb
Generated by Sphinx-Gallery
Marker edge width

Demo the marker edge widths of matplotlib’s markers. Marker face color

Demo the marker face color of matplotlib’s markers.

import matplotlib.pyplot as plt

size = 256, 16 import numpy as np

dpi = 72.0 import matplotlib.pyplot as plt
figsize= size[0] / float(dpi), size[1] / float(dpi)
fig = plt.figure(figsize=figsize, dpi=dpi) size = 256, 16
fig.patch.set_alpha(0) dpi = 72.0
plt.axes([0, 0, 1, 1], frameon=False) figsize = size[0] / float(dpi), size[1] / float(dpi)
fig = plt.figure(figsize=figsize, dpi=dpi)
for i in range(1,11): fig.patch.set_alpha(0)
plt.plot([i, ], [1, ], 's', markersize=5, plt.axes([0, 0, 1, 1], frameon=False)
markeredgewidth=1 + i/10., markeredgecolor='k', markerfacecolor='w')
plt.xlim(0, 11) for i in range(1, 11):
plt.xticks(()) r, g, b = np.random.uniform(0, 1, 3)
plt.yticks(()) plt.plot([i, ], [1, ], 's', markersize=8, markerfacecolor=(r, g, b, 1),
markeredgewidth=.1, markeredgecolor=(0, 0, 0, .5))
plt.show() plt.xlim(0, 11)
plt.xticks(())
Total running time of the script: ( 0 minutes 0.041 seconds) plt.yticks(())
plt.show()
Download Python source code: plot_mew.py
Download Jupyter notebook: plot_mew.ipynb Total running time of the script: ( 0 minutes 0.041 seconds)

Generated by Sphinx-Gallery Download Python source code: plot_mfc.py

4.7. Full code examples 156 4.7. Full code examples 157
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Download Jupyter notebook: plot_mfc.ipynb

import numpy as np
Generated by Sphinx-Gallery import matplotlib.pyplot as plt

size = 256, 16
Colormaps dpi = 72.0
figsize= size[0] / float(dpi), size[1] / float(dpi)
An example plotting the matplotlib colormaps. fig = plt.figure(figsize=figsize, dpi=dpi)
fig.patch.set_alpha(0)
plt.axes([0, 0, 1, 1], frameon=False)

plt.plot(np.arange(4), np.ones(4), color="blue", linewidth=8,

solid_capstyle='butt')

plt.plot(5 + np.arange(4), np.ones(4), color="blue", linewidth=8,

solid_capstyle='round')

plt.plot(10 + np.arange(4), np.ones(4), color="blue", linewidth=8,

solid_capstyle='projecting')

plt.xlim(0, 14)
plt.xticks(())
plt.yticks(())

plt.show()

Total running time of the script: ( 0 minutes 0.034 seconds)

Download Python source code: plot_solid_capstyle.py
Download Jupyter notebook: plot_solid_capstyle.ipynb
import numpy as np
import matplotlib.pyplot as plt Generated by Sphinx-Gallery

plt.rc('text', usetex=False)
a = np.outer(np.arange(0, 1, 0.01), np.ones(10))
Solid joint style

plt.figure(figsize=(10, 5)) An example showing the differen solid joint styles in matplotlib.
plt.subplots_adjust(top=0.8, bottom=0.05, left=0.01, right=0.99)
maps = [m for m in plt.cm.datad if not m.endswith("_r")]
maps.sort()
l = len(maps) + 1 import numpy as np
import matplotlib.pyplot as plt
for i, m in enumerate(maps):
plt.subplot(1, l, i+1) size = 256, 16
plt.axis("off") dpi = 72.0
plt.imshow(a, aspect='auto', cmap=plt.get_cmap(m), origin="lower") figsize = size[0] / float(dpi), size[1] / float(dpi)
plt.title(m, rotation=90, fontsize=10, va='bottom') fig = plt.figure(figsize=figsize, dpi=dpi)
fig.patch.set_alpha(0)
plt.show() plt.axes([0, 0, 1, 1], frameon=False)

Total running time of the script: ( 0 minutes 2.670 seconds) plt.plot(np.arange(3), [0, 1, 0], color="blue", linewidth=8,
solid_joinstyle='miter')
Download Python source code: plot_colormaps.py plt.plot(4 + np.arange(3), [0, 1, 0], color="blue", linewidth=8,
solid_joinstyle='bevel')
Download Jupyter notebook: plot_colormaps.ipynb plt.plot(8 + np.arange(3), [0, 1, 0], color="blue", linewidth=8,
Generated by Sphinx-Gallery solid_joinstyle='round')

plt.xlim(0, 12)
Solid cap style plt.ylim(-1, 2)
plt.xticks(())
An example demoing the solide cap style in matplotlib. plt.yticks(())

plt.show()

4.7. Full code examples 158 4.7. Full code examples 159
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Total running time of the script: ( 0 minutes 0.034 seconds)

Download Python source code: plot_solid_joinstyle.py plt.plot(np.arange(3), [0, 1, 0], color="blue", dashes=[12, 5], linewidth=8,
dash_joinstyle='miter')
Download Jupyter notebook: plot_solid_joinstyle.ipynb plt.plot(4 + np.arange(3), [0, 1, 0], color="blue", dashes=[12, 5],
linewidth=8, dash_joinstyle='bevel')
Generated by Sphinx-Gallery
plt.plot(8 + np.arange(3), [0, 1, 0], color="blue", dashes=[12, 5],
linewidth=8, dash_joinstyle='round')
Dash capstyle
plt.xlim(0, 12)
plt.ylim(-1, 2)
An example demoing the dash capstyle.
plt.xticks(())
plt.yticks(())

plt.show()
import numpy as np
import matplotlib.pyplot as plt
Total running time of the script: ( 0 minutes 0.034 seconds)
size = 256, 16 Download Python source code: plot_dash_joinstyle.py
dpi = 72.0
figsize = size[0] / float(dpi), size[1] / float(dpi) Download Jupyter notebook: plot_dash_joinstyle.ipynb
fig = plt.figure(figsize=figsize, dpi=dpi)
Generated by Sphinx-Gallery
fig.patch.set_alpha(0)
plt.axes([0, 0, 1, 1], frameon=False)
Linestyles
plt.plot(np.arange(4), np.ones(4), color="blue", dashes=[15, 15],
linewidth=8, dash_capstyle='butt')
Plot the different line styles.
plt.plot(5 + np.arange(4), np.ones(4), color="blue", dashes=[15, 15],
linewidth=8, dash_capstyle='round')

plt.plot(10 + np.arange(4), np.ones(4), color="blue", dashes=[15, 15],

linewidth=8, dash_capstyle='projecting')

plt.xlim(0, 14)
plt.xticks(())
plt.yticks(())

plt.show()

Total running time of the script: ( 0 minutes 0.035 seconds)

Download Python source code: plot_dash_capstyle.py
Download Jupyter notebook: plot_dash_capstyle.ipynb
Generated by Sphinx-Gallery

Dash join style

Example demoing the dash join style.

import numpy as np
import numpy as np
import matplotlib.pyplot as plt
import matplotlib.pyplot as plt
def linestyle(ls, i):
size = 256, 16
X = i * .5 * np.ones(11)
dpi = 72.0
Y = np.arange(11)
figsize= size[0] / float(dpi), size[1] / float(dpi)
plt.plot(X, Y, ls, color=(.0, .0, 1, 1), lw=3, ms=8,
fig = plt.figure(figsize=figsize, dpi=dpi)
mfc=(.75, .75, 1, 1), mec=(0, 0, 1, 1))
fig.patch.set_alpha(0)
plt.text(.5 * i, 10.25, ls, rotation=90, fontsize=15, va='bottom')
plt.axes([0, 0, 1, 1], frameon=False)

4.7. Full code examples 160 4.7. Full code examples 161
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

linestyles = ['-', '--', ':', '-.', '.', ',', 'o', '^', 'v', '<', '>', 's', markers = [0, 1, 2, 3, 4, 5, 6, 7, 'o', 'h', '_', '1', '2', '3', '4',
'+', 'x', 'd', '1', '2', '3', '4', 'h', 'p', '|', '_', 'D', 'H'] '8', 'p', '^', 'v', '<', '>', '|', 'd', ',', '+', 's', '*',
n_lines = len(linestyles) '|', 'x', 'D', 'H', '.']

size = 20 * n_lines, 300 n_markers = len(markers)

dpi = 72.0
figsize= size[0] / float(dpi), size[1] / float(dpi) size = 20 * n_markers, 300
fig = plt.figure(figsize=figsize, dpi=dpi) dpi = 72.0
plt.axes([0, 0.01, 1, .9], frameon=False) figsize= size[0] / float(dpi), size[1] / float(dpi)
fig = plt.figure(figsize=figsize, dpi=dpi)
for i, ls in enumerate(linestyles): plt.axes([0, 0.01, 1, .9], frameon=False)
linestyle(ls, i)
for i, m in enumerate(markers):
plt.xlim(-.2, .2 + .5*n_lines) marker(m, i)
plt.xticks(())
plt.yticks(()) plt.xlim(-.2, .2 + .5 * n_markers)
plt.xticks(())
plt.show() plt.yticks(())

Total running time of the script: ( 0 minutes 0.061 seconds) plt.show()

Download Python source code: plot_linestyles.py Total running time of the script: ( 0 minutes 0.070 seconds)
Download Jupyter notebook: plot_linestyles.ipynb Download Python source code: plot_markers.py
Generated by Sphinx-Gallery Download Jupyter notebook: plot_markers.ipynb
Generated by Sphinx-Gallery
Markers

Show the different markers of matplotlib. Locators for tick on axis

An example demoing different locators to position ticks on axis for matplotlib.

import numpy as np
import matplotlib.pyplot as plt

def marker(m, i):

X = i * .5 * np.ones(11)
Y = np.arange(11) import numpy as np
import matplotlib.pyplot as plt
plt.plot(X, Y, lw=1, marker=m, ms=10, mfc=(.75, .75, 1, 1),
mec=(0, 0, 1, 1))
plt.text(.5 * i, 10.25, repr(m), rotation=90, fontsize=15, va='bottom') def tickline():
plt.xlim(0, 10), plt.ylim(-1, 1), plt.yticks([])

4.7. Full code examples 162 4.7. Full code examples 163
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

ax = plt.gca()
ax.spines['right'].set_color('none')
ax.spines['left'].set_color('none')
ax.spines['top'].set_color('none')
ax.xaxis.set_ticks_position('bottom')
ax.spines['bottom'].set_position(('data',0))
ax.yaxis.set_ticks_position('none')
ax.xaxis.set_minor_locator(plt.MultipleLocator(0.1))
ax.plot(np.arange(11), np.zeros(11))
return ax

locators = [
'plt.NullLocator()',
'plt.MultipleLocator(1.0)',
'plt.FixedLocator([0, 2, 8, 9, 10])',
'plt.IndexLocator(3, 1)',
'plt.LinearLocator(5)',
'plt.LogLocator(2, [1.0])',
'plt.AutoLocator()',
]

n_locators = len(locators)

size = 512, 40 * n_locators

dpi = 72.0
figsize = size[0] / float(dpi), size[1] / float(dpi)
fig = plt.figure(figsize=figsize, dpi=dpi)
fig.patch.set_alpha(0)

for i, locator in enumerate(locators):

plt.subplot(n_locators, 1, i + 1) import numpy as np
ax = tickline() import matplotlib.pyplot as plt
ax.xaxis.set_major_locator(eval(locator))
plt.text(5, 0.3, locator[3:], ha='center') plt.subplot(1, 1, 1, polar=True)

plt.subplots_adjust(bottom=.01, top=.99, left=.01, right=.99) N = 20

plt.show() theta = np.arange(0.0, 2 * np.pi, 2 * np.pi / N)
radii = 10 * np.random.rand(N)
Total running time of the script: ( 0 minutes 0.526 seconds) width = np.pi / 4 * np.random.rand(N)
bars = plt.bar(theta, radii, width=width, bottom=0.0)
Download Python source code: plot_ticks.py
for r, bar in zip(radii, bars):
Download Jupyter notebook: plot_ticks.ipynb bar.set_facecolor(plt.cm.jet(r / 10.))
bar.set_alpha(0.5)
Generated by Sphinx-Gallery plt.gca().set_xticklabels([])
plt.gca().set_yticklabels([])

4.7.4 Code generating the summary figures with a title

plt.text(-0.2, 1.02, " Polar Axis \n",
Plotting in polar, decorated horizontalalignment='left',
verticalalignment='top',
size='xx-large',
An example showing how to plot in polar coordinnate, and some decorations.
bbox=dict(facecolor='white', alpha=1.0),
transform=plt.gca().transAxes)

plt.text(-0.2, 1.01, "\n\n Plot anything using polar axis ",

horizontalalignment='left',
verticalalignment='top',
size='large',
transform=plt.gca().transAxes)

plt.show()

4.7. Full code examples 164 4.7. Full code examples 165
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Total running time of the script: ( 0 minutes 0.049 seconds)

ax.text2D(0.05, .93, " 3D plots \n",
Download Python source code: plot_polar_ext.py horizontalalignment='left',
verticalalignment='top',
Download Jupyter notebook: plot_polar_ext.ipynb size='xx-large',
bbox=dict(facecolor='white', alpha=1.0),
Generated by Sphinx-Gallery
transform=plt.gca().transAxes)

3D plotting vignette ax.text2D(0.05, .87, " Plot 2D or 3D data",

horizontalalignment='left',
verticalalignment='top',
Demo 3D plotting with matplotlib and decorate the figure.
size='large',
transform=plt.gca().transAxes)

plt.show()

Total running time of the script: ( 0 minutes 0.066 seconds)

Download Python source code: plot_plot3d_ext.py
Download Jupyter notebook: plot_plot3d_ext.ipynb
Generated by Sphinx-Gallery

Plot example vignette

An example of plots with matplotlib, and added annotations.

import numpy as np
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D

fig = plt.figure()
ax = Axes3D(fig)
X = np.arange(-4, 4, 0.25)
Y = np.arange(-4, 4, 0.25)
X, Y = np.meshgrid(X, Y)
R = np.sqrt(X ** 2 + Y ** 2)
Z = np.sin(R)

ax.plot_surface(X, Y, Z, rstride=1, cstride=1, cmap=plt.cm.hot)

ax.contourf(X, Y, Z, zdir='z', offset=-2, cmap=plt.cm.hot)
ax.set_zlim(-2, 2)
plt.xticks(())
import numpy as np
plt.yticks(())
import matplotlib.pyplot as plt
ax.set_zticks(())
n = 256

4.7. Full code examples 166 4.7. Full code examples 167
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

X = np.linspace(0, 2, n)
Y = np.sin(2 * np.pi * X)

plt.plot (X, Y, lw=2, color='violet')

plt.xlim(-0.2, 2.2)
plt.xticks(())
plt.ylim(-1.2, 1.2)
plt.yticks(())

# Add a title and a box around it

from matplotlib.patches import FancyBboxPatch
ax = plt.gca()
ax.add_patch(FancyBboxPatch((-0.05, .87),
width=.66, height=.165, clip_on=False,
boxstyle="square,pad=0", zorder=3,
facecolor='white', alpha=1.0,
transform=plt.gca().transAxes))

plt.text(-0.05, 1.02, " Regular Plot: plt.plot(...)\n",

horizontalalignment='left',
verticalalignment='top',
size='xx-large',
transform=plt.gca().transAxes)

plt.text(-0.05, 1.01, "\n\n Plot lines and/or markers ",

horizontalalignment='left',
verticalalignment='top',
size='large',
transform=plt.gca().transAxes)

plt.show() import matplotlib.pyplot as plt

Total running time of the script: ( 0 minutes 0.036 seconds) ax = plt.subplot(2, 1, 1)

ax.set_xticklabels([])
Download Python source code: plot_plot_ext.py
ax.set_yticklabels([])
Download Jupyter notebook: plot_plot_ext.ipynb
Generated by Sphinx-Gallery # Add a title and a box around it
from matplotlib.patches import FancyBboxPatch
ax = plt.gca()
Multiple plots vignette ax.add_patch(FancyBboxPatch((-0.05, .72),
width=.66, height=.34, clip_on=False,
Demo multiple plots and style the figure. boxstyle="square,pad=0", zorder=3,
facecolor='white', alpha=1.0,
transform=plt.gca().transAxes))

plt.text(-0.05, 1.02, " Multiplot: plt.subplot(...)\n",

horizontalalignment='left',
verticalalignment='top',
size='xx-large',
transform=ax.transAxes)
plt.text(-0.05, 1.01, "\n\n Plot several plots at once ",
horizontalalignment='left',
verticalalignment='top',
size='large',
transform=ax.transAxes)

ax = plt.subplot(2, 2, 3)
ax.set_xticklabels([])
ax.set_yticklabels([])

4.7. Full code examples 168 4.7. Full code examples 169
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

ax = plt.subplot(2, 2, 4) from matplotlib.patches import FancyBboxPatch

ax.set_xticklabels([]) ax = plt.gca()
ax.set_yticklabels([]) ax.add_patch(FancyBboxPatch((-0.05, .87),
width=.66, height=.165, clip_on=False,
plt.show() boxstyle="square,pad=0", zorder=3,
facecolor='white', alpha=1.0,
Total running time of the script: ( 0 minutes 0.125 seconds) transform=plt.gca().transAxes))

Download Python source code: plot_multiplot_ext.py plt.text(-0.05, 1.02, " Box Plot: plt.boxplot(...)\n ",
horizontalalignment='left',
Download Jupyter notebook: plot_multiplot_ext.ipynb
verticalalignment='top',
Generated by Sphinx-Gallery size='xx-large',
transform=axes.transAxes)

Boxplot with matplotlib plt.text(-0.04, .98, "\n Make a box and whisker plot ",
horizontalalignment='left',
An example of doing box plots with matplotlib verticalalignment='top',
size='large',
transform=axes.transAxes)

plt.show()

Total running time of the script: ( 0 minutes 0.059 seconds)

Download Python source code: plot_boxplot_ext.py
Download Jupyter notebook: plot_boxplot_ext.ipynb
Generated by Sphinx-Gallery

Plot scatter decorated

An example showing the scatter function, with decorations.

import numpy as np
import matplotlib.pyplot as plt

fig = plt.figure(figsize=(8, 5))

axes = plt.subplot(111)

n = 5
Z = np.zeros((n, 4))
X = np.linspace(0, 2, n, endpoint=True)
Y = np.random.random((n, 4))
plt.boxplot(Y)

plt.xticks(())
plt.yticks(())

# Add a title and a box around it

4.7. Full code examples 170 4.7. Full code examples 171
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

plt.text(-0.05, 1.01, "\n\n Make a scatter plot of x versus y ",

horizontalalignment='left',
verticalalignment='top',
size='large',
transform=plt.gca().transAxes)

plt.show()

Total running time of the script: ( 0 minutes 0.040 seconds)

Download Python source code: plot_scatter_ext.py
Download Jupyter notebook: plot_scatter_ext.ipynb
Generated by Sphinx-Gallery

Pie chart vignette

Demo pie chart with matplotlib and style the figure.

import numpy as np
import matplotlib.pyplot as plt

n = 1024
X = np.random.normal(0, 1, n)
Y = np.random.normal(0, 1, n)

T = np.arctan2(Y,X)

plt.scatter(X, Y, s=75, c=T, alpha=.5)

plt.xlim(-1.5, 1.5)
plt.xticks(())
plt.ylim(-1.5, 1.5)
plt.yticks(())

# Add a title and a box around it

from matplotlib.patches import FancyBboxPatch
ax = plt.gca()
ax.add_patch(FancyBboxPatch((-0.05, .87),
width=.66, height=.165, clip_on=False, import numpy as np
boxstyle="square,pad=0", zorder=3, import matplotlib.pyplot as plt
facecolor='white', alpha=1.0,
transform=plt.gca().transAxes)) n = 20
X = np.ones(n)
plt.text(-0.05, 1.02, " Scatter Plot: plt.scatter(...)\n", X[-1] *= 2
horizontalalignment='left', plt.pie(X, explode=X*.05, colors = ['%f ' % (i/float(n)) for i in range(n)])
verticalalignment='top',
size='xx-large', fig = plt.gcf()
transform=plt.gca().transAxes) w, h = fig.get_figwidth(), fig.get_figheight()
r = h / float(w)

4.7. Full code examples 172 4.7. Full code examples 173
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

plt.xlim(-1.5, 1.5)
plt.ylim(-1.5 * r, 1.5 * r)
plt.xticks(())
plt.yticks(())

# Add a title and a box around it

from matplotlib.patches import FancyBboxPatch
ax = plt.gca()
ax.add_patch(FancyBboxPatch((-0.05, .87),
width=.66, height=.165, clip_on=False,
boxstyle="square,pad=0", zorder=3,
facecolor='white', alpha=1.0,
transform=plt.gca().transAxes))

plt.text(-0.05, 1.02, " Pie Chart: plt.pie(...)\n",

horizontalalignment='left',
verticalalignment='top',
size='xx-large',
transform=plt.gca().transAxes)

plt.text(-0.05, 1.01, "\n\n Make a pie chart of an array ",

horizontalalignment='left',
verticalalignment='top',
size='large',
transform=plt.gca().transAxes)

plt.show()

Total running time of the script: ( 0 minutes 0.049 seconds)

import numpy as np
Download Python source code: plot_pie_ext.py
import matplotlib.pyplot as plt
Download Jupyter notebook: plot_pie_ext.ipynb
n = 16
Generated by Sphinx-Gallery X = np.arange(n)
Y1 = (1 - X / float(n)) * np.random.uniform(0.5, 1.0, n)
Y2 = (1 - X / float(n)) * np.random.uniform(0.5, 1.0, n)
Bar plot advanced plt.bar(X, Y1, facecolor='#9999ff', edgecolor='white')
plt.bar(X, -Y2, facecolor='#ff9999', edgecolor='white')
An more elaborate bar plot example plt.xlim(-.5, n)
plt.xticks(())
plt.ylim(-1, 1)
plt.yticks(())

# Add a title and a box around it

from matplotlib.patches import FancyBboxPatch
ax = plt.gca()
ax.add_patch(FancyBboxPatch((-0.05, .87),
width=.66, height=.165, clip_on=False,
boxstyle="square,pad=0", zorder=3,
facecolor='white', alpha=1.0,
transform=plt.gca().transAxes))

plt.text(-0.05, 1.02, " Bar Plot: plt.bar(...)\n",

horizontalalignment='left',
verticalalignment='top',
size='xx-large',
transform=plt.gca().transAxes)

plt.text(-0.05, 1.01, "\n\n Make a bar plot with rectangles ",

4.7. Full code examples 174 4.7. Full code examples 175
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

horizontalalignment='left', plt.xlim(-1, n)
verticalalignment='top', plt.xticks(())
size='large', plt.ylim(-1, n)
transform=plt.gca().transAxes) plt.yticks(())

plt.show()
# Add a title and a box around it
Total running time of the script: ( 0 minutes 0.056 seconds) from matplotlib.patches import FancyBboxPatch
ax = plt.gca()
Download Python source code: plot_bar_ext.py ax.add_patch(FancyBboxPatch((-0.05, .87),
width=.66, height=.165, clip_on=False,
Download Jupyter notebook: plot_bar_ext.ipynb
boxstyle="square,pad=0", zorder=3,
Generated by Sphinx-Gallery facecolor='white', alpha=1.0,
transform=plt.gca().transAxes))

Plotting quiver decorated plt.text(-0.05, 1.02, " Quiver Plot: plt.quiver(...)\n",

horizontalalignment='left',
An example showing quiver with decorations. verticalalignment='top',
size='xx-large',
transform=plt.gca().transAxes)

plt.text(-0.05, 1.01, "\n\n Plot a 2-D field of arrows ",

horizontalalignment='left',
verticalalignment='top',
size='large',
transform=plt.gca().transAxes)

plt.show()

Total running time of the script: ( 0 minutes 0.083 seconds)

Download Python source code: plot_quiver_ext.py
Download Jupyter notebook: plot_quiver_ext.ipynb
Generated by Sphinx-Gallery

Imshow demo

Demoing imshow

import numpy as np
import matplotlib.pyplot as plt

n = 8
X, Y = np.mgrid[0:n, 0:n]
T = np.arctan2(Y - n/ 2., X - n / 2.)
R = 10 + np.sqrt((Y - n / 2.) ** 2 + (X - n / 2.) ** 2)
U, V = R * np.cos(T), R * np.sin(T)

plt.quiver(X, Y, U, V, R, alpha=.5)
plt.quiver(X, Y, U, V, edgecolor='k', facecolor='None', linewidth=.5)

4.7. Full code examples 176 4.7. Full code examples 177
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

plt.text(-0.05, 1.01, "\n\n Display an image to current axes ",

horizontalalignment='left',
verticalalignment='top',
family='Lint McCree Intl BB',
size='large',
transform=plt.gca().transAxes)

plt.show()

Total running time of the script: ( 0 minutes 0.150 seconds)

Download Python source code: plot_imshow_ext.py
Download Jupyter notebook: plot_imshow_ext.ipynb
Generated by Sphinx-Gallery

Display the contours of a function

An example demoing how to plot the contours of a function, with additional layout tweeks.

import numpy as np
import matplotlib.pyplot as plt

def f(x, y):

return (1 - x / 2 + x ** 5 + y ** 3) * np.exp(-x ** 2 - y ** 2)

n = 10
x = np.linspace(-3, 3, 8 * n)
y = np.linspace(-3, 3, 6 * n)
X, Y = np.meshgrid(x, y)
Z = f(X, Y)
plt.imshow(Z, interpolation='nearest', cmap='bone', origin='lower')
plt.xticks(())
plt.yticks(())

# Add a title and a box around it

from matplotlib.patches import FancyBboxPatch
ax = plt.gca()
ax.add_patch(FancyBboxPatch((-0.05, .87),
width=.66, height=.165, clip_on=False,
boxstyle="square,pad=0", zorder=3, import numpy as np
facecolor='white', alpha=1.0, import matplotlib.pyplot as plt
transform=plt.gca().transAxes))
def f(x,y):
plt.text(-0.05, 1.02, " Imshow: plt.imshow(...)\n", return (1 - x / 2 + x ** 5 + y ** 3) * np.exp(-x ** 2 - y ** 2)
horizontalalignment='left',
verticalalignment='top', n = 256
size='xx-large', x = np.linspace(-3, 3, n)
transform=plt.gca().transAxes) y = np.linspace(-3, 3, n)
X, Y = np.meshgrid(x, y)

4.7. Full code examples 178 4.7. Full code examples 179
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

plt.contourf(X, Y, f(X, Y), 8, alpha=.75, cmap=plt.cm.hot)

C = plt.contour(X, Y, f(X,Y), 8, colors='black', linewidth=.5)
plt.clabel(C, inline=1, fontsize=10)
plt.xticks(())
plt.yticks(())

# Add a title and a box around it

from matplotlib.patches import FancyBboxPatch
ax = plt.gca()
ax.add_patch(FancyBboxPatch((-0.05, .87),
width=.66, height=.165, clip_on=False,
boxstyle="square,pad=0", zorder=3,
facecolor='white', alpha=1.0,
transform=plt.gca().transAxes))

plt.text(-0.05, 1.02, " Contour Plot: plt.contour(..)\n",

horizontalalignment='left',
verticalalignment='top',
size='xx-large',
transform=plt.gca().transAxes)

plt.text(-0.05, 1.01, "\n\n Draw contour lines and filled contours ",
horizontalalignment='left',
verticalalignment='top',
size='large',
transform=plt.gca().transAxes)

plt.show()

Total running time of the script: ( 0 minutes 0.092 seconds) import matplotlib.pyplot as plt
from matplotlib.ticker import MultipleLocator
Download Python source code: plot_contour_ext.py
Download Jupyter notebook: plot_contour_ext.ipynb fig = plt.figure(figsize=(8, 6), dpi=72, facecolor="white")
axes = plt.subplot(111)
Generated by Sphinx-Gallery axes.set_xlim(0, 4)
axes.set_ylim(0, 3)

Grid elaborate axes.xaxis.set_major_locator(MultipleLocator(1.0))

axes.xaxis.set_minor_locator(MultipleLocator(0.1))
An example displaying a grid on the axes and tweaking the layout. axes.yaxis.set_major_locator(MultipleLocator(1.0))
axes.yaxis.set_minor_locator(MultipleLocator(0.1))
axes.grid(which='major', axis='x', linewidth=0.75, linestyle='-', color='0.75')
axes.grid(which='minor', axis='x', linewidth=0.25, linestyle='-', color='0.75')
axes.grid(which='major', axis='y', linewidth=0.75, linestyle='-', color='0.75')
axes.grid(which='minor', axis='y', linewidth=0.25, linestyle='-', color='0.75')
axes.set_xticklabels([])
axes.set_yticklabels([])

# Add a title and a box around it

from matplotlib.patches import FancyBboxPatch
ax = plt.gca()
ax.add_patch(FancyBboxPatch((-0.05, .87),
width=.66, height=.165, clip_on=False,
boxstyle="square,pad=0", zorder=3,
facecolor='white', alpha=1.0,
transform=plt.gca().transAxes))

plt.text(-0.05, 1.02, " Grid: plt.grid(...)\n",

horizontalalignment='left',

4.7. Full code examples 180 4.7. Full code examples 181
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

verticalalignment='top', eqs.append((r"$\frac{d\rho}{d t} + \rho \vec{v}\cdot\nabla\vec{v} = -\nabla p + \mu\nabla^2␣

size='xx-large', ,→\vec{v} + \rho \vec{g}$"))
transform=axes.transAxes) eqs.append((r"$\int_{-\infty}^\infty e^{-x^2}dx=\sqrt{\pi}$"))
eqs.append((r"$E = mc^2 = \sqrt{{m_0}^2c^4 + p^2c^2}$"))
plt.text(-0.05, 1.01, "\n\n Draw ticks and grid ", eqs.append((r"$F_G = G\frac{m_1m_2}{r^2}$"))
horizontalalignment='left',
verticalalignment='top', for i in range(24):
size='large', index = np.random.randint(0,len(eqs))
transform=axes.transAxes) eq = eqs[index]
size = np.random.uniform(12,32)
Total running time of the script: ( 0 minutes 0.044 seconds) x,y = np.random.uniform(0,1,2)
alpha = np.random.uniform(0.25,.75)
Download Python source code: plot_grid_ext.py plt.text(x, y, eq, ha='center', va='center', color="#11557c", alpha=alpha,
transform=plt.gca().transAxes, fontsize=size, clip_on=True)
Download Jupyter notebook: plot_grid_ext.ipynb
Generated by Sphinx-Gallery
# Add a title and a box around it
from matplotlib.patches import FancyBboxPatch
Text printing decorated ax = plt.gca()
ax.add_patch(FancyBboxPatch((-0.05, .87),
An example showing text printing and decorating the resulting figure. width=.66, height=.165, clip_on=False,
boxstyle="square,pad=0", zorder=3,
facecolor='white', alpha=1.0,
transform=plt.gca().transAxes))

plt.text(-0.05, 1.02, " Text: plt.text(...)\n",

horizontalalignment='left',
verticalalignment='top',
size='xx-large',
transform=plt.gca().transAxes)

plt.text(-0.05, 1.01, "\n\n Draw any kind of text ",

horizontalalignment='left',
verticalalignment='top',
size='large',
transform=plt.gca().transAxes)

plt.show()

Total running time of the script: ( 0 minutes 0.039 seconds)

Download Python source code: plot_text_ext.py
Download Jupyter notebook: plot_text_ext.ipynb
Generated by Sphinx-Gallery
Download all examples in Python source code: auto_examples_python.zip
Download all examples in Jupyter notebooks: auto_examples_jupyter.zip
Generated by Sphinx-Gallery

import numpy as np
import matplotlib.pyplot as plt

fig = plt.figure()
plt.xticks(())
plt.yticks(())

eqs = []
eqs.append((r"$W^{3\beta}_{\delta_1 \rho_1 \sigma_2} = U^{3\beta}_{\delta_1 \rho_1} + \frac{1}
,→{8 \pi 2} \int^{\alpha_2}_{\alpha_2} d \alpha^\prime_2 \left[\frac{ U^{2\beta}_{\delta_1␣
,→\rho_1} - \alpha^\prime_2U^{1\beta}_{\rho_1 \sigma_2} }{U^{0\beta}_{\rho_1 \sigma_2}}\right]$
,→"))
4.7. Full code examples 182 4.7. Full code examples 183
 Scipy lecture notes, Edition 2017.1

• Optimization and fit: scipy.optimize

• Statistics and random numbers: scipy.stats
• Numerical integration: scipy.integrate
• Fast Fourier transforms: scipy.fftpack

5
• Signal processing: scipy.signal
• Image manipulation: scipy.ndimage
• Summary exercises on scientific computing

CHAPTER
• Full code examples for the scipy chapter

Warning: This tutorial is far from an introduction to numerical computing. As enumerating the different
submodules and functions in scipy would be very boring, we concentrate instead on a few examples to give
a general idea of how to use scipy for scientific computing.

scipy is composed of task-specific sub-modules:

Scipy : high-level scientific computing scipy.cluster Vector quantization / Kmeans

scipy.constants Physical and mathematical constants
scipy.fftpack Fourier transform
scipy.integrate Integration routines
scipy.interpolate Interpolation
scipy.io Data input and output
scipy.linalg Linear algebra routines
scipy.ndimage n-dimensional image package
Authors: Gaël Varoquaux, Adrien Chauve, Andre Espaze, Emmanuelle Gouillart, Ralf Gommers
scipy.odr Orthogonal distance regression
scipy.optimize Optimization
Scipy scipy.signal Signal processing
scipy.sparse Sparse matrices
The scipy package contains various toolboxes dedicated to common issues in scientific computing. Its scipy.spatial Spatial data structures and algorithms
different submodules correspond to different applications, such as interpolation, integration, optimization, scipy.special Any special mathematical functions
image processing, statistics, special functions, etc. scipy.stats Statistics

Tip: scipy can be compared to other standard scientific-computing libraries, such as the GSL (GNU Scientific Tip: They all depend on numpy, but are mostly independent of each other. The standard way of importing
Library for C and C++), or Matlab’s toolboxes. scipy is the core package for scientific routines in Python; it is Numpy and these Scipy modules is:
meant to operate efficiently on numpy arrays, so that numpy and scipy work hand in hand.
>>> import numpy as np
Before implementing a routine, it is worth checking if the desired data processing is not already implemented >>> from scipy import stats # same for other sub-modules
in Scipy. As non-professional programmers, scientists often tend to re-invent the wheel, which leads to buggy,
non-optimal, difficult-to-share and unmaintainable code. By contrast, Scipy’s routines are optimized and The main scipy namespace mostly contains functions that are really numpy functions (try scipy.cos is
tested, and should therefore be used when possible. np.cos). Those are exposed for historical reasons; there’s no reason to use import scipy in your code.

Chapters contents
5.1 File input/output: scipy.io
• File input/output: scipy.io
• Special functions: scipy.special Matlab files: Loading and saving:

• Linear algebra operations: scipy.linalg >>> from scipy import io as spio

>>> a = np.ones((3, 3))
• Interpolation: scipy.interpolate
>>> spio.savemat('file.mat', {'a': a}) # savemat expects a dictionary
>>> data = spio.loadmat('file.mat')

184 5.1. File input/output: scipy.io 185

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

>>> data['a']
array([[ 1., 1., 1.], Tip: The scipy.linalg module provides standard linear algebra operations, relying on an underlying effi-
[ 1., 1., 1.], cient implementation (BLAS, LAPACK).
[ 1., 1., 1.]])
• The scipy.linalg.det() function computes the determinant of a square matrix:

Warning: Python / Matlab mismatches, eg matlab does not represent 1D arrays >>> from scipy import linalg
>>> arr = np.array([[1, 2],
>>> a = np.ones(3) ... [3, 4]])
>>> a >>> linalg.det(arr)
array([ 1., 1., 1.]) -2.0
>>> spio.savemat('file.mat', {'a': a}) >>> arr = np.array([[3, 2],
>>> spio.loadmat('file.mat')['a'] ... [6, 4]])
array([[ 1., 1., 1.]]) >>> linalg.det(arr)
0.0
Notice the difference? >>> linalg.det(np.ones((3, 4)))
Traceback (most recent call last):
...
ValueError: expected square matrix

• The scipy.linalg.inv() function computes the inverse of a square matrix:

Image files: Reading images: >>> arr = np.array([[1, 2],
... [3, 4]])
>>> from scipy import misc >>> iarr = linalg.inv(arr)
>>> misc.imread('fname.png') >>> iarr
array(...) array([[-2. , 1. ],
>>> # Matplotlib also has a similar function [ 1.5, -0.5]])
>>> import matplotlib.pyplot as plt >>> np.allclose(np.dot(arr, iarr), np.eye(2))
>>> plt.imread('fname.png') True
array(...)
Finally computing the inverse of a singular matrix (its determinant is zero) will raise LinAlgError:
See also:
>>> arr = np.array([[3, 2],
• Load text files: numpy.loadtxt()/numpy.savetxt()
... [6, 4]])
• Clever loading of text/csv files: numpy.genfromtxt()/numpy.recfromcsv() >>> linalg.inv(arr)
Traceback (most recent call last):
• Fast and efficient, but numpy-specific, binary format: numpy.save()/numpy.load() ...
• More advanced input/output of images in scikit-image: skimage.io ...LinAlgError: singular matrix

• More advanced operations are available, for example singular-value decomposition (SVD):

5.2 Special functions: scipy.special >>> arr = np.arange(9).reshape((3, 3)) + np.diag([1, 0, 1])
>>> uarr, spec, vharr = linalg.svd(arr)
Special functions are transcendental functions. The docstring of the scipy.special module is well-written,
so we won’t list all functions here. Frequently used ones are: The resulting array spectrum is:

• Bessel function, such as scipy.special.jn() (nth integer order Bessel function) >>> spec
array([ 14.88982544, 0.45294236, 0.29654967])
• Elliptic function (scipy.special.ellipj() for the Jacobian elliptic function, . . . )
• Gamma function: scipy.special.gamma(), also note scipy.special.gammaln() which will give the The original matrix can be re-composed by matrix multiplication of the outputs of svd with np.dot:
log of Gamma to a higher numerical precision.
>>> sarr = np.diag(spec)
• Erf, the area under a Gaussian curve: scipy.special.erf() >>> svd_mat = uarr.dot(sarr).dot(vharr)
>>> np.allclose(svd_mat, arr)
True

5.3 Linear algebra operations: scipy.linalg SVD is commonly used in statistics and signal processing. Many other standard decompositions (QR,
LU, Cholesky, Schur), as well as solvers for linear systems, are available in scipy.linalg.

5.2. Special functions: scipy.special 186 5.3. Linear algebra operations: scipy.linalg 187
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

5.4 Interpolation: scipy.interpolate 5.5.1 Curve fitting

scipy.interpolate is useful for fitting a function from experimental data and thus evaluating points where
no measure exists. The module is based on the FITPACK Fortran subroutines.
By imagining experimental data close to a sine function:

>>> measured_time = np.linspace(0, 1, 10)

>>> noise = (np.random.random(10)*2 - 1) * 1e-1
>>> measures = np.sin(2 * np.pi * measured_time) + noise

scipy.interpolate.interp1d can build a linear interpolation function:

>>> from scipy.interpolate import interp1d

>>> linear_interp = interp1d(measured_time, measures) Suppose we have data on a sine wave, with some noise:

>>> x_data = np.linspace(-5, 5, num=50)

>>> y_data = 2.9 * np.sin(1.5 * x_data) + np.random.normal(size=50)

If we know that the data lies on a sine wave, but not the amplitudes or the period, we can find those by least
squares curve fitting. First we have to define the test function to fit, here a sine with unknown amplitude and
period:

>>> def test_func(x, a, b):

... return a * np.sin(b * x)

Then the result can be evaluated at the time of

interest:

>>> interpolation_time = np.linspace(0, 1, 50)

>>> linear_results = linear_interp(interpolation_time)

A cubic interpolation can also be selected by providing the kind optional keyword argument:

>>> cubic_interp = interp1d(measured_time, measures, kind='cubic') We then use scipy.optimize.curve_fit() to find a

>>> cubic_results = cubic_interp(interpolation_time) and b:

>>> params, params_covariance = optimize.curve_fit(test_func, x_data, y_data, p0=[2, 2])

scipy.interpolate.interp2d is similar to scipy.interpolate.interp1d, but for 2-D arrays. Note that >>> print(params)
for the interp family, the interpolation points must stay within the range of given data points. See the summary [ 3.05931973 1.45754553]
exercise on Maximum wind speed prediction at the Sprogø station for a more advanced spline interpolation
example.
Exercise: Curve fitting of temperature data

5.5 Optimization and fit: scipy.optimize The temperature extremes in Alaska for each month, starting in January, are given by (in degrees
Celcius):
Optimization is the problem of finding a numerical solution to a minimization or equality. max: 17, 19, 21, 28, 33, 38, 37, 37, 31, 23, 19, 18
min: -62, -59, -56, -46, -32, -18, -9, -13, -25, -46, -52, -58

Tip: The scipy.optimize module provides algorithms for function minimization (scalar or multi- 1. Plot these temperature extremes.
dimensional), curve fitting and root finding.
2. Define a function that can describe min and max temperatures. Hint: this function has to
>>> from scipy import optimize have a period of 1 year. Hint: include a time offset.
3. Fit this function to the data with scipy.optimize.curve_fit().
4. Plot the result. Is the fit reasonable? If not, why?

5.4. Interpolation: scipy.interpolate 188 5.5. Optimization and fit: scipy.optimize 189
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

5. Is the time offset for min and max temperatures the same within the fit accuracy? >>> result.x # The coordinate of the minimum
solution array([-1.30644...])

5.5.2 Finding the minimum of a scalar function

Methods: As the function is a smooth function, gradient-descent based methods are good options. The lBFGS
algorithm is a good choice in general:

>>> optimize.minimize(f, x0=0, method="L-BFGS-B")

fun: array([-7.94582338])
hess_inv: <1x1 LbfgsInvHessProduct with dtype=float64>
jac: array([ -1.42108547e-06])
message: ...'CONVERGENCE: NORM_OF_PROJECTED_GRADIENT_<=_PGTOL'
nfev: 12
nit: 5
status: 0
success: True
x: array([-1.30644013])

Note how it cost only 12 functions evaluation above to find a good value for the minimum.

Let’s define the following function:

>>> def f(x): Global minimum: A possible issue with this approach is that, if the function has local minima, the algorithm
... return x**2 + 10*np.sin(x)
may find these local minima instead of the global minimum depending on the initial point x0:

and plot it: >>> res = optimize.minimize(f, x0=3, method="L-BFGS-B")

>>> res.x
>>> x = np.arange(-10, 10, 0.1) array([ 3.83746709])
>>> plt.plot(x, f(x))
>>> plt.show()
If we don’t know the neighborhood of the global minimum to choose the initial point, we need to resort to
costlier global optimization. To find the global minimum, we use scipy.optimize.basinhopping() (added
This function has a global minimum around -1.3 and a local minimum around 3.8.
in version 0.12.0 of Scipy). It combines a local optimizer with sampling of starting points:
Searching for minimum can be done with scipy.optimize.minimize(), given a starting point x0, it returns
>>> optimize.basinhopping(f, 0)
the location of the minimum that it has found:
nfev: 1725
minimization_failures: 0
result type fun: -7.9458233756152845
x: array([-1.30644001])
message: ['requested number of basinhopping iterations completed successfully']
The result of scipy.optimize.minimize() is a compound object comprising all information on the con-
njev: 575
vergence
nit: 100

>>> result = optimize.minimize(f, x0=0)

>>> result Note: scipy used to contain the routine anneal, it has been removed in SciPy 0.16.0.
fun: -7.9458233756...
hess_inv: array([[ 0.0858...]])
jac: array([ -1.19209...e-06])
message: 'Optimization terminated successfully.'
nfev: 18
nit: 5
njev: 6
Constraints: We can constrain the variable to the interval (0, 10) using the “bounds” argument:
status: 0
success: True
x: array([-1.30644...])

5.5. Optimization and fit: scipy.optimize 190 5.5. Optimization and fit: scipy.optimize 191
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

A list of bounds • Use numpy.meshgrid() and pylab.imshow() to find visually the regions.
• Use scipy.optimize.minimize(), optionally trying out several of its ‘methods’.
As minimize() works in general with x multidimensionsal, the “bounds” argument is a list of bound on
each dimension. How many global minima are there, and what is the function value at those points? What hap-
pens for an initial guess of (x, y) = (0, 0) ?
>>> res = optimize.minimize(f, x0=1, solution
... bounds=((0, 10), ))
>>> res.x
array([ 0.])
5.5.3 Finding the roots of a scalar function

Tip: What has happened? Why are we finding 0, which is not a mimimum of our function. To find a root, i.e. a point where f (x) = 0, of the function f above we can use scipy.optimize.root():

>>> root = optimize.root(f, x0=1) # our initial guess is 1

>>> root # The full result
Minimizing functions of several variables fjac: array([[-1.]])
fun: array([ 0.])
To minimize over several variables, the trick is to turn them into a function of a multi-dimensional variable message: 'The solution converged.'
(a vector). See for instance the exercise on 2D minimization below. nfev: 10
qtf: array([ 1.33310463e-32])
r: array([-10.])
status: 1
Note: scipy.optimize.minimize_scalar() is a function with dedicated methods to minimize functions success: True
of only one variable. x: array([ 0.])
>>> root.x # Only the root found
array([ 0.])
See also:
Finding minima of function is discussed in more details in the advanced chapter: Mathematical optimization: Note that only one root is found. Inspecting the plot of f reveals that there is a second root around -2.5. We
finding minima of functions. find the exact value of it by adjusting our initial guess:

>>> root2 = optimize.root(f, x0=-2.5)

Exercise: 2-D minimization >>> root2.x
array([-2.47948183])

Note: scipy.optimize.root() also comes with a variety of algorithms, set via the “method” argument.

The six-hump camel-

back function
x4 2
f (x, y) = (4 − 2.1x 2 + )x + x y + (4y 2 − 4)y 2
3
has multiple global and local minima. Find the global minima of this function.
Hints: Now that we have found the minima and roots of f and used curve fitting on it, we put all those results together
in a single plot:
• Variables can be restricted to −2 < x < 2 and −1 < y < 1.
See also:

5.5. Optimization and fit: scipy.optimize 192 5.5. Optimization and fit: scipy.optimize 193
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

You can find all algorithms and functions with similar functionalities in the documentation of scipy. If we know that the random process belongs to a given family of random processes, such as normal processes,
optimize. we can do a maximum-likelihood fit of the observations to estimate the parameters of the underlying distri-
bution. Here we fit a normal process to the observed data:
See the summary exercise on Non linear least squares curve fitting: application to point extraction in topo-
graphical lidar data for another, more advanced example. >>> loc, std = stats.norm.fit(samples)
>>> loc
-0.045256707...
5.6 Statistics and random numbers: scipy.stats >>> std
0.9870331586...

The module scipy.stats contains statistical tools and probabilistic descriptions of random processes. Ran-
dom number generators for various random process can be found in numpy.random. Exercise: Probability distributions

Generate 1000 random variates from a gamma distribution with a shape parameter of 1, then plot a his-
5.6.1 Distributions: histogram and probability density function togram from those samples. Can you plot the pdf on top (it should match)?

Given observations of a random process, their histogram is an estimator of the random process’s PDF (proba- Extra: the distributions have many useful methods. Explore them by reading the docstring or by using tab
bility density function): completion. Can you recover the shape parameter 1 by using the fit method on your random variates?

>>> samples = np.random.normal(size=1000)

>>> bins = np.arange(-4, 5)
>>> bins 5.6.2 Mean, median and percentiles
array([-4, -3, -2, -1, 0, 1, 2, 3, 4])
>>> histogram = np.histogram(samples, bins=bins, normed=True)[0] The mean is an estimator of the center of the distribution:
>>> bins = 0.5*(bins[1:] + bins[:-1])
>>> bins >>> np.mean(samples)
array([-3.5, -2.5, -1.5, -0.5, 0.5, 1.5, 2.5, 3.5]) -0.0452567074...
>>> from scipy import stats
>>> pdf = stats.norm.pdf(bins) # norm is a distribution object
The median another estimator of the center. It is the value with half of the observations below, and half above:
>>> plt.plot(bins, histogram) >>> np.median(samples)
[<matplotlib.lines.Line2D object at ...>] -0.0580280347...
>>> plt.plot(bins, pdf)
[<matplotlib.lines.Line2D object at ...>]

Tip: Unlike the mean, the median is not sensitive to the tails of the distribution. It is “robust”.

Exercise: Compare mean and median on samples of a Gamma distribution

Which one seems to be the best estimator of the center for the Gamma distribution?

The median is also the percentile 50, because 50% of the observation are below it:

>>> stats.scoreatpercentile(samples, 50)

-0.0580280347...

Similarly, we can calculate the percentile 90:

>>> stats.scoreatpercentile(samples, 90)

1.2315935511...

Tip: The percentile is an estimator of the CDF: cumulative distribution function.

The distribution objects

scipy.stats.norm is a distribution object: each distribution in scipy.stats is represented as an object.

Here it’s the normal distribution, and it comes with a PDF, a CDF, and much more.

5.6. Statistics and random numbers: scipy.stats 194 5.6. Statistics and random numbers: scipy.stats 195
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

5.6.3 Statistical tests scipy.integrate.odeint() solves ODE of the form:

dy/dt = rhs(y1, y2, .., t0,...)

dy
As an introduction, let us solve the ODE d t = −2y between t = 0 . . . 4, with the initial condition y(t = 0) = 1.
First the function computing the derivative of the position needs to be defined:

>>> def calc_derivative(ypos, time):

... return -2 * ypos

A statistical test is a decision indicator. For in-

stance, if we have two sets of observations, that we assume are generated from Gaussian processes, we can use
a T-test to decide whether the means of two sets of observations are significantly different:

>>> a = np.random.normal(0, 1, size=100)

>>> b = np.random.normal(1, 1, size=10)
>>> stats.ttest_ind(a, b) Then, to compute y as a function of time:
(array(-3.177574054...), 0.0019370639...)
>>> from scipy.integrate import odeint
>>> time_vec = np.linspace(0, 4, 40)
>>> y = odeint(calc_derivative, y0=1, t=time_vec)
Tip: The resulting output is composed of:
• The T statistic value: it is a number the sign of which is proportional to the difference between the two Let us integrate a more complex ODE: a damped spring-mass oscillator. The position of a mass attached to a
random processes and the magnitude is related to the significance of this difference. spring obeys the 2nd order ODE y 00 + 2εω0 y 0 + ω20 y = 0 with ω20 = k/m with k the spring constant, m the mass
and ε = c/(2mω0 ) with c the damping coefficient. We set:
• the p value: the probability of both processes being identical. If it is close to 1, the two process are almost
certainly identical. The closer it is to zero, the more likely it is that the processes have different means. >>> mass = 0.5 # kg
>>> kspring = 4 # N/m
>>> cviscous = 0.4 # N s/m
See also:
The chapter on statistics introduces much more elaborate tools for statistical testing and statistical data load- Hence:
ing and visualization outside of scipy.
>>> eps = cviscous / (2 * mass * np.sqrt(kspring/mass))
>>> omega = np.sqrt(kspring / mass)

5.7 Numerical integration: scipy.integrate The system is underdamped, as:

>>> eps < 1

5.7.1 Function integrals True
R π/2
The most generic integration routine is scipy.integrate.quad(). To compute 0 si n(t )d t : For odeint(), the 2nd order equation needs to be transformed in a system of two first-order equations for the
vector Y = (y, y 0 ): the function computes the velocity and acceleration:
>>> from scipy.integrate import quad
>>> res, err = quad(np.sin, 0, np.pi/2) >>> def calc_deri(yvec, time, eps, omega):
>>> np.allclose(res, 1) # res is the result, is should be close to 1 ... return (yvec[1], -eps * omega * yvec[1] - omega **2 * yvec[0])
True
>>> np.allclose(err, 1 - res) # err is an estimate of the err
True

Other integration schemes are available: scipy.integrate.fixed_quad(), scipy.integrate.

quadrature(), scipy.integrate.romberg(). . .

5.7.2 Integrating differential equations

scipy.integrate also features routines for integrating Ordinary Differential Equations (ODE). In particular,

5.7. Numerical integration: scipy.integrate 196 5.7. Numerical integration: scipy.integrate 197
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Integration of the system follows:

>>> time_vec = np.linspace(0, 10, 100)

>>> yinit = (1, 0)
>>> yarr = odeint(calc_deri, yinit, time_vec, args=(eps, omega))
Signal FFT

Tip: scipy.integrate.odeint() uses the LSODA (Livermore Solver for Ordinary Differential equations As the signal comes from a real function, the Fourier transform is symmetric.
with Automatic method switching for stiff and non-stiff problems), see the ODEPACK Fortran library for more The peak signal frequency can be found with freqs[power.argmax()]
details.

See also:
Partial Differental Equations
There is no Partial Differential Equations (PDE) solver in Scipy. Some Python packages for solving PDE’s are
available, such as fipy or SfePy.

5.8 Fast Fourier transforms: scipy.fftpack

The scipy.fftpack module computes fast Fourier transforms (FFTs) and offers utilities to handle them. The
main functions are:
• scipy.fftpack.fft() to compute the FFT
• scipy.fftpack.fftfreq() to generate the sampling frequencies
• scipy.fftpack.ifft() computes the inverse FFT, from frequency space to signal space Setting the Fourrier component above
this frequency to zero and inverting the FFT with scipy.fftpack.ifft(), gives a filtered signal.

Note: The code of this example can be found here

As an illustration, a (noisy) input signal (sig), and its FFT:

numpy.fft
>>> from scipy import fftpack
>>> sig_fft = fftpack.fft(sig)
Numpy also has an implementation of FFT (numpy.fft). However, the scipy one should be preferred, as it
>>> freqs = fftpack.fftfreq(sig.size, d=time_step)
uses more efficient underlying implementations.

Fully worked examples:

5.8. Fast Fourier transforms: scipy.fftpack 198 5.8. Fast Fourier transforms: scipy.fftpack 199
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

4. The spectrum consists of high and low frequency components. The noise is contained in the high-
Crude periodicity finding (link) Gaussian image blur (link)
frequency part of the spectrum, so set some of those components to zero (use array slicing).
5. Apply the inverse Fourier transform to see the resulting image.
Solution

5.9 Signal processing: scipy.signal

Tip: scipy.signal is for typical signal processing: 1D, regularly-sampled signals.

Exercise: Denoise moon landing image

Resampling scipy.signal.resample(): resample

a signal to n points using FFT.

>>> t = np.linspace(0, 5, 100)

>>> x = np.sin(t)

>>> from scipy import signal

>>> x_resampled = signal.resample(x, 25)

>>> plt.plot(t, x)
[<matplotlib.lines.Line2D object at ...>]
>>> plt.plot(t[::4], x_resampled, 'ko')
[<matplotlib.lines.Line2D object at ...>]

Tip: Notice how on the side of the window the resampling is less accurate and has a rippling effect.
This resampling is different from the interpolation provided by scipy.interpolate as it only applies to reg-
ularly sampled data.

1. Examine the provided image moonlanding.png, which is heavily contaminated with periodic noise.
In this exercise, we aim to clean up the noise using the Fast Fourier Transform.
2. Load the image using pylab.imread().
3. Find and use the 2-D FFT function in scipy.fftpack, and plot the spectrum (Fourier transform of )
the image. Do you have any trouble visualising the spectrum? If so, why?

5.8. Fast Fourier transforms: scipy.fftpack 200 5.9. Signal processing: scipy.signal 201
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

5.10 Image manipulation: scipy.ndimage

Detrending scipy.signal.detrend(): remove
linear trend from signal: scipy.ndimage provides manipulation of n-dimensional arrays as images.

>>> t = np.linspace(0, 5, 100)

>>> x = t + np.random.normal(size=100) 5.10.1 Geometrical transformations on images
>>> from scipy import signal Changing orientation, resolution, ..
>>> x_detrended = signal.detrend(x)
>>> from scipy import misc # Load an image
>>> plt.plot(t, x) >>> face = misc.face(gray=True)
[<matplotlib.lines.Line2D object at ...>]
>>> plt.plot(t, x_detrended) >>> from scipy import ndimage # Shift, roate and zoom it
[<matplotlib.lines.Line2D object at ...>] >>> shifted_face = ndimage.shift(face, (50, 50))
>>> shifted_face2 = ndimage.shift(face, (50, 50), mode='nearest')
Filtering: For non-linear filtering, scipy.signal has filtering (median filter scipy.signal.medfilt(), >>> rotated_face = ndimage.rotate(face, 30)
Wiener scipy.signal.wiener()), but we will discuss this in the image section. >>> cropped_face = face[50:-50, 50:-50]
>>> zoomed_face = ndimage.zoom(face, 2)
>>> zoomed_face.shape
Tip: scipy.signal also has a full-blown set of tools for the design of linear filter (finite and infinite response (1536, 2048)
filters), but this is out of the scope of this tutorial.

Spectral analysis: scipy.signal.spectrogram() compute a spectrogram –frequency spectrums over con-

secutive time windows–, while scipy.signal.welch() comptes a power spectrum density (PSD).

>>> plt.subplot(151)
<matplotlib.axes._subplots.AxesSubplot object at 0x...>

>>> plt.imshow(shifted_face, cmap=plt.cm.gray)

<matplotlib.image.AxesImage object at 0x...>

>>> plt.axis('off')
(-0.5, 1023.5, 767.5, -0.5)

>>> # etc.

5.10.2 Image filtering

Generate a noisy face:

5.9. Signal processing: scipy.signal 202 5.10. Image manipulation: scipy.ndimage 203
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

>>> from scipy import misc array([[False, True, False],

>>> face = misc.face(gray=True) [...True, True, True],
>>> face = face[:512, -512:] # crop out square on right [False, True, False]], dtype=bool)
>>> import numpy as np >>> el.astype(np.int)
>>> noisy_face = np.copy(face).astype(np.float) array([[0, 1, 0],
>>> noisy_face += face.std() * 0.5 * np.random.standard_normal(face.shape) [1, 1, 1],
[0, 1, 0]])
Apply a variety of filters on it:
• Erosion scipy.ndimage.binary_erosion()
>>> blurred_face = ndimage.gaussian_filter(noisy_face, sigma=3)
>>> median_face = ndimage.median_filter(noisy_face, size=5) >>> a = np.zeros((7, 7), dtype=np.int)
>>> from scipy import signal >>> a[1:6, 2:5] = 1
>>> wiener_face = signal.wiener(noisy_face, (5, 5)) >>> a
array([[0, 0, 0, 0, 0, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 0, 0, 0, 0, 0]])
>>> ndimage.binary_erosion(a).astype(a.dtype)
array([[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 1, 0, 0, 0],
[0, 0, 0, 1, 0, 0, 0],
[0, 0, 0, 1, 0, 0, 0],
Other filters in scipy.ndimage.filters and scipy.signal can be applied to images. [0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0]])
>>> #Erosion removes objects smaller than the structure
Exercise >>> ndimage.binary_erosion(a, structure=np.ones((5,5))).astype(a.dtype)
array([[0, 0, 0, 0, 0, 0, 0],
Compare histograms for the different filtered images. [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],
5.10.3 Mathematical morphology [0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0]])

Tip: Mathematical morphology stems from set theory. It characterizes and transforms geometrical structures. • Dilation scipy.ndimage.binary_dilation()
Binary (black and white) images, in particular, can be transformed using this theory: the sets to be transformed >>> a = np.zeros((5, 5))
are the sets of neighboring non-zero-valued pixels. The theory was also extended to gray-valued images. >>> a[2, 2] = 1
>>> a
array([[ 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.]])
>>> ndimage.binary_dilation(a).astype(a.dtype)
array([[ 0., 0., 0., 0., 0.],
[ 0., 0., 1., 0., 0.],
[ 0., 1., 1., 1., 0.],
[ 0., 0., 1., 0., 0.],
[ 0., 0., 0., 0., 0.]])

• Opening scipy.ndimage.binary_opening()

Mathematical-morphology operations use a structuring element in order to modify geometrical structures. >>> a = np.zeros((5, 5), dtype=np.int)
>>> a[1:4, 1:4] = 1
Let us first generate a structuring element: >>> a[4, 4] = 1
>>> a
>>> el = ndimage.generate_binary_structure(2, 1)
array([[0, 0, 0, 0, 0],
>>> el

5.10. Image manipulation: scipy.ndimage 204 5.10. Image manipulation: scipy.ndimage 205
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

[0, 1, 1, 1, 0], >>> a

[0, 1, 1, 1, 0], array([[0, 0, 0, 0, 0, 0, 0],
[0, 1, 1, 1, 0], [0, 3, 3, 3, 3, 3, 0],
[0, 0, 0, 0, 1]]) [0, 3, 3, 1, 3, 3, 0],
>>> # Opening removes small objects [0, 3, 3, 3, 3, 3, 0],
>>> ndimage.binary_opening(a, structure=np.ones((3, 3))).astype(np.int) [0, 3, 3, 3, 2, 3, 0],
array([[0, 0, 0, 0, 0], [0, 3, 3, 3, 3, 3, 0],
[0, 1, 1, 1, 0], [0, 0, 0, 0, 0, 0, 0]])
[0, 1, 1, 1, 0], >>> ndimage.grey_erosion(a, size=(3, 3))
[0, 1, 1, 1, 0], array([[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0]]) [0, 0, 0, 0, 0, 0, 0],
>>> # Opening can also smooth corners [0, 0, 1, 1, 1, 0, 0],
>>> ndimage.binary_opening(a).astype(np.int) [0, 0, 1, 1, 1, 0, 0],
array([[0, 0, 0, 0, 0], [0, 0, 3, 2, 2, 0, 0],
[0, 0, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0],
[0, 1, 1, 1, 0], [0, 0, 0, 0, 0, 0, 0]])
[0, 0, 1, 0, 0],
[0, 0, 0, 0, 0]])

5.10.4 Connected components and measurements on images

• Closing: scipy.ndimage.binary_closing()

Let us first generate a nice synthetic binary image.

Exercise
>>> x, y = np.indices((100, 100))
Check that opening amounts to eroding, then dilating. >>> sig = np.sin(2*np.pi*x/50.) * np.sin(2*np.pi*y/50.) * (1+x*y/50.**2)**2
>>> mask = sig > 1

An opening operation removes small structures, while a closing operation fills small holes. Such operations
can therefore be used to “clean” an image.

>>> a = np.zeros((50, 50))

>>> a[10:-10, 10:-10] = 1
>>> a += 0.25 * np.random.standard_normal(a.shape)
>>> mask = a>=0.5
>>> opened_mask = ndimage.binary_opening(mask)
>>> closed_mask = ndimage.binary_closing(opened_mask)

scipy.ndimage.label() assigns a different label to each connected component:

>>> labels, nb = ndimage.label(mask)

>>> nb
8

Now compute measurements on each connected component:

>>> areas = ndimage.sum(mask, labels, range(1, labels.max()+1))

>>> areas # The number of pixels in each connected component
Exercise array([ 190., 45., 424., 278., 459., 190., 549., 424.])
>>> maxima = ndimage.maximum(sig, labels, range(1, labels.max()+1))
Check that the area of the reconstructed square is smaller than the area of the initial square. >>> maxima # The maximum signal in each connected component
(The opposite would occur if the closing step was performed before the opening). array([ 1.80238238, 1.13527605, 5.51954079, 2.49611818,
6.71673619, 1.80238238, 16.76547217, 5.51954079])

For gray-valued images, eroding (resp. dilating) amounts to replacing a pixel by the minimal (resp. maximal)
value among pixels covered by the structuring element centered on the pixel of interest.

>>> a = np.zeros((7, 7), dtype=np.int)

>>> a[1:6, 1:6] = 3
>>> a[4, 4] = 2; a[2, 3] = 1

5.10. Image manipulation: scipy.ndimage 206 5.10. Image manipulation: scipy.ndimage 207
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Computing the cumulative probabilities

The annual wind speeds maxima have already been computed and saved in the numpy format in the file
examples/max-speeds.npy, thus they will be loaded by using numpy:

>>> import numpy as np

>>> max_speeds = np.load('intro/summary-exercises/examples/max-speeds.npy')
>>> years_nb = max_speeds.shape[0]

Following the cumulative probability definition p_i from the previous section, the corresponding values will
be:

>>> cprob = (np.arange(years_nb, dtype=np.float32) + 1)/(years_nb + 1)

Extract the 4th connected component, and crop the array around it:

>>> ndimage.find_objects(labels==4) and they are assumed to fit the given wind speeds:
[(slice(30L, 48L, None), slice(30L, 48L, None))]
>>> sl = ndimage.find_objects(labels==4) >>> sorted_max_speeds = np.sort(max_speeds)
>>> from matplotlib import pyplot as plt
>>> plt.imshow(sig[sl[0]])
<matplotlib.image.AxesImage object at ...> Prediction with UnivariateSpline

See the summary exercise on Image processing application: counting bubbles and unmolten grains for a more In this section the quantile function will be estimated by using the UnivariateSpline class which can
advanced example. represent a spline from points. The default behavior is to build a spline of degree 3 and points can
have different weights according to their reliability. Variants are InterpolatedUnivariateSpline and
LSQUnivariateSpline on which errors checking is going to change. In case a 2D spline is wanted, the
5.11 Summary exercises on scientific computing BivariateSpline class family is provided. All those classes for 1D and 2D splines use the FITPACK For-
tran subroutines, that’s why a lower library access is available through the splrep and splev functions for
respectively representing and evaluating a spline. Moreover interpolation functions without the use of FIT-
The summary exercises use mainly Numpy, Scipy and Matplotlib. They provide some real-life examples of sci-
PACK parameters are also provided for simpler use (see interp1d, interp2d, barycentric_interpolate
entific computing with Python. Now that the basics of working with Numpy and Scipy have been introduced,
and so on).
the interested user is invited to try these exercises.
For the Sprogø maxima wind speeds, the UnivariateSpline will be used because a spline of degree 3 seems
to correctly fit the data:
5.11.1 Maximum wind speed prediction at the Sprogø station
>>> from scipy.interpolate import UnivariateSpline
The exercise goal is to predict the maximum wind speed occurring every 50 years even if no measure exists >>> quantile_func = UnivariateSpline(cprob, sorted_max_speeds)
for such a period. The available data are only measured over 21 years at the Sprogø meteorological station
located in Denmark. First, the statistical steps will be given and then illustrated with functions from the The quantile function is now going to be evaluated from the full range of probabilities:
scipy.interpolate module. At the end the interested readers are invited to compute results from raw data and
>>> nprob = np.linspace(0, 1, 1e2)
in a slightly different approach. >>> fitted_max_speeds = quantile_func(nprob)

Statistical approach In the current model, the maximum wind speed occurring every 50 years is defined as the upper 2% quantile.
As a result, the cumulative probability value will be:
The annual maxima are supposed to fit a normal probability density function. However such function is not >>> fifty_prob = 1. - 0.02
going to be estimated because it gives a probability from a wind speed maxima. Finding the maximum wind
speed occurring every 50 years requires the opposite approach, the result needs to be found from a defined
So the storm wind speed occurring every 50 years can be guessed by:
probability. That is the quantile function role and the exercise goal will be to find it. In the current model, it is
supposed that the maximum wind speed occurring every 50 years is defined as the upper 2% quantile. >>> fifty_wind = quantile_func(fifty_prob)
>>> fifty_wind
By definition, the quantile function is the inverse of the cumulative distribution function. The latter describes
array(32.97989825...)
the probability distribution of an annual maxima. In the exercise, the cumulative probability p_i for a given
year i is defined as p_i = i/(N+1) with N = 21, the number of measured years. Thus it will be possible to
The results are now gathered on a Matplotlib figure:
calculate the cumulative probability of every measured wind speed maxima. From those experimental points,
the scipy.interpolate module will be very useful for fitting the quantile function. Finally the 50 years maxima is
going to be evaluated from the cumulative probability of the 2% quantile. Exercise with the Gumbell distribution

The interested readers are now invited to make an exercise by using the wind speeds measured over 21 years.
The measurement period is around 90 minutes (the original period was around 10 minutes but the file size
has been reduced for making the exercise setup easier). The data are stored in numpy format inside the file

5.11. Summary exercises on scientific computing 208 5.11. Summary exercises on scientific computing 209
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

examples/sprog-windspeeds.npy. Do not look at the source code for the plots until you have completed
the exercise.
• The first step will be to find the annual maxima by using numpy and plot them as a matplotlib bar figure.

Fig. 5.2: Solution: Python source file

• The second step will be to use the Gumbell distribution on cumulative probabilities p_i defined as
-log( -log(p_i) ) for fitting a linear quantile function (remember that you can define the degree
of the UnivariateSpline). Plotting the annual maxima versus the Gumbell distribution should give
you the following figure.
• The last step will be to find 34.23 m/s for the maximum wind speed occurring every 50 years.

Fig. 5.1: Solution: Python source file 5.11.2 Non linear least squares curve fitting: application to point extraction in
topographical lidar data

The goal of this exercise is to fit a model to some data. The data used in this tutorial are lidar data and are
described in details in the following introductory paragraph. If you’re impatient and want to practice now,
please skip it and go directly to Loading and visualization.

Introduction

Lidars systems are optical rangefinders that analyze property of scattered light to measure distances. Most of
them emit a short light impulsion towards a target and record the reflected signal. This signal is then processed
to extract the distance between the lidar system and the target.

5.11. Summary exercises on scientific computing 210 5.11. Summary exercises on scientific computing 211
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Topographical lidar systems are such systems embedded in airborne platforms. They measure distances be-
tween the platform and the Earth, so as to deliver information on the Earth’s topography (see1 for more details).
In this tutorial, the goal is to analyze the waveform recorded by the lidar system2 . Such a signal contains peaks
whose center and amplitude permit to compute the position and some characteristics of the hit target. When
the footprint of the laser beam is around 1m on the Earth surface, the beam can hit multiple targets during the
two-way propagation (for example the ground and the top of a tree or building). The sum of the contributions
of each target hit by the laser beam then produces a complex signal with multiple peaks, each one containing
information about one target.
One state of the art method to extract information from these data is to decompose them in a sum of Gaussian
functions where each function represents the contribution of a target hit by the laser beam.
Therefore, we use the scipy.optimize module to fit a waveform to one or a sum of Gaussian functions.

Loading and visualization

Load the first waveform using:

>>> import numpy as np

>>> waveform_1 = np.load('data/waveform_1.npy')

and visualize it:

>>> import matplotlib.pyplot as plt

>>> t = np.arange(len(waveform_1))
>>> plt.plot(t, waveform_1)
[<matplotlib.lines.Line2D object at ...>]
>>> plt.show()

1 Mallet, C. and Bretar, F. Full-Waveform Topographic Lidar: State-of-the-Art. ISPRS Journal of Photogrammetry and Remote Sensing
64(1), pp.1-16, January 2009 http://dx.doi.org/10.1016/j.isprsjprs.2008.09.007
2 The data used for this tutorial are part of the demonstration data available for the FullAnalyze software and were kindly provided by
the GIS DRAIX.

Fig. 5.3: Solution: Python source file

5.11. Summary exercises on scientific computing 212 5.11. Summary exercises on scientific computing 213
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

• coeffs[3] is σ (width)

Initial solution

An approximative initial solution that we can find from looking at the graph is for instance:

>>> x0 = np.array([3, 30, 15, 1], dtype=float)

Fit

scipy.optimize.leastsq minimizes the sum of squares of the function given as an argument. Basically, the
function to minimize is the residuals (the difference between the data and the model):

>>> def residuals(coeffs, y, t):

... return y - model(t, coeffs)

So let’s get our solution by calling scipy.optimize.leastsq() with the following arguments:
• the function to minimize
• an initial solution
• the additional arguments to pass to the function

>>> from scipy.optimize import leastsq

>>> x, flag = leastsq(residuals, x0, args=(waveform_1, t))
>>> print(x)
[ 2.70363341 27.82020742 15.47924562 3.05636228]

As you can notice, this waveform is a 80-bin-length signal with a single peak. And visualize the solution:

>>> plt.plot(t, waveform_1, t, model(t, x))

Fitting a waveform with a simple Gaussian model [<matplotlib.lines.Line2D object at ...>, <matplotlib.lines.Line2D object at ...>]
>>> plt.legend(['waveform', 'model'])
The signal is very simple and can be modeled as a single Gaussian function and an offset corresponding to the <matplotlib.legend.Legend object at ...>
>>> plt.show()
background noise. To fit the signal with the function, we must:
• define the model Remark: from scipy v0.8 and above, you should rather use scipy.optimize.curve_fit() which takes the
model and the data as arguments, so you don’t need to define the residuals any more.
• propose an initial solution
• call scipy.optimize.leastsq
Going further

Model • Try with a more complex waveform (for instance data/waveform_2.npy) that contains three significant
peaks. You must adapt the model which is now a sum of Gaussian functions instead of only one Gaussian
A Gaussian function defined by peak.
½ µ ¶ ¾
t −µ 2
B + A exp −
σ
can be defined in python by:

>>> def model(t, coeffs):

... return coeffs[0] + coeffs[1] * np.exp( - ((t-coeffs[2])/coeffs[3])**2 )

where
• coeffs[0] is B (noise)
• coeffs[1] is A (amplitude)
• coeffs[2] is µ (center)

5.11. Summary exercises on scientific computing 214 5.11. Summary exercises on scientific computing 215
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

5.11.3 Image processing application: counting bubbles and unmolten grains

• In some cases, writing an explicit function to compute the Jacobian is faster than letting leastsq esti-
mate it numerically. Create a function to compute the Jacobian of the residuals and use it as an input for
leastsq.
• When we want to detect very small peaks in the signal, or when the initial guess is too far from a good
solution, the result given by the algorithm is often not satisfying. Adding constraints to the parameters
of the model enables to overcome such limitations. An example of a priori knowledge we can add is the
sign of our variables (which are all positive).
With the following initial solution:

>>> x0 = np.array([3, 50, 20, 1], dtype=float)

compare the result of scipy.optimize.leastsq() and what you can get with scipy.optimize. Statement of the problem
fmin_slsqp() when adding boundary constraints.
1. Open the image file MV_HFV_012.jpg and display it. Browse through the keyword arguments in the doc-
string of imshow to display the image with the “right” orientation (origin in the bottom left corner, and not the
upper left corner as for standard arrays).
This Scanning Element Microscopy image shows a glass sample (light gray matrix) with some bubbles (on
black) and unmolten sand grains (dark gray). We wish to determine the fraction of the sample covered by
these three phases, and to estimate the typical size of sand grains and bubbles, their sizes, etc.
2. Crop the image to remove the lower panel with measure information.
3. Slightly filter the image with a median filter in order to refine its histogram. Check how the histogram
changes.
4. Using the histogram of the filtered image, determine thresholds that allow to define masks for sand pixels,
glass pixels and bubble pixels. Other option (homework): write a function that determines automatically the
thresholds from the minima of the histogram.
5. Display an image in which the three phases are colored with three different colors.
6. Use mathematical morphology to clean the different phases.

5.11. Summary exercises on scientific computing 216 5.11. Summary exercises on scientific computing 217
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

7. Attribute labels to all bubbles and sand grains, and remove from the sand mask grains that are smaller than
>>> filtdat = ndimage.median_filter(dat, size=(7,7))
10 pixels. To do so, use ndimage.sum or np.bincount to compute the grain sizes. >>> hi_dat = np.histogram(dat, bins=np.arange(256))
8. Compute the mean size of bubbles. >>> hi_filtdat = np.histogram(filtdat, bins=np.arange(256))

Proposed solution

5.11.4 Example of solution for the image processing exercise: unmolten grains in
glass

4. Using the histogram of the filtered image, determine thresholds that allow to define masks for sand pix-
els, glass pixels and bubble pixels. Other option (homework): write a function that determines automat-
ically the thresholds from the minima of the histogram.

>>> void = filtdat <= 50

>>> sand = np.logical_and(filtdat > 50, filtdat <= 114)
>>> glass = filtdat > 114

5. Display an image in which the three phases are colored with three different colors.

>>> phases = void.astype(np.int) + 2*glass.astype(np.int) + 3*sand.astype(np.int)

1. Open the image file MV_HFV_012.jpg and display it. Browse through the keyword arguments in the
docstring of imshow to display the image with the “right” orientation (origin in the bottom left corner,
and not the upper left corner as for standard arrays).

>>> dat = pl.imread('data/MV_HFV_012.jpg')

2. Crop the image to remove the lower panel with measure information.

>>> dat = dat[:-60]

3. Slightly filter the image with a median filter in order to refine its histogram. Check how the histogram
changes.

5.11. Summary exercises on scientific computing 218 5.11. Summary exercises on scientific computing 219
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

5.12 Full code examples for the scipy chapter

5.12.1 Finding the minimum of a smooth function

Demos various methods to find the minimum of a function.

import numpy as np
import matplotlib.pyplot as plt

def f(x):
return x**2 + 10*np.sin(x)

x = np.arange(-10, 10, 0.1)

plt.plot(x, f(x))

6. Use mathematical morphology to clean the different phases.

>>> sand_op = ndimage.binary_opening(sand, iterations=2)

7. Attribute labels to all bubbles and sand grains, and remove from the sand mask grains that are smaller
than 10 pixels. To do so, use ndimage.sum or np.bincount to compute the grain sizes.

>>> sand_labels, sand_nb = ndimage.label(sand_op)

>>> sand_areas = np.array(ndimage.sum(sand_op, sand_labels, np.arange(sand_labels.
,→max()+1)))
>>> mask = sand_areas > 100
>>> remove_small_sand = mask[sand_labels.ravel()].reshape(sand_labels.shape)

Now find the minimum with a few methods

from scipy import optimize

8. Compute the mean size of bubbles.
# The default (Nelder Mead)
>>> bubbles_labels, bubbles_nb = ndimage.label(void) print(optimize.minimize(f, x0=0))
>>> bubbles_areas = np.bincount(bubbles_labels.ravel())[1:]
>>> mean_bubble_size = bubbles_areas.mean() Out:
>>> median_bubble_size = np.median(bubbles_areas)
>>> mean_bubble_size, median_bubble_size fun: -7.945823375615215
(1699.875, 65.0) hess_inv: array([[ 0.08589237]])

5.11. Summary exercises on scientific computing 220 5.12. Full code examples for the scipy chapter 221
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

jac: array([ -1.19209290e-06])

message: 'Optimization terminated successfully.'
nfev: 18
nit: 5
njev: 6
status: 0
success: True
x: array([-1.30644012])

print(optimize.minimize(f, x0=0, method="L-BFGS-B"))

Out:

fun: array([-7.94582338])
hess_inv: <1x1 LbfgsInvHessProduct with dtype=float64>
jac: array([ -1.42108547e-06])
message: b'CONVERGENCE: NORM_OF_PROJECTED_GRADIENT_<=_PGTOL'
nfev: 12
nit: 5
status: 0
success: True
x: array([-1.30644013])

plt.show()

Total running time of the script: ( 0 minutes 0.103 seconds) Total running time of the script: ( 0 minutes 0.187 seconds)
Download Python source code: plot_optimize_example1.py Download Python source code: plot_detrend.py
Download Jupyter notebook: plot_optimize_example1.ipynb Download Jupyter notebook: plot_detrend.ipynb
Generated by Sphinx-Gallery Generated by Sphinx-Gallery

5.12.2 Detrending a signal 5.12.3 Resample a signal with scipy.signal.resample

scipy.signal.detrend() removes a linear trend. scipy.signal.resample() uses FFT to resample a 1D signal.
Generate a random signal with a trend Generate a signal with 100 data point
import numpy as np import numpy as np
t = np.linspace(0, 5, 100) t = np.linspace(0, 5, 100)
x = t + np.random.normal(size=100) x = np.sin(t)

Detrend Downsample it by a factor of 4

from scipy import signal from scipy import signal
x_detrended = signal.detrend(x) x_resampled = signal.resample(x, 25)

Plot Plot
from matplotlib import pyplot as plt from matplotlib import pyplot as plt
plt.figure(figsize=(5, 4)) plt.figure(figsize=(5, 4))
plt.plot(t, x, label="x") plt.plot(t, x, label='Original signal')
plt.plot(t, x_detrended, label="x_detrended") plt.plot(t[::4], x_resampled, 'ko', label='Resampled signal')
plt.legend(loc='best')
plt.show() plt.legend(loc='best')
plt.show()

5.12. Full code examples for the scipy chapter 222 5.12. Full code examples for the scipy chapter 223
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

def calc_derivative(ypos, time):

return -2*ypos

time_vec = np.linspace(0, 4, 40)

yvec = odeint(calc_derivative, 1, time_vec)

plt.figure(figsize=(4, 3))
plt.plot(time_vec, yvec)
plt.xlabel('t: Time')
plt.ylabel('y: Position')
plt.tight_layout()

Total running time of the script: ( 0 minutes 0.056 seconds)

Download Python source code: plot_odeint_simple.py
Download Jupyter notebook: plot_odeint_simple.ipynb
Generated by Sphinx-Gallery

5.12.5 Comparing 2 sets of samples from Gaussians

Total running time of the script: ( 0 minutes 0.037 seconds)

Download Python source code: plot_resample.py
Download Jupyter notebook: plot_resample.ipynb
Generated by Sphinx-Gallery

5.12.4 Integrating a simple ODE

Solve the ODE dy/dt = -2y between t = 0..4, with the initial condition y(t=0) = 1.

import numpy as np
from matplotlib import pyplot as plt

# Generates 2 sets of observations

samples1 = np.random.normal(0, size=1000)
samples2 = np.random.normal(1, size=1000)

# Compute a histogram of the sample

import numpy as np bins = np.linspace(-4, 4, 30)
from scipy.integrate import odeint histogram1, bins = np.histogram(samples1, bins=bins, normed=True)
from matplotlib import pyplot as plt histogram2, bins = np.histogram(samples2, bins=bins, normed=True)

5.12. Full code examples for the scipy chapter 224 5.12. Full code examples for the scipy chapter 225
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Total running time of the script: ( 0 minutes 0.038 seconds)

plt.figure(figsize=(6, 4))
plt.hist(samples1, bins=bins, normed=True, label="Samples 1") Download Python source code: plot_odeint_damped_spring_mass.py
plt.hist(samples2, bins=bins, normed=True, label="Samples 2")
plt.legend(loc='best') Download Jupyter notebook: plot_odeint_damped_spring_mass.ipynb
plt.show()
Generated by Sphinx-Gallery

Total running time of the script: ( 0 minutes 0.078 seconds)

Download Python source code: plot_t_test.py 5.12.7 Normal distribution: histogram and PDF
Download Jupyter notebook: plot_t_test.ipynb Explore the normal distribution: a histogram built from samples and the PDF (probability density function).
Generated by Sphinx-Gallery

5.12.6 Integrate the Damped spring-mass oscillator

import numpy as np
from scipy.integrate import odeint import numpy as np
from matplotlib import pyplot as plt
# Sample from a normal distribution using numpy's random number generator
mass = 0.5 # kg samples = np.random.normal(size=10000)
kspring = 4 # N/m
cviscous = 0.4 # N s/m # Compute a histogram of the sample
bins = np.linspace(-5, 5, 30)
histogram, bins = np.histogram(samples, bins=bins, normed=True)
eps = cviscous / (2 * mass * np.sqrt(kspring/mass))
omega = np.sqrt(kspring / mass) bin_centers = 0.5*(bins[1:] + bins[:-1])

# Compute the PDF on the bin centers from scipy distribution object
def calc_deri(yvec, time, eps, omega): from scipy import stats
return (yvec[1], -eps * omega * yvec[1] - omega **2 * yvec[0]) pdf = stats.norm.pdf(bin_centers)

time_vec = np.linspace(0, 10, 100) from matplotlib import pyplot as plt

yinit = (1, 0) plt.figure(figsize=(6, 4))
yarr = odeint(calc_deri, yinit, time_vec, args=(eps, omega)) plt.plot(bin_centers, histogram, label="Histogram of samples")
plt.plot(bin_centers, pdf, label="PDF")
plt.figure(figsize=(4, 3)) plt.legend()
plt.plot(time_vec, yarr[:, 0], label='y') plt.show()
plt.plot(time_vec, yarr[:, 1], label="y'")
plt.legend(loc='best')
Total running time of the script: ( 0 minutes 0.037 seconds)
plt.show()

5.12. Full code examples for the scipy chapter 226 5.12. Full code examples for the scipy chapter 227
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Download Python source code: plot_normal_distribution.py Out:

Download Jupyter notebook: plot_normal_distribution.ipynb [ 3.05931973 1.45754553]
Generated by Sphinx-Gallery
And plot the resulting curve on the data

plt.figure(figsize=(6, 4))
5.12.8 Curve fitting plt.scatter(x_data, y_data, label='Data')
plt.plot(x_data, test_func(x_data, params[0], params[1]),
Demos a simple curve fitting label='Fitted function')
First generate some data
plt.legend(loc='best')
import numpy as np
plt.show()
# Seed the random number generator for reproducibility
np.random.seed(0)

x_data = np.linspace(-5, 5, num=50)

y_data = 2.9 * np.sin(1.5 * x_data) + np.random.normal(size=50)

# And plot it
import matplotlib.pyplot as plt
plt.figure(figsize=(6, 4))
plt.scatter(x_data, y_data)

Total running time of the script: ( 0 minutes 0.071 seconds)

Download Python source code: plot_curve_fit.py
Download Jupyter notebook: plot_curve_fit.ipynb
Generated by Sphinx-Gallery

5.12.9 Spectrogram, power spectral density

Now fit a simple sine function to the data Demo spectrogram and power spectral density on a frequency chirp.

from scipy import optimize import numpy as np

from matplotlib import pyplot as plt
def test_func(x, a, b):
return a * np.sin(b * x)

params, params_covariance = optimize.curve_fit(test_func, x_data, y_data,

Generate a chirp signal
p0=[2, 2])
# Seed the random number generator
print(params) np.random.seed(0)

5.12. Full code examples for the scipy chapter 228 5.12. Full code examples for the scipy chapter 229
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

time_step = .01
time_vec = np.arange(0, 70, time_step)

# A signal with a small frequency chirp

sig = np.sin(0.5 * np.pi * time_vec * (1 + .1 * time_vec))

plt.figure(figsize=(8, 5))
plt.plot(time_vec, sig)

Compute and plot the power spectral density (PSD)

The power of the signal per frequency band

freqs, psd = signal.welch(sig)

plt.figure(figsize=(5, 4))
plt.semilogx(freqs, psd)
plt.title('PSD: power spectral density')
plt.xlabel('Frequency')
Compute and plot the spectrogram plt.ylabel('Power')
plt.tight_layout()
The spectrum of the signal on consecutive time windows

from scipy import signal

freqs, times, spectrogram = signal.spectrogram(sig)

plt.figure(figsize=(5, 4))
plt.imshow(spectrogram, aspect='auto', cmap='hot_r', origin='lower')
plt.title('Spectrogram')
plt.ylabel('Frequency band')
plt.xlabel('Time window')
plt.tight_layout()

5.12. Full code examples for the scipy chapter 230 5.12. Full code examples for the scipy chapter 231
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

5.12.10 A demo of 1D interpolation

plt.show()

Total running time of the script: ( 0 minutes 0.303 seconds) # Generate data
Download Python source code: plot_spectrogram.py import numpy as np
np.random.seed(0)
Download Jupyter notebook: plot_spectrogram.ipynb measured_time = np.linspace(0, 1, 10)
noise = 1e-1 * (np.random.random(10)*2 - 1)
Generated by Sphinx-Gallery measures = np.sin(2 * np.pi * measured_time) + noise

# Interpolate it to new time points

from scipy.interpolate import interp1d
linear_interp = interp1d(measured_time, measures)
interpolation_time = np.linspace(0, 1, 50)
linear_results = linear_interp(interpolation_time)
cubic_interp = interp1d(measured_time, measures, kind='cubic')
cubic_results = cubic_interp(interpolation_time)

# Plot the data and the interpolation

from matplotlib import pyplot as plt
plt.figure(figsize=(6, 4))
plt.plot(measured_time, measures, 'o', ms=6, label='measures')
plt.plot(interpolation_time, linear_results, label='linear interp')
plt.plot(interpolation_time, cubic_results, label='cubic interp')
plt.legend()
plt.show()

Total running time of the script: ( 0 minutes 0.039 seconds)

Download Python source code: plot_interpolation.py
Download Jupyter notebook: plot_interpolation.ipynb
Generated by Sphinx-Gallery

5.12. Full code examples for the scipy chapter 232 5.12. Full code examples for the scipy chapter 233
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

5.12.11 Demo mathematical morphology Generated by Sphinx-Gallery

A basic demo of binary opening and closing.

5.12.12 Plot geometrical transformations on images

Demo geometrical transformations of images.

# Load some data

# Generate some binary data from scipy import misc
import numpy as np face = misc.face(gray=True)
np.random.seed(0)
a = np.zeros((50, 50)) # Apply a variety of transformations
a[10:-10, 10:-10] = 1 from scipy import ndimage
a += 0.25 * np.random.standard_normal(a.shape) from matplotlib import pyplot as plt
mask = a>=0.5 shifted_face = ndimage.shift(face, (50, 50))
shifted_face2 = ndimage.shift(face, (50, 50), mode='nearest')
# Apply mathematical morphology rotated_face = ndimage.rotate(face, 30)
from scipy import ndimage cropped_face = face[50:-50, 50:-50]
opened_mask = ndimage.binary_opening(mask) zoomed_face = ndimage.zoom(face, 2)
closed_mask = ndimage.binary_closing(opened_mask) zoomed_face.shape

# Plot plt.figure(figsize=(15, 3))

from matplotlib import pyplot as plt plt.subplot(151)
plt.imshow(shifted_face, cmap=plt.cm.gray)
plt.figure(figsize=(12, 3.5)) plt.axis('off')
plt.subplot(141)
plt.imshow(a, cmap=plt.cm.gray) plt.subplot(152)
plt.axis('off') plt.imshow(shifted_face2, cmap=plt.cm.gray)
plt.title('a') plt.axis('off')

plt.subplot(142) plt.subplot(153)
plt.imshow(mask, cmap=plt.cm.gray) plt.imshow(rotated_face, cmap=plt.cm.gray)
plt.axis('off') plt.axis('off')
plt.title('mask')
plt.subplot(154)
plt.subplot(143) plt.imshow(cropped_face, cmap=plt.cm.gray)
plt.imshow(opened_mask, cmap=plt.cm.gray) plt.axis('off')
plt.axis('off')
plt.title('opened_mask') plt.subplot(155)
plt.imshow(zoomed_face, cmap=plt.cm.gray)
plt.subplot(144) plt.axis('off')
plt.imshow(closed_mask, cmap=plt.cm.gray)
plt.title('closed_mask') plt.subplots_adjust(wspace=.05, left=.01, bottom=.01, right=.99, top=.99)
plt.axis('off')
plt.show()
plt.subplots_adjust(wspace=.05, left=.01, bottom=.01, right=.99, top=.99)
Total running time of the script: ( 0 minutes 0.916 seconds)
plt.show()
Download Python source code: plot_image_transform.py
Total running time of the script: ( 0 minutes 0.133 seconds)
Download Jupyter notebook: plot_image_transform.ipynb
Download Python source code: plot_mathematical_morpho.py
Generated by Sphinx-Gallery
Download Jupyter notebook: plot_mathematical_morpho.ipynb

5.12. Full code examples for the scipy chapter 234 5.12. Full code examples for the scipy chapter 235
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

5.12.13 Demo connected components

Extracting and labeling connected components in a 2D array

import numpy as np
from matplotlib import pyplot as plt

Generate some binary data

np.random.seed(0)
x, y = np.indices((100, 100))
sig = np.sin(2*np.pi*x/50.) * np.sin(2*np.pi*y/50.) * (1+x*y/50.**2)**2
mask = sig > 1

plt.figure(figsize=(7, 3.5))
plt.subplot(1, 2, 1)
plt.imshow(sig)
plt.axis('off')
plt.title('sig')

plt.subplot(1, 2, 2)
plt.imshow(mask, cmap=plt.cm.gray)
plt.axis('off')
plt.title('mask')
plt.subplots_adjust(wspace=.05, left=.01, bottom=.01, right=.99, top=.9) Extract the 4th connected component, and crop the array around it

sl = ndimage.find_objects(labels==4)
plt.figure(figsize=(3.5, 3.5))
plt.imshow(sig[sl[0]])
plt.title('Cropped connected component')
plt.axis('off')

plt.subplots_adjust(wspace=.05, left=.01, bottom=.01, right=.99, top=.9)

plt.show()

Label connected components

from scipy import ndimage

labels, nb = ndimage.label(mask)

plt.figure(figsize=(3.5, 3.5))
plt.imshow(labels)
plt.title('label')
plt.axis('off')

plt.subplots_adjust(wspace=.05, left=.01, bottom=.01, right=.99, top=.9)

Total running time of the script: ( 0 minutes 0.138 seconds)

Download Python source code: plot_connect_measurements.py
Download Jupyter notebook: plot_connect_measurements.ipynb

5.12. Full code examples for the scipy chapter 236 5.12. Full code examples for the scipy chapter 237
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Generated by Sphinx-Gallery
ax.plot(xmins, f(xmins), 'go', label="Minima")

# Plot the roots

5.12.14 Minima and roots of a function roots = np.array([root.x, root2.x])
ax.plot(roots, f(roots), 'kv', label="Roots")
Demos finding minima and roots of a function.
# Decorate the figure
ax.legend(loc='best')
Define the function ax.set_xlabel('x')
ax.set_ylabel('f(x)')
import numpy as np ax.axhline(0, color='gray')
plt.show()
x = np.arange(-10, 10, 0.1)
def f(x):
return x**2 + 10*np.sin(x)

Find minima

from scipy import optimize

# Global optimization
grid = (-10, 10, 0.1)
xmin_global = optimize.brute(f, (grid, ))
print("Global minima found %s " % xmin_global)

# Constrain optimization
xmin_local = optimize.fminbound(f, 0, 10)
print("Local minimum found %s " % xmin_local)

Out:

Global minima found [-1.30641113]

Local minimum found 3.8374671195

Root finding

root = optimize.root(f, 1) # our initial guess is 1 Total running time of the script: ( 0 minutes 0.043 seconds)
print("First root found %s " % root.x)
root2 = optimize.root(f, -2.5) Download Python source code: plot_optimize_example2.py
print("Second root found %s " % root2.x)
Download Jupyter notebook: plot_optimize_example2.ipynb
Out: Generated by Sphinx-Gallery

First root found [ 0.]

Second root found [-2.47948183] 5.12.15 Optimization of a two-parameter function

import numpy as np
Plot function, minima, and roots

import matplotlib.pyplot as plt # Define the function that we are interested in

fig = plt.figure(figsize=(6, 4)) def sixhump(x):
ax = fig.add_subplot(111) return ((4 - 2.1*x[0]**2 + x[0]**4 / 3.) * x[0]**2 + x[0] * x[1]
+ (-4 + 4*x[1]**2) * x[1] **2)
# Plot the function
ax.plot(x, f(x), 'b-', label="f(x)") # Make a grid to evaluate the function (for plotting)
x = np.linspace(-2, 2)
# Plot the minima y = np.linspace(-1, 1)
xmins = np.array([xmin_global[0], xmin_local]) xg, yg = np.meshgrid(x, y)

5.12. Full code examples for the scipy chapter 238 5.12. Full code examples for the scipy chapter 239
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

A 2D image plot of the function

Simple visualization in 2D

import matplotlib.pyplot as plt

plt.figure()
plt.imshow(sixhump([xg, yg]), extent=[-2, 2, -1, 1])
plt.colorbar()

Find the minima

from scipy import optimize

x_min = optimize.minimize(sixhump, x0=[0, 0])

plt.figure()
# Show the function in 2D
plt.imshow(sixhump([xg, yg]), extent=[-2, 2, -1, 1])
A 3D surface plot of the function plt.colorbar()
# And the minimum that we've found:
from mpl_toolkits.mplot3d import Axes3D plt.scatter(x_min.x[0], x_min.x[1])
fig = plt.figure()
ax = fig.add_subplot(111, projection='3d') plt.show()
surf = ax.plot_surface(xg, yg, sixhump([xg, yg]), rstride=1, cstride=1,
cmap=plt.cm.jet, linewidth=0, antialiased=False)

ax.set_xlabel('x')
ax.set_ylabel('y')
ax.set_zlabel('f(x, y)')
ax.set_title('Six-hump Camelback function')

5.12. Full code examples for the scipy chapter 240 5.12. Full code examples for the scipy chapter 241
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

from scipy import signal

from matplotlib import pyplot as plt

import numpy as np
noisy_face = np.copy(face).astype(np.float)
noisy_face += face.std() * 0.5 * np.random.standard_normal(face.shape)
blurred_face = ndimage.gaussian_filter(noisy_face, sigma=3)
median_face = ndimage.median_filter(noisy_face, size=5)
wiener_face = signal.wiener(noisy_face, (5, 5))

plt.figure(figsize=(12, 3.5))
plt.subplot(141)
plt.imshow(noisy_face, cmap=plt.cm.gray)
plt.axis('off')
plt.title('noisy')

plt.subplot(142)
plt.imshow(blurred_face, cmap=plt.cm.gray)
plt.axis('off')
plt.title('Gaussian filter')

plt.subplot(143)
plt.imshow(median_face, cmap=plt.cm.gray)
plt.axis('off')
plt.title('median filter')

plt.subplot(144)
plt.imshow(wiener_face, cmap=plt.cm.gray)
plt.title('Wiener filter')
plt.axis('off')

Total running time of the script: ( 0 minutes 0.236 seconds) plt.subplots_adjust(wspace=.05, left=.01, bottom=.01, right=.99, top=.99)

Download Python source code: plot_2d_minimization.py plt.show()

Download Jupyter notebook: plot_2d_minimization.ipynb
Total running time of the script: ( 0 minutes 0.420 seconds)
Generated by Sphinx-Gallery
Download Python source code: plot_image_filters.py
Download Jupyter notebook: plot_image_filters.ipynb
5.12.16 Plot filtering on images
Generated by Sphinx-Gallery
Demo filtering for denoising of images.

5.12.17 Plotting and manipulating FFTs for filtering

Plot the power of the FFT of a signal and inverse FFT back to reconstruct a signal.
This example demonstrate scipy.fftpack.fft(), scipy.fftpack.fftfreq() and scipy.fftpack.
ifft(). It implements a basic filter that is very suboptimal, and should not be used.
import numpy as np
from scipy import fftpack
from matplotlib import pyplot as plt

Generate the signal

# Load some data
from scipy import misc
face = misc.face(gray=True) # Seed the random number generator
face = face[:512, -512:] # crop out square on right np.random.seed(1234)

# Apply a variety of filters time_step = 0.02

from scipy import ndimage period = 5.

5.12. Full code examples for the scipy chapter 242 5.12. Full code examples for the scipy chapter 243
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

freqs = sample_freq[pos_mask]
time_vec = np.arange(0, 20, time_step) peak_freq = freqs[power[pos_mask].argmax()]
sig = (np.sin(2 * np.pi / period * time_vec)
+ 0.5 * np.random.randn(time_vec.size)) # Check that it does indeed correspond to the frequency that we generate
# the signal with
plt.figure(figsize=(6, 5)) np.allclose(peak_freq, 1./period)
plt.plot(time_vec, sig, label='Original signal')
# An inner plot to show the peak frequency
axes = plt.axes([0.55, 0.3, 0.3, 0.5])
plt.title('Peak frequency')
plt.plot(freqs[:8], power[:8])
plt.setp(axes, yticks=[])

# scipy.signal.find_peaks_cwt can also be used for more advanced

# peak detection

Compute and plot the power

# The FFT of the signal

sig_fft = fftpack.fft(sig)

# And the power (sig_fft is of complex dtype)

power = np.abs(sig_fft)
Remove all the high frequencies
# The corresponding frequencies
sample_freq = fftpack.fftfreq(sig.size, d=time_step)
We now remove all the high frequencies and transform back from frequencies to signal.
# Plot the FFT power
high_freq_fft = sig_fft.copy()
plt.figure(figsize=(6, 5))
high_freq_fft[np.abs(sample_freq) > peak_freq] = 0
plt.plot(sample_freq, power)
filtered_sig = fftpack.ifft(high_freq_fft)
plt.xlabel('Frequency [Hz]')
plt.ylabel('plower')
plt.figure(figsize=(6, 5))
plt.plot(time_vec, sig, label='Original signal')
# Find the peak frequency: we can focus on only the positive frequencies
plt.plot(time_vec, filtered_sig, linewidth=3, label='Filtered signal')
pos_mask = np.where(sample_freq > 0)
plt.xlabel('Time [s]')

5.12. Full code examples for the scipy chapter 244 5.12. Full code examples for the scipy chapter 245
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Load the data

plt.ylabel('Amplitude')

plt.legend(loc='best') import numpy as np

data = np.loadtxt('../../../../data/populations.txt')
years = data[:, 0]
populations = data[:, 1:]

Plot the data

import matplotlib.pyplot as plt

plt.figure()
plt.plot(years, populations * 1e-3)
plt.xlabel('Year')
plt.ylabel('Population number ($\cdot10^3$)')
plt.legend(['hare', 'lynx', 'carrot'], loc=1)

Note This is actually a bad way of creating a filter: such brutal cut-off in frequency space does not control
distorsion on the signal.
Filters should be created using the scipy filter design code

plt.show()

Total running time of the script: ( 0 minutes 0.142 seconds)

Download Python source code: plot_fftpack.py
Download Jupyter notebook: plot_fftpack.ipynb
Generated by Sphinx-Gallery

Plot its periods

5.12.18 Solutions of the exercises for scipy
from scipy import fftpack
Crude periodicity finding
ft_populations = fftpack.fft(populations, axis=0)
Discover the periods in evolution of animal populations (../../data/populations.txt) frequencies = fftpack.fftfreq(populations.shape[0], years[1] - years[0])
periods = 1 / frequencies

plt.figure()
plt.plot(periods, abs(ft_populations) * 1e-3, 'o')

5.12. Full code examples for the scipy chapter 246 5.12. Full code examples for the scipy chapter 247
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

plt.xlim(0, 22)
plt.xlabel('Period') import matplotlib.pyplot as plt
plt.ylabel('Power ($\cdot10^3$)') months = np.arange(12)
plt.plot(months, temp_max, 'ro')
plt.show() plt.plot(months, temp_min, 'bo')
plt.xlabel('Month')
plt.ylabel('Min and max temperature')

There’s probably a period of around 10 years (obvious from the plot), but for this crude a method, there’s not
enough data to say much more.
Total running time of the script: ( 0 minutes 0.073 seconds)
Fitting it to a periodic function
Download Python source code: plot_periodicity_finder.py
Download Jupyter notebook: plot_periodicity_finder.ipynb from scipy import optimize
def yearly_temps(times, avg, ampl, time_offset):
Generated by Sphinx-Gallery return (avg
+ ampl * np.cos((times + time_offset) * 2 * np.pi / times.max()))
Curve fitting: temperature as a function of month of the year res_max, cov_max = optimize.curve_fit(yearly_temps, months,
temp_max, [20, 10, 0])
We have the min and max temperatures in Alaska for each months of the year. We would like to find a function res_min, cov_min = optimize.curve_fit(yearly_temps, months,
to describe this yearly evolution. temp_min, [-40, 20, 0])
For this, we will fit a periodic function.

Plotting the fit

The data
days = np.linspace(0, 12, num=365)
import numpy as np
plt.figure()
temp_max = np.array([17, 19, 21, 28, 33, 38, 37, 37, 31, 23, 19, 18]) plt.plot(months, temp_max, 'ro')
temp_min = np.array([-62, -59, -56, -46, -32, -18, -9, -13, -25, -46, -52, -58]) plt.plot(days, yearly_temps(days, *res_max), 'r-')

5.12. Full code examples for the scipy chapter 248 5.12. Full code examples for the scipy chapter 249
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

The original image

plt.plot(months, temp_min, 'bo')
plt.plot(days, yearly_temps(days, *res_min), 'b-')
plt.xlabel('Month') # read image
plt.ylabel('Temperature ($^\circ$C)') img = plt.imread('../../../../data/elephant.png')
plt.figure()
plt.show() plt.imshow(img)

Total running time of the script: ( 0 minutes 0.072 seconds)

Download Python source code: plot_curvefit_temperature_data.py Prepare an Gaussian convolution kernel
Download Jupyter notebook: plot_curvefit_temperature_data.ipynb
# First a 1-D Gaussian
Generated by Sphinx-Gallery t = np.linspace(-10, 10, 30)
bump = np.exp(-0.1*t**2)
bump /= np.trapz(bump) # normalize the integral to 1
Simple image blur by convolution with a Gaussian kernel
# make a 2-D kernel out of it
Blur an an image (../../../../data/elephant.png) using a Gaussian kernel. kernel = bump[:, np.newaxis] * bump[np.newaxis, :]
Convolution is easy to perform with FFT: convolving two signals boils down to multiplying their FFTs (and
performing an inverse FFT).
Implement convolution via FFT
import numpy as np
from scipy import fftpack
import matplotlib.pyplot as plt # Padded fourier transform, with the same shape as the image
# We use :func:`scipy.signal.fftpack.fft2` to have a 2D FFT
kernel_ft = fftpack.fft2(kernel, shape=img.shape[:2], axes=(0, 1))

# convolve
img_ft = fftpack.fft2(img, axes=(0, 1))
# the 'newaxis' is to match to color direction

5.12. Full code examples for the scipy chapter 250 5.12. Full code examples for the scipy chapter 251
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

img2_ft = kernel_ft[:, :, np.newaxis] * img_ft

img2 = fftpack.ifft2(img2_ft, axes=(0, 1)).real

# clip values to range

img2 = np.clip(img2, 0, 1)

# plot output
plt.figure()
plt.imshow(img2)

Note that we still have a decay to zero at the border of the image. Using scipy.ndimage.gaussian_filter()
would get rid of this artifact

plt.show()

Total running time of the script: ( 0 minutes 0.195 seconds)

Download Python source code: plot_image_blur.py
Download Jupyter notebook: plot_image_blur.ipynb
Generated by Sphinx-Gallery
Further exercise (only if you are familiar with this stuff):
A “wrapped border” appears in the upper left and top edges of the image. This is because the padding is not
done correctly, and does not take the kernel size into account (so the convolution “flows out of bounds of the Image denoising by FFT
image”). Try to remove this artifact.
Denoise an image (../../../../data/moonlanding.png) by implementing a blur with an FFT.
Implements, via FFT, the following convolution:
A function to do it: scipy.signal.fftconvolve()
Z
f 1 (t ) = d t 0 K (t − t 0 ) f 0 (t 0 )
The above exercise was only for didactic reasons: there exists a function in scipy that will do this
for us, and probably do a better job: scipy.signal.fftconvolve()
f˜1 (ω) = K̃ (ω) f˜0 (ω)
from scipy import signal
# mode='same' is there to enforce the same output shape as input arrays
# (ie avoid border effects) Read and plot the image
img3 = signal.fftconvolve(img, kernel[:, :, np.newaxis], mode='same')
plt.figure()
plt.imshow(img3) import numpy as np
import matplotlib.pyplot as plt

im = plt.imread('../../../../data/moonlanding.png').astype(float)

5.12. Full code examples for the scipy chapter 252 5.12. Full code examples for the scipy chapter 253
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

plt.figure()
plt.imshow(im, plt.cm.gray)
plt.title('Original image')

Filter in FFT

# In the lines following, we'll make a copy of the original spectrum and
# truncate coefficients.

Compute the 2d FFT of the input image # Define the fraction of coefficients (in each direction) we keep
keep_fraction = 0.1
from scipy import fftpack # Call ff a copy of the original transform. Numpy arrays have a copy
im_fft = fftpack.fft2(im) # method for this purpose.
im_fft2 = im_fft.copy()
# Show the results
# Set r and c to be the number of rows and columns of the array.
def plot_spectrum(im_fft): r, c = im_fft2.shape
from matplotlib.colors import LogNorm
# A logarithmic colormap # Set to zero all rows with indices between r*keep_fraction and
plt.imshow(np.abs(im_fft), norm=LogNorm(vmin=5)) # r*(1-keep_fraction):
plt.colorbar() im_fft2[int(r*keep_fraction):int(r*(1-keep_fraction))] = 0
plt.figure() # Similarly with the columns:
plot_spectrum(im_fft) im_fft2[:, int(c*keep_fraction):int(c*(1-keep_fraction))] = 0
plt.title('Fourier transform')
plt.figure()
plot_spectrum(im_fft2)
plt.title('Filtered Spectrum')

5.12. Full code examples for the scipy chapter 254 5.12. Full code examples for the scipy chapter 255
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Reconstruct the final image Easier and better: scipy.ndimage.gaussian_filter()

# Reconstruct the denoised image from the filtered spectrum, keep only the Implementing filtering directly with FFTs is tricky and time consuming. We can use the Gaussian
# real part for display. filter from scipy.ndimage
im_new = fftpack.ifft2(im_fft2).real
from scipy import ndimage
plt.figure() im_blur = ndimage.gaussian_filter(im, 4)
plt.imshow(im_new, plt.cm.gray)
plt.title('Reconstructed Image') plt.figure()
plt.imshow(im_blur, plt.cm.gray)
plt.title('Blurred image')

plt.show()

5.12. Full code examples for the scipy chapter 256 5.12. Full code examples for the scipy chapter 257
 Scipy lecture notes, Edition 2017.1

CHAPTER 6
Getting help and finding documentation

Total running time of the script: ( 0 minutes 0.381 seconds) Author: Emmanuelle Gouillart

Download Python source code: plot_fft_image_denoise.py Rather than knowing all functions in Numpy and Scipy, it is important to find rapidly information throughout
the documentation and the available help. Here are some ways to get information:
Download Jupyter notebook: plot_fft_image_denoise.ipynb
• In Ipython, help function opens the docstring of the function. Only type the beginning of the func-
Generated by Sphinx-Gallery
tion’s name and use tab completion to display the matching functions.
Download all examples in Python source code: auto_examples_python.zip
In [204]: help np.v
Download all examples in Jupyter notebooks: auto_examples_jupyter.zip np.vander np.vdot np.version np.void0 np.vstack
np.var np.vectorize np.void np.vsplit
Generated by Sphinx-Gallery
See also: In [204]: help np.vander

References to go further In Ipython it is not possible to open a separated window for help and documentation; how-
• Some chapters of the advanced and the packages and applications parts of the scipy lectures ever one can always open a second Ipython shell just to display help and docstrings. . .

• The scipy cookbook

• Numpy’s and Scipy’s documentations can be browsed online on http://docs.scipy.org/doc. The search
button is quite useful inside the reference documentation of the two packages (http://docs.scipy.org/
doc/numpy/reference/ and http://docs.scipy.org/doc/scipy/reference/).

5.12. Full code examples for the scipy chapter 258 259
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Tutorials on various topics as well as the complete API with all docstrings are found on this website.
remove(path)
os.removedirs
removedirs(path)
os.rmdir
rmdir(path)
os.unlink
unlink(path)
os.walk
Directory tree generator.

• If everything listed above fails (and Google doesn’t have the answer). . . don’t despair! Write to the
mailing-list suited to your problem: you should have a quick answer if you describe your problem well.
Experts on scientific python often give very enlightening explanations on the mailing-list.
• Numpy’s and Scipy’s documentation is enriched and updated on a regular basis by users on a wiki http:
//docs.scipy.org/doc/numpy/. As a result, some docstrings are clearer or more detailed on the wiki, – Numpy discussion (numpy-discussion@scipy.org): all about numpy arrays, manipulating them,
and you may want to read directly the documentation on the wiki instead of the official documentation indexation questions, etc.
website. Note that anyone can create an account on the wiki and write better documentation; this is an
– SciPy Users List (scipy-user@scipy.org): scientific computing with Python, high-level data process-
easy way to contribute to an open-source project and improve the tools you are using!
ing, in particular with the scipy package.
• Scipy central http://central.scipy.org/ gives recipes on many common problems frequently encoun-
– matplotlib-users@lists.sourceforge.net for plotting with matplotlib.

tered, such as fitting data points, solving ODE, etc.

• Matplotlib’s website http://matplotlib.org/ features a very nice gallery with a large number of plots, each
of them shows both the source code and the resulting plot. This is very useful for learning by example.
More standard documentation is also available.
Finally, two more “technical” possibilities are useful as well:
• In Ipython, the magical function %psearch search for objects matching patterns. This is useful if, for
example, one does not know the exact name of a function.

In [3]: import numpy as np

In [4]: %psearch np.diag*
np.diag
np.diagflat
np.diagonal

• numpy.lookfor looks for keywords inside the docstrings of specified modules.

In [45]: numpy.lookfor('convolution')
Search results for 'convolution'
--------------------------------
numpy.convolve
Returns the discrete, linear convolution of two one-dimensional
sequences.
numpy.bartlett
Return the Bartlett window.
numpy.correlate
Discrete, linear correlation of two 1-dimensional sequences.
In [46]: numpy.lookfor('remove', module='os')
Search results for 'remove'
---------------------------
os.remove

260 261
 Scipy lecture notes, Edition 2017.1

This part of the Scipy lecture notes is dedicated to advanced usage. It strives to educate the proficient Python
coder to be an expert and tackles various specific topics.

Part II

Advanced topics

262 263
 Scipy lecture notes, Edition 2017.1

– Decorators implemented as classes and as functions

– Copying the docstring and other attributes of the original function
– Examples in the standard library
– Deprecation of functions

7
– A while-loop removing decorator
– A plugin registration system
• Context managers

CHAPTER
– Catching exceptions
– Using generators to define context managers

7.1 Iterators, generator expressions and generators

7.1.1 Iterators
Advanced Python Constructs
Simplicity

Duplication of effort is wasteful, and replacing the various home-grown approaches with a standard feature
usually ends up making things more readable, and interoperable as well.
Guido van Rossum — Adding Optional Static Typing to Python

Author Zbigniew J˛edrzejewski-Szmek An iterator is an object adhering to the iterator protocol — basically this means that it has a next method,
This section covers some features of the Python language which can be considered advanced — in the sense which, when called, returns the next item in the sequence, and when there’s nothing to return, raises the
that not every language has them, and also in the sense that they are more useful in more complicated pro- StopIteration exception.
grams or libraries, but not in the sense of being particularly specialized, or particularly complicated. An iterator object allows to loop just once. It holds the state (position) of a single iteration, or from the other
It is important to underline that this chapter is purely about the language itself — about features supported side, each loop over a sequence requires a single iterator object. This means that we can iterate over the same
through special syntax complemented by functionality of the Python stdlib, which could not be implemented sequence more than once concurrently. Separating the iteration logic from the sequence allows us to have
through clever external modules. more than one way of iteration.

The process of developing the Python programming language, its syntax, is very transparent; proposed Calling the __iter__ method on a container to create an iterator object is the most straightforward way to get
changes are evaluated from various angles and discussed via Python Enhancement Proposals — PEPs. As a hold of an iterator. The iter function does that for us, saving a few keystrokes.
result, features described in this chapter were added after it was shown that they indeed solve real problems >>> nums = [1, 2, 3] # note that ... varies: these are different objects
and that their use is as simple as possible. >>> iter(nums)
<...iterator object at ...>
>>> nums.__iter__()
Chapter contents
<...iterator object at ...>
>>> nums.__reversed__()
• Iterators, generator expressions and generators <...reverseiterator object at ...>
– Iterators
>>> it = iter(nums)
– Generator expressions >>> next(it)
1
– Generators
>>> next(it)
– Bidirectional communication 2
>>> next(it)
– Chaining generators 3
>>> next(it)
• Decorators
Traceback (most recent call last):
– Replacing or tweaking the original object File "<stdin>", line 1, in <module>
StopIteration

264 7.1. Iterators, generator expressions and generators 265

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

When used in a loop, StopIteration is swallowed and causes the loop to finish. But with explicit invocation, of a generator function creates a generator object, adhering to the iterator protocol. As with normal function
we can see that once the iterator is exhausted, accessing it raises an exception. invocations, concurrent and recursive invocations are allowed.
Using the for..in loop also uses the __iter__ method. This allows us to transparently start the iteration over a When next is called, the function is executed until the first yield. Each encountered yield statement gives a
sequence. But if we already have the iterator, we want to be able to use it in an for loop in the same way. In value becomes the return value of next. After executing the yield statement, the execution of this function is
order to achieve this, iterators in addition to next are also required to have a method called __iter__ which suspended.
returns the iterator (self).
>>> def f():
Support for iteration is pervasive in Python: all sequences and unordered containers in the standard library ... yield 1
allow this. The concept is also stretched to other things: e.g. file objects support iteration over lines. ... yield 2
>>> f()
>>> f = open('/etc/fstab') <generator object f at 0x...>
>>> f is f.__iter__() >>> gen = f()
True >>> next(gen)
1
The file is an iterator itself and it’s __iter__ method doesn’t create a separate object: only a single thread of >>> next(gen)
sequential access is allowed. 2
>>> next(gen)
Traceback (most recent call last):
7.1.2 Generator expressions File "<stdin>", line 1, in <module>
StopIteration
A second way in which iterator objects are created is through generator expressions, the basis for list compre-
hensions. To increase clarity, a generator expression must always be enclosed in parentheses or an expression. Let’s go over the life of the single invocation of the generator function.
If round parentheses are used, then a generator iterator is created. If rectangular parentheses are used, the pro-
>>> def f():
cess is short-circuited and we get a list. ... print("-- start --")
... yield 3
>>> (i for i in nums)
... print("-- middle --")
<generator object <genexpr> at 0x...>
... yield 4
>>> [i for i in nums]
... print("-- finished --")
[1, 2, 3]
>>> gen = f()
>>> list(i for i in nums)
>>> next(gen)
[1, 2, 3]
-- start --
3
The list comprehension syntax also extends to dictionary and set comprehensions. A set is created when the >>> next(gen)
generator expression is enclosed in curly braces. A dict is created when the generator expression contains -- middle --
“pairs” of the form key:value: 4
>>> next(gen)
>>> {i for i in range(3)} -- finished --
set([0, 1, 2]) Traceback (most recent call last):
>>> {i:i**2 for i in range(3)} ...
{0: 0, 1: 1, 2: 4} StopIteration

One gotcha should be mentioned: in old Pythons the index variable (i) would leak, and in versions >= 3 this is Contrary to a normal function, where executing f() would immediately cause the first print to be executed,
fixed. gen is assigned without executing any statements in the function body. Only when gen.next() is invoked by
next, the statements up to the first yield are executed. The second next prints -- middle -- and execution
halts on the second yield. The third next prints -- finished -- and falls of the end of the function. Since
7.1.3 Generators no yield was reached, an exception is raised.
What happens with the function after a yield, when the control passes to the caller? The state of each generator
Generators is stored in the generator object. From the point of view of the generator function, is looks almost as if it was
running in a separate thread, but this is just an illusion: execution is strictly single-threaded, but the interpreter
A generator is a function that produces a sequence of results instead of a single value. keeps and restores the state in between the requests for the next value.

David Beazley — A Curious Course on Coroutines and Concurrency Why are generators useful? As noted in the parts about iterators, a generator function is just a different way to
create an iterator object. Everything that can be done with yield statements, could also be done with next
methods. Nevertheless, using a function and having the interpreter perform its magic to create an iterator
A third way to create iterator objects is to call a generator function. A generator is a function containing the
has advantages. A function can be much shorter than the definition of a class with the required next and
keyword yield. It must be noted that the mere presence of this keyword completely changes the nature of the
__iter__ methods. What is more important, it is easier for the author of the generator to understand the state
function: this yield statement doesn’t have to be invoked, or even reachable, but causes the function to be
which is kept in local variables, as opposed to instance attributes, which have to be used to pass data between
marked as a generator. When a normal function is called, the instructions contained in the body start to be
consecutive invocations of next on an iterator object.
executed. When a generator is called, the execution stops before the first instruction in the body. An invocation
A broader question is why are iterators useful? When an iterator is used to power a loop, the loop becomes very

7.1. Iterators, generator expressions and generators 266 7.1. Iterators, generator expressions and generators 267
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

simple. The code to initialise the state, to decide if the loop is finished, and to find the next value is extracted
1
into a separate place. This highlights the body of the loop — the interesting part. In addition, it is possible to >>> it.throw(IndexError)
reuse the iterator code in other places. --yield raised IndexError()--
--yielding 2--
2
7.1.4 Bidirectional communication >>> it.close()
--closing--
Each yield statement causes a value to be passed to the caller. This is the reason for the introduction of
generators by PEP 255 (implemented in Python 2.2). But communication in the reverse direction is also useful.
One obvious way would be some external state, either a global variable or a shared mutable object. Direct next or __next__?
communication is possible thanks to PEP 342 (implemented in 2.5). It is achieved by turning the previously
boring yield statement into an expression. When the generator resumes execution after a yield statement, In Python 2.x, the iterator method to retrieve the next value is called next. It is invoked implicitly through
the caller can call a method on the generator object to either pass a value into the generator, which then is the global function next, which means that it should be called __next__. Just like the global function iter
returned by the yield statement, or a different method to inject an exception into the generator. calls __iter__. This inconsistency is corrected in Python 3.x, where it.next becomes it.__next__. For
other generator methods — send and throw — the situation is more complicated, because they are not
The first of the new methods is send(value), which is similar to next(), but passes value into the generator
called implicitly by the interpreter. Nevertheless, there’s a proposed syntax extension to allow continue
to be used for the value of the yield expression. In fact, g.next() and g.send(None) are equivalent.
to take an argument which will be passed to send of the loop’s iterator. If this extension is accepted, it’s
The second of the new methods is throw(type, value=None, traceback=None) which is equivalent to: likely that gen.send will become gen.__send__. The last of generator methods, close, is pretty obviously
named incorrectly, because it is already invoked implicitly.
raise type, value, traceback

at the point of the yield statement.

Unlike raise (which immediately raises an exception from the current execution point), throw() first resumes 7.1.5 Chaining generators
the generator, and only then raises the exception. The word throw was picked because it is suggestive of putting
the exception in another location, and is associated with exceptions in other languages.
Note: This is a preview of PEP 380 (not yet implemented, but accepted for Python 3.3).
What happens when an exception is raised inside the generator? It can be either raised explicitly or when
executing some statements or it can be injected at the point of a yield statement by means of the throw()
method. In either case, such an exception propagates in the standard manner: it can be intercepted by an Let’s say we are writing a generator and we want to yield a number of values generated by a second generator,
except or finally clause, or otherwise it causes the execution of the generator function to be aborted and a subgenerator. If yielding of values is the only concern, this can be performed without much difficulty using
propagates in the caller. a loop such as

For completeness’ sake, it’s worth mentioning that generator iterators also have a close() method, which can subgen = some_other_generator()
be used to force a generator that would otherwise be able to provide more values to finish immediately. It for v in subgen:
allows the generator __del__ method to destroy objects holding the state of generator. yield v

Let’s define a generator which just prints what is passed in through send and throw. However, if the subgenerator is to interact properly with the caller in the case of calls to send(), throw()
and close(), things become considerably more difficult. The yield statement has to be guarded by a
>>> import itertools
>>> def g(): try..except..finally structure similar to the one defined in the previous section to “debug” the generator func-
... print('--start--') tion. Such code is provided in PEP 380#id13, here it suffices to say that new syntax to properly yield from a
... for i in itertools.count(): subgenerator is being introduced in Python 3.3:
... print('--yielding %i --' % i)
... try: yield from some_other_generator()
... ans = yield i
... except GeneratorExit: This behaves like the explicit loop above, repeatedly yielding values from some_other_generator until it is
... print('--closing--') exhausted, but also forwards send, throw and close to the subgenerator.
... raise
... except Exception as e:
...
...
print('--yield raised %r --' % e)
else:
7.2 Decorators
... print('--yield returned %s --' % ans)

>>> it = g() Summary

>>> next(it)
--start-- This amazing feature appeared in the language almost apologetically and with concern that it might not be
--yielding 0--
that useful.
0
>>> it.send(11) Bruce Eckel — An Introduction to Python Decorators
--yield returned 11--
--yielding 1--

7.1. Iterators, generator expressions and generators 268 7.2. Decorators 269
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Since functions and classes are objects, they can be passed around. Since they are mutable objects, they can 7.2.2 Decorators implemented as classes and as functions
be modified. The act of altering a function or class object after it has been constructed but before is is bound
to its name is called decorating. The only requirement on decorators is that they can be called with a single argument. This means that deco-
rators can be implemented as normal functions, or as classes with a __call__ method, or in theory, even as
There are two things hiding behind the name “decorator” — one is the function which does the work of deco-
lambda functions.
rating, i.e. performs the real work, and the other one is the expression adhering to the decorator syntax, i.e. an
at-symbol and the name of the decorating function. Let’s compare the function and class approaches. The decorator expression (the part after @) can be either just
a name, or a call. The bare-name approach is nice (less to type, looks cleaner, etc.), but is only possible when
Function can be decorated by using the decorator syntax for functions:
no arguments are needed to customise the decorator. Decorators written as functions can be used in those
@decorator # · two cases:
def function(): # ¶
pass >>> def simple_decorator(function):
... print("doing decoration")
... return function
• A function is defined in the standard way. ¶
>>> @simple_decorator
• An expression starting with @ placed before the function definition is the decorator ·. The part after @ ... def function():
must be a simple expression, usually this is just the name of a function or class. This part is evaluated ... print("inside function")
first, and after the function defined below is ready, the decorator is called with the newly defined function doing decoration
>>> function()
object as the single argument. The value returned by the decorator is attached to the original name of
inside function
the function.
Decorators can be applied to functions and to classes. For classes the semantics are identical — the original >>> def decorator_with_arguments(arg):
class definition is used as an argument to call the decorator and whatever is returned is assigned under the ... print("defining the decorator")
original name. ... def _decorator(function):
... # in this inner function, arg is available too
Before the decorator syntax was implemented (PEP 318), it was possible to achieve the same effect by assign- ... print("doing decoration, %r " % arg)
ing the function or class object to a temporary variable and then invoking the decorator explicitly and then ... return function
assigning the return value to the name of the function. This sounds like more typing, and it is, and also the ... return _decorator
name of the decorated function doubling as a temporary variable must be used at least three times, which is >>> @decorator_with_arguments("abc")
prone to errors. Nevertheless, the example above is equivalent to: ... def function():
... print("inside function")
def function(): # ¶ defining the decorator
pass doing decoration, 'abc'
function = decorator(function) # · >>> function()
inside function
Decorators can be stacked — the order of application is bottom-to-top, or inside-out. The semantics are such
that the originally defined function is used as an argument for the first decorator, whatever is returned by The two trivial decorators above fall into the category of decorators which return the original function. If they
the first decorator is used as an argument for the second decorator, . . . , and whatever is returned by the last were to return a new function, an extra level of nestedness would be required. In the worst case, three levels of
decorator is attached under the name of the original function. nested functions.

The decorator syntax was chosen for its readability. Since the decorator is specified before the header of the >>> def replacing_decorator_with_args(arg):
function, it is obvious that its is not a part of the function body and its clear that it can only operate on the ... print("defining the decorator")
whole function. Because the expression is prefixed with @ is stands out and is hard to miss (“in your face”, ... def _decorator(function):
according to the PEP :) ). When more than one decorator is applied, each one is placed on a separate line in an ... # in this inner function, arg is available too
... print("doing decoration, %r " % arg)
easy to read way.
... def _wrapper(*args, **kwargs):
... print("inside wrapper, %r %r " % (args, kwargs))
... return function(*args, **kwargs)
7.2.1 Replacing or tweaking the original object ... return _wrapper
... return _decorator
Decorators can either return the same function or class object or they can return a completely different ob- >>> @replacing_decorator_with_args("abc")
ject. In the first case, the decorator can exploit the fact that function and class objects are mutable and add ... def function(*args, **kwargs):
attributes, e.g. add a docstring to a class. A decorator might do something useful even without modifying ... print("inside function, %r %r " % (args, kwargs))
the object, for example register the decorated class in a global registry. In the second case, virtually anything is ... return 14
possible: when something different is substituted for the original function or class, the new object can be com- defining the decorator
pletely different. Nevertheless, such behaviour is not the purpose of decorators: they are intended to tweak doing decoration, 'abc'
the decorated object, not do something unpredictable. Therefore, when a function is “decorated” by replacing >>> function(11, 12)
inside wrapper, (11, 12) {}
it with a different function, the new function usually calls the original function, after doing some preparatory
inside function, (11, 12) {}
work. Likewise, when a class is “decorated” by replacing if with a new class, the new class is usually derived
14
from the original class. When the purpose of the decorator is to do something “every time”, like to log every call
to a decorated function, only the second type of decorators can be used. On the other hand, if the first type is
The _wrapper function is defined to accept all positional and keyword arguments. In general we cannot know
sufficient, it is better to use it, because it is simpler.
what arguments the decorated function is supposed to accept, so the wrapper function just passes everything

7.2. Decorators 270 7.2. Decorators 271

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

to the wrapped function. One unfortunate consequence is that the apparent argument list is misleading. 7.2.3 Copying the docstring and other attributes of the original function
Compared to decorators defined as functions, complex decorators defined as classes are simpler. When an
When a new function is returned by the decorator to replace the original function, an unfortunate conse-
object is created, the __init__ method is only allowed to return None, and the type of the created object
quence is that the original function name, the original docstring, the original argument list are lost. Those
cannot be changed. This means that when a decorator is defined as a class, it doesn’t make much sense to use
attributes of the original function can partially be “transplanted” to the new function by setting __doc__ (the
the argument-less form: the final decorated object would just be an instance of the decorating class, returned
docstring), __module__ and __name__ (the full name of the function), and __annotations__ (extra informa-
by the constructor call, which is not very useful. Therefore it’s enough to discuss class-based decorators where
tion about arguments and the return value of the function available in Python 3). This can be done automati-
arguments are given in the decorator expression and the decorator __init__ method is used for decorator
cally by using functools.update_wrapper.
construction.

>>> class decorator_class(object):

functools.update_wrapper(wrapper, wrapped)
... def __init__(self, arg):
... # this method is called in the decorator expression
... print("in decorator init, %s " % arg) “Update a wrapper function to look like the wrapped function.”
... self.arg = arg >>> import functools
... def __call__(self, function): >>> def replacing_decorator_with_args(arg):
... # this method is called to do the job ... print("defining the decorator")
... print("in decorator call, %s " % self.arg) ... def _decorator(function):
... return function ... print("doing decoration, %r " % arg)
>>> deco_instance = decorator_class('foo') ... def _wrapper(*args, **kwargs):
in decorator init, foo ... print("inside wrapper, %r %r " % (args, kwargs))
>>> @deco_instance ... return function(*args, **kwargs)
... def function(*args, **kwargs): ... return functools.update_wrapper(_wrapper, function)
... print("in function, %s %s " % (args, kwargs)) ... return _decorator
in decorator call, foo >>> @replacing_decorator_with_args("abc")
>>> function() ... def function():
in function, () {} ... "extensive documentation"
... print("inside function")
Contrary to normal rules (PEP 8) decorators written as classes behave more like functions and therefore their ... return 14
name often starts with a lowercase letter. defining the decorator
doing decoration, 'abc'
In reality, it doesn’t make much sense to create a new class just to have a decorator which returns the original >>> function
function. Objects are supposed to hold state, and such decorators are more useful when the decorator returns <function function at 0x...>
a new object. >>> print(function.__doc__)
extensive documentation
>>> class replacing_decorator_class(object):
... def __init__(self, arg):
... # this method is called in the decorator expression One important thing is missing from the list of attributes which can be copied to the replacement func-
... print("in decorator init, %s " % arg) tion: the argument list. The default values for arguments can be modified through the __defaults__,
... self.arg = arg
__kwdefaults__ attributes, but unfortunately the argument list itself cannot be set as an attribute. This
... def __call__(self, function):
means that help(function) will display a useless argument list which will be confusing for the user of the
... # this method is called to do the job
... print("in decorator call, %s " % self.arg) function. An effective but ugly way around this problem is to create the wrapper dynamically, using eval. This
... self.function = function can be automated by using the external decorator module. It provides support for the decorator decorator,
... return self._wrapper which takes a wrapper and turns it into a decorator which preserves the function signature.
... def _wrapper(self, *args, **kwargs):
To sum things up, decorators should always use functools.update_wrapper or some other means of copy-
... print("in the wrapper, %s %s " % (args, kwargs))
ing function attributes.
... return self.function(*args, **kwargs)
>>> deco_instance = replacing_decorator_class('foo')
in decorator init, foo
>>> @deco_instance 7.2.4 Examples in the standard library
... def function(*args, **kwargs):
... print("in function, %s %s " % (args, kwargs)) First, it should be mentioned that there’s a number of useful decorators available in the standard library. There
in decorator call, foo are three decorators which really form a part of the language:
>>> function(11, 12)
in the wrapper, (11, 12) {} • classmethod causes a method to become a “class method”, which means that it can be invoked without
in function, (11, 12) {} creating an instance of the class. When a normal method is invoked, the interpreter inserts the instance
object as the first positional parameter, self. When a class method is invoked, the class itself is given as
A decorator like this can do pretty much anything, since it can modify the original function object and mangle the first parameter, often called cls.
the arguments, call the original function or not, and afterwards mangle the return value. Class methods are still accessible through the class’ namespace, so they don’t pollute the module’s
namespace. Class methods can be used to provide alternative constructors:

7.2. Decorators 272 7.2. Decorators 273

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

To make everything crystal clear, let’s define a “debug” example:

class Array(object):
def __init__(self, data): >>> class D(object):
self.data = data ... @property
... def a(self):
@classmethod ... print("getting 1")
def fromfile(cls, file): ... return 1
data = numpy.load(file) ... @a.setter
return cls(data) ... def a(self, value):
... print("setting %r " % value)
This is cleaner then using a multitude of flags to __init__. ... @a.deleter
... def a(self):
• staticmethod is applied to methods to make them “static”, i.e. basically a normal function, but acces-
... print("deleting")
sible through the class namespace. This can be useful when the function is only needed inside this class >>> D.a
(its name would then be prefixed with _), or when we want the user to think of the method as connected <property object at 0x...>
to the class, despite an implementation which doesn’t require this. >>> D.a.fget
<function ...>
• property is the pythonic answer to the problem of getters and setters. A method decorated with
>>> D.a.fset
property becomes a getter which is automatically called on attribute access.
<function ...>
>>> class A(object): >>> D.a.fdel
... @property <function ...>
... def a(self): >>> d = D() # ... varies, this is not the same `a` function
... "an important attribute" >>> d.a
... return "a value" getting 1
>>> A.a 1
<property object at 0x...> >>> d.a = 2
>>> A().a setting 2
'a value' >>> del d.a
deleting
>>> d.a
In this example, A.a is an read-only attribute. It is also documented: help(A) includes the docstring for getting 1
attribute a taken from the getter method. Defining a as a property allows it to be a calculated on the fly, 1
and has the side effect of making it read-only, because no setter is defined.
To have a setter and a getter, two methods are required, obviously. Since Python 2.6 the following syntax Properties are a bit of a stretch for the decorator syntax. One of the premises of the decorator syntax —
is preferred: that the name is not duplicated — is violated, but nothing better has been invented so far. It is just good
style to use the same name for the getter, setter, and deleter methods.
class Rectangle(object):
def __init__(self, edge): Some newer examples include:
self.edge = edge • functools.lru_cache memoizes an arbitrary function maintaining a limited cache of argu-
ments:answer pairs (Python 3.2)
@property
def area(self): • functools.total_ordering is a class decorator which fills in missing ordering methods (__lt__,
"""Computed area. __gt__, __le__, . . . ) based on a single available one (Python 2.7).
Setting this updates the edge length to the proper value.
""" 7.2.5 Deprecation of functions
return self.edge**2
Let’s say we want to print a deprecation warning on stderr on the first invocation of a function we don’t like
@area.setter
def area(self, area): anymore. If we don’t want to modify the function, we can use a decorator:
self.edge = area ** 0.5
class deprecated(object):
"""Print a deprecation warning once on first use of the function.
The way that this works, is that the property decorator replaces the getter method with a property
object. This object in turn has three methods, getter, setter, and deleter, which can be used as >>> @deprecated() # doctest: +SKIP
decorators. Their job is to set the getter, setter and deleter of the property object (stored as attributes ... def f():
fget, fset, and fdel). The getter can be set like in the example above, when creating the object. When ... pass
defining the setter, we already have the property object under area, and we add the setter to it by using >>> f() # doctest: +SKIP
the setter method. All this happens when we are creating the class. f is deprecated
"""
Afterwards, when an instance of the class has been created, the property object is special. When the def __call__(self, func):
interpreter executes attribute access, assignment, or deletion, the job is delegated to the methods of the self.func = func
property object. self.count = 0

7.2. Decorators 274 7.2. Decorators 275

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

return self._wrapper
7.2.7 A plugin registration system
def _wrapper(self, *args, **kwargs):
self.count += 1 This is a class decorator which doesn’t modify the class, but just puts it in a global registry. It falls into the
if self.count == 1: category of decorators returning the original object:
print self.func.__name__, 'is deprecated'
return self.func(*args, **kwargs) class WordProcessor(object):
PLUGINS = []
def process(self, text):
It can also be implemented as a function:
for plugin in self.PLUGINS:
def deprecated(func): text = plugin().cleanup(text)
"""Print a deprecation warning once on first use of the function. return text

>>> @deprecated # doctest: +SKIP @classmethod

... def f(): def plugin(cls, plugin):
... pass cls.PLUGINS.append(plugin)
>>> f() # doctest: +SKIP
f is deprecated @WordProcessor.plugin
""" class CleanMdashesExtension(object):
count = [0] def cleanup(self, text):
def wrapper(*args, **kwargs): return text.replace('&mdash;', u'\N{em dash}')
count[0] += 1
if count[0] == 1: Here we use a decorator to decentralise the registration of plugins. We call our decorator with a noun, instead
print func.__name__, 'is deprecated' of a verb, because we use it to declare that our class is a plugin for WordProcessor. Method plugin simply
return func(*args, **kwargs) appends the class to the list of plugins.
return wrapper
A word about the plugin itself: it replaces HTML entity for em-dash with a real Unicode em-dash character.
It exploits the unicode literal notation to insert a character by using its name in the unicode database (“EM
DASH”). If the Unicode character was inserted directly, it would be impossible to distinguish it from an en-
7.2.6 A while-loop removing decorator dash in the source of a program.

Let’s say we have function which returns a lists of things, and this list created by running a loop. If we don’t See also:
know how many objects will be needed, the standard way to do this is something like: More examples and reading
def find_answers(): • PEP 318 (function and method decorator syntax)
answers = []
while True: • PEP 3129 (class decorator syntax)
ans = look_for_next_answer() • http://wiki.python.org/moin/PythonDecoratorLibrary
if ans is None:
break • https://docs.python.org/dev/library/functools.html
answers.append(ans)
• http://pypi.python.org/pypi/decorator
return answers
• Bruce Eckel
This is fine, as long as the body of the loop is fairly compact. Once it becomes more complicated, as often
– Decorators I: Introduction to Python Decorators
happens in real code, this becomes pretty unreadable. We could simplify this by using yield statements, but
then the user would have to explicitly call list(find_answers()). – Python Decorators II: Decorator Arguments
We can define a decorator which constructs the list for us: – Python Decorators III: A Decorator-Based Build System

def vectorized(generator_func):
def wrapper(*args, **kwargs):
return list(generator_func(*args, **kwargs)) 7.3 Context managers
return functools.update_wrapper(wrapper, generator_func)
A context manager is an object with __enter__ and __exit__ methods which can be used in the with state-
Our function then becomes: ment:
@vectorized with manager as var:
def find_answers(): do_something(var)
while True:
ans = look_for_next_answer()
is in the simplest case equivalent to
if ans is None:
break
yield ans

7.2. Decorators 276 7.3. Context managers 277

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

• contextlib.closing å the same as the example above, call close

var = manager.__enter__()
try: • parallel programming
do_something(var)
finally: – concurrent.futures.ThreadPoolExecutor å invoke in parallel then kill thread pool (py >= 3.2)
manager.__exit__()
– concurrent.futures.ProcessPoolExecutor å invoke in parallel then kill process pool (py >=
3.2)
In other words, the context manager protocol defined in PEP 343 permits the extraction of the boring part of a
try..except..finally structure into a separate class leaving only the interesting do_something block. – nogil å solve the GIL problem temporarily (cython only :( )

1. The __enter__ method is called first. It can return a value which will be assigned to var. The as-part is
optional: if it isn’t present, the value returned by __enter__ is simply ignored. 7.3.1 Catching exceptions
2. The block of code underneath with is executed. Just like with try clauses, it can either execute success-
fully to the end, or it can break, continue‘ or return, or it can throw an exception. Either way, after the When an exception is thrown in the with-block, it is passed as arguments to __exit__. Three arguments are
block is finished, the __exit__ method is called. If an exception was thrown, the information about the used, the same as returned by sys.exc_info(): type, value, traceback. When no exception is thrown, None is
exception is passed to __exit__, which is described below in the next subsection. In the normal case, used for all three arguments. The context manager can “swallow” the exception by returning a true value from
exceptions can be ignored, just like in a finally clause, and will be rethrown after __exit__ is finished. __exit__. Exceptions can be easily ignored, because if __exit__ doesn’t use return and just falls of the end,
None is returned, a false value, and therefore the exception is rethrown after __exit__ is finished.
Let’s say we want to make sure that a file is closed immediately after we are done writing to it:
The ability to catch exceptions opens interesting possibilities. A classic example comes from unit-tests — we
>>> class closing(object): want to make sure that some code throws the right kind of exception:
... def __init__(self, obj):
... self.obj = obj class assert_raises(object):
... def __enter__(self): # based on pytest and unittest.TestCase
... return self.obj def __init__(self, type):
... def __exit__(self, *args): self.type = type
... self.obj.close() def __enter__(self):
>>> with closing(open('/tmp/file', 'w')) as f: pass
... f.write('the contents\n') def __exit__(self, type, value, traceback):
if type is None:
Here we have made sure that the f.close() is called when the with block is exited. Since closing files is such raise AssertionError('exception expected')
if issubclass(type, self.type):
a common operation, the support for this is already present in the file class. It has an __exit__ method
return True # swallow the expected exception
which calls close and can be used as a context manager itself:
raise AssertionError('wrong exception type')
>>> with open('/tmp/file', 'a') as f:
... f.write('more contents\n') with assert_raises(KeyError):
{}['foo']
The common use for try..finally is releasing resources. Various different cases are implemented similarly:
in the __enter__ phase the resource is acquired, in the __exit__ phase it is released, and the exception, if
thrown, is propagated. As with files, there’s often a natural operation to perform after the object has been used 7.3.2 Using generators to define context managers
and it is most convenient to have the support built in. With each release, Python provides support in more
places: When discussing generators, it was said that we prefer generators to iterators implemented as classes because
they are shorter, sweeter, and the state is stored as local, not instance, variables. On the other hand, as de-
• all file-like objects:
scribed in Bidirectional communication, the flow of data between the generator and its caller can be bidirec-
– file å automatically closed tional. This includes exceptions, which can be thrown into the generator. We would like to implement context
managers as special generator functions. In fact, the generator protocol was designed to support this use case.
– fileinput, tempfile (py >= 3.2)
@contextlib.contextmanager
– bz2.BZ2File, gzip.GzipFile, tarfile.TarFile, zipfile.ZipFile
def some_generator(<arguments>):
– ftplib, nntplib å close connection (py >= 3.2 or 3.3) <setup>
try:
• locks yield <value>
– multiprocessing.RLock å lock and unlock finally:
<cleanup>
– multiprocessing.Semaphore
– memoryview å automatically release (py >= 3.2 and 2.7) The contextlib.contextmanager helper takes a generator and turns it into a context manager. The gener-
ator has to obey some rules which are enforced by the wrapper function — most importantly it must yield
• decimal.localcontext å modify precision of computations temporarily exactly once. The part before the yield is executed from __enter__, the block of code protected by the con-
• _winreg.PyHKEY å open and close hive key text manager is executed when the generator is suspended in yield, and the rest is executed in __exit__. If
an exception is thrown, the interpreter hands it to the wrapper through __exit__ arguments, and the wrap-
• warnings.catch_warnings å kill warnings temporarily per function then throws it at the point of the yield statement. Through the use of generators, the context
manager is shorter and simpler.

7.3. Context managers 278 7.3. Context managers 279

 Scipy lecture notes, Edition 2017.1

Let’s rewrite the closing example as a generator:

@contextlib.contextmanager
def closing(obj):
try:
yield obj

8
finally:
obj.close()

Let’s rewrite the assert_raises example as a generator:

@contextlib.contextmanager
def assert_raises(type):
try:
CHAPTER
yield
except type:
return
except Exception as value:
raise AssertionError('wrong exception type')
else:
raise AssertionError('exception expected')

Here we use a decorator to turn generator functions into context managers!

Advanced NumPy

Author: Pauli Virtanen

NumPy is at the base of Python’s scientific stack of tools. Its purpose to implement efficient operations on
many items in a block of memory. Understanding how it works in detail helps in making efficient use of its
flexibility, taking useful shortcuts.
This section covers:
• Anatomy of NumPy arrays, and its consequences. Tips and tricks.
• Universal functions: what, why, and what to do if you want a new one.
• Integration with other tools: NumPy offers several ways to wrap any data in an ndarray, without unnec-
essary copies.
• Recently added features, and what’s in them: PEP 3118 buffers, generalized ufuncs, . . .

Prerequisites

• NumPy
• Cython
• Pillow (Python imaging library, used in a couple of examples)

Chapter contents

• Life of ndarray
– It’s. . .

7.3. Context managers 280 281

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

– Block of memory • how to interpret an element

– Data types
– Indexing scheme: strides
– Findings in dissection
• Universal functions
– What they are?
– Exercise: building an ufunc from scratch
– Solution: building an ufunc from scratch
– Generalized ufuncs
• Interoperability features
– Sharing multidimensional, typed data
typedef struct PyArrayObject {
– The old buffer protocol PyObject_HEAD
– The old buffer protocol
/* Block of memory */
– Array interface protocol char *data;

• Array siblings: chararray, maskedarray, matrix /* Data type descriptor */

– chararray: vectorized string operations PyArray_Descr *descr;

– masked_array missing data /* Indexing scheme */

int nd;
– recarray: purely convenience npy_intp *dimensions;
– matrix: convenience? npy_intp *strides;

• Summary /* Other stuff */

PyObject *base;
• Contributing to NumPy/Scipy
int flags;
– Why PyObject *weakreflist;
} PyArrayObject;
– Reporting bugs
– Contributing to documentation
– Contributing features 8.1.2 Block of memory
– How to help, in general >>> x = np.array([1, 2, 3], dtype=np.int32)
>>> x.data
<... at ...>
>>> str(x.data)
Tip: In this section, numpy will be imported as follows:
'\x01\x00\x00\x00\x02\x00\x00\x00\x03\x00\x00\x00'
>>> import numpy as np
Memory address of the data:

>>> x.__array_interface__['data'][0]
64803824
8.1 Life of ndarray
The whole __array_interface__:

8.1.1 It’s. . . >>> x.__array_interface__

{'data': (35828928, False),
'descr': [('', '<i4')],
ndarray = 'shape': (4,),
block of memory + indexing scheme + data type descriptor 'strides': None,
'typestr': '<i4',
• raw data 'version': 3}
• how to locate an element
Reminder: two ndarrays may share the same memory:

8.1. Life of ndarray 282 8.1. Life of ndarray 283

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

>>> x = np.array([1, 2, 3, 4]) chunk_id "RIFF"

>>> y = x[:-1]
chunk_size 4-byte unsigned little-endian integer
>>> x[0] = 9
>>> y
format "WAVE"
array([9, 2, 3]) fmt_id "fmt "
fmt_size 4-byte unsigned little-endian integer
Memory does not need to be owned by an ndarray: audio_fmt 2-byte unsigned little-endian integer
num_channels 2-byte unsigned little-endian integer
>>> x = b'1234' # The 'b' is for "bytes", necessary in Python 3 sample_rate 4-byte unsigned little-endian integer
byte_rate 4-byte unsigned little-endian integer
x is a string (in Python 3 a bytes), we can represent its data as an array of ints: block_align 2-byte unsigned little-endian integer
bits_per_sample 2-byte unsigned little-endian integer
>>> y = np.frombuffer(x, dtype=np.int8)
>>> y.data
data_id "data"
<... at ...> data_size 4-byte unsigned little-endian integer
>>> y.base is x
True • 44-byte block of raw data (in the beginning of the file)
• . . . followed by data_size bytes of actual sound data.
>>> y.flags
C_CONTIGUOUS : True The .wav file header as a NumPy structured data type:
F_CONTIGUOUS : True
OWNDATA : False >>> wav_header_dtype = np.dtype([
WRITEABLE : False ... ("chunk_id", (bytes, 4)), # flexible-sized scalar type, item size 4
ALIGNED : True ... ("chunk_size", "<u4"), # little-endian unsigned 32-bit integer
UPDATEIFCOPY : False ... ("format", "S4"), # 4-byte string
... ("fmt_id", "S4"),
... ("fmt_size", "<u4"),
The owndata and writeable flags indicate status of the memory block.
... ("audio_fmt", "<u2"), #
... ("num_channels", "<u2"), # .. more of the same ...
... ("sample_rate", "<u4"), #
8.1.3 Data types ... ("byte_rate", "<u4"),
... ("block_align", "<u2"),
The descriptor ... ("bits_per_sample", "<u2"),
... ("data_id", ("S1", (2, 2))), # sub-array, just for fun!
dtype describes a single item in the array: ... ("data_size", "u4"),
... #
... # the sound data itself cannot be represented here:
type scalar type of the data, one of:
... # it does not have a fixed size
int8, int16, float64, et al. (fixed size) ... ])
str, unicode, void (flexible size)
itemsize size of the data block See also:
byte- byte order: big-endian > / little-endian < / not applicable |
order wavreader.py
fields sub-dtypes, if it’s a structured data type
>>> wav_header_dtype['format']
shape shape of the array, if it’s a sub-array dtype('S4')
>>> wav_header_dtype.fields
>>> np.dtype(int).type dict_proxy({'block_align': (dtype('uint16'), 32), 'format': (dtype('S4'), 8), 'data_id':␣
,→(dtype(('S1', (2, 2))), 36), 'fmt_id': (dtype('S4'), 12), 'byte_rate': (dtype('uint32'), 28),
<type 'numpy.int64'>
,→ 'chunk_id': (dtype('S4'), 0), 'num_channels': (dtype('uint16'), 22), 'sample_rate': (dtype(
>>> np.dtype(int).itemsize
,→'uint32'), 24), 'bits_per_sample': (dtype('uint16'), 34), 'chunk_size': (dtype('uint32'), 4),
8
,→ 'fmt_size': (dtype('uint32'), 16), 'data_size': (dtype('uint32'), 40), 'audio_fmt': (dtype(
>>> np.dtype(int).byteorder
,→'uint16'), 20)})
'='
>>> wav_header_dtype.fields['format']
(dtype('S4'), 8)

Example: reading .wav files • The first element is the sub-dtype in the structured data, corresponding to the name format

The .wav file header: • The second one is its offset (in bytes) from the beginning of the item

8.1. Life of ndarray 284 8.1. Life of ndarray 285

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Exercise – scalars can “lose” to arrays in some situations

• Casting in general copies data:
Mini-exercise, make a “sparse” dtype by using offsets, and only some of the fields:
>>> x = np.array([1, 2, 3, 4], dtype=np.float)
>>> wav_header_dtype = np.dtype(dict(
>>> x
... names=['format', 'sample_rate', 'data_id'],
array([ 1., 2., 3., 4.])
... offsets=[offset_1, offset_2, offset_3], # counted from start of structure in bytes
>>> y = x.astype(np.int8)
... formats=list of dtypes for each of the fields,
>>> y
... ))
array([1, 2, 3, 4], dtype=int8)
and use that to read the sample rate, and data_id (as sub-array). >>> y + 1
array([2, 3, 4, 5], dtype=int8)
>>> y + 256
>>> f = open('data/test.wav', 'r') array([257, 258, 259, 260], dtype=int16)
>>> wav_header = np.fromfile(f, dtype=wav_header_dtype, count=1) >>> y + 256.0
>>> f.close() array([ 257., 258., 259., 260.])
>>> print(wav_header) >>> y + np.array([256], dtype=np.int32)
[ ('RIFF', 17402L, 'WAVE', 'fmt ', 16L, 1, 1, 16000L, 32000L, 2, 16, [['d', 'a'], ['t', 'a']],␣ array([257, 258, 259, 260], dtype=int32)
,→17366L)]
>>> wav_header['sample_rate'] • Casting on setitem: dtype of the array is not changed on item assignment:
array([16000], dtype=uint32)
>>> y[:] = y + 1.5
Let’s try accessing the sub-array: >>> y
array([2, 3, 4, 5], dtype=int8)
>>> wav_header['data_id']
array([[['d', 'a'],
['t', 'a']]], Note: Exact rules: see numpy documentation
dtype='|S1')
>>> wav_header.shape
(1,)
>>> wav_header['data_id'].shape Re-interpretation / viewing
(1, 2, 2)

• Data block in memory (4 bytes)

When accessing sub-arrays, the dimensions get added to the end!

0x01 || 0x02 || 0x03 || 0x04

Note: There are existing modules such as wavfile, audiolab, etc. for loading sound data. . .
– 4 of uint8, OR,
– 4 of int8, OR,
Casting and re-interpretation/views – 2 of int16, OR,

casting – 1 of int32, OR,

• on assignment – 1 of float32, OR,

• on array construction – ...

• on arithmetic How to switch from one to another?

• etc. 1. Switch the dtype:

• and manually: .astype(dtype) >>> x = np.array([1, 2, 3, 4], dtype=np.uint8)

>>> x.dtype = "<i2"
data re-interpretation >>> x
• manually: .view(dtype) array([ 513, 1027], dtype=int16)
>>> 0x0201, 0x0403
(513, 1027)
Casting
0x01 0x02 || 0x03 0x04
• Casting in arithmetic, in nutshell:
– only type (not value!) of operands matters
Note: little-endian: least significant byte is on the left in memory
– largest “safe” type able to represent both is picked

8.1. Life of ndarray 286 8.1. Life of ndarray 287

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

2. Create a new view:

>>> y = np.array([[1, 3], [2, 4]], dtype=np.uint8).transpose()
>>> y = x.view("<i4") >>> x = y.copy()
>>> y >>> x
array([67305985], dtype=int32) array([[1, 2],
>>> 0x04030201 [3, 4]], dtype=uint8)
67305985 >>> y
array([[1, 2],
[3, 4]], dtype=uint8)
0x01 0x02 0x03 0x04 >>> x.view(np.int16)
array([[ 513],
[1027]], dtype=int16)
Note: >>> 0x0201, 0x0403
(513, 1027)
• .view() makes views, does not copy (or alter) the memory block >>> y.view(np.int16)
array([[ 769, 1026]], dtype=int16)
• only changes the dtype (and adjusts array shape):
• What happened?
>>> x[1] = 5
>>> y • . . . we need to look into what x[0,1] actually means
array([328193], dtype=int32)
>>> y.base is x >>> 0x0301, 0x0402
True (769, 1026)

8.1.4 Indexing scheme: strides

Mini-exercise: data re-interpretation
Main point
See also:
view-colors.py The question:

You have RGBA data in an array: >>> x = np.array([[1, 2, 3],

... [4, 5, 6],
>>> x = np.zeros((10, 10, 4), dtype=np.int8) ... [7, 8, 9]], dtype=np.int8)
>>> x[:, :, 0] = 1 >>> str(x.data)
>>> x[:, :, 1] = 2 '\x01\x02\x03\x04\x05\x06\x07\x08\t'
>>> x[:, :, 2] = 3
>>> x[:, :, 3] = 4 At which byte in ``x.data`` does the item ``x[1, 2]`` begin?

where the last three dimensions are the R, B, and G, and alpha channels. The answer (in NumPy)
How to make a (10, 10) structured array with field names ‘r’, ‘g’, ‘b’, ‘a’ without copying data? • strides: the number of bytes to jump to find the next element

>>> y = ... • 1 stride per dimension

>>> x.strides
>>> assert (y['r'] == 1).all()
(3, 1)
>>> assert (y['g'] == 2).all()
>>> byte_offset = 3*1 + 1*2 # to find x[1, 2]
>>> assert (y['b'] == 3).all()
>>> x.flat[byte_offset]
>>> assert (y['a'] == 4).all()
6
>>> x[1, 2]
Solution 6
>>> y = x.view([('r', 'i1'),
- simple, **flexible**
... ('g', 'i1'),
... ('b', 'i1'),
... ('a', 'i1')]
... )[:, :, 0] C and Fortran order

>>> x = np.array([[1, 2, 3],

Warning: Another array taking exactly 4 bytes of memory: ... [4, 5, 6]], dtype=np.int16, order='C')
>>> x.strides
(6, 2)

8.1. Life of ndarray 288 8.1. Life of ndarray 289

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

>>> str(x.data) >>> y.strides

'\x01\x00\x02\x00\x03\x00\x04\x00\x05\x00\x06\x00' (-4,)

• Need to jump 6 bytes to find the next row >>> y = x[2:]

>>> y.__array_interface__['data'][0] - x.__array_interface__['data'][0]
• Need to jump 2 bytes to find the next column 8

>>> y = np.array(x, order='F') >>> x = np.zeros((10, 10, 10), dtype=np.float)

>>> y.strides >>> x.strides
(2, 4) (800, 80, 8)
>>> str(y.data) >>> x[::2,::3,::4].strides
'\x01\x00\x04\x00\x02\x00\x05\x00\x03\x00\x06\x00' (1600, 240, 32)

• Need to jump 2 bytes to find the next row • Similarly, transposes never make copies (it just swaps strides):
• Need to jump 4 bytes to find the next column
>>> x = np.zeros((10, 10, 10), dtype=np.float)
• Similarly to higher dimensions: >>> x.strides
(800, 80, 8)
– C: last dimensions vary fastest (= smaller strides) >>> x.T.strides
(8, 80, 800)
– F: first dimensions vary fastest
shape = (d 1 , d 2 , ..., d n ) But: not all reshaping operations can be represented by playing with strides:
strides = (s 1 , s 2 , ..., s n )
>>> a = np.arange(6, dtype=np.int8).reshape(3, 2)
s Cj = d j +1 d j +2 ...d n × itemsize >>> b = a.T
>>> b.strides
s Fj = d 1 d 2 ...d j −1 × itemsize (1, 2)

Note: Now we can understand the behavior of .view(): So far, so good. However:

>>> y = np.array([[1, 3], [2, 4]], dtype=np.uint8).transpose() >>> str(a.data)

>>> x = y.copy() '\x00\x01\x02\x03\x04\x05'
>>> b
Transposition does not affect the memory layout of the data, only strides array([[0, 2, 4],
[1, 3, 5]], dtype=int8)
>>> x.strides >>> c = b.reshape(3*2)
(2, 1) >>> c
>>> y.strides array([0, 2, 4, 1, 3, 5], dtype=int8)
(1, 2)
Here, there is no way to represent the array c given one stride and the block of memory for a. Therefore, the
>>> str(x.data) reshape operation needs to make a copy here.
'\x01\x02\x03\x04'
>>> str(y.data)
Example: fake dimensions with strides
'\x01\x03\x02\x04'

Stride manipulation
• the results are different when interpreted as 2 of int16
• .copy() creates new arrays in the C order (by default)
>>> from numpy.lib.stride_tricks import as_strided
>>> help(as_strided)
as_strided(x, shape=None, strides=None)
Make an ndarray from the given array with the given shape and strides
Slicing with integers

• Everything can be represented by changing only shape, strides, and possibly adjusting the data
Warning: as_strided does not check that you stay inside the memory block bounds. . .
pointer!
• Never makes copies of the data
>>> x = np.array([1, 2, 3, 4], dtype=np.int16)
>>> x = np.array([1, 2, 3, 4, 5, 6], dtype=np.int32) >>> as_strided(x, strides=(2*2, ), shape=(2, ))
>>> y = x[::-1] array([1, 3], dtype=int16)
>>> y >>> x[::2]
array([6, 5, 4, 3, 2, 1], dtype=int32) array([1, 3], dtype=int16)

8.1. Life of ndarray 290 8.1. Life of ndarray 291

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

See also: • Internally, array broadcasting is indeed implemented using 0-strides.

stride-fakedims.py
Exercise More tricks: diagonals

array([1, 2, 3, 4], dtype=np.int8) See also:

-> array([[1, 2, 3, 4], stride-diagonals.py

[1, 2, 3, 4], Challenge
[1, 2, 3, 4]], dtype=np.int8)
• Pick diagonal entries of the matrix: (assume C memory order):
using only as_strided.:
>>> x = np.array([[1, 2, 3],
Hint: byte_offset = stride[0]*index[0] + stride[1]*index[1] + ... ... [4, 5, 6],
... [7, 8, 9]], dtype=np.int32)
Spoiler
>>> x_diag = as_strided(x, shape=(3,), strides=(???,))
Stride can also be 0:
• Pick the first super-diagonal entries [2, 6].
>>> x = np.array([1, 2, 3, 4], dtype=np.int8)
>>> y = as_strided(x, strides=(0, 1), shape=(3, 4)) • And the sub-diagonals?
>>> y
array([[1, 2, 3, 4], (Hint to the last two: slicing first moves the point where striding starts from.)
[1, 2, 3, 4], Solution
[1, 2, 3, 4]], dtype=int8)
>>> y.base.base is x Pick diagonals:
True
>>> x_diag = as_strided(x, shape=(3, ), strides=((3+1)*x.itemsize, ))
>>> x_diag
array([1, 5, 9], dtype=int32)
Broadcasting
Slice first, to adjust the data pointer:
• Doing something useful with it: outer product of [1, 2, 3, 4] and [5, 6, 7]
>>> as_strided(x[0, 1:], shape=(2, ), strides=((3+1)*x.itemsize, ))
>>> x = np.array([1, 2, 3, 4], dtype=np.int16) array([2, 6], dtype=int32)
>>> x2 = as_strided(x, strides=(0, 1*2), shape=(3, 4))
>>> x2 >>> as_strided(x[1:, 0], shape=(2, ), strides=((3+1)*x.itemsize, ))
array([[1, 2, 3, 4], array([4, 8], dtype=int32)
[1, 2, 3, 4],
[1, 2, 3, 4]], dtype=int16)

Note: Using np.diag

>>> y = np.array([5, 6, 7], dtype=np.int16)
>>> y2 = as_strided(y, strides=(1*2, 0), shape=(3, 4)) >>> y = np.diag(x, k=1)
>>> y2 >>> y
array([[5, 5, 5, 5], array([2, 6], dtype=int32)
[6, 6, 6, 6],
[7, 7, 7, 7]], dtype=int16) However,

>>> x2 * y2 >>> y.flags.owndata

array([[ 5, 10, 15, 20], False
[ 6, 12, 18, 24],
[ 7, 14, 21, 28]], dtype=int16) Note This behavior has changed: before numpy 1.9, np.diag would make a copy.

See also:
. . . seems somehow familiar . . .
stride-diagonals.py

>>> x = np.array([1, 2, 3, 4], dtype=np.int16) Challenge

>>> y = np.array([5, 6, 7], dtype=np.int16)
Compute the tensor trace:
>>> x[np.newaxis,:] * y[:,np.newaxis]
array([[ 5, 10, 15, 20],
[ 6, 12, 18, 24],
[ 7, 14, 21, 28]], dtype=int16)

8.1. Life of ndarray 292 8.1. Life of ndarray 293

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Example: inplace operations (caveat emptor)

>>> x = np.arange(5*5*5*5).reshape(5, 5, 5, 5)
>>> s = 0
>>> for i in range(5): • Sometimes,
... for j in range(5):
>>> a -= b
... s += x[j, i, j, i]

by striding, and using sum() on the result. is not the same as

>>> y = as_strided(x, shape=(5, 5), strides=(TODO, TODO)) >>> a -= b.copy()

>>> s2 = ...
>>> assert s == s2 >>> x = np.array([[1, 2], [3, 4]])
>>> x -= x.transpose()
Solution >>> x
array([[ 0, -1],
>>> y = as_strided(x, shape=(5, 5), strides=((5*5*5 + 5)*x.itemsize, [ 4, 0]])
... (5*5 + 1)*x.itemsize))
>>> s2 = y.sum()
>>> y = np.array([[1, 2], [3, 4]])
>>> y -= y.T.copy()
>>> y
CPU cache effects array([[ 0, -1],
[ 1, 0]])
Memory layout can affect performance:
• x and x.transpose() share data
In [1]: x = np.zeros((20000,))
• x -= x.transpose() modifies the data element-by-element. . .
In [2]: y = np.zeros((20000*67,))[::67] • because x and x.transpose() have different striding, modified data re-appears on the RHS
In [3]: x.shape, y.shape
((20000,), (20000,))
8.1.5 Findings in dissection
In [4]: %timeit x.sum()
100000 loops, best of 3: 0.180 ms per loop

In [5]: %timeit y.sum()

100000 loops, best of 3: 2.34 ms per loop

In [6]: x.strides, y.strides

((8,), (536,))

Smaller strides are faster?

• memory block: may be shared, .base, .data

• data type descriptor: structured data, sub-arrays, byte order, casting, viewing, .astype(), .view()
• strided indexing: strides, C/F-order, slicing w/ integers, as_strided, broadcasting, stride tricks, diag,
CPU cache coherence

• CPU pulls data from main memory to its cache in blocks

8.2 Universal functions
• If many array items consecutively operated on fit in a single block (small stride):
– ⇒ fewer transfers needed
8.2.1 What they are?
– ⇒ faster
• Ufunc performs and elementwise operation on all elements of an array.
See also:
Examples:
numexpr is designed to mitigate cache effects in array computing.

8.1. Life of ndarray 294 8.2. Universal functions 295

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

np.add, np.subtract, scipy.special.*, ... PyUfunc_f_f float elementwise_func(float input_1)

PyUfunc_ff_f float elementwise_func(float input_1, float input_2)
• Automatically support: broadcasting, casting, . . . PyUfunc_d_d double elementwise_func(double input_1)
• The author of an ufunc only has to supply the elementwise operation, NumPy takes care of the rest. PyUfunc_dd_d double elementwise_func(double input_1, double input_2)
PyUfunc_D_D elementwise_func(npy_cdouble *input, npy_cdouble* output)
• The elementwise operation needs to be implemented in C (or, e.g., Cython)
PyUfunc_DD_D elementwise_func(npy_cdouble *in1, npy_cdouble *in2, npy_cdouble*
out)
Parts of an Ufunc
• Only elementwise_func needs to be supplied
1. Provided by user • . . . except when your elementwise function is not in one of the above forms

void ufunc_loop(void **args, int *dimensions, int *steps, void *data)

{ 8.2.2 Exercise: building an ufunc from scratch
/*
* int8 output = elementwise_function(int8 input_1, int8 input_2)
The Mandelbrot fractal is defined by the iteration
*
* This function must compute the ufunc for many values at once,
* in the way shown below. z ← z2 + c
*/
char *input_1 = (char*)args[0]; where c = x + i y is a complex number. This iteration is repeated – if z stays finite no matter how long the
char *input_2 = (char*)args[1]; iteration runs, c belongs to the Mandelbrot set.
char *output = (char*)args[2];
• Make ufunc called mandel(z0, c) that computes:
int i;
z = z0
for (i = 0; i < dimensions[0]; ++i) { for k in range(iterations):
*output = elementwise_function(*input_1, *input_2); z = z*z + c
input_1 += steps[0];
input_2 += steps[1];
say, 100 iterations or until z.real**2 + z.imag**2 > 1000. Use it to determine which c are in the
output += steps[2];
} Mandelbrot set.
} • Our function is a simple one, so make use of the PyUFunc_* helpers.

2. The NumPy part, built by • Write it in Cython

See also:
char types[3]
mandel.pyx, mandelplot.py
types[0] = NPY_BYTE /* type of first input arg */
types[1] = NPY_BYTE /* type of second input arg */ #
types[2] = NPY_BYTE /* type of third input arg */ # Fix the parts marked by TODO
#
PyObject *python_ufunc = PyUFunc_FromFuncAndData(
ufunc_loop, #
NULL, # Compile this file by (Cython >= 0.12 required because of the complex vars)
types, #
1, /* ntypes */ # cython mandel.pyx
2, /* num_inputs */ # python setup.py build_ext -i
1, /* num_outputs */ #
identity_element, # and try it out with, in this directory,
name, #
docstring, # >>> import mandel
unused) # >>> mandel.mandel(0, 1 + 2j)
#
• A ufunc can also support multiple different input-output type combinations. #

# The elementwise function

Making it easier # ------------------------

cdef void mandel_single_point(double complex *z_in,

3. ufunc_loop is of very generic form, and NumPy provides pre-made ones
double complex *c_in,
double complex *z_out) nogil:
#

8.2. Universal functions 296 8.2. Universal functions 297

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

# The Mandelbrot iteration PyUFunc_GG_G)

#

# # Required module initialization

# Some points of note: # ------------------------------
#
# - It's *NOT* allowed to call any Python functions here. import_array()
# import_ufunc()
# The Ufunc loop runs with the Python Global Interpreter Lock released.
# Hence, the ``nogil``.
# # The actual ufunc declaration
# - And so all local variables must be declared with ``cdef`` # ----------------------------
#
# - Note also that this function receives *pointers* to the data cdef PyUFuncGenericFunction loop_func[1]
# cdef char input_output_types[3]
cdef void *elementwise_funcs[1]
cdef double complex z = z_in[0]
cdef double complex c = c_in[0] #
cdef int k # the integer we use in the for loop # Reminder: some pre-made Ufunc loops:
#
# # ================ =======================================================
# TODO: write the Mandelbrot iteration for one point here, # ``PyUfunc_f_f`` ``float elementwise_func(float input_1)``
# as you would write it in Python. # ``PyUfunc_ff_f`` ``float elementwise_func(float input_1, float input_2)``
# # ``PyUfunc_d_d`` ``double elementwise_func(double input_1)``
# Say, use 100 as the maximum number of iterations, and 1000 # ``PyUfunc_dd_d`` ``double elementwise_func(double input_1, double input_2)``
# as the cutoff for z.real**2 + z.imag**2. # ``PyUfunc_D_D`` ``elementwise_func(complex_double *input, complex_double* complex_double)``
# # ``PyUfunc_DD_D`` ``elementwise_func(complex_double *in1, complex_double *in2, complex_
,→double* out)``
TODO: mandelbrot iteration should go here # ================ =======================================================
#
# Return the answer for this point # The full list is above.
z_out[0] = z #
#
# Type codes:
# Boilerplate Cython definitions #
# # NPY_BOOL, NPY_BYTE, NPY_UBYTE, NPY_SHORT, NPY_USHORT, NPY_INT, NPY_UINT,
# Pulls definitions from the Numpy C headers. # NPY_LONG, NPY_ULONG, NPY_LONGLONG, NPY_ULONGLONG, NPY_FLOAT, NPY_DOUBLE,
# ------------------------------------------- # NPY_LONGDOUBLE, NPY_CFLOAT, NPY_CDOUBLE, NPY_CLONGDOUBLE, NPY_DATETIME,
# NPY_TIMEDELTA, NPY_OBJECT, NPY_STRING, NPY_UNICODE, NPY_VOID
from numpy cimport import_array, import_ufunc #
from numpy cimport (PyUFunc_FromFuncAndData,
PyUFuncGenericFunction) loop_func[0] = ... TODO: suitable PyUFunc_* ...
from numpy cimport NPY_CDOUBLE, NP_DOUBLE, NPY_LONG input_output_types[0] = ... TODO ...
... TODO: fill in rest of input_output_types ...
# Import all pre-defined loop functions
# you won't need all of them - keep the relevant ones # This thing is passed as the ``data`` parameter for the generic
# PyUFunc_* loop, to let it know which function it should call.
from numpy cimport ( elementwise_funcs[0] = <void*>mandel_single_point
PyUFunc_f_f_As_d_d,
PyUFunc_d_d, # Construct the ufunc:
PyUFunc_f_f,
PyUFunc_g_g, mandel = PyUFunc_FromFuncAndData(
PyUFunc_F_F_As_D_D, loop_func,
PyUFunc_F_F, elementwise_funcs,
PyUFunc_D_D, input_output_types,
PyUFunc_G_G, 1, # number of supported input types
PyUFunc_ff_f_As_dd_d, TODO, # number of input args
PyUFunc_ff_f, TODO, # number of output args
PyUFunc_dd_d, 0, # `identity` element, never mind this
PyUFunc_gg_g, "mandel", # function name
PyUFunc_FF_F_As_DD_D, "mandel(z, c) -> computes z*z + c", # docstring
PyUFunc_DD_D, 0 # unused
PyUFunc_FF_F, )

8.2. Universal functions 298 8.2. Universal functions 299

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

# Boilerplate Cython definitions

Reminder: some pre-made Ufunc loops: #
# Pulls definitions from the Numpy C headers.
# -------------------------------------------
PyUfunc_f_f float elementwise_func(float input_1)
PyUfunc_ff_f float elementwise_func(float input_1, float input_2) from numpy cimport import_array, import_ufunc
PyUfunc_d_d double elementwise_func(double input_1) from numpy cimport (PyUFunc_FromFuncAndData,
PyUfunc_dd_d double elementwise_func(double input_1, double input_2) PyUFuncGenericFunction)
PyUfunc_D_D elementwise_func(complex_double *input, complex_double* output) from numpy cimport NPY_CDOUBLE
PyUfunc_DD_D elementwise_func(complex_double *in1, complex_double *in2, from numpy cimport PyUFunc_DD_D
complex_double* out)
# Required module initialization
# ------------------------------
Type codes:
import_array()
NPY_BOOL, NPY_BYTE, NPY_UBYTE, NPY_SHORT, NPY_USHORT, NPY_INT, NPY_UINT, import_ufunc()
NPY_LONG, NPY_ULONG, NPY_LONGLONG, NPY_ULONGLONG, NPY_FLOAT, NPY_DOUBLE,
NPY_LONGDOUBLE, NPY_CFLOAT, NPY_CDOUBLE, NPY_CLONGDOUBLE, NPY_DATETIME,
NPY_TIMEDELTA, NPY_OBJECT, NPY_STRING, NPY_UNICODE, NPY_VOID # The actual ufunc declaration
# ----------------------------

8.2.3 Solution: building an ufunc from scratch cdef PyUFuncGenericFunction loop_func[1]

cdef char input_output_types[3]
cdef void *elementwise_funcs[1]
# The elementwise function
# ------------------------ loop_func[0] = PyUFunc_DD_D

cdef void mandel_single_point(double complex *z_in, input_output_types[0] = NPY_CDOUBLE

double complex *c_in, input_output_types[1] = NPY_CDOUBLE
double complex *z_out) nogil: input_output_types[2] = NPY_CDOUBLE
#
# The Mandelbrot iteration elementwise_funcs[0] = <void*>mandel_single_point
#
mandel = PyUFunc_FromFuncAndData(
# loop_func,
# Some points of note: elementwise_funcs,
# input_output_types,
# - It's *NOT* allowed to call any Python functions here. 1, # number of supported input types
# 2, # number of input args
# The Ufunc loop runs with the Python Global Interpreter Lock released. 1, # number of output args
# Hence, the ``nogil``. 0, # `identity` element, never mind this
# "mandel", # function name
# - And so all local variables must be declared with ``cdef`` "mandel(z, c) -> computes iterated z*z + c", # docstring
# 0 # unused
# - Note also that this function receives *pointers* to the data; )
# the "traditional" solution to passing complex variables around
#
"""
cdef double complex z = z_in[0] Plot Mandelbrot
cdef double complex c = c_in[0] ================
cdef int k # the integer we use in the for loop
Plot the Mandelbrot ensemble.
# Straightforward iteration
"""
for k in range(100):
z = z*z + c import numpy as np
if z.real**2 + z.imag**2 > 1000: import mandel
break x = np.linspace(-1.7, 0.6, 1000)
y = np.linspace(-1.4, 1.4, 1000)
# Return the answer for this point c = x[None,:] + 1j*y[:,None]
z_out[0] = z z = mandel.mandel(c, c)

import matplotlib.pyplot as plt

8.2. Universal functions 300 8.2. Universal functions 301

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

plt.imshow(abs(z)**2 < 1000, extent=[-1.7, 0.6, -1.4, 1.4]) 2, # number of input args
plt.gray() 1, # number of output args
plt.show() 0, # `identity` element, never mind this
"mandel", # function name
"mandel(z, c) -> computes iterated z*z + c", # docstring
0 # unused
)

8.2.4 Generalized ufuncs

ufunc
output = elementwise_function(input)
Both output and input can be a single array element only.
generalized ufunc
output and input can be arrays with a fixed number of dimensions
For example, matrix trace (sum of diag elements):
Note: Most of the boilerplate could be automated by these Cython modules:
input shape = (n, n)
http://wiki.cython.org/MarkLodato/CreatingUfuncs output shape = () i.e. scalar

(n, n) -> ()

Several accepted input types Matrix product:

input_1 shape = (m, n)

E.g. supporting both single- and double-precision versions input_2 shape = (n, p)
output shape = (m, p)
cdef void mandel_single_point(double complex *z_in,
double complex *c_in,
(m, n), (n, p) -> (m, p)
double complex *z_out) nogil:
...
• This is called the “signature” of the generalized ufunc
cdef void mandel_single_point_singleprec(float complex *z_in, • The dimensions on which the g-ufunc acts, are “core dimensions”
float complex *c_in,
float complex *z_out) nogil:
... Status in NumPy
cdef PyUFuncGenericFunction loop_funcs[2]
cdef char input_output_types[3*2] • g-ufuncs are in NumPy already . . .
cdef void *elementwise_funcs[1*2]
• new ones can be created with PyUFunc_FromFuncAndDataAndSignature
loop_funcs[0] = PyUFunc_DD_D • most linear-algebra functions are implemented as g-ufuncs to enable working with stacked arrays:
input_output_types[0] = NPY_CDOUBLE
input_output_types[1] = NPY_CDOUBLE >>> import numpy as np
input_output_types[2] = NPY_CDOUBLE >>> np.linalg.det(np.random.rand(3, 5, 5))
elementwise_funcs[0] = <void*>mandel_single_point array([ 0.00965823, -0.13344729, 0.04583961])
>>> np.linalg._umath_linalg.det.signature
loop_funcs[1] = PyUFunc_FF_F '(m,m)->()'
input_output_types[3] = NPY_CFLOAT
input_output_types[4] = NPY_CFLOAT • we also ship with a few g-ufuncs for testing, ATM:
input_output_types[5] = NPY_CFLOAT
elementwise_funcs[1] = <void*>mandel_single_point_singleprec >>> import numpy.core.umath_tests as ut
>>> ut.matrix_multiply.signature
mandel = PyUFunc_FromFuncAndData( '(m,n),(n,p)->(m,p)'
loop_func,
elementwise_funcs, >>> x = np.ones((10, 2, 4))
input_output_types, >>> y = np.ones((10, 4, 5))
2, # number of supported input types <----------------

8.2. Universal functions 302 8.2. Universal functions 303

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

>>> ut.matrix_multiply(x, y).shape

8.3.2 The old buffer protocol
(10, 2, 5)
• Only 1-D buffers
• in both examples the last two dimensions became core dimensions, and are modified as per the signature • No data type information
• otherwise, the g-ufunc operates “elementwise” • C-level interface; PyBufferProcs tp_as_buffer in the type object
• matrix multiplication this way could be useful for operating on many small matrices at once • But it’s integrated into Python (e.g. strings support it)
Mini-exercise using Pillow (Python Imaging Library):
Generalized ufunc loop
See also:

Matrix multiplication (m,n),(n,p) -> (m,p) pilbuffer.py

void gufunc_loop(void **args, int *dimensions, int *steps, void *data) >>> from PIL import Image
{ >>> data = np.zeros((200, 200, 4), dtype=np.int8)
char *input_1 = (char*)args[0]; /* these are as previously */ >>> data[:, :] = [255, 0, 0, 255] # Red
char *input_2 = (char*)args[1]; >>> # In PIL, RGBA images consist of 32-bit integers whose bytes are [RR,GG,BB,AA]
char *output = (char*)args[2]; >>> data = data.view(np.int32).squeeze()
>>> img = Image.frombuffer("RGBA", (200, 200), data, "raw", "RGBA", 0, 1)
int input_1_stride_m = steps[3]; /* strides for the core dimensions */ >>> img.save('test.png')
int input_1_stride_n = steps[4]; /* are added after the non-core */
int input_2_strides_n = steps[5]; /* steps */ Q:
int input_2_strides_p = steps[6];
Check what happens if data is now modified, and img saved again.
int output_strides_n = steps[7];
int output_strides_p = steps[8];

int m = dimension[1]; /* core dimensions are added after */

8.3.3 The old buffer protocol
int n = dimension[2]; /* the main dimension; order as in */
int p = dimension[3]; /* signature */ """
From buffer
int i; ============

for (i = 0; i < dimensions[0]; ++i) { Show how to exchange data between numpy and a library that only knows
matmul_for_strided_matrices(input_1, input_2, output, the buffer interface.
strides for each array...); """

input_1 += steps[0]; import numpy as np

input_2 += steps[1]; import Image
output += steps[2];
} # Let's make a sample image, RGBA format
}
x = np.zeros((200, 200, 4), dtype=np.int8)

x[:,:,0] = 254 # red

8.3 Interoperability features x[:,:,3] = 255 # opaque

data = x.view(np.int32) # Check that you understand why this is OK!

8.3.1 Sharing multidimensional, typed data
img = Image.frombuffer("RGBA", (200, 200), data)
Suppose you img.save('test.png')

1. Write a library than handles (multidimensional) binary data, #

# Modify the original data, and save again.
2. Want to make it easy to manipulate the data with NumPy, or whatever other library, #
3. . . . but would not like to have NumPy as a dependency. # It turns out that PIL, which knows next to nothing about Numpy,
# happily shares the same data.
Currently, 3 solutions: #
1. the “old” buffer interface
x[:,:,1] = 254
2. the array interface img.save('test2.png')

3. the “new” buffer interface (PEP 3118)

8.3. Interoperability features 304 8.3. Interoperability features 305

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

>>> from PIL import Image

>>> img = Image.open('data/test.png')
>>> img.__array_interface__
{'data': ...,
'shape': (200, 200, 4),
'typestr': '|u1'}
>>> x = np.asarray(img)
>>> x.shape
(200, 200, 4)

Note: A more C-friendly variant of the array interface is also defined.

8.4 Array siblings: chararray, maskedarray, matrix

8.4.1 chararray: vectorized string operations

>>> x = np.array(['a', ' bbb', ' ccc']).view(np.chararray)

>>> x.lstrip(' ')
chararray(['a', 'bbb', 'ccc'],
dtype='...')
>>> x.upper()
chararray(['A', ' BBB', ' CCC'],
dtype='...')

Note: .view() has a second meaning: it can make an ndarray an instance of a specialized ndarray subclass

8.4.2 masked_array missing data

Masked arrays are arrays that may have missing or invalid entries.

8.3.4 Array interface protocol For example, suppose we have an array where the fourth entry is invalid:

>>> x = np.array([1, 2, 3, -99, 5])

• Multidimensional buffers
• Data type information present One way to describe this is to create a masked array:

• NumPy-specific approach; slowly deprecated (but not going away) >>> mx = np.ma.masked_array(x, mask=[0, 0, 0, 1, 0])
>>> mx
• Not integrated in Python otherwise masked_array(data = [1 2 3 -- 5],
See also: mask = [False False False True False],
fill_value = 999999)
Documentation: http://docs.scipy.org/doc/numpy/reference/arrays.interface.html
Masked mean ignores masked data:
>>> x = np.array([[1, 2], [3, 4]])
>>> x.__array_interface__ >>> mx.mean()
{'data': (171694552, False), # memory address of data, is readonly? 2.75
'descr': [('', '<i4')], # data type descriptor >>> np.mean(mx)
'typestr': '<i4', # same, in another form 2.75
'strides': None, # strides; or None if in C-order
'shape': (2, 2),
'version': 3,
} Warning: Not all NumPy functions respect masks, for instance np.dot, so check the return types.

::
The masked_array returns a view to the original array:

8.3. Interoperability features 306 8.4. Array siblings: chararray, maskedarray, matrix 307
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

>>> mx[1] = 9 Example: Masked statistics

>>> x
array([ 1, 9, 3, -99, 5]) Canadian rangers were distracted when counting hares and lynxes in 1903-1910 and 1917-1918, and got
the numbers are wrong. (Carrot farmers stayed alert, though.) Compute the mean populations over time,
ignoring the invalid numbers.
The mask >>> data = np.loadtxt('data/populations.txt')
>>> populations = np.ma.masked_array(data[:,1:])
You can modify the mask by assigning: >>> year = data[:, 0]

>>> mx[1] = np.ma.masked >>> bad_years = (((year >= 1903) & (year <= 1910))
>>> mx ... | ((year >= 1917) & (year <= 1918)))
masked_array(data = [1 -- 3 -- 5], >>> # '&' means 'and' and '|' means 'or'
mask = [False True False True False], >>> populations[bad_years, 0] = np.ma.masked
fill_value = 999999) >>> populations[bad_years, 1] = np.ma.masked

The mask is cleared on assignment: >>> populations.mean(axis=0)

masked_array(data = [40472.72727272727 18627.272727272728 42400.0],
>>> mx[1] = 9 mask = [False False False],
>>> mx fill_value = 1e+20)
masked_array(data = [1 9 3 -- 5],
mask = [False False False True False], >>> populations.std(axis=0)
fill_value = 999999) masked_array(data = [21087.656489006717 15625.799814240254 3322.5062255844787],
mask = [False False False],
The mask is also available directly: fill_value = 1e+20)

>>> mx.mask Note that Matplotlib knows about masked arrays:

array([False, False, False, True, False], dtype=bool)
>>> plt.plot(year, populations, 'o-')
[<matplotlib.lines.Line2D object at ...>, ...]
The masked entries can be filled with a given value to get an usual array back:

>>> x2 = mx.filled(-1)
>>> x2
array([ 1, 9, 3, -1, 5])

The mask can also be cleared:

>>> mx.mask = np.ma.nomask

>>> mx
masked_array(data = [1 9 3 -99 5],
mask = [False False False False False],
fill_value = 999999)

Domain-aware functions

The masked array package also contains domain-aware functions:

>>> np.ma.log(np.array([1, 2, -1, -2, 3, -5])) 8.4.3 recarray: purely convenience

masked_array(data = [0.0 0.6931471805599453 -- -- 1.0986122886681098 --],
mask = [False False True True False True],
fill_value = 1e+20) >>> arr = np.array([('a', 1), ('b', 2)], dtype=[('x', 'S1'), ('y', int)])
>>> arr2 = arr.view(np.recarray)
>>> arr2.x
chararray(['a', 'b'],
Note: Streamlined and more seamless support for dealing with missing data in arrays is making its way into
dtype='|S1')
NumPy 1.7. Stay tuned! >>> arr2.y
array([1, 2])

8.4. Array siblings: chararray, maskedarray, matrix 308 8.4. Array siblings: chararray, maskedarray, matrix 309
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

8.4.4 matrix: convenience? it fails with a cryptic error message::

• always 2-D >>> np.random.permutation(12)

• * is the matrix product, not the elementwise one array([11, 5, 8, 4, 6, 1, 9, 3, 7, 2, 10, 0])
>>> np.random.permutation(12.)
>>> np.matrix([[1, 0], [0, 1]]) * np.matrix([[1, 2], [3, 4]]) Traceback (most recent call last):
matrix([[1, 2], File "<stdin>", line 1, in <module>
[3, 4]]) File "mtrand.pyx", line 3311, in mtrand.RandomState.permutation
File "mtrand.pyx", line 3254, in mtrand.RandomState.shuffle
TypeError: len() of unsized object

8.5 Summary This also happens with long arguments, and so

np.random.permutation(X.shape[0]) where X is an array fails on 64
bit windows (where shape is a tuple of longs).
• Anatomy of the ndarray: data, dtype, strides.
It would be great if it could cast to integer or at least raise a
• Universal functions: elementwise operations, how to make new ones proper error for non-integer types.
• Ndarray subclasses
I'm using NumPy 1.4.1, built from the official tarball, on Windows
• Various buffer interfaces for integration with other tools 64 with Visual studio 2008, on Python.org 64-bit Python.
• Recent additions: PEP 3118, generalized ufuncs
0. What are you trying to do?
1. Small code snippet reproducing the bug (if possible)
8.6 Contributing to NumPy/Scipy • What actually happens
• What you’d expect
Get this tutorial: http://www.euroscipy.org/talk/882
2. Platform (Windows / Linux / OSX, 32/64 bits, x86/PPC, . . . )

8.6.1 Why 3. Version of NumPy/Scipy

>>> print(np.__version__)
• “There’s a bug?” 1...
• “I don’t understand what this is supposed to do?”
Check that the following is what you expect
• “I have this fancy code. Would you like to have it?”
>>> print(np.__file__)
• “I’d like to help! What can I do?” /...

In case you have old/broken NumPy installations lying around.

8.6.2 Reporting bugs
If unsure, try to remove existing NumPy installations, and reinstall. . .
• Bug tracker (prefer this)
– http://projects.scipy.org/numpy 8.6.3 Contributing to documentation
– http://projects.scipy.org/scipy
1. Documentation editor
– Click the “Register” link to get an account
• http://docs.scipy.org/doc/numpy
• Mailing lists ( scipy.org/Mailing_Lists )
• Registration
– If you’re unsure
– Register an account
– No replies in a week or so? Just file a bug ticket.
– Subscribe to scipy-dev mailing list (subscribers-only)

Good bug report – Problem with mailing lists: you get mail

* But: you can turn mail delivery off

Title: numpy.random.permutations fails for non-integer arguments
* “change your subscription options”, at the bottom of
I'm trying to generate random permutations, using numpy.random.permutations http://mail.scipy.org/mailman/listinfo/scipy-dev

When calling numpy.random.permutation with non-integer arguments – Send a mail @ scipy-dev mailing list; ask for activation:

8.5. Summary 310 8.6. Contributing to NumPy/Scipy 311

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

– API docs: improvements to docstrings

To: scipy-dev@scipy.org
* Know some Scipy module well?
Hi,
– User guide
I'd like to edit NumPy/Scipy docstrings. My account is XXXXX
* Needs to be done eventually.
Cheers, * Want to think? Come up with a Table of Contents
N. N.
http://scipy.org/Developer_Zone/UG_Toc
• Check the style guide: • Ask on communication channels:
– http://docs.scipy.org/doc/numpy/ – numpy-discussion list
– Don’t be intimidated; to fix a small thing, just fix it – scipy-dev list
• Edit
2. Edit sources and send patches (as for bugs)
3. Complain on the mailing list

8.6.4 Contributing features

0. Ask on mailing list, if unsure where it should go

1. Write a patch, add an enhancement ticket on the bug tracket
2. OR, create a Git branch implementing the feature + add enhancement ticket.
• Especially for big/invasive additions
• http://projects.scipy.org/numpy/wiki/GitMirror
• http://www.spheredev.org/wiki/Git_for_the_lazy

# Clone numpy repository

git clone --origin svn http://projects.scipy.org/git/numpy.git numpy
cd numpy

# Create a feature branch

git checkout -b name-of-my-feature-branch svn/trunk

<edit stuff>

git commit -a

• Create account on https://github.com (or anywhere)

• Create a new repository @ Github
• Push your work to github

git remote add github git@github:USERNAME/REPOSITORYNAME.git

git push github name-of-my-feature-branch

8.6.5 How to help, in general

• Bug fixes always welcome!

– What irks you most
– Browse the tracker
• Documentation work

8.6. Contributing to NumPy/Scipy 312 8.6. Contributing to NumPy/Scipy 313

 Scipy lecture notes, Edition 2017.1

– Debugger commands and interaction

• Debugging segmentation faults using gdb

9
9.1 Avoiding bugs

9.1.1 Coding best practices to avoid getting in trouble

CHAPTER Brian Kernighan

“Everyone knows that debugging is twice as hard as writing a program in the first place. So if you’re as clever
as you can be when you write it, how will you ever debug it?”

• We all write buggy code. Accept it. Deal with it.

• Write your code with testing and debugging in mind.

Debugging code • Keep It Simple, Stupid (KISS).

– What is the simplest thing that could possibly work?
• Don’t Repeat Yourself (DRY).
– Every piece of knowledge must have a single, unambiguous, authoritative representation within a
system.
– Constants, algorithms, etc. . .

Author: Gaël Varoquaux • Try to limit interdependencies of your code. (Loose Coupling)

This section explores tools to understand better your code base: debugging, to find and fix bugs. • Give your variables, functions and modules meaningful names (not mathematics names)

It is not specific to the scientific Python community, but the strategies that we will employ are tailored to its
needs. 9.1.2 pyflakes: fast static analysis

They are several static analysis tools in Python; to name a few:

Prerequisites
• pylint
• Numpy
• pychecker
• IPython
• pyflakes
• nosetests
• flake8
• pyflakes
Here we focus on pyflakes, which is the simplest tool.
• gdb for the C-debugging part.
• Fast, simple
• Detects syntax errors, missing imports, typos on names.
Chapter contents Another good recommendation is the flake8 tool which is a combination of pyflakes and pep8. Thus, in addi-
tion to the types of errors that pyflakes catches, flake8 detects violations of the recommendation in PEP8 style
• Avoiding bugs guide.
– Coding best practices to avoid getting in trouble Integrating pyflakes (or flake8) in your editor or IDE is highly recommended, it does yield productivity gains.
– pyflakes: fast static analysis
• Debugging workflow Running pyflakes on the current edited file

• Using the Python debugger You can bind a key to run pyflakes in the current buffer.
– Invoking the debugger • In kate Menu: ‘settings -> configure kate

314 9.1. Avoiding bugs 315

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

– In plugins enable ‘external tools’

– In external Tools’, add pyflakes:

kdialog --title "pyflakes %f ilename" --msgbox "$(pyflakes %f ilename)"

• In TextMate
Menu: TextMate -> Preferences -> Advanced -> Shell variables, add a shell variable:

TM_PYCHECKER = /Library/Frameworks/Python.framework/Versions/Current/bin/pyflakes

Then Ctrl-Shift-V is binded to a pyflakes report

• In vim In your .vimrc (binds F5 to pyflakes):
• In emacs Use the flymake mode with pyflakes, documented on http://www.plope.com/Members/
autocmd FileType python let &mp = 'echo "*** running % ***" ; pyflakes %' chrism/flymake-mode : add the following to your .emacs file:
autocmd FileType tex,mp,rst,python imap <Esc>[15~ <C-O>:make!^M
autocmd FileType tex,mp,rst,python map <Esc>[15~ :make!^M (when (load "flymake" t)
autocmd FileType tex,mp,rst,python set autowrite (defun flymake-pyflakes-init ()
(let* ((temp-file (flymake-init-create-temp-buffer-copy
• In emacs In your .emacs (binds F5 to pyflakes): 'flymake-create-temp-inplace))
(local-file (file-relative-name
(defun pyflakes-thisfile () (interactive) temp-file
(compile (format "pyflakes %s " (buffer-file-name))) (file-name-directory buffer-file-name))))
) (list "pyflakes" (list local-file))))

(define-minor-mode pyflakes-mode (add-to-list 'flymake-allowed-file-name-masks

"Toggle pyflakes mode. '("\\.py\\'" flymake-pyflakes-init)))
With no argument, this command toggles the mode.
Non-null prefix argument turns on the mode. (add-hook 'find-file-hook 'flymake-find-file-hook)
Null prefix argument turns off the mode."
;; The initial value.
nil
;; The indicator for the mode line. 9.2 Debugging workflow
" Pyflakes"
;; The minor mode bindings.
'( ([f5] . pyflakes-thisfile) ) If you do have a non trivial bug, this is when debugging strategies kick in. There is no silver bullet. Yet, strategies
) help:

(add-hook 'python-mode-hook (lambda () (pyflakes-mode t))) For debugging a given problem, the favorable situation is when the problem is isolated in a
small number of lines of code, outside framework or application code, with short modify-run-
fail cycles

A type-as-go spell-checker like integration 1. Make it fail reliably. Find a test case that makes the code fail every time.
2. Divide and Conquer. Once you have a failing test case, isolate the failing code.
• In vim
• Which module.
– Use the pyflakes.vim plugin:
• Which function.
1. download the zip file from http://www.vim.org/scripts/script.php?script_id=2441
• Which line of code.
2. extract the files in ~/.vim/ftplugin/python
=> isolate a small reproducible failure: a test case
3. make sure your vimrc has filetype plugin indent on
3. Change one thing at a time and re-run the failing test case.
4. Use the debugger to understand what is going wrong.
5. Take notes and be patient. It may take a while.

Note: Once you have gone through this process: isolated a tight piece of code reproducing the bug and fix the
bug using this piece of code, add the corresponding code to your test suite.
– Alternatively: use the syntastic plugin. This can be configured to use flake8 too and also handles
on-the-fly checking for many other languages.

9.1. Avoiding bugs 316 9.2. Debugging workflow 317

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

9.3 Using the Python debugger 6

ipdb> list
The python debugger, pdb: https://docs.python.org/library/pdb.html, allows you to inspect your code inter-
1 """Small snippet to raise an IndexError."""
actively.
2
Specifically it allows you to: 3 def index_error():
4 lst = list('foobar')
• View the source code. ----> 5 print lst[len(lst)]
6
• Walk up and down the call stack.
7 if __name__ == '__main__':
• Inspect values of variables. 8 index_error()
9
• Modify values of variables.
• Set breakpoints. ipdb> len(lst)
6
ipdb> print lst[len(lst)-1]
print r
ipdb> quit
Yes, print statements do work as a debugging tool. However to inspect runtime, it is often more efficient to
use the debugger. In [3]:

Post-mortem debugging without IPython

9.3.1 Invoking the debugger
In some situations you cannot use IPython, for instance to debug a script that wants to be called from the
Ways to launch the debugger: command line. In this case, you can call the script with python -m pdb script.py:

1. Postmortem, launch debugger after module errors. $ python -m pdb index_error.py

> /home/varoquau/dev/scipy-lecture-notes/advanced/optimizing/index_error.py(1)<module>()
2. Launch the module with the debugger. -> """Small snippet to raise an IndexError."""
3. Call the debugger inside the module (Pdb) continue
Traceback (most recent call last):
File "/usr/lib/python2.6/pdb.py", line 1296, in main
Postmortem pdb._runscript(mainpyfile)
File "/usr/lib/python2.6/pdb.py", line 1215, in _runscript
self.run(statement)
Situation: You’re working in IPython and you get a traceback.
File "/usr/lib/python2.6/bdb.py", line 372, in run
Here we debug the file index_error.py. When running it, an IndexError is raised. Type %debug and drop exec cmd in globals, locals
into the debugger. File "<string>", line 1, in <module>
File "index_error.py", line 8, in <module>
In [1]: %run index_error.py index_error()
--------------------------------------------------------------------------- File "index_error.py", line 5, in index_error
IndexError Traceback (most recent call last) print lst[len(lst)]
/home/varoquau/dev/scipy-lecture-notes/advanced/debugging/index_error.py in <module>() IndexError: list index out of range
6 Uncaught exception. Entering post mortem debugging
7 if __name__ == '__main__': Running 'cont' or 'step' will restart the program
----> 8 index_error() > /home/varoquau/dev/scipy-lecture-notes/advanced/optimizing/index_error.py(5)index_error()
9 -> print lst[len(lst)]
(Pdb)
/home/varoquau/dev/scipy-lecture-notes/advanced/debugging/index_error.py in index_error()
3 def index_error():
4 lst = list('foobar')
----> 5 print lst[len(lst)] Step-by-step execution
6
7 if __name__ == '__main__':
Situation: You believe a bug exists in a module but are not sure where.
IndexError: list index out of range For instance we are trying to debug wiener_filtering.py. Indeed the code runs, but the filtering does not
work well.
In [2]: %debug
> /home/varoquau/dev/scipy-lecture-notes/advanced/debugging/index_error.py(5)index_error() • Run the script in IPython with the debugger using %run -d wiener_filtering.p :
4 lst = list('foobar')
In [1]: %run -d wiener_filtering.py
----> 5 print lst[len(lst)]
*** Blank or comment

9.3. Using the Python debugger 318 9.3. Using the Python debugger 319
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

*** Blank or comment ipdb> print l_var.min()

*** Blank or comment 0
Breakpoint 1 at /home/varoquau/dev/scipy-lecture-notes/advanced/optimizing/wiener_
,→filtering.py:4 Oh dear, nothing but integers, and 0 variation. Here is our bug, we are doing integer arithmetic.
NOTE: Enter 'c' at the ipdb> prompt to start your script.
> <string>(1)<module>()
Raising exception on numerical errors
• Set a break point at line 34 using b 34:
When we run the wiener_filtering.py file, the following warnings are raised:
ipdb> n
> /home/varoquau/dev/scipy-lecture-notes/advanced/optimizing/wiener_filtering.py(4) In [2]: %run wiener_filtering.py
,→<module>()
wiener_filtering.py:40: RuntimeWarning: divide by zero encountered in divide
3 noise_level = (1 - noise/l_var )
1---> 4 import numpy as np
We can turn these warnings in exception, which enables us to do post-mortem debugging on them, and find
5 import scipy as sp
our problem more quickly:
ipdb> b 34 In [3]: np.seterr(all='raise')
Breakpoint 2 at /home/varoquau/dev/scipy-lecture-notes/advanced/optimizing/wiener_ Out[3]: {'divide': 'print', 'invalid': 'print', 'over': 'print', 'under': 'ignore'}
,→filtering.py:34 In [4]: %run wiener_filtering.py
---------------------------------------------------------------------------
• Continue execution to next breakpoint with c(ont(inue)): FloatingPointError Traceback (most recent call last)
/home/esc/anaconda/lib/python2.7/site-packages/IPython/utils/py3compat.pyc in execfile(fname,
ipdb> c ,→ *where)
> /home/varoquau/dev/scipy-lecture-notes/advanced/optimizing/wiener_filtering. 176 else:
,→py(34)iterated_wiener() 177 filename = fname
33 """ --> 178 __builtin__.execfile(filename, *where)
2--> 34 noisy_img = noisy_img
35 denoised_img = local_mean(noisy_img, size=size) /home/esc/physique-cuso-python-2013/scipy-lecture-notes/advanced/debugging/wiener_filtering.
,→py in <module>()

• Step into code with n(ext) and s(tep): next jumps to the next statement in the current execution 55 pl.matshow(noisy_face[cut], cmap=pl.cm.gray)
context, while step will go across execution contexts, i.e. enable exploring inside function calls: 56
---> 57 denoised_face = iterated_wiener(noisy_face)
ipdb> s 58 pl.matshow(denoised_face[cut], cmap=pl.cm.gray)
> /home/varoquau/dev/scipy-lecture-notes/advanced/optimizing/wiener_filtering. 59
,→py(35)iterated_wiener()
2 34 noisy_img = noisy_img /home/esc/physique-cuso-python-2013/scipy-lecture-notes/advanced/debugging/wiener_filtering.
---> 35 denoised_img = local_mean(noisy_img, size=size) ,→py in iterated_wiener(noisy_img, size)

36 l_var = local_var(noisy_img, size=size) 38 res = noisy_img - denoised_img

39 noise = (res**2).sum()/res.size
ipdb> n ---> 40 noise_level = (1 - noise/l_var )
> /home/varoquau/dev/scipy-lecture-notes/advanced/optimizing/wiener_filtering. 41 noise_level[noise_level<0] = 0
,→py(36)iterated_wiener() 42 denoised_img += noise_level*res
35 denoised_img = local_mean(noisy_img, size=size)
---> 36 l_var = local_var(noisy_img, size=size) FloatingPointError: divide by zero encountered in divide
37 for i in range(3):

• Step a few lines and explore the local variables:

Other ways of starting a debugger
ipdb> n
> /home/varoquau/dev/scipy-lecture-notes/advanced/optimizing/wiener_filtering.
• Raising an exception as a poor man break point
,→py(37)iterated_wiener()
36 l_var = local_var(noisy_img, size=size) If you find it tedious to note the line number to set a break point, you can simply raise an exception at
---> 37 for i in range(3): the point that you want to inspect and use IPython’s %debug. Note that in this case you cannot step or
38 res = noisy_img - denoised_img continue the execution.
ipdb> print l_var
[[5868 5379 5316 ..., 5071 4799 5149] • Debugging test failures using nosetests
[5013 363 437 ..., 346 262 4355]
[5379 410 344 ..., 392 604 3377] You can run nosetests --pdb to drop in post-mortem debugging on exceptions, and nosetests
..., --pdb-failure to inspect test failures using the debugger.
[ 435 362 308 ..., 275 198 1632] In addition, you can use the IPython interface for the debugger in nose by installing the nose plugin
[ 548 392 290 ..., 248 263 1653]
ipdbplugin. You can than pass --ipdb and --ipdb-failure options to nosetests.
[ 466 789 736 ..., 1835 1725 1940]]

9.3. Using the Python debugger 320 9.3. Using the Python debugger 321
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

• Calling the debugger explicitly

Insert the following line where you want to drop in the debugger: Undocumented commands:
======================
import pdb; pdb.set_trace() retval rv

Warning: When running nosetests, the output is captured, and thus it seems that the debugger does not
work. Simply run the nosetests with the -s flag.
9.4 Debugging segmentation faults using gdb

If you have a segmentation fault, you cannot debug it with pdb, as it crashes the Python interpreter before it
Graphical debuggers and alternatives can drop in the debugger. Similarly, if you have a bug in C code embedded in Python, pdb is useless. For this
we turn to the gnu debugger, gdb, available on Linux.
• For stepping through code and inspecting variables, you might find it more convenient to use a graph- Before we start with gdb, let us add a few Python-specific tools to it. For this we add a few macros to our ~/.
ical debugger such as winpdb. gdbinit. The optimal choice of macro depends on your Python version and your gdb version. I have added a
• Alternatively, pudb is a good semi-graphical debugger with a text user interface in the console. simplified version in gdbinit, but feel free to read DebuggingWithGdb.

• Also, the pydbgr project is probably worth looking at. To debug with gdb the Python script segfault.py, we can run the script in gdb as follows

$ gdb python
...
(gdb) run segfault.py
9.3.2 Debugger commands and interaction Starting program: /usr/bin/python segfault.py
[Thread debugging using libthread_db enabled]

l(list) Lists the code at the current position Program received signal SIGSEGV, Segmentation fault.
u(p) Walk up the call stack _strided_byte_copy (dst=0x8537478 "\360\343G", outstrides=4, src=
d(own) Walk down the call stack 0x86c0690 <Address 0x86c0690 out of bounds>, instrides=32, N=3,
n(ext) Execute the next line (does not go down in new functions) elsize=4)
s(tep) Execute the next statement (goes down in new functions) at numpy/core/src/multiarray/ctors.c:365
bt Print the call stack 365 _FAST_MOVE(Int32);
a Print the local variables (gdb)
!command Execute the given Python command (by opposition to pdb commands
We get a segfault, and gdb captures it for post-mortem debugging in the C level stack (not the Python call
stack). We can debug the C call stack using gdb’s commands:

Warning: Debugger commands are not Python code (gdb) up

#1 0x004af4f5 in _copy_from_same_shape (dest=<value optimized out>,
You cannot name the variables the way you want. For instance, if in you cannot override the variables in src=<value optimized out>, myfunc=0x496780 <_strided_byte_copy>,
the current frame with the same name: use different names than your local variable when typing code in swap=0)
the debugger. at numpy/core/src/multiarray/ctors.c:748
748 myfunc(dit->dataptr, dest->strides[maxaxis],

As you can see, right now, we are in the C code of numpy. We would like to know what is the Python code that
Getting help when in the debugger
triggers this segfault, so we go up the stack until we hit the Python execution loop:
Type h or help to access the interactive help: (gdb) up
#8 0x080ddd23 in call_function (f=
ipdb> help Frame 0x85371ec, for file /home/varoquau/usr/lib/python2.6/site-packages/numpy/core/
,→arrayprint.py, line 156, in _leading_trailing (a=<numpy.ndarray at remote 0x85371b0>, _nc=
Documented commands (type help <topic>): ,→<module at remote 0xb7f93a64>), throwflag=0)
======================================== at ../Python/ceval.c:3750
EOF bt cont enable jump pdef r tbreak w 3750 ../Python/ceval.c: No such file or directory.
a c continue exit l pdoc restart u whatis in ../Python/ceval.c
alias cl d h list pinfo return unalias where
args clear debug help n pp run unt (gdb) up
b commands disable ignore next q s until #9 PyEval_EvalFrameEx (f=
break condition down j p quit step up Frame 0x85371ec, for file /home/varoquau/usr/lib/python2.6/site-packages/numpy/core/
,→arrayprint.py, line 156, in _leading_trailing (a=<numpy.ndarray at remote 0x85371b0>, _nc=
Miscellaneous help topics: ,→<module at remote 0xb7f93a64>), throwflag=0)
========================== at ../Python/ceval.c:2412
exec pdb

9.3. Using the Python debugger 322 9.4. Debugging segmentation faults using gdb 323
 Scipy lecture notes, Edition 2017.1

2412 in ../Python/ceval.c
(gdb)

Once we are in the Python execution loop, we can use our special Python helper function. For instance we can
find the corresponding Python code:

10
(gdb) pyframe
/home/varoquau/usr/lib/python2.6/site-packages/numpy/core/arrayprint.py (158): _leading_
,→trailing
(gdb)

This is numpy code, we need to go up until we find code that we have written:
CHAPTER
(gdb) up
...
(gdb) up
#34 0x080dc97a in PyEval_EvalFrameEx (f=
Frame 0x82f064c, for file segfault.py, line 11, in print_big_array (small_array=<numpy.
,→ndarray at remote 0x853ecf0>, big_array=<numpy.ndarray at remote 0x853ed20>), throwflag=0)␣
,→at ../Python/ceval.c:1630
1630 ../Python/ceval.c: No such file or directory.

Optimizing code
in ../Python/ceval.c
(gdb) pyframe
segfault.py (12): print_big_array

The corresponding code is:

def make_big_array(small_array):
big_array = stride_tricks.as_strided(small_array,
shape=(2e6, 2e6), strides=(32, 32))
return big_array

def print_big_array(small_array): Donald Knuth

big_array = make_big_array(small_array)
“Premature optimization is the root of all evil”
Thus the segfault happens when printing big_array[-10:]. The reason is simply that big_array has been
allocated with its end outside the program memory.
Author: Gaël Varoquaux
This chapter deals with strategies to make Python code go faster.
Note: For a list of Python-specific commands defined in the gdbinit, read the source of this file.

Prerequisites

• line_profiler
Wrap up exercise

The following script is well documented and hopefully legible. It seeks to answer a problem of actual interest
for numerical computing, but it does not work. . . Can you debug it? Chapters contents

Python source code: to_debug.py • Optimization workflow

• Profiling Python code
– Timeit
– Profiler
– Line-profiler
• Making code go faster
– Algorithmic optimization

* Example of the SVD

9.4. Debugging segmentation faults using gdb 324 325

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

• Writing faster numerical code from scipy import linalg

– Additional Links
from ica import fastica

def test():
10.1 Optimization workflow data = np.random.random((5000, 100))
u, s, v = linalg.svd(data)
pca = np.dot(u[:, :10].T, data)
1. Make it work: write the code in a simple legible ways.
results = fastica(pca.T, whiten=False)
2. Make it work reliably: write automated test cases, make really sure that your algorithm is right and that
if you break it, the tests will capture the breakage. if __name__ == '__main__':
test()
3. Optimize the code by profiling simple use-cases to find the bottlenecks and speeding up these bottle-
neck, finding a better algorithm or implementation. Keep in mind that a trade off should be found be-
tween profiling on a realistic example and the simplicity and speed of execution of the code. For efficient Note: This is a combination of two unsupervised learning techniques, principal component analysis (PCA)
work, it is best to work with profiling runs lasting around 10s. and independent component analysis (ICA). PCA is a technique for dimensionality reduction, i.e. an algorithm
to explain the observed variance in your data using less dimensions. ICA is a source seperation technique,
for example to unmix multiple signals that have been recorded through multiple sensors. Doing a PCA first
10.2 Profiling Python code and then an ICA can be useful if you have more sensors than signals. For more information see: the FastICA
example from scikits-learn.

No optimization without measuring! To run it, you also need to download the ica module. In IPython we can time the script:

• Measure: profiling, timing In [1]: %run -t demo.py

• You’ll have surprises: the fastest code is not always what you think IPython CPU timings (estimated):
User : 14.3929 s.
System: 0.256016 s.

10.2.1 Timeit and profile it:

In IPython, use timeit (https://docs.python.org/library/timeit.html) to time elementary operations: In [2]: %run -p demo.py

In [1]: import numpy as np 916 function calls in 14.551 CPU seconds

In [2]: a = np.arange(1000) Ordered by: internal time

In [3]: %timeit a ** 2 ncalls tottime percall cumtime percall filename:lineno (function)

100000 loops, best of 3: 5.73 us per loop 1 14.457 14.457 14.479 14.479 decomp.py:849 (svd)
1 0.054 0.054 0.054 0.054 {method 'random_sample' of 'mtrand.RandomState'␣
In [4]: %timeit a ** 2.1 ,→objects}
1000 loops, best of 3: 154 us per loop 1 0.017 0.017 0.021 0.021 function_base.py:645 (asarray_chkfinite)
54 0.011 0.000 0.011 0.000 {numpy.core._dotblas.dot}
In [5]: %timeit a * a 2 0.005 0.002 0.005 0.002 {method 'any' of 'numpy.ndarray' objects}
100000 loops, best of 3: 5.56 us per loop 6 0.001 0.000 0.001 0.000 ica.py:195 (gprime)
6 0.001 0.000 0.001 0.000 ica.py:192 (g)
Use this to guide your choice between strategies. 14 0.001 0.000 0.001 0.000 {numpy.linalg.lapack_lite.dsyevd}
19 0.001 0.000 0.001 0.000 twodim_base.py:204 (diag)
1 0.001 0.001 0.008 0.008 ica.py:69 (_ica_par)
Note: For long running calls, using %time instead of %timeit; it is less precise but faster 1 0.001 0.001 14.551 14.551 {execfile}
107 0.000 0.000 0.001 0.000 defmatrix.py:239 (__array_finalize__)
7 0.000 0.000 0.004 0.001 ica.py:58 (_sym_decorrelation)
7 0.000 0.000 0.002 0.000 linalg.py:841 (eigh)
10.2.2 Profiler 172 0.000 0.000 0.000 0.000 {isinstance}
1 0.000 0.000 14.551 14.551 demo.py:1 (<module>)
Useful when you have a large program to profile, for example the following file: 29 0.000 0.000 0.000 0.000 numeric.py:180 (asarray)
35 0.000 0.000 0.000 0.000 defmatrix.py:193 (__new__)
# For this example to run, you also need the 'ica.py' file 35 0.000 0.000 0.001 0.000 defmatrix.py:43 (asmatrix)
21 0.000 0.000 0.001 0.000 defmatrix.py:287 (__mul__)
import numpy as np 41 0.000 0.000 0.000 0.000 {numpy.core.multiarray.zeros}

10.1. Optimization workflow 326 10.2. Profiling Python code 327

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

28 0.000 0.000 0.000 0.000 {method 'transpose' of 'numpy.ndarray' objects}

10.3.1 Algorithmic optimization
1 0.000 0.000 0.008 0.008 ica.py:97 (fastica)
... The first thing to look for is algorithmic optimization: are there ways to compute less, or better?
For a high-level view of the problem, a good understanding of the maths behind the algorithm helps. However,
Clearly the svd (in decomp.py) is what takes most of our time, a.k.a. the bottleneck. We have to find a way to it is not uncommon to find simple changes, like moving computation or memory allocation outside a for
make this step go faster, or to avoid this step (algorithmic optimization). Spending time on the rest of the code loop, that bring in big gains.
is useless.

Example of the SVD

Profiling outside of IPython, running ‘‘cProfile‘‘
In both examples above, the SVD - Singular Value Decomposition - is what takes most of the time. Indeed, the
Similar profiling can be done outside of IPython, simply calling the built-in Python profilers cProfile and
computational cost of this algorithm is roughly n 3 in the size of the input matrix.
profile.
However, in both of these example, we are not using all the output of the SVD, but only the first few rows of
$ python -m cProfile -o demo.prof demo.py
its first return argument. If we use the svd implementation of scipy, we can ask for an incomplete version of
Using the -o switch will output the profiler results to the file demo.prof to view with an external tool. This the SVD. Note that implementations of linear algebra in scipy are richer then those in numpy and should be
can be useful if you wish to process the profiler output with a visualization tool. preferred.

In [3]: %timeit np.linalg.svd(data)

1 loops, best of 3: 14.5 s per loop
10.2.3 Line-profiler In [4]: from scipy import linalg

The profiler tells us which function takes most of the time, but not where it is called. In [5]: %timeit linalg.svd(data)
1 loops, best of 3: 14.2 s per loop
For this, we use the line_profiler: in the source file, we decorate a few functions that we want to inspect with
@profile (no need to import it)
In [6]: %timeit linalg.svd(data, full_matrices=False)
@profile 1 loops, best of 3: 295 ms per loop
def test():
data = np.random.random((5000, 100)) In [7]: %timeit np.linalg.svd(data, full_matrices=False)
u, s, v = linalg.svd(data) 1 loops, best of 3: 293 ms per loop
pca = np.dot(u[:, :10], data)
results = fastica(pca.T, whiten=False) We can then use this insight to optimize the previous code:

def test():
Then we run the script using the kernprof.py program, with switches -l, --line-by-line and -v, --view
data = np.random.random((5000, 100))
to use the line-by-line profiler and view the results in addition to saving them:
u, s, v = linalg.svd(data, full_matrices=False)
$ kernprof.py -l -v demo.py pca = np.dot(u[:, :10].T, data)
results = fastica(pca.T, whiten=False)
Wrote profile results to demo.py.lprof
Timer unit: 1e-06 s In [1]: import demo

File: demo.py In [2]: %timeit demo.

Function: test at line 5 demo.fastica demo.np demo.prof.pdf demo.py demo.pyc
Total time: 14.2793 s demo.linalg demo.prof demo.prof.png demo.py.lprof demo.test

Line # Hits Time Per Hit % Time Line Contents In [2]: %timeit demo.test()
=========== ============ ===== ========= ======= ==== ======== ica.py:65: RuntimeWarning: invalid value encountered in sqrt
5 @profile W = (u * np.diag(1.0/np.sqrt(s)) * u.T) * W # W = (W * W.T) ^{-1/2} * W
6 def test(): 1 loops, best of 3: 17.5 s per loop
7 1 19015 19015.0 0.1 data = np.random.random((5000, 100))
8 1 14242163 14242163.0 99.7 u, s, v = linalg.svd(data) In [3]: import demo_opt
9 1 10282 10282.0 0.1 pca = np.dot(u[:10, :], data)
10 1 7799 7799.0 0.1 results = fastica(pca.T, whiten=False) In [4]: %timeit demo_opt.test()
1 loops, best of 3: 208 ms per loop
The SVD is taking all the time. We need to optimise this line.
Real incomplete SVDs, e.g. computing only the first 10 eigenvectors, can be computed with arpack, available
in scipy.sparse.linalg.eigsh.
10.3 Making code go faster

Once we have identified the bottlenecks, we need to make the corresponding code go faster.

10.3. Making code go faster 328 10.3. Making code go faster 329
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Computational linear algebra

In [4]: c.strides
For certain algorithms, many of the bottlenecks will be linear algebra computations. In this case, using Out[4]: (80000, 8)
the right function to solve the right problem is key. For instance, an eigenvalue problem with a symmetric
matrix is easier to solve than with a general matrix. Also, most often, you can avoid inverting a matrix and This is the reason why Fortran ordering or C ordering may make a big difference on operations:
use a less costly (and more numerically stable) operation.
In [5]: a = np.random.rand(20, 2**18)
Know your computational linear algebra. When in doubt, explore scipy.linalg, and use %timeit to try
out different alternatives on your data. In [6]: b = np.random.rand(20, 2**18)

In [7]: %timeit np.dot(b, a.T)

1 loops, best of 3: 194 ms per loop
10.4 Writing faster numerical code In [8]: c = np.ascontiguousarray(a.T)

A complete discussion on advanced use of numpy is found in chapter Advanced NumPy, or in the article The In [9]: %timeit np.dot(b, c)
NumPy array: a structure for efficient numerical computation by van der Walt et al. Here we discuss only some 10 loops, best of 3: 84.2 ms per loop
commonly encountered tricks to make code faster.
Note that copying the data to work around this effect may not be worth it:
• Vectorizing for loops
In [10]: %timeit c = np.ascontiguousarray(a.T)
Find tricks to avoid for loops using numpy arrays. For this, masks and indices arrays can be useful.
10 loops, best of 3: 106 ms per loop
• Broadcasting
Using numexpr can be useful to automatically optimize code for such effects.
Use broadcasting to do operations on arrays as small as possible before combining them.
• Use compiled code
• In place operations
The last resort, once you are sure that all the high-level optimizations have been explored, is to transfer
In [1]: a = np.zeros(1e7) the hot spots, i.e. the few lines or functions in which most of the time is spent, to compiled code. For
compiled code, the preferred option is to use Cython: it is easy to transform exiting Python code in
In [2]: %timeit global a ; a = 0*a
compiled code, and with a good use of the numpy support yields efficient code on numpy arrays, for
10 loops, best of 3: 111 ms per loop
instance by unrolling loops.
In [3]: %timeit global a ; a *= 0
10 loops, best of 3: 48.4 ms per loop
Warning: For all the above: profile and time your choices. Don’t base your optimization on theoretical
considerations.
note: we need global a in the timeit so that it work, as it is assigning to a, and thus considers it as a local
variable.
• Be easy on the memory: use views, and not copies
10.4.1 Additional Links
Copying big arrays is as costly as making simple numerical operations on them:
• If you need to profile memory usage, you could try the memory_profiler
In [1]: a = np.zeros(1e7)
• If you need to profile down into C extensions, you could try using gperftools from Python with yep.
In [2]: %timeit a.copy()
• If you would like to track performace of your code across time, i.e. as you make new commits to your
10 loops, best of 3: 124 ms per loop
repository, you could try: vbench
In [3]: %timeit a + 1 • If you need some interactive visualization why not try RunSnakeRun
10 loops, best of 3: 112 ms per loop

• Beware of cache effects

Memory access is cheaper when it is grouped: accessing a big array in a continuous way is much faster
than random access. This implies amongst other things that smaller strides are faster (see CPU cache
effects):

In [1]: c = np.zeros((1e4, 1e4), order='C')

In [2]: %timeit c.sum(axis=0)

1 loops, best of 3: 3.89 s per loop

In [3]: %timeit c.sum(axis=1)

1 loops, best of 3: 188 ms per loop

10.4. Writing faster numerical code 330 10.4. Writing faster numerical code 331
 Scipy lecture notes, Edition 2017.1

>>> plt.plot(x, 8.0 * (x**2) / 1e6, lw=5)

[<matplotlib.lines.Line2D object at ...>]
>>> plt.xlabel('size n')
<matplotlib.text.Text object at ...>
>>> plt.ylabel('memory [MB]')

11
<matplotlib.text.Text object at ...>

11.1.2 Sparse Matrices vs. Sparse Matrix Storage Schemes

• sparse matrix is a matrix, which is almost empty

CHAPTER
• storing all the zeros is wasteful -> store only nonzero items
• think compression
• pros: huge memory savings
• cons: depends on actual storage scheme, (*) usually does not hold

11.1.3 Typical Applications

Sparse Matrices in SciPy • solution of partial differential equations (PDEs)
– the finite element method
– mechanical engineering, electrotechnics, physics, . . .
• graph theory
– nonzero at (i, j) means that node i is connected to node j

Author: Robert Cimrman • ...

11.1.4 Prerequisites

• numpy

11.1 Introduction • scipy

• matplotlib (optional)
(dense) matrix is: • ipython (the enhancements come handy)
• mathematical object
• data structure for storing a 2D array of values 11.1.5 Sparsity Structure Visualization
important features:
• spy() from matplotlib
• memory allocated once for all items
• example plots:
– usually a contiguous chunk, think NumPy ndarray
• fast access to individual items (*)

11.1.1 Why Sparse Matrices?

• the memory, that grows like n**2

• small example (double precision matrix):

>>> import numpy as np

>>> import matplotlib.pyplot as plt
>>> x = np.linspace(0, 1e6, 10)

332 11.1. Introduction 333

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

• each suitable for some tasks

• many employ sparsetools C++ module by Nathan Bell
• assume the following is imported:

>>> import numpy as np

>>> import scipy.sparse as sps
>>> import matplotlib.pyplot as plt

• warning for NumPy users:

– the multiplication with ‘*’ is the matrix multiplication (dot product)
– not part of NumPy!

* passing a sparse matrix object to NumPy functions expecting ndarray/matrix does not
work

11.2.1 Common Methods

• all scipy.sparse classes are subclasses of spmatrix

– default implementation of arithmetic operations

* always converts to CSR

* subclasses override for efficiency
– shape, data type set/get
– nonzero indices
– format conversion, interaction with NumPy (toarray(), todense())
– ...
• attributes:
– mtx.A - same as mtx.toarray()
– mtx.T - transpose (same as mtx.transpose())
– mtx.H - Hermitian (conjugate) transpose
– mtx.real - real part of complex matrix
– mtx.imag - imaginary part of complex matrix
– mtx.size - the number of nonzeros (same as self.getnnz())
– mtx.shape - the number of rows and columns (tuple)
• data usually stored in NumPy arrays
11.2 Storage Schemes
11.2.2 Sparse Matrix Classes
• seven sparse matrix types in scipy.sparse:
1. csc_matrix: Compressed Sparse Column format Diagonal Format (DIA)
2. csr_matrix: Compressed Sparse Row format
• very simple scheme
3. bsr_matrix: Block Sparse Row format
• diagonals in dense NumPy array of shape (n_diag, length)
4. lil_matrix: List of Lists format
– fixed length -> waste space a bit when far from main diagonal
5. dok_matrix: Dictionary of Keys format
– subclass of _data_matrix (sparse matrix classes with .data attribute)
6. coo_matrix: COOrdinate format (aka IJV, triplet format)
• offset for each diagonal
7. dia_matrix: DIAgonal format
– 0 is the main diagonal

11.2. Storage Schemes 334 11.2. Storage Schemes 335

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

– negative offset = below

(1, 0) 5
– positive offset = above (2, 1) 6
(3, 2) 7
• fast matrix * vector (sparsetools) (0, 2) 11
(1, 3) 12
• fast and easy item-wise operations
>>> mtx.todense()
– manipulate data array directly (fast NumPy machinery) matrix([[ 1, 0, 11, 0],
[ 5, 2, 0, 12],
• constructor accepts: [ 0, 6, 3, 0],
– dense matrix (array) [ 0, 0, 7, 4]])

– sparse matrix • explanation with a scheme:

– shape tuple (create empty matrix)
offset: row
– (data, offsets) tuple
2: 9
• no slicing, no individual item access 1: --10------
0: 1 . 11 .
• use:
-1: 5 2 . 12
– rather specialized -2: . 6 3 .
-3: . . 7 4
– solving PDEs by finite differences ---------8
– with an iterative solver
• matrix-vector multiplication

Examples >>> vec = np.ones((4, ))

>>> vec
array([ 1., 1., 1., 1.])
• create some DIA matrices: >>> mtx * vec
array([ 12., 19., 9., 11.])
>>> data = np.array([[1, 2, 3, 4]]).repeat(3, axis=0)
>>> mtx.toarray() * vec
>>> data
array([[ 1., 0., 11., 0.],
array([[1, 2, 3, 4],
[ 5., 2., 0., 12.],
[1, 2, 3, 4],
[ 0., 6., 3., 0.],
[1, 2, 3, 4]])
[ 0., 0., 7., 4.]])
>>> offsets = np.array([0, -1, 2])
>>> mtx = sparse.dia_matrix((data, offsets), shape=(4, 4))
>>> mtx
<4x4 sparse matrix of type '<... 'numpy.int64'>' List of Lists Format (LIL)
with 9 stored elements (3 diagonals) in DIAgonal format>
>>> mtx.todense() • row-based linked list
matrix([[1, 0, 3, 0],
[1, 2, 0, 4], – each row is a Python list (sorted) of column indices of non-zero elements
[0, 2, 3, 0],
– rows stored in a NumPy array (dtype=np.object)
[0, 0, 3, 4]])
– non-zero values data stored analogously
>>> data = np.arange(12).reshape((3, 4)) + 1
>>> data • efficient for constructing sparse matrices incrementally
array([[ 1, 2, 3, 4], • constructor accepts:
[ 5, 6, 7, 8],
[ 9, 10, 11, 12]]) – dense matrix (array)
>>> mtx = sparse.dia_matrix((data, offsets), shape=(4, 4))
>>> mtx.data – sparse matrix
array([[ 1, 2, 3, 4], – shape tuple (create empty matrix)
[ 5, 6, 7, 8],
[ 9, 10, 11, 12]]...) • flexible slicing, changing sparsity structure is efficient
>>> mtx.offsets
• slow arithmetics, slow column slicing due to being row-based
array([ 0, -1, 2], dtype=int32)
>>> print(mtx) • use:
(0, 0) 1
(1, 1) 2 – when sparsity pattern is not known apriori or changes
(2, 2) 3
– example: reading a sparse matrix from a text file
(3, 3) 4

11.2. Storage Schemes 336 11.2. Storage Schemes 337

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Examples

• create an empty LIL matrix:

Dictionary of Keys Format (DOK)
>>> mtx = sparse.lil_matrix((4, 5))
• subclass of Python dict
• prepare random data: – keys are (row, column) index tuples (no duplicate entries allowed)
>>> from numpy.random import rand – values are corresponding non-zero values
>>> data = np.round(rand(2, 3))
>>> data • efficient for constructing sparse matrices incrementally
array([[ 1., 1., 1.],
• constructor accepts:
[ 1., 0., 1.]])
– dense matrix (array)
• assign the data using fancy indexing:
– sparse matrix
>>> mtx[:2, [1, 2, 3]] = data – shape tuple (create empty matrix)
>>> mtx
<4x5 sparse matrix of type '<... 'numpy.float64'>' • efficient O(1) access to individual elements
with 5 stored elements in LInked List format>
>>> print(mtx) • flexible slicing, changing sparsity structure is efficient
(0, 1) 1.0 • can be efficiently converted to a coo_matrix once constructed
(0, 2) 1.0
(0, 3) 1.0 • slow arithmetics (for loops with dict.iteritems())
(1, 1) 1.0
• use:
(1, 3) 1.0
>>> mtx.todense() – when sparsity pattern is not known apriori or changes
matrix([[ 0., 1., 1., 1., 0.],
[ 0., 1., 0., 1., 0.],
[ 0., 0., 0., 0., 0.], Examples
[ 0., 0., 0., 0., 0.]])
>>> mtx.toarray()
array([[ 0., 1., 1., 1., 0.], • create a DOK matrix element by element:
[ 0., 1., 0., 1., 0.],
>>> mtx = sparse.dok_matrix((5, 5), dtype=np.float64)
[ 0., 0., 0., 0., 0.],
>>> mtx
[ 0., 0., 0., 0., 0.]])
<5x5 sparse matrix of type '<... 'numpy.float64'>'
with 0 stored elements in Dictionary Of Keys format>
• more slicing and indexing: >>> for ir in range(5):
... for ic in range(5):
>>> mtx = sparse.lil_matrix([[0, 1, 2, 0], [3, 0, 1, 0], [1, 0, 0, 1]])
... mtx[ir, ic] = 1.0 * (ir != ic)
>>> mtx.todense()
>>> mtx
matrix([[0, 1, 2, 0],
<5x5 sparse matrix of type '<... 'numpy.float64'>'
[3, 0, 1, 0],
with 20 stored elements in Dictionary Of Keys format>
[1, 0, 0, 1]]...)
>>> mtx.todense()
>>> print(mtx)
matrix([[ 0., 1., 1., 1., 1.],
(0, 1) 1
[ 1., 0., 1., 1., 1.],
(0, 2) 2
[ 1., 1., 0., 1., 1.],
(1, 0) 3
[ 1., 1., 1., 0., 1.],
(1, 2) 1
[ 1., 1., 1., 1., 0.]])
(2, 0) 1
(2, 3) 1
>>> mtx[:2, :] • slicing and indexing:
<2x4 sparse matrix of type '<... 'numpy.int64'>'
>>> mtx[1, 1]
with 4 stored elements in LInked List format>
0.0
>>> mtx[:2, :].todense()
>>> mtx[1, 1:3]
matrix([[0, 1, 2, 0],
<1x2 sparse matrix of type '<... 'numpy.float64'>'
[3, 0, 1, 0]]...)
with 1 stored elements in Dictionary Of Keys format>
>>> mtx[1:2, [0,2]].todense()
>>> mtx[1, 1:3].todense()
matrix([[3, 1]]...)
matrix([[ 0., 1.]])
>>> mtx.todense()
>>> mtx[[2,1], 1:3].todense()
matrix([[0, 1, 2, 0],
matrix([[ 1., 0.],
[3, 0, 1, 0],
[ 0., 1.]])
[1, 0, 0, 1]]...)

11.2. Storage Schemes 338 11.2. Storage Schemes 339

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Coordinate Format (COO)

>>> row = np.array([0, 0, 1, 3, 1, 0, 0])
>>> col = np.array([0, 2, 1, 3, 1, 0, 0])
• also known as the ‘ijv’ or ‘triplet’ format >>> data = np.array([1, 1, 1, 1, 1, 1, 1])
– three NumPy arrays: row, col, data >>> mtx = sparse.coo_matrix((data, (row, col)), shape=(4, 4))
>>> mtx.todense()
– data[i] is value at (row[i], col[i]) position matrix([[3, 0, 1, 0],
[0, 2, 0, 0],
– permits duplicate entries [0, 0, 0, 0],
– subclass of _data_matrix (sparse matrix classes with .data attribute) [0, 0, 0, 1]])

• fast format for constructing sparse matrices • no slicing. . . :

• constructor accepts:
>>> mtx[2, 3]
– dense matrix (array) Traceback (most recent call last):
...
– sparse matrix TypeError: 'coo_matrix' object ...
– shape tuple (create empty matrix)
– (data, ij) tuple
Compressed Sparse Row Format (CSR)
• very fast conversion to and from CSR/CSC formats
• row oriented
• fast matrix * vector (sparsetools)
– three NumPy arrays: indices, indptr, data
• fast and easy item-wise operations
– manipulate data array directly (fast NumPy machinery) * indices is array of column indices

• no slicing, no arithmetics (directly) * data is array of corresponding nonzero values

• use: * indptr points to row starts in indices and data

– facilitates fast conversion among sparse formats * length is n_row + 1, last item = number of values = length of both indices and data

– when converting to other format (usually CSR or CSC), duplicate entries are summed together * nonzero values of the i-th row are data[indptr[i]:indptr[i+1]] with column indices in-
dices[indptr[i]:indptr[i+1]]
* facilitates efficient construction of finite element matrices
* item (i, j) can be accessed as data[indptr[i]+k], where k is position of j in in-
dices[indptr[i]:indptr[i+1]]
Examples – subclass of _cs_matrix (common CSR/CSC functionality)

• create empty COO matrix: * subclass of _data_matrix (sparse matrix classes with .data attribute)
• fast matrix vector products and other arithmetics (sparsetools)
>>> mtx = sparse.coo_matrix((3, 4), dtype=np.int8)
>>> mtx.todense() • constructor accepts:
matrix([[0, 0, 0, 0],
– dense matrix (array)
[0, 0, 0, 0],
[0, 0, 0, 0]], dtype=int8) – sparse matrix
– shape tuple (create empty matrix)
• create using (data, ij) tuple:
– (data, ij) tuple
>>> row = np.array([0, 3, 1, 0])
>>> col = np.array([0, 3, 1, 2]) – (data, indices, indptr) tuple
>>> data = np.array([4, 5, 7, 9])
>>> mtx = sparse.coo_matrix((data, (row, col)), shape=(4, 4)) • efficient row slicing, row-oriented operations
>>> mtx • slow column slicing, expensive changes to the sparsity structure
<4x4 sparse matrix of type '<... 'numpy.int64'>'
with 4 stored elements in COOrdinate format> • use:
>>> mtx.todense()
– actual computations (most linear solvers support this format)
matrix([[4, 0, 9, 0],
[0, 7, 0, 0],
[0, 0, 0, 0],
[0, 0, 0, 5]])

• duplicates entries are summed together:

11.2. Storage Schemes 340 11.2. Storage Schemes 341

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Examples • fast matrix vector products and other arithmetics (sparsetools)

• constructor accepts:
• create empty CSR matrix:
– dense matrix (array)
>>> mtx = sparse.csr_matrix((3, 4), dtype=np.int8)
– sparse matrix
>>> mtx.todense()
matrix([[0, 0, 0, 0], – shape tuple (create empty matrix)
[0, 0, 0, 0],
[0, 0, 0, 0]], dtype=int8) – (data, ij) tuple
– (data, indices, indptr) tuple
• create using (data, ij) tuple:
• efficient column slicing, column-oriented operations
>>> row = np.array([0, 0, 1, 2, 2, 2])
>>> col = np.array([0, 2, 2, 0, 1, 2]) • slow row slicing, expensive changes to the sparsity structure
>>> data = np.array([1, 2, 3, 4, 5, 6]) • use:
>>> mtx = sparse.csr_matrix((data, (row, col)), shape=(3, 3))
>>> mtx – actual computations (most linear solvers support this format)
<3x3 sparse matrix of type '<... 'numpy.int64'>'
with 6 stored elements in Compressed Sparse Row format>
>>> mtx.todense() Examples
matrix([[1, 0, 2],
[0, 0, 3], • create empty CSC matrix:
[4, 5, 6]]...)
>>> mtx.data >>> mtx = sparse.csc_matrix((3, 4), dtype=np.int8)
array([1, 2, 3, 4, 5, 6]...) >>> mtx.todense()
>>> mtx.indices matrix([[0, 0, 0, 0],
array([0, 2, 2, 0, 1, 2], dtype=int32) [0, 0, 0, 0],
>>> mtx.indptr [0, 0, 0, 0]], dtype=int8)
array([0, 2, 3, 6], dtype=int32)
• create using (data, ij) tuple:
• create using (data, indices, indptr) tuple:
>>> row = np.array([0, 0, 1, 2, 2, 2])
>>> data = np.array([1, 2, 3, 4, 5, 6]) >>> col = np.array([0, 2, 2, 0, 1, 2])
>>> indices = np.array([0, 2, 2, 0, 1, 2]) >>> data = np.array([1, 2, 3, 4, 5, 6])
>>> indptr = np.array([0, 2, 3, 6]) >>> mtx = sparse.csc_matrix((data, (row, col)), shape=(3, 3))
>>> mtx = sparse.csr_matrix((data, indices, indptr), shape=(3, 3)) >>> mtx
>>> mtx.todense() <3x3 sparse matrix of type '<... 'numpy.int64'>'
matrix([[1, 0, 2], with 6 stored elements in Compressed Sparse Column format>
[0, 0, 3], >>> mtx.todense()
[4, 5, 6]]) matrix([[1, 0, 2],
[0, 0, 3],
[4, 5, 6]]...)
Compressed Sparse Column Format (CSC) >>> mtx.data
array([1, 4, 5, 2, 3, 6]...)
>>> mtx.indices
• column oriented
array([0, 2, 2, 0, 1, 2], dtype=int32)
– three NumPy arrays: indices, indptr, data >>> mtx.indptr
array([0, 2, 3, 6], dtype=int32)
* indices is array of row indices
* data is array of corresponding nonzero values • create using (data, indices, indptr) tuple:

* indptr points to column starts in indices and data >>> data = np.array([1, 4, 5, 2, 3, 6])
>>> indices = np.array([0, 2, 2, 0, 1, 2])
* length is n_col + 1, last item = number of values = length of both indices and data >>> indptr = np.array([0, 2, 3, 6])
>>> mtx = sparse.csc_matrix((data, indices, indptr), shape=(3, 3))
* nonzero values of the i-th column are data[indptr[i]:indptr[i+1]] with row indices in-
dices[indptr[i]:indptr[i+1]] >>> mtx.todense()
matrix([[1, 0, 2],
* item (i, j) can be accessed as data[indptr[j]+k], where k is position of i in in- [0, 0, 3],
dices[indptr[j]:indptr[j+1]] [4, 5, 6]])

– subclass of _cs_matrix (common CSR/CSC functionality)

* subclass of _data_matrix (sparse matrix classes with .data attribute)

11.2. Storage Schemes 342 11.2. Storage Schemes 343

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Block Compressed Row Format (BSR)

>>> row = np.array([0, 0, 1, 2, 2, 2])
>>> col = np.array([0, 2, 2, 0, 1, 2])
• basically a CSR with dense sub-matrices of fixed shape instead of scalar items >>> data = np.array([1, 2, 3, 4, 5, 6])
– block size (R, C) must evenly divide the shape of the matrix (M, N) >>> mtx = sparse.bsr_matrix((data, (row, col)), shape=(3, 3))
>>> mtx
– three NumPy arrays: indices, indptr, data <3x3 sparse matrix of type '<... 'numpy.int64'>'
with 6 stored elements (blocksize = 1x1) in Block Sparse Row format>
* indices is array of column indices for each block >>> mtx.todense()
matrix([[1, 0, 2],
* data is array of corresponding nonzero values of shape (nnz, R, C)
[0, 0, 3],
* ... [4, 5, 6]]...)
>>> mtx.data
– subclass of _cs_matrix (common CSR/CSC functionality)
array([[[1]],
* subclass of _data_matrix (sparse matrix classes with .data attribute)
[[2]],
• fast matrix vector products and other arithmetics (sparsetools)
[[3]],
• constructor accepts:
– dense matrix (array) [[4]],

– sparse matrix [[5]],

– shape tuple (create empty matrix)
[[6]]]...)
– (data, ij) tuple >>> mtx.indices
array([0, 2, 2, 0, 1, 2], dtype=int32)
– (data, indices, indptr) tuple >>> mtx.indptr
• many arithmetic operations considerably more efficient than CSR for sparse matrices with dense sub- array([0, 2, 3, 6], dtype=int32)
matrices
• create using (data, indices, indptr) tuple with (2, 2) block size:
• use:
>>> indptr = np.array([0, 2, 3, 6])
– like CSR >>> indices = np.array([0, 2, 2, 0, 1, 2])
– vector-valued finite element discretizations >>> data = np.array([1, 2, 3, 4, 5, 6]).repeat(4).reshape(6, 2, 2)
>>> mtx = sparse.bsr_matrix((data, indices, indptr), shape=(6, 6))
>>> mtx.todense()
Examples matrix([[1, 1, 0, 0, 2, 2],
[1, 1, 0, 0, 2, 2],
[0, 0, 0, 0, 3, 3],
• create empty BSR matrix with (1, 1) block size (like CSR. . . ): [0, 0, 0, 0, 3, 3],
[4, 4, 5, 5, 6, 6],
>>> mtx = sparse.bsr_matrix((3, 4), dtype=np.int8)
[4, 4, 5, 5, 6, 6]])
>>> mtx
>>> data
<3x4 sparse matrix of type '<... 'numpy.int8'>'
array([[[1, 1],
with 0 stored elements (blocksize = 1x1) in Block Sparse Row format>
[1, 1]],
>>> mtx.todense()
matrix([[0, 0, 0, 0],
[[2, 2],
[0, 0, 0, 0],
[2, 2]],
[0, 0, 0, 0]], dtype=int8)
[[3, 3],
• create empty BSR matrix with (3, 2) block size: [3, 3]],
>>> mtx = sparse.bsr_matrix((3, 4), blocksize=(3, 2), dtype=np.int8)
[[4, 4],
>>> mtx
[4, 4]],
<3x4 sparse matrix of type '<... 'numpy.int8'>'
with 0 stored elements (blocksize = 3x2) in Block Sparse Row format>
[[5, 5],
>>> mtx.todense()
[5, 5]],
matrix([[0, 0, 0, 0],
[0, 0, 0, 0],
[[6, 6],
[0, 0, 0, 0]], dtype=int8)
[6, 6]]])

– a bug?
• create using (data, ij) tuple with (1, 1) block size (like CSR. . . ):

11.2. Storage Schemes 344 11.2. Storage Schemes 345

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

11.2.3 Summary – check-out the new scikits.suitesparse by Nathaniel Smith

Table 11.1: Summary of storage schemes. Examples

format matrix get fancy set item fancy solvers note
* vector item get set • import the whole module, and see its docstring:
DIA sparsetools. . . . iterative has data array, specialized >>> from scipy.sparse.linalg import dsolve
LIL via CSR yes yes yes yes iterative arithmetics via CSR, incre- >>> help(dsolve)
mental construction
DOK python yes one yes yes iterative O(1) item access, incremental • both superlu and umfpack can be used (if the latter is installed) as follows:
axis construction
only – prepare a linear system:
COO sparsetools. . . . iterative has data array, facilitates fast >>> import numpy as np
conversion >>> from scipy import sparse
CSR sparsetoolsyes yes slow . any has data array, fast row-wise >>> mtx = sparse.spdiags([[1, 2, 3, 4, 5], [6, 5, 8, 9, 10]], [0, 1], 5, 5)
ops >>> mtx.todense()
CSC sparsetoolsyes yes slow . any has data array, fast column- matrix([[ 1, 5, 0, 0, 0],
wise ops [ 0, 2, 8, 0, 0],
BSR sparsetools. . . . specializedhas data array, specialized [ 0, 0, 3, 9, 0],
[ 0, 0, 0, 4, 10],
[ 0, 0, 0, 0, 5]])
>>> rhs = np.array([1, 2, 3, 4, 5], dtype=np.float32)
11.3 Linear System Solvers
– solve as single precision real:
• sparse matrix/eigenvalue problem solvers live in scipy.sparse.linalg >>> mtx1 = mtx.astype(np.float32)
• the submodules: >>> x = dsolve.spsolve(mtx1, rhs, use_umfpack=False)
>>> print(x)
– dsolve: direct factorization methods for solving linear systems [ 106. -21. 5.5 -1.5 1. ]
>>> print("Error: %s " % (mtx1 * x - rhs))
– isolve: iterative methods for solving linear systems Error: [ 0. 0. 0. 0. 0.]
– eigen: sparse eigenvalue problem solvers
– solve as double precision real:
• all solvers are accessible from:
>>> mtx2 = mtx.astype(np.float64)
>>> import scipy.sparse.linalg as spla >>> x = dsolve.spsolve(mtx2, rhs, use_umfpack=True)
>>> spla.__all__ >>> print(x)
['LinearOperator', 'Tester', 'arpack', 'aslinearoperator', 'bicg', [ 106. -21. 5.5 -1.5 1. ]
'bicgstab', 'cg', 'cgs', 'csc_matrix', 'csr_matrix', 'dsolve', >>> print("Error: %s " % (mtx2 * x - rhs))
'eigen', 'eigen_symmetric', 'factorized', 'gmres', 'interface', Error: [ 0. 0. 0. 0. 0.]
'isolve', 'iterative', 'lgmres', 'linsolve', 'lobpcg', 'lsqr',
'minres', 'np', 'qmr', 'speigs', 'spilu', 'splu', 'spsolve', 'svd',
– solve as single precision complex:
'test', 'umfpack', 'use_solver', 'utils', 'warnings']
>>> mtx1 = mtx.astype(np.complex64)
>>> x = dsolve.spsolve(mtx1, rhs, use_umfpack=False)
11.3.1 Sparse Direct Solvers >>> print(x)
[ 106.0+0.j -21.0+0.j 5.5+0.j -1.5+0.j 1.0+0.j]
>>> print("Error: %s " % (mtx1 * x - rhs))
• default solver: SuperLU 4.0 Error: [ 0.+0.j 0.+0.j 0.+0.j 0.+0.j 0.+0.j]
– included in SciPy
– solve as double precision complex:
– real and complex systems
>>> mtx2 = mtx.astype(np.complex128)
– both single and double precision
>>> x = dsolve.spsolve(mtx2, rhs, use_umfpack=True)
• optional: umfpack >>> print(x)
[ 106.0+0.j -21.0+0.j 5.5+0.j -1.5+0.j 1.0+0.j]
– real and complex systems >>> print("Error: %s " % (mtx2 * x - rhs))
– double precision only Error: [ 0.+0.j 0.+0.j 0.+0.j 0.+0.j 0.+0.j]

– recommended for performance """

– wrappers now live in scikits.umfpack Solve a linear system

11.3. Linear System Solvers 346 11.3. Linear System Solvers 347
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

M [{sparse matrix, dense matrix, LinearOperator}] Preconditioner for A. The preconditioner should ap-
=======================
proximate the inverse of A. Effective preconditioning dramatically improves the rate of conver-
Construct a 1000x1000 lil_matrix and add some values to it, convert it gence, which implies that fewer iterations are needed to reach a given error tolerance.
to CSR format and solve A x = b for x:and solve a linear system with a callback [function] User-supplied function to call after each iteration. It is called as callback(xk), where
direct solver.
xk is the current solution vector.
"""
import numpy as np
import scipy.sparse as sps LinearOperator Class
from matplotlib import pyplot as plt
from scipy.sparse.linalg.dsolve import linsolve
from scipy.sparse.linalg.interface import LinearOperator
rand = np.random.rand
• common interface for performing matrix vector products
mtx = sps.lil_matrix((1000, 1000), dtype=np.float64)
mtx[0, :100] = rand(100) • useful abstraction that enables using dense and sparse matrices within the solvers, as well as matrix-free
mtx[1, 100:200] = mtx[0, :100] solutions
mtx.setdiag(rand(1000)) • has shape and matvec() (+ some optional parameters)
plt.clf() • example:
plt.spy(mtx, marker='.', markersize=2)
plt.show() >>> import numpy as np
>>> from scipy.sparse.linalg import LinearOperator
mtx = mtx.tocsr() >>> def mv(v):
rhs = rand(1000) ... return np.array([2*v[0], 3*v[1]])
...
x = linsolve.spsolve(mtx, rhs) >>> A = LinearOperator((2, 2), matvec=mv)
>>> A
print('rezidual: %r ' % np.linalg.norm(mtx * x - rhs)) <2x2 _CustomLinearOperator with dtype=float64>
>>> A.matvec(np.ones(2))
array([ 2., 3.])
• examples/direct_solve.py
>>> A * np.ones(2)
array([ 2., 3.])
11.3.2 Iterative Solvers

• the isolve module contains the following solvers: A Few Notes on Preconditioning

– bicg (BIConjugate Gradient) • problem specific

– bicgstab (BIConjugate Gradient STABilized) • often hard to develop
– cg (Conjugate Gradient) - symmetric positive definite matrices only • if not sure, try ILU
– cgs (Conjugate Gradient Squared) – available in dsolve as spilu()
– gmres (Generalized Minimal RESidual)
– minres (MINimum RESidual) 11.3.3 Eigenvalue Problem Solvers
– qmr (Quasi-Minimal Residual)
The eigen module

Common Parameters • arpack * a collection of Fortran77 subroutines designed to solve large scale eigenvalue problems

• mandatory: • lobpcg (Locally Optimal Block Preconditioned Conjugate Gradient Method) * works very well in combi-
nation with PyAMG * example by Nathan Bell:
A [{sparse matrix, dense matrix, LinearOperator}] The N-by-N matrix of the linear system.
"""
b [{array, matrix}] Right hand side of the linear system. Has shape (N,) or (N,1). Compute eigenvectors and eigenvalues using a preconditioned eigensolver
• optional: ========================================================================

x0 [{array, matrix}] Starting guess for the solution. In this example Smoothed Aggregation (SA) is used to precondition
the LOBPCG eigensolver on a two-dimensional Poisson problem with
tol [float] Relative tolerance to achieve before terminating. Dirichlet boundary conditions.
maxiter [integer] Maximum number of iterations. Iteration will stop after maxiter steps even if the spec- """
ified tolerance has not been achieved.

11.3. Linear System Solvers 348 11.3. Linear System Solvers 349
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

import scipy
from scipy.sparse.linalg import lobpcg

from pyamg import smoothed_aggregation_solver

from pyamg.gallery import poisson

N = 100
K = 9
A = poisson((N,N), format='csr')

# create the AMG hierarchy

ml = smoothed_aggregation_solver(A)

# initial approximation to the K eigenvectors

X = scipy.rand(A.shape[0], K)

# preconditioner based on ml
M = ml.aspreconditioner()

# compute eigenvalues and eigenvectors with LOBPCG

W,V = lobpcg(A, X, M=M, tol=1e-8, largest=False)

#plot the eigenvectors

import pylab

pylab.figure(figsize=(9,9))

for i in range(K):
pylab.subplot(3, 3, i+1)
pylab.title('Eigenvector %d ' % i)
pylab.pcolor(V[:,i].reshape(N,N))
pylab.axis('equal')
pylab.axis('off') 11.4 Other Interesting Packages
pylab.show()
• PyAMG
– examples/pyamg_with_lobpcg.py
– algebraic multigrid solvers
• example by Nils Wagner:
– https://github.com/pyamg/pyamg
– examples/lobpcg_sakurai.py
• Pysparse
• output:
– own sparse matrix classes
$ python examples/lobpcg_sakurai.py
Results by LOBPCG for n=2500 – matrix and eigenvalue problem solvers
– http://pysparse.sourceforge.net/
[ 0.06250083 0.06250028 0.06250007]

Exact eigenvalues

[ 0.06250005 0.0625002 0.06250044]

Elapsed time 7.01

11.3. Linear System Solvers 350 11.4. Other Interesting Packages 351
 Scipy lecture notes, Edition 2017.1

• Input/Output, displaying images

• Basic manipulations: cropping, flipping, rotating, . . .
• Image filtering: denoising, sharpening
• Image segmentation: labeling pixels corresponding to different objects

12
• Classification
• Feature extraction
• Registration
• ...
CHAPTER
Chapters contents

• Opening and writing to image files

• Displaying images
• Basic manipulations

Image manipulation and processing

– Statistical information
– Geometrical transformations

using Numpy and Scipy

• Image filtering
– Blurring/smoothing
– Sharpening
– Denoising
– Mathematical morphology
• Feature extraction
Authors: Emmanuelle Gouillart, Gaël Varoquaux – Edge detection
This section addresses basic image manipulation and processing using the core scientific modules NumPy and – Segmentation
SciPy. Some of the operations covered by this tutorial may be useful for other kinds of multidimensional array
• Measuring objects properties: ndimage.measurements
processing than image processing. In particular, the submodule scipy.ndimage provides functions operating
on n-dimensional NumPy arrays. • Full code examples
See also: • Examples for the image processing chapter
For more advanced image processing and image-specific routines, see the tutorial Scikit-image: image pro-
cessing, dedicated to the skimage module.

12.1 Opening and writing to image files

Image = 2-D numerical array

(or 3-D: CT, MRI, 2D + time; 4-D, . . . ) Writing an array to a file:

Here, image == Numpy array np.array from scipy import misc

f = misc.face()
misc.imsave('face.png', f) # uses the Image module (PIL)
Tools used in this tutorial:
import matplotlib.pyplot as plt
• numpy: basic array manipulation
plt.imshow(f)
• scipy: scipy.ndimage submodule dedicated to image processing (n-dimensional images). See the plt.show()
documentation:

>>> from scipy import ndimage

Common tasks in image processing:

352 12.1. Opening and writing to image files 353

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

>>> for i in range(10):

... im = np.random.randint(0, 256, 10000).reshape((100, 100))
... misc.imsave('random_%02d .png' % i, im)
>>> from glob import glob
>>> filelist = glob('random*.png')
>>> filelist.sort()

12.2 Displaying images

Use matplotlib and imshow to display an image inside a matplotlib figure:

>>> f = misc.face(gray=True) # retrieve a grayscale image

>>> import matplotlib.pyplot as plt
>>> plt.imshow(f, cmap=plt.cm.gray)
<matplotlib.image.AxesImage object at 0x...>

Increase contrast by setting min and max values:

>>> plt.imshow(f, cmap=plt.cm.gray, vmin=30, vmax=200)

<matplotlib.image.AxesImage object at 0x...>
>>> # Remove axes and ticks
>>> plt.axis('off')
(-0.5, 1023.5, 767.5, -0.5)

Draw contour lines:

>>> plt.contour(f, [50, 200])

<matplotlib.contour.QuadContourSet ...>
Creating a numpy array from an image file:

>>> from scipy import misc

>>> face = misc.face()
>>> misc.imsave('face.png', face) # First we need to create the PNG file

>>> face = misc.imread('face.png')

>>> type(face)
<... 'numpy.ndarray'>
>>> face.shape, face.dtype
((768, 1024, 3), dtype('uint8'))

dtype is uint8 for 8-bit images (0-255)

Opening raw files (camera, 3-D images)

>>> face.tofile('face.raw') # Create raw file

>>> face_from_raw = np.fromfile('face.raw', dtype=np.uint8)
>>> face_from_raw.shape For fine inspection of intensity variations, use interpolation='nearest':
(2359296,)
>>> face_from_raw.shape = (768, 1024, 3) >>> plt.imshow(f[320:340, 510:530], cmap=plt.cm.gray)
<matplotlib.image.AxesImage object at 0x...>
>>> plt.imshow(f[320:340, 510:530], cmap=plt.cm.gray, interpolation='nearest')
Need to know the shape and dtype of the image (how to separate data bytes). <matplotlib.image.AxesImage object at 0x...>
For large data, use np.memmap for memory mapping:
See also:
>>> face_memmap = np.memmap('face.raw', dtype=np.uint8, shape=(768, 1024, 3))
3-D visualization: Mayavi
(data are read from the file, and not loaded into memory) See 3D plotting with Mayavi.
Working on a list of image files • Image plane widgets
• Isosurfaces

12.1. Opening and writing to image files 354 12.2. Displaying images 355
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

• ... >>> face = misc.face(gray=True)

>>> face[0, 40]
127
>>> # Slicing
>>> face[10:13, 20:23]
array([[141, 153, 145],
[133, 134, 125],
[ 96, 92, 94]], dtype=uint8)
>>> face[100:120] = 255
>>>
>>> lx, ly = face.shape
>>> X, Y = np.ogrid[0:lx, 0:ly]
>>> mask = (X - lx / 2) ** 2 + (Y - ly / 2) ** 2 > lx * ly / 4
>>> # Masks
>>> face[mask] = 0
>>> # Fancy indexing
>>> face[range(400), range(400)] = 255

12.3 Basic manipulations

Images are arrays: use the whole numpy machinery.

12.3. Basic manipulations 356 12.3. Basic manipulations 357

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

12.3.1 Statistical information 12.4 Image filtering

>>> face = misc.face(gray=True)
Local filters: replace the value of pixels by a function of the values of neighboring pixels.
>>> face.mean()
113.48026784261067 Neighbourhood: square (choose size), disk, or more complicated structuring element.
>>> face.max(), face.min()
(250, 0)

np.histogram

Exercise

• Open as an array the scikit-image logo (http://scikit-image.org/_static/img/logo.png), or an image

that you have on your computer. 12.4.1 Blurring/smoothing
• Crop a meaningful part of the image, for example the python circle in the logo.
Gaussian filter from scipy.ndimage:
• Display the image array using matplotlib. Change the interpolation method and zoom to see the
>>> from scipy import misc
difference.
>>> face = misc.face(gray=True)
• Transform your image to greyscale >>> blurred_face = ndimage.gaussian_filter(face, sigma=3)
>>> very_blurred = ndimage.gaussian_filter(face, sigma=5)
• Increase the contrast of the image by changing its minimum and maximum values. Optional: use
scipy.stats.scoreatpercentile (read the docstring!) to saturate 5% of the darkest pixels and 5% Uniform filter
of the lightest pixels.
>>> local_mean = ndimage.uniform_filter(face, size=11)
• Save the array to two different file formats (png, jpg, tiff)

12.3.2 Geometrical transformations

>>> face = misc.face(gray=True)

>>> lx, ly = face.shape
>>> # Cropping 12.4.2 Sharpening
>>> crop_face = face[lx // 4: - lx // 4, ly // 4: - ly // 4]
>>> # up <-> down flip Sharpen a blurred image:
>>> flip_ud_face = np.flipud(face)
>>> # rotation >>> from scipy import misc
>>> rotate_face = ndimage.rotate(face, 45) >>> face = misc.face(gray=True).astype(float)
>>> rotate_face_noreshape = ndimage.rotate(face, 45, reshape=False) >>> blurred_f = ndimage.gaussian_filter(face, 3)

increase the weight of edges by adding an approximation of the Laplacian:

>>> filter_blurred_f = ndimage.gaussian_filter(blurred_f, 1)

>>> alpha = 30
>>> sharpened = blurred_f + alpha * (blurred_f - filter_blurred_f)

12.4.3 Denoising

Noisy face:

12.3. Basic manipulations 358 12.4. Image filtering 359

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Other rank filter: ndimage.maximum_filter, ndimage.percentile_filter

Other local non-linear filters: Wiener (scipy.signal.wiener), etc.
Non-local filters

Exercise: denoising

• Create a binary image (of 0s and 1s) with several objects (circles, ellipses, squares, or random shapes).
• Add some noise (e.g., 20% of noise)
• Try two different denoising methods for denoising the image: gaussian filtering and median filtering.
• Compare the histograms of the two different denoised images. Which one is the closest to the his-
togram of the original (noise-free) image?

>>> from scipy import misc

>>> f = misc.face(gray=True) See also:
>>> f = f[230:290, 220:320]
More denoising filters are available in skimage.denoising, see the Scikit-image: image processing tutorial.
>>> noisy = f + 0.4 * f.std() * np.random.random(f.shape)

A Gaussian filter smoothes the noise out. . . and the edges as well: 12.4.4 Mathematical morphology
>>> gauss_denoised = ndimage.gaussian_filter(noisy, 2)
See wikipedia for a definition of mathematical morphology.
Most local linear isotropic filters blur the image (ndimage.uniform_filter) Probe an image with a simple shape (a structuring element), and modify this image according to how the
A median filter preserves better the edges: shape locally fits or misses the image.
Structuring element:
>>> med_denoised = ndimage.median_filter(noisy, 3)
>>> el = ndimage.generate_binary_structure(2, 1)
>>> el
array([[False, True, False],
[ True, True, True],
[False, True, False]], dtype=bool)
>>> el.astype(np.int)
array([[0, 1, 0],
[1, 1, 1],
[0, 1, 0]])

Median filter: better result for straight boundaries (low curvature):

>>> im = np.zeros((20, 20))

>>> im[5:-5, 5:-5] = 1
>>> im = ndimage.distance_transform_bf(im)
>>> im_noise = im + 0.2 * np.random.randn(*im.shape)
>>> im_med = ndimage.median_filter(im_noise, 3)

Erosion = minimum filter. Replace the value of a pixel by the minimal value covered by the structuring ele-
ment.:

>>> a = np.zeros((7,7), dtype=np.int)

>>> a[1:6, 2:5] = 1
>>> a
array([[0, 0, 0, 0, 0, 0, 0],
[0, 0, 1, 1, 1, 0, 0],

12.4. Image filtering 360 12.4. Image filtering 361

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

[0, 0, 1, 1, 1, 0, 0], >>> square = np.zeros((16, 16))

[0, 0, 1, 1, 1, 0, 0], >>> square[4:-4, 4:-4] = 1
[0, 0, 1, 1, 1, 0, 0], >>> dist = ndimage.distance_transform_bf(square)
[0, 0, 1, 1, 1, 0, 0], >>> dilate_dist = ndimage.grey_dilation(dist, size=(3, 3), \
[0, 0, 0, 0, 0, 0, 0]]) ... structure=np.ones((3, 3)))
>>> ndimage.binary_erosion(a).astype(a.dtype)
array([[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 1, 0, 0, 0],
[0, 0, 0, 1, 0, 0, 0],
[0, 0, 0, 1, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0]])
>>> #Erosion removes objects smaller than the structure
>>> ndimage.binary_erosion(a, structure=np.ones((5,5))).astype(a.dtype)
array([[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0], Opening: erosion + dilation:
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0], >>> a = np.zeros((5,5), dtype=np.int)
[0, 0, 0, 0, 0, 0, 0], >>> a[1:4, 1:4] = 1; a[4, 4] = 1
[0, 0, 0, 0, 0, 0, 0], >>> a
[0, 0, 0, 0, 0, 0, 0]]) array([[0, 0, 0, 0, 0],
[0, 1, 1, 1, 0],
[0, 1, 1, 1, 0],
[0, 1, 1, 1, 0],
[0, 0, 0, 0, 1]])
>>> # Opening removes small objects
>>> ndimage.binary_opening(a, structure=np.ones((3,3))).astype(np.int)
array([[0, 0, 0, 0, 0],
[0, 1, 1, 1, 0],
[0, 1, 1, 1, 0],
[0, 1, 1, 1, 0],
[0, 0, 0, 0, 0]])
>>> # Opening can also smooth corners
>>> ndimage.binary_opening(a).astype(np.int)
array([[0, 0, 0, 0, 0],
[0, 0, 1, 0, 0],
Dilation: maximum filter: [0, 1, 1, 1, 0],
[0, 0, 1, 0, 0],
>>> a = np.zeros((5, 5)) [0, 0, 0, 0, 0]])
>>> a[2, 2] = 1
>>> a Application: remove noise:
array([[ 0., 0., 0., 0., 0.],
[ 0., 0., 0., 0., 0.], >>> square = np.zeros((32, 32))
[ 0., 0., 1., 0., 0.], >>> square[10:-10, 10:-10] = 1
[ 0., 0., 0., 0., 0.], >>> np.random.seed(2)
[ 0., 0., 0., 0., 0.]]) >>> x, y = (32*np.random.random((2, 20))).astype(np.int)
>>> ndimage.binary_dilation(a).astype(a.dtype) >>> square[x, y] = 1
array([[ 0., 0., 0., 0., 0.],
[ 0., 0., 1., 0., 0.], >>> open_square = ndimage.binary_opening(square)
[ 0., 1., 1., 1., 0.],
[ 0., 0., 1., 0., 0.], >>> eroded_square = ndimage.binary_erosion(square)
[ 0., 0., 0., 0., 0.]]) >>> reconstruction = ndimage.binary_propagation(eroded_square, mask=square)

Also works for grey-valued images:

>>> np.random.seed(2)
>>> im = np.zeros((64, 64))
>>> x, y = (63*np.random.random((2, 8))).astype(np.int)
>>> im[x, y] = np.arange(8)

>>> bigger_points = ndimage.grey_dilation(im, size=(5, 5), structure=np.ones((5, 5)))

12.4. Image filtering 362 12.4. Image filtering 363

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Closing: dilation + erosion

Many other mathematical morphology operations: hit and miss transform, tophat, etc.

12.5 Feature extraction

12.5.1 Edge detection

Synthetic data:

>>> im = np.zeros((256, 256))

>>> im[64:-64, 64:-64] = 1
>>>
>>> im = ndimage.rotate(im, 15, mode='constant')
>>> im = ndimage.gaussian_filter(im, 8)
>>> # Remove small white regions
Use a gradient operator (Sobel) to find high intensity variations: >>> open_img = ndimage.binary_opening(binary_img)
>>> # Remove small black hole
>>> sx = ndimage.sobel(im, axis=0, mode='constant') >>> close_img = ndimage.binary_closing(open_img)
>>> sy = ndimage.sobel(im, axis=1, mode='constant')
>>> sob = np.hypot(sx, sy)

Exercise

Check that reconstruction operations (erosion + propagation) produce a better result than opening/closing:
>>> eroded_img = ndimage.binary_erosion(binary_img)
12.5.2 Segmentation >>> reconstruct_img = ndimage.binary_propagation(eroded_img, mask=binary_img)
>>> tmp = np.logical_not(reconstruct_img)
• Histogram-based segmentation (no spatial information) >>> eroded_tmp = ndimage.binary_erosion(tmp)
>>> reconstruct_final = np.logical_not(ndimage.binary_propagation(eroded_tmp, mask=tmp))
>>> n = 10 >>> np.abs(mask - close_img).mean()
>>> l = 256 0.00727836...
>>> im = np.zeros((l, l)) >>> np.abs(mask - reconstruct_final).mean()
>>> np.random.seed(1) 0.00059502...
>>> points = l*np.random.random((2, n**2))
>>> im[(points[0]).astype(np.int), (points[1]).astype(np.int)] = 1
>>> im = ndimage.gaussian_filter(im, sigma=l/(4.*n))
Exercise
>>> mask = (im > im.mean()).astype(np.float)
>>> mask += 0.1 * im Check how a first denoising step (e.g. with a median filter) modifies the histogram, and check that the
>>> img = mask + 0.2*np.random.randn(*mask.shape) resulting histogram-based segmentation is more accurate.

>>> hist, bin_edges = np.histogram(img, bins=60)

>>> bin_centers = 0.5*(bin_edges[:-1] + bin_edges[1:]) See also:

>>> binary_img = img > 0.5

More advanced segmentation algorithms are found in the scikit-image: see Scikit-image: image processing.
See also:
Use mathematical morphology to clean up the result:
Other Scientific Packages provide algorithms that can be useful for image processing. In this example, we use

12.5. Feature extraction 364 12.5. Feature extraction 365

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

the spectral clustering function of the scikit-learn in order to segment glued objects.
>>> points = l*np.random.random((2, n**2))
>>> from sklearn.feature_extraction import image >>> im[(points[0]).astype(np.int), (points[1]).astype(np.int)] = 1
>>> from sklearn.cluster import spectral_clustering >>> im = ndimage.gaussian_filter(im, sigma=l/(4.*n))
>>> mask = im > im.mean()
>>> l = 100
>>> x, y = np.indices((l, l)) • Analysis of connected components

>>> center1 = (28, 24) Label connected components: ndimage.label:

>>> center2 = (40, 50)
>>> label_im, nb_labels = ndimage.label(mask)
>>> center3 = (67, 58)
>>> nb_labels # how many regions?
>>> center4 = (24, 70)
28
>>> radius1, radius2, radius3, radius4 = 16, 14, 15, 14
>>> plt.imshow(label_im)
<matplotlib.image.AxesImage object at 0x...>
>>> circle1 = (x - center1[0])**2 + (y - center1[1])**2 < radius1**2
>>> circle2 = (x - center2[0])**2 + (y - center2[1])**2 < radius2**2
>>> circle3 = (x - center3[0])**2 + (y - center3[1])**2 < radius3**2
>>> circle4 = (x - center4[0])**2 + (y - center4[1])**2 < radius4**2

>>> # 4 circles
>>> img = circle1 + circle2 + circle3 + circle4
>>> mask = img.astype(bool)
>>> img = img.astype(float)

>>> img += 1 + 0.2*np.random.randn(*img.shape)

>>> # Convert the image into a graph with the value of the gradient on
>>> # the edges.
>>> graph = image.img_to_graph(img, mask=mask)

>>> # Take a decreasing function of the gradient: we take it weakly

>>> # dependant from the gradient the segmentation is close to a voronoi
>>> graph.data = np.exp(-graph.data/graph.data.std()) Compute size, mean_value, etc. of each region:

>>> labels = spectral_clustering(graph, n_clusters=4, eigen_solver='arpack') >>> sizes = ndimage.sum(mask, label_im, range(nb_labels + 1))
>>> label_im = -np.ones(mask.shape) >>> mean_vals = ndimage.sum(im, label_im, range(1, nb_labels + 1))
>>> label_im[mask] = labels
Clean up small connect components:

>>> mask_size = sizes < 1000

>>> remove_pixel = mask_size[label_im]
>>> remove_pixel.shape
(256, 256)
>>> label_im[remove_pixel] = 0
>>> plt.imshow(label_im)
<matplotlib.image.AxesImage object at 0x...>

Now reassign labels with np.searchsorted:

>>> labels = np.unique(label_im)

>>> label_im = np.searchsorted(labels, label_im)

Find region of interest enclosing object:

>>> slice_x, slice_y = ndimage.find_objects(label_im==4)[0]

>>> roi = im[slice_x, slice_y]
>>> plt.imshow(roi)
12.6 Measuring objects properties: ndimage.measurements <matplotlib.image.AxesImage object at 0x...>

Other spatial measures: ndimage.center_of_mass, ndimage.maximum_position, etc.

Synthetic data:
Can be used outside the limited scope of segmentation applications.
>>> n = 10
>>> l = 256 Example: block mean:
>>> im = np.zeros((l, l))

12.6. Measuring objects properties: ndimage.measurements 366 12.6. Measuring objects properties: ndimage.measurements 367
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

>>> from scipy import misc

>>> f = misc.face(gray=True)
>>> sx, sy = f.shape
>>> X, Y = np.ogrid[0:sx, 0:sy]
>>> regions = (sy//6) * (X//4) + (Y//6) # note that we use broadcasting
>>> block_mean = ndimage.mean(f, labels=regions, index=np.arange(1,
... regions.max() +1))
>>> block_mean.shape = (sx // 4, sy // 6)

When regions are regular blocks, it is more efficient to use stride tricks (Example: fake dimensions with strides).
Non-regularly-spaced blocks: radial mean:

>>> sx, sy = f.shape

>>> X, Y = np.ogrid[0:sx, 0:sy]
>>> r = np.hypot(X - sx/2, Y - sy/2)
>>> rbin = (20* r/r.max()).astype(np.int)
>>> radial_mean = ndimage.mean(f, labels=rbin, index=np.arange(1, rbin.max() +1))

• Other measures
Correlation function, Fourier/wavelet spectrum, etc.
One example with mathematical morphology: granulometry

>>> def disk_structure(n):

... struct = np.zeros((2 * n + 1, 2 * n + 1))
... x, y = np.indices((2 * n + 1, 2 * n + 1))
... mask = (x - n)**2 + (y - n)**2 <= n**2
... struct[mask] = 1
... return struct.astype(np.bool)
...
>>>
>>> def granulometry(data, sizes=None):
... s = max(data.shape)
... if sizes is None:
... sizes = range(1, s/2, 2)
... granulo = [ndimage.binary_opening(data, \
... structure=disk_structure(n)).sum() for n in sizes]

12.6. Measuring objects properties: ndimage.measurements 368 12.6. Measuring objects properties: ndimage.measurements 369
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

12.7 Full code examples

12.8 Examples for the image processing chapter

12.8.1 Displaying a Racoon Face

Small example to plot a racoon face.

... return granulo

...
>>>
>>> np.random.seed(1)
>>> n = 10
>>> l = 256
>>> im = np.zeros((l, l))
>>> points = l*np.random.random((2, n**2))
>>> im[(points[0]).astype(np.int), (points[1]).astype(np.int)] = 1
>>> im = ndimage.gaussian_filter(im, sigma=l/(4.*n))
>>>
>>> mask = im > im.mean()
>>>
>>> granulo = granulometry(mask, sizes=np.arange(2, 19, 4))

from scipy import misc

f = misc.face()
misc.imsave('face.png', f) # uses the Image module (PIL)

import matplotlib.pyplot as plt

plt.imshow(f)
plt.show()

Total running time of the script: ( 0 minutes 0.426 seconds)

Download Python source code: plot_face.py
Download Jupyter notebook: plot_face.ipynb
Generated by Sphinx-Gallery

12.8.2 Image interpolation

The example demonstrates image interpolation on a Racoon face.

12.6. Measuring objects properties: ndimage.measurements 370 12.7. Full code examples 371
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

import numpy as np
import scipy.misc import scipy
import matplotlib.pyplot as plt import scipy.misc
import matplotlib.pyplot as plt
f = scipy.misc.face(gray=True)
face = scipy.misc.face(gray=True)
plt.figure(figsize=(8, 4)) face[10:13, 20:23]
face[100:120] = 255
plt.subplot(1, 2, 1)
plt.imshow(f[320:340, 510:530], cmap=plt.cm.gray) lx, ly = face.shape
plt.axis('off') X, Y = np.ogrid[0:lx, 0:ly]
mask = (X - lx/2)**2 + (Y - ly/2)**2 > lx*ly/4
plt.subplot(1, 2, 2) face[mask] = 0
plt.imshow(f[320:340, 510:530], cmap=plt.cm.gray, interpolation='nearest') face[range(400), range(400)] = 255
plt.axis('off')
plt.figure(figsize=(3, 3))
plt.subplots_adjust(wspace=0.02, hspace=0.02, top=1, bottom=0, left=0, right=1) plt.axes([0, 0, 1, 1])
plt.show() plt.imshow(face, cmap=plt.cm.gray)
plt.axis('off')
Total running time of the script: ( 0 minutes 0.254 seconds)
plt.show()
Download Python source code: plot_interpolation_face.py
Download Jupyter notebook: plot_interpolation_face.ipynb Total running time of the script: ( 0 minutes 0.171 seconds)

Generated by Sphinx-Gallery Download Python source code: plot_numpy_array.py

Download Jupyter notebook: plot_numpy_array.ipynb
12.8.3 Image manipulation and numpy arrays Generated by Sphinx-Gallery

This example shows how to do image manipulation using common numpy arrays tricks.
12.8.4 Radial mean

This example shows how to do a radial mean with scikit-image.

12.8. Examples for the image processing chapter 372 12.8. Examples for the image processing chapter 373
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

12.8.5 Plot the block mean of an image

An example showing how to use broad-casting to plot the mean of blocks of an image.

import numpy as np
import scipy
from scipy import ndimage
import matplotlib.pyplot as plt
import numpy as np
f = scipy.misc.face(gray=True) import scipy.misc
sx, sy = f.shape from scipy import ndimage
X, Y = np.ogrid[0:sx, 0:sy] import matplotlib.pyplot as plt

f = scipy.misc.face(gray=True)
r = np.hypot(X - sx/2, Y - sy/2) sx, sy = f.shape
X, Y = np.ogrid[0:sx, 0:sy]
rbin = (20* r/r.max()).astype(np.int)
radial_mean = ndimage.mean(f, labels=rbin, index=np.arange(1, rbin.max() +1)) regions = sy//6 * (X//4) + Y//6
block_mean = ndimage.mean(f, labels=regions,
plt.figure(figsize=(5, 5)) index=np.arange(1, regions.max() +1))
plt.axes([0, 0, 1, 1]) block_mean.shape = (sx//4, sy//6)
plt.imshow(rbin, cmap=plt.cm.spectral)
plt.axis('off') plt.figure(figsize=(5, 5))
plt.imshow(block_mean, cmap=plt.cm.gray)
plt.show() plt.axis('off')

Total running time of the script: ( 0 minutes 0.186 seconds) plt.show()

Download Python source code: plot_radial_mean.py Total running time of the script: ( 0 minutes 0.164 seconds)
Download Jupyter notebook: plot_radial_mean.ipynb Download Python source code: plot_block_mean.py
Generated by Sphinx-Gallery Download Jupyter notebook: plot_block_mean.ipynb

12.8. Examples for the image processing chapter 374 12.8. Examples for the image processing chapter 375
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Generated by Sphinx-Gallery

12.8.6 Display a Racoon Face

An example that displays a racoon face with matplotlib.

import scipy
from scipy import ndimage
import matplotlib.pyplot as plt

f = scipy.misc.face(gray=True).astype(float)
blurred_f = ndimage.gaussian_filter(f, 3)

filter_blurred_f = ndimage.gaussian_filter(blurred_f, 1)
import scipy.misc
import matplotlib.pyplot as plt alpha = 30
sharpened = blurred_f + alpha * (blurred_f - filter_blurred_f)
f = scipy.misc.face(gray=True)
plt.figure(figsize=(12, 4))
plt.figure(figsize=(10, 3.6))
plt.subplot(131)
plt.subplot(131) plt.imshow(f, cmap=plt.cm.gray)
plt.imshow(f, cmap=plt.cm.gray) plt.axis('off')
plt.subplot(132)
plt.subplot(132) plt.imshow(blurred_f, cmap=plt.cm.gray)
plt.imshow(f, cmap=plt.cm.gray, vmin=30, vmax=200) plt.axis('off')
plt.axis('off') plt.subplot(133)
plt.imshow(sharpened, cmap=plt.cm.gray)
plt.subplot(133) plt.axis('off')
plt.imshow(f, cmap=plt.cm.gray)
plt.contour(f, [50, 200]) plt.tight_layout()
plt.axis('off') plt.show()

plt.subplots_adjust(wspace=0, hspace=0., top=0.99, bottom=0.01, left=0.05, Total running time of the script: ( 0 minutes 0.355 seconds)
right=0.99)
plt.show() Download Python source code: plot_sharpen.py
Download Jupyter notebook: plot_sharpen.ipynb
Total running time of the script: ( 0 minutes 0.403 seconds)
Generated by Sphinx-Gallery
Download Python source code: plot_display_face.py
Download Jupyter notebook: plot_display_face.ipynb
12.8.8 Blurring of images
Generated by Sphinx-Gallery
An example showing various processes that blur an image.

12.8.7 Image sharpening

This example shows how to sharpen an image in noiseless situation by applying the filter inverse to the blur.

12.8. Examples for the image processing chapter 376 12.8. Examples for the image processing chapter 377
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

import scipy.misc import numpy as np

from scipy import ndimage from scipy import ndimage
import matplotlib.pyplot as plt import matplotlib.pyplot as plt

face = scipy.misc.face(gray=True) np.random.seed(1)

blurred_face = ndimage.gaussian_filter(face, sigma=3) n = 10
very_blurred = ndimage.gaussian_filter(face, sigma=5) l = 256
local_mean = ndimage.uniform_filter(face, size=11) im = np.zeros((l, l))
points = l*np.random.random((2, n**2))
plt.figure(figsize=(9, 3)) im[(points[0]).astype(np.int), (points[1]).astype(np.int)] = 1
plt.subplot(131) im = ndimage.gaussian_filter(im, sigma=l/(4.*n))
plt.imshow(blurred_face, cmap=plt.cm.gray)
plt.axis('off') mask = im > im.mean()
plt.subplot(132)
plt.imshow(very_blurred, cmap=plt.cm.gray) label_im, nb_labels = ndimage.label(mask)
plt.axis('off')
plt.subplot(133) plt.figure(figsize=(9,3))
plt.imshow(local_mean, cmap=plt.cm.gray)
plt.axis('off') plt.subplot(131)
plt.imshow(im)
plt.subplots_adjust(wspace=0, hspace=0., top=0.99, bottom=0.01, plt.axis('off')
left=0.01, right=0.99) plt.subplot(132)
plt.imshow(mask, cmap=plt.cm.gray)
plt.show() plt.axis('off')
plt.subplot(133)
Total running time of the script: ( 0 minutes 0.309 seconds) plt.imshow(label_im, cmap=plt.cm.spectral)
plt.axis('off')
Download Python source code: plot_blur.py
plt.subplots_adjust(wspace=0.02, hspace=0.02, top=1, bottom=0, left=0, right=1)
Download Jupyter notebook: plot_blur.ipynb
plt.show()
Generated by Sphinx-Gallery
Total running time of the script: ( 0 minutes 0.106 seconds)

12.8.9 Synthetic data Download Python source code: plot_synthetic_data.py

Download Jupyter notebook: plot_synthetic_data.ipynb
The example generates and displays simple synthetic data.
Generated by Sphinx-Gallery

12.8.10 Opening, erosion, and propagation

This example shows simple operations of mathematical morphology.

12.8. Examples for the image processing chapter 378 12.8. Examples for the image processing chapter 379
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

import numpy as np
import scipy
import scipy.misc
import numpy as np from scipy import ndimage
from scipy import ndimage import matplotlib.pyplot as plt
import matplotlib.pyplot as plt
f = scipy.misc.face(gray=True)
square = np.zeros((32, 32)) f = f[230:290, 220:320]
square[10:-10, 10:-10] = 1
np.random.seed(2) noisy = f + 0.4*f.std()*np.random.random(f.shape)
x, y = (32*np.random.random((2, 20))).astype(np.int)
square[x, y] = 1 gauss_denoised = ndimage.gaussian_filter(noisy, 2)
med_denoised = ndimage.median_filter(noisy, 3)
open_square = ndimage.binary_opening(square)

eroded_square = ndimage.binary_erosion(square) plt.figure(figsize=(12,2.8))

reconstruction = ndimage.binary_propagation(eroded_square, mask=square)
plt.subplot(131)
plt.figure(figsize=(9.5, 3)) plt.imshow(noisy, cmap=plt.cm.gray, vmin=40, vmax=220)
plt.subplot(131) plt.axis('off')
plt.imshow(square, cmap=plt.cm.gray, interpolation='nearest') plt.title('noisy', fontsize=20)
plt.axis('off') plt.subplot(132)
plt.subplot(132) plt.imshow(gauss_denoised, cmap=plt.cm.gray, vmin=40, vmax=220)
plt.imshow(open_square, cmap=plt.cm.gray, interpolation='nearest') plt.axis('off')
plt.axis('off') plt.title('Gaussian filter', fontsize=20)
plt.subplot(133) plt.subplot(133)
plt.imshow(reconstruction, cmap=plt.cm.gray, interpolation='nearest') plt.imshow(med_denoised, cmap=plt.cm.gray, vmin=40, vmax=220)
plt.axis('off') plt.axis('off')
plt.title('Median filter', fontsize=20)
plt.subplots_adjust(wspace=0, hspace=0.02, top=0.99, bottom=0.01, left=0.01, right=0.99)
plt.show() plt.subplots_adjust(wspace=0.02, hspace=0.02, top=0.9, bottom=0, left=0,
right=1)
plt.show()
Total running time of the script: ( 0 minutes 0.100 seconds)
Download Python source code: plot_propagation.py Total running time of the script: ( 0 minutes 0.228 seconds)
Download Jupyter notebook: plot_propagation.ipynb Download Python source code: plot_face_denoise.py
Generated by Sphinx-Gallery Download Jupyter notebook: plot_face_denoise.ipynb
Generated by Sphinx-Gallery
12.8.11 Image denoising
12.8.12 Find the bounding box of an object
This example demoes image denoising on a Racoon face.
This example shows how to extract the bounding box of the largest object

12.8. Examples for the image processing chapter 380 12.8. Examples for the image processing chapter 381
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

import numpy as np
from scipy import ndimage
import matplotlib.pyplot as plt

np.random.seed(1)
n = 10
l = 256
import numpy as np
im = np.zeros((l, l))
from scipy import ndimage
points = l*np.random.random((2, n**2))
import matplotlib.pyplot as plt
im[(points[0]).astype(np.int), (points[1]).astype(np.int)] = 1
im = ndimage.gaussian_filter(im, sigma=l/(4.*n))
np.random.seed(1)
n = 10
mask = im > im.mean()
l = 256
im = np.zeros((l, l))
label_im, nb_labels = ndimage.label(mask)
points = l*np.random.random((2, n**2))
im[(points[0]).astype(np.int), (points[1]).astype(np.int)] = 1
# Find the largest connected component
im = ndimage.gaussian_filter(im, sigma=l/(4.*n))
sizes = ndimage.sum(mask, label_im, range(nb_labels + 1))
mask_size = sizes < 1000
mask = im > im.mean()
remove_pixel = mask_size[label_im]
label_im[remove_pixel] = 0
label_im, nb_labels = ndimage.label(mask)
labels = np.unique(label_im)
label_im = np.searchsorted(labels, label_im)
sizes = ndimage.sum(mask, label_im, range(nb_labels + 1))
mask_size = sizes < 1000
# Now that we have only one connected component, extract it's bounding box
remove_pixel = mask_size[label_im]
slice_x, slice_y = ndimage.find_objects(label_im==4)[0]
label_im[remove_pixel] = 0
roi = im[slice_x, slice_y]
labels = np.unique(label_im)
label_clean = np.searchsorted(labels, label_im)
plt.figure(figsize=(4, 2))
plt.axes([0, 0, 1, 1])
plt.imshow(roi)
plt.figure(figsize=(6 ,3))
plt.axis('off')
plt.subplot(121)
plt.show()
plt.imshow(label_im, cmap=plt.cm.spectral)
plt.axis('off')
Total running time of the script: ( 0 minutes 0.102 seconds) plt.subplot(122)
Download Python source code: plot_find_object.py plt.imshow(label_clean, vmax=nb_labels, cmap=plt.cm.spectral)
plt.axis('off')
Download Jupyter notebook: plot_find_object.ipynb
plt.subplots_adjust(wspace=0.01, hspace=0.01, top=1, bottom=0, left=0, right=1)
Generated by Sphinx-Gallery plt.show()

12.8.13 Measurements from images Total running time of the script: ( 0 minutes 0.073 seconds)
Download Python source code: plot_measure_data.py
This examples shows how to measure quantities from various images.
Download Jupyter notebook: plot_measure_data.ipynb
Generated by Sphinx-Gallery

12.8. Examples for the image processing chapter 382 12.8. Examples for the image processing chapter 383
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

12.8.14 Geometrical transformations 12.8.15 Denoising an image with the median filter

This examples demos some simple geometrical transformations on a Racoon face. This example shows the original image, the noisy image, the denoised one (with the median filter) and the
difference between the two.

import numpy as np
import scipy.misc
from scipy import ndimage
import matplotlib.pyplot as plt

face = scipy.misc.face(gray=True)
import numpy as np
lx, ly = face.shape
from scipy import ndimage
# Cropping
import matplotlib.pyplot as plt
crop_face = face[lx//4:-lx//4, ly//4:-ly//4]
# up <-> down flip
im = np.zeros((20, 20))
flip_ud_face = np.flipud(face)
im[5:-5, 5:-5] = 1
# rotation
im = ndimage.distance_transform_bf(im)
rotate_face = ndimage.rotate(face, 45)
im_noise = im + 0.2*np.random.randn(*im.shape)
rotate_face_noreshape = ndimage.rotate(face, 45, reshape=False)
im_med = ndimage.median_filter(im_noise, 3)
plt.figure(figsize=(12.5, 2.5))
plt.figure(figsize=(16, 5))
plt.subplot(151)
plt.subplot(141)
plt.imshow(face, cmap=plt.cm.gray)
plt.imshow(im, interpolation='nearest')
plt.axis('off')
plt.axis('off')
plt.subplot(152)
plt.title('Original image', fontsize=20)
plt.imshow(crop_face, cmap=plt.cm.gray)
plt.subplot(142)
plt.axis('off')
plt.imshow(im_noise, interpolation='nearest', vmin=0, vmax=5)
plt.subplot(153)
plt.axis('off')
plt.imshow(flip_ud_face, cmap=plt.cm.gray)
plt.title('Noisy image', fontsize=20)
plt.axis('off')
plt.subplot(143)
plt.subplot(154)
plt.imshow(im_med, interpolation='nearest', vmin=0, vmax=5)
plt.imshow(rotate_face, cmap=plt.cm.gray)
plt.axis('off')
plt.axis('off')
plt.title('Median filter', fontsize=20)
plt.subplot(155)
plt.subplot(144)
plt.imshow(rotate_face_noreshape, cmap=plt.cm.gray)
plt.imshow(np.abs(im - im_med), cmap=plt.cm.hot, interpolation='nearest')
plt.axis('off')
plt.axis('off')
plt.title('Error', fontsize=20)
plt.subplots_adjust(wspace=0.02, hspace=0.3, top=1, bottom=0.1, left=0,
right=1)
plt.subplots_adjust(wspace=0.02, hspace=0.02, top=0.9, bottom=0, left=0,
plt.show()
right=1)

Total running time of the script: ( 0 minutes 0.545 seconds) plt.show()

Download Python source code: plot_geom_face.py
Total running time of the script: ( 0 minutes 0.132 seconds)
Download Jupyter notebook: plot_geom_face.ipynb
Download Python source code: plot_denoising.py
Generated by Sphinx-Gallery
Download Jupyter notebook: plot_denoising.ipynb
Generated by Sphinx-Gallery

12.8. Examples for the image processing chapter 384 12.8. Examples for the image processing chapter 385
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

12.8.16 Histogram segmentation Download Jupyter notebook: plot_histo_segmentation.ipynb

Generated by Sphinx-Gallery
This example does simple histogram analysis to perform segmentation.

12.8.17 Finding edges with Sobel filters

The Sobel filter is one of the simplest way of finding edges.

import numpy as np
from scipy import ndimage
import matplotlib.pyplot as plt
import numpy as np
from scipy import ndimage
np.random.seed(1)
import matplotlib.pyplot as plt
n = 10
l = 256
im = np.zeros((256, 256))
im = np.zeros((l, l))
im[64:-64, 64:-64] = 1
points = l*np.random.random((2, n**2))
im[(points[0]).astype(np.int), (points[1]).astype(np.int)] = 1
im = ndimage.rotate(im, 15, mode='constant')
im = ndimage.gaussian_filter(im, sigma=l/(4.*n))
im = ndimage.gaussian_filter(im, 8)
mask = (im > im.mean()).astype(np.float)
sx = ndimage.sobel(im, axis=0, mode='constant')
sy = ndimage.sobel(im, axis=1, mode='constant')
mask += 0.1 * im
sob = np.hypot(sx, sy)
img = mask + 0.2*np.random.randn(*mask.shape)
plt.figure(figsize=(16, 5))
plt.subplot(141)
hist, bin_edges = np.histogram(img, bins=60)
plt.imshow(im, cmap=plt.cm.gray)
bin_centers = 0.5*(bin_edges[:-1] + bin_edges[1:])
plt.axis('off')
plt.title('square', fontsize=20)
binary_img = img > 0.5
plt.subplot(142)
plt.imshow(sx)
plt.figure(figsize=(11,4))
plt.axis('off')
plt.title('Sobel (x direction)', fontsize=20)
plt.subplot(131)
plt.subplot(143)
plt.imshow(img)
plt.imshow(sob)
plt.axis('off')
plt.axis('off')
plt.subplot(132)
plt.title('Sobel filter', fontsize=20)
plt.plot(bin_centers, hist, lw=2)
plt.axvline(0.5, color='r', ls='--', lw=2)
im += 0.07*np.random.random(im.shape)
plt.text(0.57, 0.8, 'histogram', fontsize=20, transform = plt.gca().transAxes)
plt.yticks([])
sx = ndimage.sobel(im, axis=0, mode='constant')
plt.subplot(133)
sy = ndimage.sobel(im, axis=1, mode='constant')
plt.imshow(binary_img, cmap=plt.cm.gray, interpolation='nearest')
sob = np.hypot(sx, sy)
plt.axis('off')
plt.subplot(144)
plt.subplots_adjust(wspace=0.02, hspace=0.3, top=1, bottom=0.1, left=0, right=1)
plt.imshow(sob)
plt.show()
plt.axis('off')
plt.title('Sobel for noisy image', fontsize=20)
Total running time of the script: ( 0 minutes 0.109 seconds)
Download Python source code: plot_histo_segmentation.py

12.8. Examples for the image processing chapter 386 12.8. Examples for the image processing chapter 387
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

plt.subplots_adjust(wspace=0.02, hspace=0.02, top=1, bottom=0, left=0, right=0.9) right=1)

plt.show()
plt.show()
Total running time of the script: ( 0 minutes 0.246 seconds)
Total running time of the script: ( 0 minutes 0.158 seconds)
Download Python source code: plot_face_tv_denoise.py
Download Python source code: plot_find_edges.py
Download Jupyter notebook: plot_face_tv_denoise.ipynb
Download Jupyter notebook: plot_find_edges.ipynb
Generated by Sphinx-Gallery
Generated by Sphinx-Gallery

12.8.19 Greyscale dilation

12.8.18 Total Variation denoising
This example illustrates greyscale mathematical morphology.
This example demoes Total-Variation (TV) denoising on a Racoon face.

import numpy as np
import numpy as np from scipy import ndimage
import scipy import matplotlib.pyplot as plt
import scipy.misc
import matplotlib.pyplot as plt im = np.zeros((64, 64))
try: np.random.seed(2)
from skimage.restoration import denoise_tv_chambolle x, y = (63*np.random.random((2, 8))).astype(np.int)
except ImportError: im[x, y] = np.arange(8)
# skimage < 0.12
from skimage.filters import denoise_tv_chambolle bigger_points = ndimage.grey_dilation(im, size=(5, 5), structure=np.ones((5, 5)))

f = scipy.misc.face(gray=True) square = np.zeros((16, 16))

f = f[230:290, 220:320] square[4:-4, 4:-4] = 1
dist = ndimage.distance_transform_bf(square)
noisy = f + 0.4*f.std()*np.random.random(f.shape) dilate_dist = ndimage.grey_dilation(dist, size=(3, 3), \
structure=np.ones((3, 3)))
tv_denoised = denoise_tv_chambolle(noisy, weight=10)
plt.figure(figsize=(12.5, 3))
plt.subplot(141)
plt.figure(figsize=(12, 2.8)) plt.imshow(im, interpolation='nearest', cmap=plt.cm.spectral)
plt.axis('off')
plt.subplot(131) plt.subplot(142)
plt.imshow(noisy, cmap=plt.cm.gray, vmin=40, vmax=220) plt.imshow(bigger_points, interpolation='nearest', cmap=plt.cm.spectral)
plt.axis('off') plt.axis('off')
plt.title('noisy', fontsize=20) plt.subplot(143)
plt.subplot(132) plt.imshow(dist, interpolation='nearest', cmap=plt.cm.spectral)
plt.imshow(tv_denoised, cmap=plt.cm.gray, vmin=40, vmax=220) plt.axis('off')
plt.axis('off') plt.subplot(144)
plt.title('TV denoising', fontsize=20) plt.imshow(dilate_dist, interpolation='nearest', cmap=plt.cm.spectral)
plt.axis('off')
tv_denoised = denoise_tv_chambolle(noisy, weight=50)
plt.subplot(133) plt.subplots_adjust(wspace=0, hspace=0.02, top=0.99, bottom=0.01, left=0.01, right=0.99)
plt.imshow(tv_denoised, cmap=plt.cm.gray, vmin=40, vmax=220) plt.show()
plt.axis('off')
plt.title('(more) TV denoising', fontsize=20)
Total running time of the script: ( 0 minutes 0.133 seconds)
plt.subplots_adjust(wspace=0.02, hspace=0.02, top=0.9, bottom=0, left=0, Download Python source code: plot_greyscale_dilation.py

12.8. Examples for the image processing chapter 388 12.8. Examples for the image processing chapter 389
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Download Jupyter notebook: plot_greyscale_dilation.ipynb

plt.subplots_adjust(wspace=0.02, hspace=0.3, top=1, bottom=0.1, left=0, right=1)
Generated by Sphinx-Gallery
plt.show()

12.8.20 Cleaning segmentation with mathematical morphology Total running time of the script: ( 0 minutes 0.146 seconds)
Download Python source code: plot_clean_morpho.py
An example showing how to clean segmentation with mathematical morphology: removing small regions and
holes. Download Jupyter notebook: plot_clean_morpho.ipynb
Generated by Sphinx-Gallery

12.8.21 Segmentation with Gaussian mixture models

This example performs a Gaussian mixture model analysis of the image histogram to find the right thresholds
for separating foreground from background.

import numpy as np
from scipy import ndimage
import matplotlib.pyplot as plt

np.random.seed(1)
n = 10
l = 256
im = np.zeros((l, l))
points = l*np.random.random((2, n**2))
im[(points[0]).astype(np.int), (points[1]).astype(np.int)] = 1
im = ndimage.gaussian_filter(im, sigma=l/(4.*n))

mask = (im > im.mean()).astype(np.float)

import numpy as np
from scipy import ndimage
img = mask + 0.3*np.random.randn(*mask.shape) import matplotlib.pyplot as plt
from sklearn.mixture import GMM
binary_img = img > 0.5
np.random.seed(1)
# Remove small white regions n = 10
open_img = ndimage.binary_opening(binary_img) l = 256
# Remove small black hole im = np.zeros((l, l))
close_img = ndimage.binary_closing(open_img) points = l*np.random.random((2, n**2))
im[(points[0]).astype(np.int), (points[1]).astype(np.int)] = 1
plt.figure(figsize=(12, 3)) im = ndimage.gaussian_filter(im, sigma=l/(4.*n))

l = 128 mask = (im > im.mean()).astype(np.float)

plt.subplot(141)
plt.imshow(binary_img[:l, :l], cmap=plt.cm.gray) img = mask + 0.3*np.random.randn(*mask.shape)
plt.axis('off')
plt.subplot(142) hist, bin_edges = np.histogram(img, bins=60)
plt.imshow(open_img[:l, :l], cmap=plt.cm.gray) bin_centers = 0.5*(bin_edges[:-1] + bin_edges[1:])
plt.axis('off')
plt.subplot(143) classif = GMM(n_components=2)
plt.imshow(close_img[:l, :l], cmap=plt.cm.gray) classif.fit(img.reshape((img.size, 1)))
plt.axis('off')
plt.subplot(144) threshold = np.mean(classif.means_)
plt.imshow(mask[:l, :l], cmap=plt.cm.gray) binary_img = img > threshold
plt.contour(close_img[:l, :l], [0.5], linewidths=2, colors='r')
plt.axis('off')
plt.figure(figsize=(11,4))

12.8. Examples for the image processing chapter 390 12.8. Examples for the image processing chapter 391
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

local_maxi = peak_local_max(
plt.subplot(131) distance, indices=False, footprint=np.ones((3, 3)), labels=image)
plt.imshow(img) markers = ndimage.label(local_maxi)[0]
plt.axis('off') labels = watershed(-distance, markers, mask=image)
plt.subplot(132)
plt.plot(bin_centers, hist, lw=2) plt.figure(figsize=(9, 3.5))
plt.axvline(0.5, color='r', ls='--', lw=2) plt.subplot(131)
plt.text(0.57, 0.8, 'histogram', fontsize=20, transform = plt.gca().transAxes) plt.imshow(image, cmap='gray', interpolation='nearest')
plt.yticks([]) plt.axis('off')
plt.subplot(133) plt.subplot(132)
plt.imshow(binary_img, cmap=plt.cm.gray, interpolation='nearest') plt.imshow(-distance, interpolation='nearest')
plt.axis('off') plt.axis('off')
plt.subplot(133)
plt.subplots_adjust(wspace=0.02, hspace=0.3, top=1, bottom=0.1, left=0, right=1) plt.imshow(labels, cmap='spectral', interpolation='nearest')
plt.show() plt.axis('off')

Total running time of the script: ( 0 minutes 0.463 seconds) plt.subplots_adjust(hspace=0.01, wspace=0.01, top=1, bottom=0, left=0,
right=1)
Download Python source code: plot_GMM.py plt.show()
Download Jupyter notebook: plot_GMM.ipynb
Total running time of the script: ( 0 minutes 0.127 seconds)
Generated by Sphinx-Gallery
Download Python source code: plot_watershed_segmentation.py
Download Jupyter notebook: plot_watershed_segmentation.ipynb
12.8.22 Watershed segmentation
Generated by Sphinx-Gallery
This example shows how to do segmentation with watershed.

12.8.23 Granulometry

This example performs a simple granulometry analysis.

import numpy as np
from skimage.morphology import watershed
from skimage.feature import peak_local_max
import numpy as np
import matplotlib.pyplot as plt
from scipy import ndimage
from scipy import ndimage
import matplotlib.pyplot as plt
# Generate an initial image with two overlapping circles
def disk_structure(n):
x, y = np.indices((80, 80))
struct = np.zeros((2 * n + 1, 2 * n + 1))
x1, y1, x2, y2 = 28, 28, 44, 52
x, y = np.indices((2 * n + 1, 2 * n + 1))
r1, r2 = 16, 20
mask = (x - n)**2 + (y - n)**2 <= n**2
mask_circle1 = (x - x1) ** 2 + (y - y1) ** 2 < r1 ** 2
struct[mask] = 1
mask_circle2 = (x - x2) ** 2 + (y - y2) ** 2 < r2 ** 2
return struct.astype(np.bool)
image = np.logical_or(mask_circle1, mask_circle2)
# Now we want to separate the two objects in image
# Generate the markers as local maxima of the distance
def granulometry(data, sizes=None):
# to the background
s = max(data.shape)
distance = ndimage.distance_transform_edt(image)
if sizes is None:

12.8. Examples for the image processing chapter 392 12.8. Examples for the image processing chapter 393
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

sizes = range(1, s/2, 2) circle1 = (x - center1[0])**2 + (y - center1[1])**2 < radius1**2

granulo = [ndimage.binary_opening(data, \ circle2 = (x - center2[0])**2 + (y - center2[1])**2 < radius2**2
structure=disk_structure(n)).sum() for n in sizes] circle3 = (x - center3[0])**2 + (y - center3[1])**2 < radius3**2
return granulo circle4 = (x - center4[0])**2 + (y - center4[1])**2 < radius4**2

4 circles
np.random.seed(1)
n = 10 img = circle1 + circle2 + circle3 + circle4
l = 256 mask = img.astype(bool)
im = np.zeros((l, l)) img = img.astype(float)
points = l*np.random.random((2, n**2))
im[(points[0]).astype(np.int), (points[1]).astype(np.int)] = 1 img += 1 + 0.2*np.random.randn(*img.shape)
im = ndimage.gaussian_filter(im, sigma=l/(4.*n))
# Convert the image into a graph with the value of the gradient on the
mask = im > im.mean() # edges.
graph = image.img_to_graph(img, mask=mask)
granulo = granulometry(mask, sizes=np.arange(2, 19, 4))
# Take a decreasing function of the gradient: we take it weakly
plt.figure(figsize=(6, 2.2)) # dependant from the gradient the segmentation is close to a voronoi
graph.data = np.exp(-graph.data / graph.data.std())
plt.subplot(121)
plt.imshow(mask, cmap=plt.cm.gray) # Force the solver to be arpack, since amg is numerically
opened = ndimage.binary_opening(mask, structure=disk_structure(10)) # unstable on this example
opened_more = ndimage.binary_opening(mask, structure=disk_structure(14)) labels = spectral_clustering(graph, n_clusters=4)
plt.contour(opened, [0.5], colors='b', linewidths=2) label_im = -np.ones(mask.shape)
plt.contour(opened_more, [0.5], colors='r', linewidths=2) label_im[mask] = labels
plt.axis('off')
plt.subplot(122) plt.figure(figsize=(6, 3))
plt.plot(np.arange(2, 19, 4), granulo, 'ok', ms=8) plt.subplot(121)
plt.imshow(img, cmap=plt.cm.spectral, interpolation='nearest')
plt.axis('off')
plt.subplots_adjust(wspace=0.02, hspace=0.15, top=0.95, bottom=0.15, left=0, right=0.95) plt.subplot(122)
plt.show() plt.imshow(label_im, cmap=plt.cm.spectral, interpolation='nearest')
plt.axis('off')
Total running time of the script: ( 0 minutes 0.318 seconds)
plt.subplots_adjust(wspace=0, hspace=0., top=0.99, bottom=0.01, left=0.01, right=0.99)
Download Python source code: plot_granulo.py plt.show()
Download Jupyter notebook: plot_granulo.ipynb
Generated by Sphinx-Gallery

12.8.24 Segmentation with spectral clustering

This example uses spectral clustering to do segmentation.

import numpy as np
import matplotlib.pyplot as plt

from sklearn.feature_extraction import image

from sklearn.cluster import spectral_clustering

l = 100
x, y = np.indices((l, l))

center1 = (28, 24)

center2 = (40, 50)
center3 = (67, 58)
center4 = (24, 70) Total running time of the script: ( 0 minutes 0.406 seconds)

radius1, radius2, radius3, radius4 = 16, 14, 15, 14 Download Python source code: plot_spectral_clustering.py
Download Jupyter notebook: plot_spectral_clustering.ipynb

12.8. Examples for the image processing chapter 394 12.8. Examples for the image processing chapter 395
 Scipy lecture notes, Edition 2017.1

Generated by Sphinx-Gallery
Download all examples in Python source code: auto_examples_python.zip
Download all examples in Jupyter notebooks: auto_examples_jupyter.zip
Generated by Sphinx-Gallery

See also:
More on image-processing:
• The chapter on Scikit-image
CHAPTER 13
• Other, more powerful and complete modules: OpenCV (Python bindings), CellProfiler, ITK with Python
bindings

Mathematical optimization: finding

minima of functions

Authors: Gaël Varoquaux

Mathematical optimization deals with the problem of finding numerically minimums (or maximums or zeros)
of a function. In this context, the function is called cost function, or objective function, or energy.
Here, we are interested in using scipy.optimize for black-box optimization: we do not rely on the mathe-
matical expression of the function that we are optimizing. Note that this expression can often be used for more
efficient, non black-box, optimization.

Prerequisites

• Numpy
• Scipy
• Matplotlib

See also:
References
Mathematical optimization is very . . . mathematical. If you want performance, it really pays to read the books:
• Convex Optimization by Boyd and Vandenberghe (pdf available free online).
• Numerical Optimization, by Nocedal and Wright. Detailed reference on gradient descent methods.
• Practical Methods of Optimization by Fletcher: good at hand-waving explanations.

12.8. Examples for the image processing chapter 396 397

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Chapters contents 13.1.1 Convex versus non-convex optimization

• Knowing your problem

– Convex versus non-convex optimization
– Smooth and non-smooth problems
– Noisy versus exact cost functions
– Constraints
• A review of the different optimizers
– Getting started: 1D optimization
– Gradient based methods
– Newton and quasi-newton methods
• Full code examples
• Examples for the mathematical optimization chapter
A convex function: A non-convex function
– Gradient-less methods • f is above all its tangents.
– Global optimizers • equivalently, for two point A, B, f(C) lies below
the segment [f(A), f(B])], if A < C < B
• Practical guide to optimization with scipy
– Choosing a method
Optimizing convex functions is easy. Optimizing non-convex functions can be very hard.
– Making your optimizer faster
– Computing gradients
Note: It can be proven that for a convex function a local minimum is also a global minimum. Then, in some
– Synthetic exercices sense, the minimum is unique.
• Special case: non-linear least-squares
– Minimizing the norm of a vector function
13.1.2 Smooth and non-smooth problems
– Curve fitting
• Optimization with constraints
– Box bounds
– General constraints
• Full code examples
• Examples for the mathematical optimization chapter

13.1 Knowing your problem

Not all optimization problems are equal. Knowing your problem enables you to choose the right tool.

Dimensionality of the problem

A smooth function: A non-smooth function
The scale of an optimization problem is pretty much set by the dimensionality of the problem, i.e. the num- The gradient is defined everywhere, and is a continuous function
ber of scalar variables on which the search is performed.
Optimizing smooth functions is easier (true in the context of black-box optimization, otherwise Linear Pro-
gramming is an example of methods which deal very efficiently with piece-wise linear functions).

13.1. Knowing your problem 398 13.1. Knowing your problem 399
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

13.1.3 Noisy versus exact cost functions >>> from scipy import optimize
>>> def f(x):
... return -np.exp(-(x - 0.7)**2)
>>> result = optimize.minimize_scalar(f)
>>> result.success # check if solver was successful
True
>>> x_min = result.x
>>> x_min
0.699999999...
>>> x_min - 0.7
-2.160590595323697e-10

Noisy (blue) and non-noisy (green) functions

Brent’s method on a quadratic function: it converges in
3 iterations, as the quadratic approximation is then ex-
Noisy gradients act.

Many optimization methods rely on gradients of the objective function. If the gradient function is not given,
they are computed numerically, which induces errors. In such situation, even if the objective function is not
noisy, a gradient-based optimization may be a noisy optimization.

Brent’s method on a non-convex function: note that the

fact that the optimizer avoided the local minimum is a
13.1.4 Constraints matter of luck.

Note: You can use different solvers using the parameter method.

Note: scipy.optimize.minimize_scalar() can also be used for optimization constrained to an interval

using the parameter bounds.

13.2.2 Gradient based methods

Some intuitions about gradient descent

Optimizations under constraints
Here we focus on intuitions, not code. Code will follow.
Here:
−1 < x 1 < 1 Gradient descent basically consists in taking small steps in the direction of the gradient, that is the direction of
−1 < x 2 < 1 the steepest descent.

13.2 A review of the different optimizers

13.2.1 Getting started: 1D optimization

Let’s get started by finding the minimum of the scalar function f (x) = exp[(x − 0.7)2 ]. scipy.optimize.
minimize_scalar() uses Brent’s method to find the minimum of a function:

13.2. A review of the different optimizers 400 13.2. A review of the different optimizers 401
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Table 13.1: Fixed step gradient descent Table 13.2: Adaptive step gradient descent

A well-conditioned quadratic function.

A well-conditioned quadratic
An ill-conditioned quadratic function. function.
The core problem of gradient-methods on ill-conditioned prob-
lems is that the gradient tends not to point in the direction of the
minimum.

We can see that very anisotropic (ill-conditioned) functions are harder to optimize.

Take home message: conditioning number and preconditioning

If you know natural scaling for your variables, prescale them so that they behave similarly. This is related to An ill-conditioned quadratic
preconditioning. function.

Also, it clearly can be advantageous to take bigger steps. This is done in gradient descent code using a line
search.

An ill-conditioned non-
quadratic function.

An ill-conditioned very non-

quadratic function.

The more a function looks like a quadratic function (elliptic iso-curves), the easier it is to optimize.

Conjugate gradient descent

The gradient descent algorithms above are toys not to be used on real problems.
As can be seen from the above experiments, one of the problems of the simple gradient descent algorithms, is
that it tends to oscillate across a valley, each time following the direction of the gradient, that makes it cross
the valley. The conjugate gradient solves this problem by adding a friction term: each step depends on the two
last values of the gradient and sharp turns are reduced.

13.2. A review of the different optimizers 402 13.2. A review of the different optimizers 403
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

13.2.3 Newton and quasi-newton methods

Table 13.3: Conjugate gradient descent
Newton methods: using the Hessian (2nd differential)

Newton methods use a local quadratic approximation to compute the jump direction. For this purpose, they
rely on the 2 first derivative of the function: the gradient and the Hessian.

An ill-conditioned non-
quadratic function. An ill-conditioned quadratic function:
Note that, as the quadratic approximation is exact, the Newton
method is blazing fast

An ill-conditioned non-quadratic function:

Here we are optimizing a Gaussian, which is always below its
quadratic approximation. As a result, the Newton method over-
An ill-conditioned very non- shoots and leads to oscillations.
quadratic function.

scipy provides scipy.optimize.minimize() to find the minimum of scalar functions of one or more vari-
ables. The simple conjugate gradient method can be used by setting the parameter method to CG
An ill-conditioned very non-quadratic function:
>>> def f(x): # The rosenbrock function
... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2
In scipy, you can use the Newton method by setting method to Newton-CG in scipy.optimize.minimize().
>>> optimize.minimize(f, [2, -1], method="CG")
Here, CG refers to the fact that an internal inversion of the Hessian is performed by conjugate gradient
fun: 1.6503...e-11
jac: array([ -6.1534...e-06, 2.5380...e-07]) >>> def f(x): # The rosenbrock function
message: ...'Optimization terminated successfully.' ... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2
nfev: 108 >>> def jacobian(x):
nit: 13 ... return np.array((-2*.5*(1 - x[0]) - 4*x[0]*(x[1] - x[0]**2), 2*(x[1] - x[0]**2)))
njev: 27 >>> optimize.minimize(f, [2,-1], method="Newton-CG", jac=jacobian)
status: 0 fun: 1.5601...e-15
success: True jac: array([ 1.0575...e-07, -7.4832...e-08])
x: array([ 0.99999..., 0.99998...]) message: ...'Optimization terminated successfully.'
nfev: 11
Gradient methods need the Jacobian (gradient) of the function. They can compute it numerically, but will nhev: 0
perform better if you can pass them the gradient: nit: 10
njev: 52
>>> def jacobian(x): status: 0
... return np.array((-2*.5*(1 - x[0]) - 4*x[0]*(x[1] - x[0]**2), 2*(x[1] - x[0]**2))) success: True
>>> optimize.minimize(f, [2, 1], method="CG", jac=jacobian) x: array([ 0.99999..., 0.99999...])
fun: 2.957...e-14
jac: array([ 7.1825...e-07, -2.9903...e-07])
Note that compared to a conjugate gradient (above), Newton’s method has required less function evaluations,
message: 'Optimization terminated successfully.'
but more gradient evaluations, as it uses it to approximate the Hessian. Let’s compute the Hessian and pass it
nfev: 16
nit: 8 to the algorithm:
njev: 16
>>> def hessian(x): # Computed with sympy
status: 0
... return np.array(((1 - 4*x[1] + 12*x[0]**2, -4*x[0]), (-4*x[0], 2)))
success: True
>>> optimize.minimize(f, [2,-1], method="Newton-CG", jac=jacobian, hess=hessian)
x: array([ 1.0000..., 1.0000...])
fun: 1.6277...e-15
jac: array([ 1.1104...e-07, -7.7809...e-08])
Note that the function has only been evaluated 27 times, compared to 108 without the gradient. message: ...'Optimization terminated successfully.'
nfev: 11
nhev: 10
nit: 10

13.2. A review of the different optimizers 404 13.2. A review of the different optimizers 405
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

njev: 20 pl.clf()
status: 0
success: True pl.plot(x_, f(x_) + .2*np.random.normal(size=31), linewidth=2)
x: array([ 0.99999..., 0.99999...]) pl.plot(x, f(x), linewidth=2)

pl.ylim(ymin=-1.3)
Note: At very high-dimension, the inversion of the Hessian can be costly and unstable (large scale > 250). pl.axis('off')
pl.tight_layout()
pl.show()

Note: Newton optimizers should not to be confused with Newton’s root finding method, based on the same Total running time of the script: ( 0 minutes 0.053 seconds)
principles, scipy.optimize.newton().
Download Python source code: plot_noisy.py
Download Jupyter notebook: plot_noisy.ipynb
Quasi-Newton methods: approximating the Hessian on the fly Generated by Sphinx-Gallery

BFGS: BFGS (Broyden-Fletcher-Goldfarb-Shanno algorithm) refines at each step an approximation of the Hes-
sian. 13.4.2 Smooth vs non-smooth

Draws a figure to explain smooth versus non smooth optimization.

13.3 Full code examples

13.4 Examples for the mathematical optimization chapter

13.4.1 Noisy optimization problem

•
Draws a figure explaining noisy vs non-noisy optimization

import numpy as np
import pylab as pl

x = np.linspace(-1.5, 1.5, 101)

# A smooth function
pl.figure(1, figsize=(3, 2.5))
pl.clf()

pl.plot(x, np.sqrt(.2 + x**2), linewidth=2)

import numpy as np pl.text(-1, 0, '$f$', size=20)
import pylab as pl
pl.ylim(ymin=-.2)
np.random.seed(0) pl.axis('off')
pl.tight_layout()
x = np.linspace(-5, 5, 101)
x_ = np.linspace(-5, 5, 31) # A non-smooth function
pl.figure(2, figsize=(3, 2.5))
def f(x): pl.clf()
return -np.exp(-x**2) pl.plot(x, np.abs(x), linewidth=2)
pl.text(-1, 0, '$f$', size=20)
# A smooth function
pl.figure(1, figsize=(3, 2.5)) pl.ylim(ymin=-.2)

13.3. Full code examples 406 13.4. Examples for the mathematical optimization chapter 407
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

pl.axis('off') # `params` vector

pl.tight_layout() params, params_cov = optimize.curve_fit(f, x, y)
pl.show()
# plot the data and the fitted curve
Total running time of the script: ( 0 minutes 0.402 seconds) t = np.linspace(0, 3, 1000)

Download Python source code: plot_smooth.py pl.figure(1)

pl.clf()
Download Jupyter notebook: plot_smooth.ipynb
pl.plot(x, y, 'bx')
Generated by Sphinx-Gallery pl.plot(t, f(t, *params), 'r-')
pl.show()

13.4.3 Curve fitting Total running time of the script: ( 0 minutes 0.035 seconds)
Download Python source code: plot_curve_fit.py
A curve fitting example
Download Jupyter notebook: plot_curve_fit.ipynb
Generated by Sphinx-Gallery

13.4.4 Convex function

A figure showing the definition of a convex function

import numpy as np
import pylab as pl

x = np.linspace(-1, 2)

pl.figure(1, figsize=(3, 2.5))

pl.clf()
import numpy as np
from scipy import optimize # A convex function
import pylab as pl pl.plot(x, x**2, linewidth=2)
pl.text(-.7, -.6**2, '$f$', size=20)
np.random.seed(0)
# The tangent in one point
# Our test function pl.plot(x, 2*x - 1)
def f(t, omega, phi): pl.plot(1, 1, 'k+')
return np.cos(omega * t + phi) pl.text(.3, -.75, "Tangent to $f$", size=15)
pl.text(1, 1 - .5, 'C', size=15)
# Our x and y data
x = np.linspace(0, 3, 50) # Convexity as barycenter
y = f(x, 1.5, 1) + .1*np.random.normal(size=50) pl.plot([.35, 1.85], [.35**2, 1.85**2])
pl.plot([.35, 1.85], [.35**2, 1.85**2], 'k+')
# Fit the model: the parameters omega and phi can be found in the

13.4. Examples for the mathematical optimization chapter 408 13.4. Examples for the mathematical optimization chapter 409
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

pl.text(.35 - .2, .35**2 + .1, 'A', size=15) pl.figure(1)

pl.text(1.85 - .2, 1.85**2, 'B', size=15) pl.clf()
X, Y = np.mgrid[-1.5:1.5:100j, -1.1:1.1:100j]
pl.ylim(ymin=-1) pl.imshow(f([X, Y]).T, cmap=pl.cm.gray_r, extent=[-1.5, 1.5, -1.1, 1.1],
pl.axis('off') origin='lower')
pl.tight_layout() pl.contour(X, Y, f([X, Y]), cmap=pl.cm.gnuplot)

# Convexity as barycenter # Plot the gradient

pl.figure(2, figsize=(3, 2.5)) dX, dY = g_prime([.1*X[::5, ::5], Y[::5, ::5]])
pl.clf() # Adjust for our preconditioning
pl.plot(x, x**2 + np.exp(-5*(x - .5)**2), linewidth=2) dX *= .1
pl.text(-.7, -.6**2, '$f$', size=20) pl.quiver(X[::5, ::5], Y[::5, ::5], dX, dY, color='.5')

pl.ylim(ymin=-1) # Plot our solution

pl.axis('off') pl.plot(x_min[0], x_min[1], 'r+', markersize=15)
pl.tight_layout()
pl.show() pl.show()

Total running time of the script: ( 0 minutes 0.112 seconds)

Download Python source code: plot_convex.py
Download Jupyter notebook: plot_convex.ipynb
Generated by Sphinx-Gallery

13.4.5 Finding a minimum in a flat neighborhood

An excercise of finding minimum. This excercise is hard because the function is very flat around the minimum
(all its derivatives are zero). Thus gradient information is unreliable.
The function admits a minimum in [0, 0]. The challenge is to get within 1e-7 of this minimum, starting at x0 =
[1, 1].
The solution that we adopt here is to give up on using gradient or information based on local differences, and
to rely on the Powell algorithm. With 162 function evaluations, we get to 1e-8 of the solution.
•
import numpy as np
from scipy import optimize
import pylab as pl

def f(x):
return np.exp(-1/(.01*x[0]**2 + x[1]**2))

# A well-conditionned version of f:
def g(x):
return f([10*x[0], x[1]])

# The gradient of g. We won't use it here for the optimization.

def g_prime(x):
r = np.sqrt(x[0]**2 + x[1]**2)
return 2/r**3*g(x)*x/r

result = optimize.minimize(g, [1, 1], method="Powell", tol=1e-10)

x_min = result.x
•
Some pretty plotting
Total running time of the script: ( 0 minutes 0.160 seconds)
pl.figure(0)
Download Python source code: plot_exercise_flat_minimum.py
pl.clf()
t = np.linspace(-1.1, 1.1, 100) Download Jupyter notebook: plot_exercise_flat_minimum.ipynb
pl.plot(t, f([0, t]))
Generated by Sphinx-Gallery

13.4. Examples for the mathematical optimization chapter 410 13.4. Examples for the mathematical optimization chapter 411
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

13.4.6 Optimization with constraints def constraint(x):

return np.atleast_1d(1.5 - np.sum(np.abs(x)))
An example showing how to do optimization with general constraints using SLSQP and cobyla.
optimize.minimize(f, np.array([0, 0]), method="SLSQP",
constraints={"fun": constraint, "type": "ineq"})

accumulated = np.array(accumulator)
pl.plot(accumulated[:, 0], accumulated[:, 1])

pl.show()

Total running time of the script: ( 0 minutes 0.061 seconds)

Download Python source code: plot_non_bounds_constraints.py
Download Jupyter notebook: plot_non_bounds_constraints.ipynb
Generated by Sphinx-Gallery

13.4.7 Brent’s method

import numpy as np
import pylab as pl Illustration of 1D optimization: Brent’s method
from scipy import optimize

x, y = np.mgrid[-2.03:4.2:.04, -1.6:3.2:.04]
x = x.T
y = y.T

pl.figure(1, figsize=(3, 2.5))

pl.clf() •
pl.axes([0, 0, 1, 1])

contours = pl.contour(np.sqrt((x - 3)**2 + (y - 2)**2),

extent=[-2.03, 4.2, -1.6, 3.2],
cmap=pl.cm.gnuplot)
pl.clabel(contours,
inline=1,
fmt='%1.1f ',
fontsize=14)
pl.plot([-1.5, 0, 1.5, 0, -1.5], •
[ 0, 1.5, 0, -1.5, 0], 'k', linewidth=2)
pl.fill_between([ -1.5, 0, 1.5],
[ 0, -1.5, 0],
[ 0, 1.5, 0],
color='.8')
pl.axvline(0, color='k')
pl.axhline(0, color='k')
•
pl.text(-.9, 2.8, '$x_2$', size=20)
pl.text(3.6, -.6, '$x_1$', size=20)
pl.axis('tight')
pl.axis('off')

# And now plot the optimization path

accumulator = list()

def f(x):
# Store the list of function calls •
accumulator.append(x)
return np.sqrt((x[0] - 3)**2 + (x[1] - 2)**2) Out:

13.4. Examples for the mathematical optimization chapter 412 13.4. Examples for the mathematical optimization chapter 413
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Generated by Sphinx-Gallery
Converged at 6
Converged at 23
13.4.8 Constraint optimization: visualizing the geometry

A small figure explaining optimization with constraints

import numpy as np
import pylab as pl
from scipy import optimize

x = np.linspace(-1, 3, 100)
x_0 = np.exp(-1) •

def f(x):
return (x - x_0)**2 + epsilon*np.exp(-5*(x - .5 - x_0)**2)

for epsilon in (0, 1):

pl.figure(figsize=(3, 2.5))
pl.axes([0, 0, 1, 1])
•
# A convex function
pl.plot(x, f(x), linewidth=2) import numpy as np
import pylab as pl
# Apply brent method. To have access to the iteration, do this in an from scipy import optimize
# artificial way: allow the algorithm to iter only once
all_x = list() x, y = np.mgrid[-2.9:5.8:.05, -2.5:5:.05]
all_y = list() x = x.T
for iter in range(30): y = y.T
result = optimize.minimize_scalar(f, bracket=(-5, 2.9, 4.5), method="Brent",
options={"maxiter": iter}, tol=np.finfo(1.).eps) for i in (1, 2):
if result.success: # Create 2 figure: only the second one will have the optimization
print('Converged at ', iter) # path
break pl.figure(i, figsize=(3, 2.5))
pl.clf()
this_x = result.x pl.axes([0, 0, 1, 1])
all_x.append(this_x)
all_y.append(f(this_x)) contours = pl.contour(np.sqrt((x - 3)**2 + (y - 2)**2),
if iter < 6: extent=[-3, 6, -2.5, 5],
pl.text(this_x - .05*np.sign(this_x) - .05, cmap=pl.cm.gnuplot)
f(this_x) + 1.2*(.3 - iter % 2), iter + 1, pl.clabel(contours,
size=12) inline=1,
fmt='%1.1f ',
pl.plot(all_x[:10], all_y[:10], 'k+', markersize=12, markeredgewidth=2) fontsize=14)
pl.plot([-1.5, -1.5, 1.5, 1.5, -1.5],
pl.plot(all_x[-1], all_y[-1], 'rx', markersize=12) [-1.5, 1.5, 1.5, -1.5, -1.5], 'k', linewidth=2)
pl.axis('off') pl.fill_between([ -1.5, 1.5],
pl.ylim(ymin=-1, ymax=8) [ -1.5, -1.5],
[ 1.5, 1.5],
pl.figure(figsize=(4, 3)) color='.8')
pl.semilogy(np.abs(all_y - all_y[-1]), linewidth=2) pl.axvline(0, color='k')
pl.ylabel('Error on f(x)') pl.axhline(0, color='k')
pl.xlabel('Iteration')
pl.tight_layout() pl.text(-.9, 4.4, '$x_2$', size=20)
pl.text(5.6, -.6, '$x_1$', size=20)
pl.show() pl.axis('equal')
pl.axis('off')
Total running time of the script: ( 0 minutes 0.450 seconds)
# And now plot the optimization path
Download Python source code: plot_1d_optim.py accumulator = list()
Download Jupyter notebook: plot_1d_optim.ipynb

13.4. Examples for the mathematical optimization chapter 414 13.4. Examples for the mathematical optimization chapter 415
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

def f(x):
# Store the list of function calls pl.figure(1, figsize=(10, 4))
accumulator.append(x) pl.clf()
return np.sqrt((x[0] - 3)**2 + (x[1] - 2)**2)
colors = pl.cm.Spectral(np.linspace(0, 1, n_dims))[:, :3]

# We don't use the gradient, as with the gradient, L-BFGS is too fast, method_names = list(list(results.values())[0]['Rosenbrock '].keys())
# and finds the optimum without showing us a pretty path method_names.sort(key=lambda x: x[::-1], reverse=True)
def f_prime(x):
r = np.sqrt((x[0] - 3)**2 + (x[0] - 2)**2) for n_dim_index, ((n_dim, n_dim_bench), color) in enumerate(
return np.array(((x[0] - 3)/r, (x[0] - 2)/r)) zip(sorted(results.items()), colors)):
for (cost_name, cost_bench), symbol in zip(sorted(n_dim_bench.items()),
optimize.minimize(f, np.array([0, 0]), method="L-BFGS-B", symbols):
bounds=((-1.5, 1.5), (-1.5, 1.5))) for method_index, method_name, in enumerate(method_names):
this_bench = cost_bench[method_name]
accumulated = np.array(accumulator) bench = np.mean(this_bench)
pl.plot(accumulated[:, 0], accumulated[:, 1]) pl.semilogy([method_index + .1*n_dim_index, ], [bench, ],
marker=symbol, color=color)
pl.show()
# Create a legend for the problem type
Total running time of the script: ( 0 minutes 0.124 seconds) for cost_name, symbol in zip(sorted(n_dim_bench.keys()),
symbols):
Download Python source code: plot_constraints.py pl.semilogy([-10, ], [0, ], symbol, color='.5',
label=cost_name)
Download Jupyter notebook: plot_constraints.ipynb
Generated by Sphinx-Gallery pl.xticks(np.arange(n_methods), method_names, size=11)
pl.xlim(-.2, n_methods - .5)
pl.legend(loc='best', numpoints=1, handletextpad=0, prop=dict(size=12),
13.4.9 Plotting the comparison of optimizers frameon=False)
pl.ylabel('# function calls (a.u.)')
Plots the results from the comparison of optimizers.
# Create a second legend for the problem dimensionality
pl.twinx()

for n_dim, color in zip(sorted(results.keys()), colors):

pl.plot([-10, ], [0, ], 'o', color=color,
label='# dim: %i ' % n_dim)
pl.legend(loc=(.47, .07), numpoints=1, handletextpad=0, prop=dict(size=12),
frameon=False, ncol=2)
pl.xlim(-.2, n_methods - .5)

pl.xticks(np.arange(n_methods), method_names)
pl.yticks(())

pl.tight_layout()
pl.show()

Total running time of the script: ( 0 minutes 0.537 seconds)

Download Python source code: plot_compare_optimizers.py
import pickle
import sys Download Jupyter notebook: plot_compare_optimizers.ipynb
Generated by Sphinx-Gallery
import numpy as np
import pylab as pl

results = pickle.load(open( 13.4.10 Alternating optimization

'helper/compare_optimizers_py%s .pkl' % sys.version_info[0],
'rb')) The challenge here is that Hessian of the problem is a very ill-conditioned matrix. This can easily be seen, as
n_methods = len(list(results.values())[0]['Rosenbrock ']) the Hessian of the first term in simply 2*np.dot(K.T, K). Thus the conditioning of the problem can be judged
n_dims = len(results) from looking at the conditioning of K.

symbols = 'o>*Ds'

13.4. Examples for the mathematical optimization chapter 416 13.4. Examples for the mathematical optimization chapter 417
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

import time time.time() - t0, np.sqrt(np.sum((x_l_bfgs - x_ref)**2)),

f(x_l_bfgs) - f_ref))
import numpy as np
from scipy import optimize t0 = time.time()
import pylab as pl x_newton = optimize.minimize(f, K[0], jac=f_prime, hess=hessian, method="Newton-CG").x
print(" Newton: time %.2f s, x error %.2f , f error %.2f " % (
np.random.seed(0) time.time() - t0, np.sqrt(np.sum((x_newton - x_ref)**2)),
f(x_newton) - f_ref))
K = np.random.normal(size=(100, 100))
pl.show()
def f(x):
return np.sum((np.dot(K, x - 1))**2) + np.sum(x**2)**2

def f_prime(x):
return 2*np.dot(np.dot(K.T, K), x - 1) + 4*np.sum(x**2)*x

def hessian(x):
H = 2*np.dot(K.T, K) + 4*2*x*x[:, np.newaxis]
return H + 4*np.eye(H.shape[0])*np.sum(x**2)

Some pretty plotting

pl.figure(1)
pl.clf()
Z = X, Y = np.mgrid[-1.5:1.5:100j, -1.1:1.1:100j]
# Complete in the additional dimensions with zeros
Z = np.reshape(Z, (2, -1)).copy()
Z.resize((100, Z.shape[-1]))
Z = np.apply_along_axis(f, 0, Z)
Z = np.reshape(Z, X.shape)
pl.imshow(Z.T, cmap=pl.cm.gray_r, extent=[-1.5, 1.5, -1.1, 1.1],
origin='lower')
pl.contour(X, Y, Z, cmap=pl.cm.gnuplot)

# A reference but slow solution:

t0 = time.time()
x_ref = optimize.minimize(f, K[0], method="Powell").x
print(' Powell: time %.2f s' % (time.time() - t0))
f_ref = f(x_ref)

# Compare different approaches

t0 = time.time() Out:
x_bfgs = optimize.minimize(f, K[0], method="BFGS").x
print(' BFGS: time %.2f s, x error %.2f , f error %.2f ' % (time.time() - t0, Powell: time 0.29s
np.sqrt(np.sum((x_bfgs - x_ref)**2)), f(x_bfgs) - f_ref)) BFGS: time 0.78s, x error 0.02, f error -0.02
L-BFGS: time 0.06s, x error 0.02, f error -0.02
t0 = time.time() BFGS w f': time 0.05s, x error 0.02, f error -0.02
x_l_bfgs = optimize.minimize(f, K[0], method="L-BFGS-B").x L-BFGS w f': time 0.00s, x error 0.02, f error -0.02
print(' L-BFGS: time %.2f s, x error %.2f , f error %.2f ' % (time.time() - t0, Newton: time 0.01s, x error 0.02, f error -0.02
np.sqrt(np.sum((x_l_bfgs - x_ref)**2)), f(x_l_bfgs) - f_ref))
Total running time of the script: ( 0 minutes 1.620 seconds)
Download Python source code: plot_exercise_ill_conditioned.py
t0 = time.time()
x_bfgs = optimize.minimize(f, K[0], jac=f_prime, method="BFGS").x Download Jupyter notebook: plot_exercise_ill_conditioned.ipynb
print(" BFGS w f': time %.2f s, x error %.2f , f error %.2f " % (
time.time() - t0, np.sqrt(np.sum((x_bfgs - x_ref)**2)), Generated by Sphinx-Gallery
f(x_bfgs) - f_ref))

t0 = time.time() 13.4.11 Gradient descent

x_l_bfgs = optimize.minimize(f, K[0], jac=f_prime, method="L-BFGS-B").x
print("L-BFGS w f': time %.2f s, x error %.2f , f error %.2f " % ( An example demoing gradient descent by creating figures that trace the evolution of the optimizer.

13.4. Examples for the mathematical optimization chapter 418 13.4. Examples for the mathematical optimization chapter 419
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

import numpy as np return gradient_descent(x0, f, f_prime, adaptative=True)

import pylab as pl
from scipy import optimize
def conjugate_gradient(x0, f, f_prime, hessian=None):
import sys, os all_x_i = [x0[0]]
sys.path.append(os.path.abspath('helper')) all_y_i = [x0[1]]
from cost_functions import mk_quad, mk_gauss, rosenbrock,\ all_f_i = [f(x0)]
rosenbrock_prime, rosenbrock_hessian, LoggingFunction,\ def store(X):
CountingFunction x, y = X
all_x_i.append(x)
x_min, x_max = -1, 2 all_y_i.append(y)
y_min, y_max = 2.25/3*x_min - .2, 2.25/3*x_max - .2 all_f_i.append(f(X))
optimize.minimize(f, x0, jac=f_prime, method="CG", callback=store, options={"gtol": 1e-12})
A formatter to print values on contours return all_x_i, all_y_i, all_f_i

def super_fmt(value):
if value > 1: def newton_cg(x0, f, f_prime, hessian):
if np.abs(int(value) - value) < .1: all_x_i = [x0[0]]
out = '$10^{%.1i }$' % value all_y_i = [x0[1]]
else: all_f_i = [f(x0)]
out = '$10^{%.1f }$' % value def store(X):
else: x, y = X
value = np.exp(value - .01) all_x_i.append(x)
if value > .1: all_y_i.append(y)
out = '%1.1f ' % value all_f_i.append(f(X))
elif value > .01: optimize.minimize(f, x0, method="Newton-CG", jac=f_prime, hess=hessian, callback=store,␣
out = '%.2f ' % value ,→options={"xtol": 1e-12})
else: return all_x_i, all_y_i, all_f_i
out = '%.2e ' % value
return out
def bfgs(x0, f, f_prime, hessian=None):
A gradient descent algorithm do not use: its a toy, use scipy’s optimize.fmin_cg all_x_i = [x0[0]]
all_y_i = [x0[1]]
def gradient_descent(x0, f, f_prime, hessian=None, adaptative=False): all_f_i = [f(x0)]
x_i, y_i = x0 def store(X):
all_x_i = list() x, y = X
all_y_i = list() all_x_i.append(x)
all_f_i = list() all_y_i.append(y)
all_f_i.append(f(X))
for i in range(1, 100): optimize.minimize(f, x0, method="BFGS", jac=f_prime, callback=store, options={"gtol": 1e-
all_x_i.append(x_i) ,→12})
all_y_i.append(y_i) return all_x_i, all_y_i, all_f_i
all_f_i.append(f([x_i, y_i]))
dx_i, dy_i = f_prime(np.asarray([x_i, y_i]))
if adaptative: def powell(x0, f, f_prime, hessian=None):
# Compute a step size using a line_search to satisfy the Wolf all_x_i = [x0[0]]
# conditions all_y_i = [x0[1]]
step = optimize.line_search(f, f_prime, all_f_i = [f(x0)]
np.r_[x_i, y_i], -np.r_[dx_i, dy_i], def store(X):
np.r_[dx_i, dy_i], c2=.05) x, y = X
step = step[0] all_x_i.append(x)
if step is None: all_y_i.append(y)
step = 0 all_f_i.append(f(X))
else: optimize.minimize(f, x0, method="Powell", callback=store, options={"ftol": 1e-12})
step = 1 return all_x_i, all_y_i, all_f_i
x_i += - step*dx_i
y_i += - step*dy_i
if np.abs(all_f_i[-1]) < 1e-16: def nelder_mead(x0, f, f_prime, hessian=None):
break all_x_i = [x0[0]]
return all_x_i, all_y_i, all_f_i all_y_i = [x0[1]]
all_f_i = [f(x0)]
def store(X):
def gradient_descent_adaptative(x0, f, f_prime, hessian=None): x, y = X

13.4. Examples for the mathematical optimization chapter 420 13.4. Examples for the mathematical optimization chapter 421
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

all_x_i.append(x) log_z = np.log(z + .01)

all_y_i.append(y) pl.imshow(log_z,
all_f_i.append(f(X)) extent=[x_min, x_max, y_min, y_max],
optimize.minimize(f, x0, method="Nelder-Mead", callback=store, options={"ftol": 1e-12}) cmap=pl.cm.gray_r, origin='lower',
return all_x_i, all_y_i, all_f_i vmax=log_z.min() + 1.5*log_z.ptp())
contours = pl.contour(log_z,
Run different optimizers on these problems levels=levels.get(f, None),
extent=[x_min, x_max, y_min, y_max],
levels = dict() cmap=pl.cm.gnuplot, origin='lower')
levels[f] = contours.levels
for index, ((f, f_prime, hessian), optimizer) in enumerate(( pl.clabel(contours, inline=1,
(mk_quad(.7), gradient_descent), fmt=super_fmt, fontsize=14)
(mk_quad(.7), gradient_descent_adaptative),
(mk_quad(.02), gradient_descent), pl.plot(all_x_i, all_y_i, 'b-', linewidth=2)
(mk_quad(.02), gradient_descent_adaptative), pl.plot(all_x_i, all_y_i, 'k+')
(mk_gauss(.02), gradient_descent_adaptative),
((rosenbrock, rosenbrock_prime, rosenbrock_hessian), pl.plot(logging_f.all_x_i, logging_f.all_y_i, 'k.', markersize=2)
gradient_descent_adaptative),
(mk_gauss(.02), conjugate_gradient), pl.plot([0], [0], 'rx', markersize=12)
((rosenbrock, rosenbrock_prime, rosenbrock_hessian),
conjugate_gradient),
(mk_quad(.02), newton_cg), pl.xticks(())
(mk_gauss(.02), newton_cg), pl.yticks(())
((rosenbrock, rosenbrock_prime, rosenbrock_hessian), pl.xlim(x_min, x_max)
newton_cg), pl.ylim(y_min, y_max)
(mk_quad(.02), bfgs), pl.draw()
(mk_gauss(.02), bfgs),
((rosenbrock, rosenbrock_prime, rosenbrock_hessian), pl.figure(index + 100, figsize=(4, 3))
bfgs), pl.clf()
(mk_quad(.02), powell), pl.semilogy(np.maximum(np.abs(all_f_i), 1e-30), linewidth=2,
(mk_gauss(.02), powell), label='# iterations')
((rosenbrock, rosenbrock_prime, rosenbrock_hessian), pl.ylabel('Error on f(x)')
powell), pl.semilogy(logging_f.counts,
(mk_gauss(.02), nelder_mead), np.maximum(np.abs(logging_f.all_f_i), 1e-30),
((rosenbrock, rosenbrock_prime, rosenbrock_hessian), linewidth=2, color='g', label='# function calls')
nelder_mead), pl.legend(loc='upper right', frameon=True, prop=dict(size=11),
)): borderaxespad=0, handlelength=1.5, handletextpad=.5)
pl.tight_layout()
# Compute a gradient-descent pl.draw()
x_i, y_i = 1.6, 1.1
counting_f_prime = CountingFunction(f_prime)
counting_hessian = CountingFunction(hessian)
logging_f = LoggingFunction(f, counter=counting_f_prime.counter)
all_x_i, all_y_i, all_f_i = optimizer(np.array([x_i, y_i]),
logging_f, counting_f_prime,
hessian=counting_hessian)

# Plot the contour plot •

if not max(all_y_i) < y_max:
x_min *= 1.2
x_max *= 1.2
y_min *= 1.2
y_max *= 1.2
x, y = np.mgrid[x_min:x_max:100j, y_min:y_max:100j]
x = x.T
y = y.T •

pl.figure(index, figsize=(3, 2.5))

pl.clf()
pl.axes([0, 0, 1, 1])

X = np.concatenate((x[np.newaxis, ...], y[np.newaxis, ...]), axis=0)

z = np.apply_along_axis(f, 0, X)

13.4. Examples for the mathematical optimization chapter 422 13.4. Examples for the mathematical optimization chapter 423
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

• •

• •

• •

• •

• •

• •

• •

13.4. Examples for the mathematical optimization chapter 424 13.4. Examples for the mathematical optimization chapter 425
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

•
•

13.4. Examples for the mathematical optimization chapter 426 13.4. Examples for the mathematical optimization chapter 427
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

• •

• •

• •

• •
Total running time of the script: ( 0 minutes 12.552 seconds)
Download Python source code: plot_gradient_descent.py
Download Jupyter notebook: plot_gradient_descent.ipynb
Generated by Sphinx-Gallery
Download all examples in Python source code: auto_examples_python.zip
• Download all examples in Jupyter notebooks: auto_examples_jupyter.zip
Generated by Sphinx-Gallery

13.4. Examples for the mathematical optimization chapter 428 13.4. Examples for the mathematical optimization chapter 429
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

13.4.12 Gradient-less methods

A shooting method: the Powell algorithm

Almost a gradient approach

An ill-conditioned quadratic function:
On a exactly quadratic function, BFGS is not as fast as Newton’s
method, but still very fast.

An ill-conditioned quadratic function:

An ill-conditioned non-quadratic function: Powell’s method isn’t too sensitive to local ill-
Here BFGS does better than Newton, as its empirical estimate conditionning in low dimensions
of the curvature is better than that given by the Hessian.

An ill-conditioned very non-quadratic function: An ill-conditioned very non-quadratic function:

>>> def f(x): # The rosenbrock function

... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2 Simplex method: the Nelder-Mead
>>> def jacobian(x):
... return np.array((-2*.5*(1 - x[0]) - 4*x[0]*(x[1] - x[0]**2), 2*(x[1] - x[0]**2))) The Nelder-Mead algorithms is a generalization of dichotomy approaches to high-dimensional spaces. The
>>> optimize.minimize(f, [2, -1], method="BFGS", jac=jacobian) algorithm works by refining a simplex, the generalization of intervals and triangles to high-dimensional spaces,
fun: 2.6306...e-16 to bracket the minimum.
hess_inv: array([[ 0.99986..., 2.0000...],
[ 2.0000..., 4.498...]]) Strong points: it is robust to noise, as it does not rely on computing gradients. Thus it can work on functions
jac: array([ 6.7089...e-08, -3.2222...e-08]) that are not locally smooth such as experimental data points, as long as they display a large-scale bell-shape
message: ...'Optimization terminated successfully.' behavior. However it is slower than gradient-based methods on smooth, non-noisy functions.
nfev: 10
nit: 8
njev: 10
status: 0
success: True
x: array([ 1. , 0.99999...])

L-BFGS: Limited-memory BFGS Sits between BFGS and conjugate gradient: in very high dimensions (> 250)
the Hessian matrix is too costly to compute and invert. L-BFGS keeps a low-rank version. In addition, box
bounds are also supported by L-BFGS-B:
An ill-conditioned non-
>>> def f(x): # The rosenbrock function
quadratic function:
... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2
>>> def jacobian(x):
... return np.array((-2*.5*(1 - x[0]) - 4*x[0]*(x[1] - x[0]**2), 2*(x[1] - x[0]**2)))
>>> optimize.minimize(f, [2, 2], method="L-BFGS-B", jac=jacobian)
fun: 1.4417...e-15
hess_inv: <2x2 LbfgsInvHessProduct with dtype=float64>
jac: array([ 1.0233...e-07, -2.5929...e-08])
message: ...'CONVERGENCE: NORM_OF_PROJECTED_GRADIENT_<=_PGTOL'
nfev: 17
nit: 16
status: 0 An ill-conditioned very non-
success: True quadratic function:
x: array([ 1.0000..., 1.0000...])
Using the Nelder-Mead solver in scipy.optimize.minimize():

13.4. Examples for the mathematical optimization chapter 430 13.4. Examples for the mathematical optimization chapter 431
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

• In general, prefer BFGS or L-BFGS, even if you have to approximate numerically gra-
>>> def f(x): # The rosenbrock function
... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2 dients. These are also the default if you omit the parameter method - depending if the
>>> optimize.minimize(f, [2, -1], method="Nelder-Mead") problem has constraints or bounds
final_simplex: (array([[ 1.0000..., 1.0000...], • On well-conditioned problems, Powell and Nelder-Mead, both gradient-free methods,
[ 0.99998... , 0.99996... ],
work well in high dimension, but they collapse for ill-conditioned problems.
[ 1.0000..., 1.0000... ]]), array([ 1.1152...e-10, 1.5367...e-10, 4.9883...e-10]))
fun: 1.1152...e-10 With knowledge of the gradient
message: ...'Optimization terminated successfully.'
nfev: 111 • BFGS or L-BFGS.
nit: 58 • Computational overhead of BFGS is larger than that L-BFGS, itself larger than that of
status: 0
conjugate gradient. On the other side, BFGS usually needs less function evaluations than
success: True
CG. Thus conjugate gradient method is better than BFGS at optimizing computationally
x: array([ 1.0000..., 1.0000...])
cheap functions.
With the Hessian
13.4.13 Global optimizers • If you can compute the Hessian, prefer the Newton method (Newton-CG or TCG).

If your problem does not admit a unique local minimum (which can be hard to test unless the function is If you have noisy measurements
convex), and you do not have prior information to initialize the optimization close to the solution, you may • Use Nelder-Mead or Powell.
need a global optimizer.

13.5.2 Making your optimizer faster

Brute force: a grid search
• Choose the right method (see above), do compute analytically the gradient and Hessian, if you can.
scipy.optimize.brute() evaluates the function on a given grid of parameters and returns the parameters
corresponding to the minimum value. The parameters are specified with ranges given to numpy.mgrid. By • Use preconditionning when possible.
default, 20 steps are taken in each direction: • Choose your initialization points wisely. For instance, if you are running many similar optimizations,
>>> def f(x): # The rosenbrock function warm-restart one with the results of another.
... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2 • Relax the tolerance if you don’t need precision using the parameter tol.
>>> optimize.brute(f, ((-1, 2), (-1, 2)))
array([ 1.0000..., 1.0000...])
13.5.3 Computing gradients

13.5 Practical guide to optimization with scipy Computing gradients, and even more Hessians, is very tedious but worth the effort. Symbolic computation
with Sympy may come in handy.

13.5.1 Choosing a method Warning: A very common source of optimization not converging well is human error in the computation
of the gradient. You can use scipy.optimize.check_grad() to check that your gradient is correct. It
All methods are exposed as the method argument of scipy.optimize.minimize().
returns the norm of the different between the gradient given, and a gradient computed numerically:
>>> optimize.check_grad(f, jacobian, [2, -1])
2.384185791015625e-07

See also scipy.optimize.approx_fprime() to find your errors.

Without knowledge of the gradient

13.5. Practical guide to optimization with scipy 432 13.5. Practical guide to optimization with scipy 433
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

13.5.4 Synthetic exercices

Exercice: A simple (?) quadratic function

Optimize the following function, using K[0] as a starting point:

np.random.seed(0)
K = np.random.normal(size=(100, 100)) 13.6 Special case: non-linear least-squares
def f(x):
return np.sum((np.dot(K, x - 1))**2) + np.sum(x**2)**2 13.6.1 Minimizing the norm of a vector function
Time your approach. Find the fastest approach. Why is BFGS not working well? Least square problems, minimizing the norm of a vector function, have a specific structure that can be used in
the Levenberg–Marquardt algorithm implemented in scipy.optimize.leastsq().
Lets try to minimize the norm of the following vectorial function:
Exercice: A locally flat minimum
>>> def f(x):
Consider the function exp(-1/(.1*x**2 + y**2). This function admits a minimum in (0, 0). Starting from an ... return np.arctan(x) - np.arctan(np.linspace(0, 1, len(x)))
initialization at (1, 1), try to get within 1e-8 of this minimum point.
>>> x0 = np.zeros(10)
>>> optimize.leastsq(f, x0)
(array([ 0. , 0.11111111, 0.22222222, 0.33333333, 0.44444444,
0.55555556, 0.66666667, 0.77777778, 0.88888889, 1. ]), 2)

This took 67 function evaluations (check it with ‘full_output=1’). What if we compute the norm ourselves and
use a good generic optimizer (BFGS):

>>> def g(x):

... return np.sum(f(x)**2)
>>> optimize.minimize(g, x0, method="BFGS")
fun: 2.6940...e-11
hess_inv: array([[...
...
...]])
jac: array([
...
...])
message: ...'Optimization terminated successfully.'
nfev: 144
nit: 11
njev: 12
status: 0
success: True
x: array([ -7.3845...e-09, 1.1111...e-01, 2.2222...e-01,
3.3333...e-01, 4.4444...e-01, 5.5555...e-01,

13.5. Practical guide to optimization with scipy 434 13.6. Special case: non-linear least-squares 435
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

6.6666...e-01, 7.7777...e-01, 8.8889...e-01, 13.7 Optimization with constraints

1.0000...e+00])

13.7.1 Box bounds

BFGS needs more function calls, and gives a less precise result.
Box bounds correspond to limiting each of the individual parameters of the optimization. Note that some
Note: leastsq is interesting compared to BFGS only if the dimensionality of the output vector is large, and problems that are not originally written as box bounds can be rewritten as such via change of variables. Both
larger than the number of parameters to optimize. scipy.optimize.minimize_scalar() and scipy.optimize.minimize() support bound constraints with
the parameter bounds:

>>> def f(x):

Warning: If the function is linear, this is a linear-algebra problem, and should be solved with scipy. ... return np.sqrt((x[0] - 3)**2 + (x[1] - 2)**2)
linalg.lstsq(). >>> optimize.minimize(f, np.array([0, 0]), bounds=((-1.5, 1.5), (-1.5, 1.5)))
fun: 1.5811...
hess_inv: <2x2 LbfgsInvHessProduct with dtype=float64>
jac: array([-0.94868..., -0.31622...])
13.6.2 Curve fitting message: ...'CONVERGENCE: NORM_OF_PROJECTED_GRADIENT_<=_PGTOL'
nfev: 9
nit: 2
status: 0
success: True
x: array([ 1.5, 1.5])

Least square problems occur often when

fitting a non-linear to data. While it is possible to construct our optimization problem ourselves, scipy provides 13.7.2 General constraints
a helper function for this purpose: scipy.optimize.curve_fit():
Equality and inequality constraints specified as functions: f (x) = 0 and g (x) < 0.
>>> def f(t, omega, phi):
... return np.cos(omega * t + phi) • scipy.optimize.fmin_slsqp() Sequential least square programming: equality and inequality con-

>>> x = np.linspace(0, 3, 50)

>>> y = f(x, 1.5, 1) + .1*np.random.normal(size=50)

>>> optimize.curve_fit(f, x, y)
(array([ 1.5185..., 0.92665...]), array([[ 0.00037..., -0.00056...],
[-0.0005..., 0.00123...]]))

Exercise

Do the same with omega = 3. What is the difficulty? straints:

>>> def f(x):

... return np.sqrt((x[0] - 3)**2 + (x[1] - 2)**2)

>>> def constraint(x):

... return np.atleast_1d(1.5 - np.sum(np.abs(x)))

>>> x0 = np.array([0, 0])

13.6. Special case: non-linear least-squares 436 13.7. Optimization with constraints 437
 Scipy lecture notes, Edition 2017.1

>>> optimize.minimize(f, x0, constraints={"fun": constraint, "type": "ineq"})

fun: 2.4748...
jac: array([-0.70708..., -0.70712...])
message: ...'Optimization terminated successfully.'
nfev: 20

14
nit: 5
njev: 5
status: 0
success: True
x: array([ 1.2500..., 0.2499...])

Warning: The above problem is known as the Lasso problem in statistics, and there exist very efficient
CHAPTER
solvers for it (for instance in scikit-learn). In general do not use generic solvers when specific ones exist.

Lagrange multipliers

If you are ready to do a bit of math, many constrained optimization problems can be converted to non-
constrained optimization problems using a mathematical trick known as Lagrange multipliers.

Interfacing with C
13.8 Full code examples

13.9 Examples for the mathematical optimization chapter

Download all examples in Python source code: auto_examples_python.zip Author: Valentin Haenel
Download all examples in Jupyter notebooks: auto_examples_jupyter.zip This chapter contains an introduction to the many different routes for making your native code (primarily
Generated by Sphinx-Gallery C/C++) available from Python, a process commonly referred to wrapping. The goal of this chapter is to give
you a flavour of what technologies exist and what their respective merits and shortcomings are, so that you
can select the appropriate one for your specific needs. In any case, once you do start wrapping, you almost
certainly will want to consult the respective documentation for your selected technique.

Chapters contents

• Introduction
• Python-C-Api
• Ctypes
• SWIG
• Cython
• Summary
• Further Reading and References
• Exercises

14.1 Introduction

This chapter covers the following techniques:

• Python-C-Api

13.8. Full code examples 438 439

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

• Ctypes • High maintenance cost

• SWIG (Simplified Wrapper and Interface Generator) • No forward compatibility across Python versions as C-API changes
• Cython • Reference count bugs are easy to create and very hard to track down.
These four techniques are perhaps the most well known ones, of which Cython is probably the most advanced
one and the one you should consider using first. The others are also important, if you want to understand the Note: The Python-C-Api example here serves mainly for didactic reasons. Many of the other techniques
wrapping problem from different angles. Having said that, there are other alternatives out there, but having actually depend on this, so it is good to have a high-level understanding of how it works. In 99% of the use-
understood the basics of the ones above, you will be in a position to evaluate the technique of your choice to cases you will be better off, using an alternative technique.
see if it fits your needs.
The following criteria may be useful when evaluating a technology:
Note: Since refernce counting bugs are easy to create and hard to track down, anyone really needing to use
• Are additional libraries required?
the Python C-API should read the section about objects, types and reference counts from the official python
• Is the code autogenerated? documentation. Additionally, there is a tool by the name of cpychecker which can help discover common
errors with reference counting.
• Does it need to be compiled?
• Is there good support for interacting with Numpy arrays?
• Does it support C++? 14.2.1 Example
Before you set out, you should consider your use case. When interfacing with native code, there are usually
two use-cases that come up: The following C-extension module, make the cos function from the standard math library available to Python:

• Existing code in C/C++ that needs to be leveraged, either because it already exists, or because it is faster. /* Example of wrapping cos function from math.h with the Python-C-API. */

• Python code too slow, push inner loops to native code # include <Python.h>
Each technology is demonstrated by wrapping the cos function from math.h. While this is a mostly a trivial # include <math.h>
example, it should serve us well to demonstrate the basics of the wrapping solution. Since each technique
/* wrapped cosine function */
also includes some form of Numpy support, this is also demonstrated using an example where the cosine is
static PyObject* cos_func(PyObject* self, PyObject* args)
computed on some kind of array. {
Last but not least, two small warnings: double value;
double answer;
• All of these techniques may crash (segmentation fault) the Python interpreter, which is (usually) due to
bugs in the C code. /* parse the input, from python float to c double */
if (!PyArg_ParseTuple(args, "d", &value))
• All the examples have been done on Linux, they should be possible on other operating systems. return NULL;
• You will need a C compiler for most of the examples. /* if the above function returns -1, an appropriate Python exception will
* have been set, and the function simply returns NULL
*/

14.2 Python-C-Api /* call cos from libm */

answer = cos(value);
The Python-C-API is the backbone of the standard Python interpreter (a.k.a CPython). Using this API it is
/* construct the output from cos, from c double to python float */
possible to write Python extension module in C and C++. Obviously, these extension modules can, by virtue of
return Py_BuildValue("f", answer);
language compatibility, call any function written in C or C++. }
When using the Python-C-API, one usually writes much boilerplate code, first to parse the arguments that were
given to a function, and later to construct the return type. /* define functions in module */
static PyMethodDef CosMethods[] =
Advantages {
{"cos_func", cos_func, METH_VARARGS, "evaluate the cosine"},
• Requires no additional libraries {NULL, NULL, 0, NULL}
• Lots of low-level control };

• Entirely usable from C++ # if PY_MAJOR_VERSION >= 3

/* module initialization */
Disadvantages
/* Python version 3*/
• May require a substantial amount of effort static struct PyModuleDef cModPyDem =
{
• Much overhead in the code PyModuleDef_HEAD_INIT,
"cos_module", "Some documentation",
• Must be compiled
-1,

14.2. Python-C-Api 440 14.2. Python-C-Api 441

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

CosMethods
}; Note: In Python 3, the filename for compiled modules includes metadata on the Python interpreter (see PEP
3149) and is thus longer. The import statement is not affected by this.
PyMODINIT_FUNC
PyInit_cos_module(void)
{ In [1]: import cos_module
return PyModule_Create(&cModPyDem);
} In [2]: cos_module?
Type: module
# else String Form:<module 'cos_module' from 'cos_module.so'>
File: /home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/python_c_api/
,→cos_module.so
/* module initialization */
/* Python version 2 */ Docstring: <no docstring>
PyMODINIT_FUNC
initcos_module(void) In [3]: dir(cos_module)
{ Out[3]: ['__doc__', '__file__', '__name__', '__package__', 'cos_func']
(void) Py_InitModule("cos_module", CosMethods);
} In [4]: cos_module.cos_func(1.0)
Out[4]: 0.5403023058681398
# endif
In [5]: cos_module.cos_func(0.0)
Out[5]: 1.0
As you can see, there is much boilerplate, both to «massage» the arguments and return types into place and for
the module initialisation. Although some of this is amortised, as the extension grows, the boilerplate required In [6]: cos_module.cos_func(3.14159265359)
for each function(s) remains. Out[7]: -1.0
The standard python build system distutils supports compiling C-extensions from a setup.py, which is
rather convenient: Now let’s see how robust this is:

from distutils.core import setup, Extension In [10]: cos_module.cos_func('foo')

---------------------------------------------------------------------------
# define the extension module TypeError Traceback (most recent call last)
cos_module = Extension('cos_module', sources=['cos_module.c']) <ipython-input-10-11bee483665d> in <module>()
----> 1 cos_module.cos_func('foo')
# run the setup
setup(ext_modules=[cos_module]) TypeError: a float is required

This can be compiled:

14.2.2 Numpy Support
$ cd advanced/interfacing_with_c/python_c_api

$ ls Analog to the Python-C-API, Numpy, which is itself implemented as a C-extension, comes with the Numpy-C-
cos_module.c setup.py API. This API can be used to create and manipulate Numpy arrays from C, when writing a custom C-extension.
See also: Advanced NumPy.
$ python setup.py build_ext --inplace
running build_ext
Note: If you do ever need to use the Numpy C-API refer to the documentation about Arrays and Iterators.
building 'cos_module' extension
creating build
creating build/temp.linux-x86_64-2.7 The following example shows how to pass Numpy arrays as arguments to functions and how to iterate over
gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes - Numpy arrays using the (old) Numpy-C-API. It simply takes an array as argument applies the cosine function
,→fPIC -I/home/esc/anaconda/include/python2.7 -c cos_module.c -o build/temp.linux-x86_64-2.7/
from the math.h and returns a resulting new array.
,→cos_module.o
gcc -pthread -shared build/temp.linux-x86_64-2.7/cos_module.o -L/home/esc/anaconda/lib - /* Example of wrapping the cos function from math.h using the Numpy-C-API. */
,→lpython2.7 -o /home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/python_c_
,→api/cos_module.so
# include <Python.h>
# include <numpy/arrayobject.h>
$ ls # include <math.h>
build/ cos_module.c cos_module.so setup.py
/* wrapped cosine function */
• build_ext is to build extension modules static PyObject* cos_func_np(PyObject* self, PyObject* args)
{
• --inplace will output the compiled extension module into the current directory
The file cos_module.so contains the compiled extension, which we can now load in the IPython interpreter: PyArrayObject *in_array;

14.2. Python-C-Api 442 14.2. Python-C-Api 443

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

PyObject *out_array; /* module initialization */

NpyIter *in_iter; PyMODINIT_FUNC
NpyIter *out_iter; initcos_module_np(void)
NpyIter_IterNextFunc *in_iternext; {
NpyIter_IterNextFunc *out_iternext; (void) Py_InitModule("cos_module_np", CosMethods);
/* IMPORTANT: this must be called */
/* parse single numpy array argument */ import_array();
if (!PyArg_ParseTuple(args, "O!", &PyArray_Type, &in_array)) }
return NULL;
To compile this we can use distutils again. However we need to be sure to include the Numpy headers by using
/* construct the output array, like the input array */ :func:numpy.get_include.
out_array = PyArray_NewLikeArray(in_array, NPY_ANYORDER, NULL, 0);
if (out_array == NULL) from distutils.core import setup, Extension
return NULL; import numpy

/* create the iterators */ # define the extension module

in_iter = NpyIter_New(in_array, NPY_ITER_READONLY, NPY_KEEPORDER, cos_module_np = Extension('cos_module_np', sources=['cos_module_np.c'],
NPY_NO_CASTING, NULL); include_dirs=[numpy.get_include()])
if (in_iter == NULL)
goto fail; # run the setup
setup(ext_modules=[cos_module_np])
out_iter = NpyIter_New((PyArrayObject *)out_array, NPY_ITER_READWRITE,
NPY_KEEPORDER, NPY_NO_CASTING, NULL); To convince ourselves if this does actually works, we run the following test script:
if (out_iter == NULL) {
NpyIter_Deallocate(in_iter); import cos_module_np
goto fail; import numpy as np
} import pylab

in_iternext = NpyIter_GetIterNext(in_iter, NULL); x = np.arange(0, 2 * np.pi, 0.1)

out_iternext = NpyIter_GetIterNext(out_iter, NULL); y = cos_module_np.cos_func_np(x)
if (in_iternext == NULL || out_iternext == NULL) { pylab.plot(x, y)
NpyIter_Deallocate(in_iter); pylab.show()
NpyIter_Deallocate(out_iter);
goto fail;
And this should result in the following figure:
}
double ** in_dataptr = (double **) NpyIter_GetDataPtrArray(in_iter);
double ** out_dataptr = (double **) NpyIter_GetDataPtrArray(out_iter);

/* iterate over the arrays */

do {
**out_dataptr = cos(**in_dataptr);
} while(in_iternext(in_iter) && out_iternext(out_iter));

/* clean up and return the result */

NpyIter_Deallocate(in_iter);
NpyIter_Deallocate(out_iter);
Py_INCREF(out_array);
return out_array;

/* in case bad things happen */

fail: 14.3 Ctypes
Py_XDECREF(out_array);
return NULL;
} Ctypes is a foreign function library for Python. It provides C compatible data types, and allows calling functions
in DLLs or shared libraries. It can be used to wrap these libraries in pure Python.
/* define functions in module */
Advantages
static PyMethodDef CosMethods[] =
{ • Part of the Python standard library
{"cos_func_np", cos_func_np, METH_VARARGS,
"evaluate the cosine on a numpy array"}, • Does not need to be compiled
{NULL, NULL, 0, NULL}
• Wrapping code entirely in Python
};
Disadvantages

14.2. Python-C-Api 444 14.3. Ctypes 445

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

• Requires code to be wrapped to be available as a shared library (roughly speaking *.dll in Windows
*.so in Linux and *.dylib in Mac OSX.) In [4]: cos_module.cos_func(1.0)
• No good support for C++ Out[4]: 0.5403023058681398

In [5]: cos_module.cos_func(0.0)
14.3.1 Example Out[5]: 1.0

In [6]: cos_module.cos_func(3.14159265359)
As advertised, the wrapper code is in pure Python.
Out[6]: -1.0
""" Example of wrapping cos function from math.h using ctypes. """
As with the previous example, this code is somewhat robust, although the error message is not quite as helpful,
import ctypes since it does not tell us what the type should be.

# find and load the library In [7]: cos_module.cos_func('foo')

---------------------------------------------------------------------------
# OSX or linux ArgumentError Traceback (most recent call last)
from ctypes.util import find_library <ipython-input-7-11bee483665d> in <module>()
libm = ctypes.cdll.LoadLibrary(find_library('m')) ----> 1 cos_module.cos_func('foo')

# Windows /home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/ctypes/cos_module.py in␣

# from ctypes import windll ,→cos_func(arg)

# libm = cdll.msvcrt 12 def cos_func(arg):

13 ''' Wrapper for cos from math.h '''
---> 14 return libm.cos(arg)
# set the argument type
libm.cos.argtypes = [ctypes.c_double] ArgumentError: argument 1: <type 'exceptions.TypeError'>: wrong type
# set the return type
libm.cos.restype = ctypes.c_double
14.3.2 Numpy Support
def cos_func(arg):
''' Wrapper for cos from math.h ''' Numpy contains some support for interfacing with ctypes. In particular there is support for exporting certain
return libm.cos(arg) attributes of a Numpy array as ctypes data-types and there are functions to convert from C arrays to Numpy
arrays and back.
• Finding and loading the library may vary depending on your operating system, check the documentation For more information, consult the corresponding section in the Numpy Cookbook and the API documentation
for details for numpy.ndarray.ctypes and numpy.ctypeslib.
• This may be somewhat deceptive, since the math library exists in compiled form on the system already. For the following example, let’s consider a C function in a library that takes an input and an output array,
If you were to wrap a in-house library, you would have to compile it first, which may or may not require computes the cosine of the input array and stores the result in the output array.
some additional effort.
The library consists of the following header file (although this is not strictly needed for this example, we list it
We may now use this, as before: for completeness):
In [1]: import cos_module
void cos_doubles(double * in_array, double * out_array, int size);
In [2]: cos_module?
Type: module The function implementation resides in the following C source file:
String Form:<module 'cos_module' from 'cos_module.py'>
# include <math.h>
File: /home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/ctypes/cos_
,→module.py
/* Compute the cosine of each element in in_array, storing the result in
Docstring: <no docstring>
* out_array. */
void cos_doubles(double * in_array, double * out_array, int size){
In [3]: dir(cos_module)
int i;
Out[3]:
for(i=0;i<size;i++){
['__builtins__',
out_array[i] = cos(in_array[i]);
'__doc__',
}
'__file__',
}
'__name__',
'__package__',
'cos_func', And since the library is pure C, we can’t use distutils to compile it, but must use a combination of make and
'ctypes', gcc:
'find_library',
'libm']

14.3. Ctypes 446 14.3. Ctypes 447

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

m.PHONY : clean
cos_doubles.cos_doubles_func(x, y)
libcos_doubles.so : cos_doubles.o pylab.plot(x, y)
gcc -shared -Wl,-soname,libcos_doubles.so -o libcos_doubles.so cos_doubles.o pylab.show()

cos_doubles.o : cos_doubles.c
gcc -c -fPIC cos_doubles.c -o cos_doubles.o

clean :
-rm -vf libcos_doubles.so cos_doubles.o cos_doubles.pyc

We can then compile this (on Linux) into the shared library libcos_doubles.so:

$ ls
cos_doubles.c cos_doubles.h cos_doubles.py makefile test_cos_doubles.py
$ make
gcc -c -fPIC cos_doubles.c -o cos_doubles.o
gcc -shared -Wl,-soname,libcos_doubles.so -o libcos_doubles.so cos_doubles.o
$ ls
cos_doubles.c cos_doubles.o libcos_doubles.so* test_cos_doubles.py
cos_doubles.h cos_doubles.py makefile 14.4 SWIG
Now we can proceed to wrap this library via ctypes with direct support for (certain kinds of) Numpy arrays: SWIG, the Simplified Wrapper Interface Generator, is a software development tool that connects programs
""" Example of wrapping a C library function that accepts a C double array as written in C and C++ with a variety of high-level programming languages, including Python. The important
input using the numpy.ctypeslib. """ thing with SWIG is, that it can autogenerate the wrapper code for you. While this is an advantage in terms of
development time, it can also be a burden. The generated file tend to be quite large and may not be too human
import numpy as np readable and the multiple levels of indirection which are a result of the wrapping process, may be a bit tricky
import numpy.ctypeslib as npct to understand.
from ctypes import c_int

# input type for the cos_doubles function Note: The autogenerated C code uses the Python-C-Api.
# must be a double array, with single dimension that is contiguous
array_1d_double = npct.ndpointer(dtype=np.double, ndim=1, flags='CONTIGUOUS')
Advantages
# load the library, using numpy mechanisms • Can automatically wrap entire libraries given the headers
libcd = npct.load_library("libcos_doubles", ".")
• Works nicely with C++
# setup the return types and argument types Disadvantages
libcd.cos_doubles.restype = None
libcd.cos_doubles.argtypes = [array_1d_double, array_1d_double, c_int] • Autogenerates enormous files
• Hard to debug if something goes wrong
def cos_doubles_func(in_array, out_array): • Steep learning curve
return libcd.cos_doubles(in_array, out_array, len(in_array))

• Note the inherent limitation of contiguous single dimensional Numpy arrays, since the C functions re- 14.4.1 Example
quires this kind of buffer.
• Also note that the output array must be preallocated, for example with numpy.zeros() and the function Let’s imagine that our cos function lives in a cos_module which has been written in c and consists of the
will write into it’s buffer. source file cos_module.c:

• Although the original signature of the cos_doubles function is ARRAY, ARRAY, int the final # include <math.h>
cos_doubles_func takes only two Numpy arrays as arguments.
double cos_func(double arg){
And, as before, we convince ourselves that it worked: return cos(arg);
}
import numpy as np
import pylab
and the header file cos_module.h:
import cos_doubles
double cos_func(double arg);
x = np.arange(0, 2 * np.pi, 0.1)
y = np.empty_like(x)

14.3. Ctypes 448 14.4. SWIG 449

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

And our goal is to expose the cos_func to Python. To achieve this with SWIG, we must write an interface file
In [2]: cos_module?
which contains the instructions for SWIG.
Type: module
/* Example of wrapping cos function from math.h using SWIG. */ String Form:<module 'cos_module' from 'cos_module.py'>
File: /home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/swig/cos_
,→module.py
%module cos_module
%{ Docstring: <no docstring>
/* the resulting C file should be built as a python extension */
# define SWIG_FILE_WITH_INIT In [3]: dir(cos_module)
/* Includes the header in the wrapper code */ Out[3]:
# include "cos_module.h" ['__builtins__',
%} '__doc__',
/* Parse the header file to generate wrappers */ '__file__',
%include "cos_module.h" '__name__',
'__package__',
'_cos_module',
As you can see, not too much code is needed here. For this simple example it is enough to simply include '_newclass',
the header file in the interface file, to expose the function to Python. However, SWIG does allow for more fine '_object',
grained inclusion/exclusion of functions found in header files, check the documentation for details. '_swig_getattr',
'_swig_property',
Generating the compiled wrappers is a two stage process:
'_swig_repr',
1. Run the swig executable on the interface file to generate the files cos_module_wrap.c, which is the '_swig_setattr',
source file for the autogenerated Python C-extension and cos_module.py, which is the autogenerated '_swig_setattr_nondynamic',
pure python module. 'cos_func']

2. Compile the cos_module_wrap.c into the _cos_module.so. Luckily, distutils knows how to handle In [4]: cos_module.cos_func(1.0)
SWIG interface files, so that our setup.py is simply: Out[4]: 0.5403023058681398

from distutils.core import setup, Extension In [5]: cos_module.cos_func(0.0)

Out[5]: 1.0
setup(ext_modules=[Extension("_cos_module",
sources=["cos_module.c", "cos_module.i"])]) In [6]: cos_module.cos_func(3.14159265359)
Out[6]: -1.0
$ cd advanced/interfacing_with_c/swig
Again we test for robustness, and we see that we get a better error message (although, strictly speaking in
$ ls Python there is no double type):
cos_module.c cos_module.h cos_module.i setup.py
In [7]: cos_module.cos_func('foo')
$ python setup.py build_ext --inplace ---------------------------------------------------------------------------
running build_ext TypeError Traceback (most recent call last)
building '_cos_module' extension <ipython-input-7-11bee483665d> in <module>()
swigging cos_module.i to cos_module_wrap.c ----> 1 cos_module.cos_func('foo')
swig -python -o cos_module_wrap.c cos_module.i
creating build TypeError: in method 'cos_func', argument 1 of type 'double'
creating build/temp.linux-x86_64-2.7
gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -
,→fPIC -I/home/esc/anaconda/include/python2.7 -c cos_module.c -o build/temp.linux-x86_64-2.7/
,→cos_module.o
14.4.2 Numpy Support
gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -
,→fPIC -I/home/esc/anaconda/include/python2.7 -c cos_module_wrap.c -o build/temp.linux-x86_64- Numpy provides support for SWIG with the numpy.i file. This interface file defines various so-called typemaps
,→2.7/cos_module_wrap.o which support conversion between Numpy arrays and C-Arrays. In the following example we will take a quick
gcc -pthread -shared build/temp.linux-x86_64-2.7/cos_module.o build/temp.linux-x86_64-2.7/cos_ look at how such typemaps work in practice.
,→module_wrap.o -L/home/esc/anaconda/lib -lpython2.7 -o /home/esc/git-working/scipy-lecture-
,→notes/advanced/interfacing_with_c/swig/_cos_module.so
We have the same cos_doubles function as in the ctypes example:

void cos_doubles(double * in_array, double * out_array, int size);

$ ls
build/ cos_module.c cos_module.h cos_module.i cos_module.py _cos_module.so* cos_module_
,→wrap.c setup.py # include <math.h>

We can now load and execute the cos_module as we have done in the previous examples: /* Compute the cosine of each element in in_array, storing the result in
* out_array. */
In [1]: import cos_module void cos_doubles(double * in_array, double * out_array, int size){
int i;

14.4. SWIG 450 14.4. SWIG 451

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

for(i=0;i<size;i++){ swigging cos_doubles.i to cos_doubles_wrap.c

out_array[i] = cos(in_array[i]); swig -python -o cos_doubles_wrap.c cos_doubles.i
} cos_doubles.i:24: Warning(490): Fragment 'NumPy_Backward_Compatibility' not found.
} cos_doubles.i:24: Warning(490): Fragment 'NumPy_Backward_Compatibility' not found.
cos_doubles.i:24: Warning(490): Fragment 'NumPy_Backward_Compatibility' not found.
This is wrapped as cos_doubles_func using the following SWIG interface file: creating build
creating build/temp.linux-x86_64-2.7
/* Example of wrapping a C function that takes a C double array as input using gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -
* numpy typemaps for SWIG. */ ,→fPIC -I/home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include -I/home/esc/
,→anaconda/include/python2.7 -c cos_doubles.c -o build/temp.linux-x86_64-2.7/cos_doubles.o
%module cos_doubles gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -
%{ ,→fPIC -I/home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include -I/home/esc/
/* the resulting C file should be built as a python extension */ ,→anaconda/include/python2.7 -c cos_doubles_wrap.c -o build/temp.linux-x86_64-2.7/cos_doubles_
# define SWIG_FILE_WITH_INIT ,→wrap.o
/* Includes the header in the wrapper code */ In file included from /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/
# include "cos_doubles.h" ,→ndarraytypes.h:1722,
%} from /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/
,→ndarrayobject.h:17,
/* include the numpy typemaps */ from /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/
%include "numpy.i" ,→arrayobject.h:15,
/* need this for correct module initialization */ from cos_doubles_wrap.c:2706:
%init %{ /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/npy_deprecated_api.
import_array(); ,→h:11:2: warning: #warning "Using deprecated NumPy API, disable it by #defining NPY_NO_
%} ,→DEPRECATED_API NPY_1_7_API_VERSION"
gcc -pthread -shared build/temp.linux-x86_64-2.7/cos_doubles.o build/temp.linux-x86_64-2.7/cos_
/* typemaps for the two arrays, the second will be modified in-place */ ,→doubles_wrap.o -L/home/esc/anaconda/lib -lpython2.7 -o /home/esc/git-working/scipy-lecture-
%apply (double* IN_ARRAY1, int DIM1) {(double * in_array, int size_in)} ,→notes/advanced/interfacing_with_c/swig_numpy/_cos_doubles.so
%apply (double* INPLACE_ARRAY1, int DIM1) {(double * out_array, int size_out)} $ ls
build/ cos_doubles.h cos_doubles.py cos_doubles_wrap.c setup.py
/* Wrapper for cos_doubles that massages the types */ cos_doubles.c cos_doubles.i _cos_doubles.so* numpy.i test_cos_doubles.py
%inline %{
/* takes as input two numpy arrays */ And, as before, we convince ourselves that it worked:
void cos_doubles_func(double * in_array, int size_in, double * out_array, int size_out) {
/* calls the original funcion, providing only the size of the first */ import numpy as np
cos_doubles(in_array, out_array, size_in); import pylab
} import cos_doubles
%}
x = np.arange(0, 2 * np.pi, 0.1)
• To use the Numpy typemaps, we need include the numpy.i file. y = np.empty_like(x)

• Observe the call to import_array() which we encountered already in the Numpy-C-API example. cos_doubles.cos_doubles_func(x, y)
pylab.plot(x, y)
• Since the type maps only support the signature ARRAY, SIZE we need to wrap the cos_doubles as
pylab.show()
cos_doubles_func which takes two arrays including sizes as input.
• As opposed to the simple SWIG example, we don’t include the cos_doubles.h header, There is nothing
there that we wish to expose to Python since we expose the functionality through cos_doubles_func.
And, as before we can use distutils to wrap this:

from distutils.core import setup, Extension

import numpy

setup(ext_modules=[Extension("_cos_doubles",
sources=["cos_doubles.c", "cos_doubles.i"],
include_dirs=[numpy.get_include()])])

As previously, we need to use include_dirs to specify the location.

$ ls
cos_doubles.c cos_doubles.h cos_doubles.i
$ python setup.py build_ext -i
numpy.i setup.py test_cos_doubles.py
14.5 Cython
running build_ext
building '_cos_doubles' extension Cython is both a Python-like language for writing C-extensions and an advanced compiler for this language.

14.4. SWIG 452 14.5. Cython 453

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

The Cython language is a superset of Python, which comes with additional constructs that allow you call C
$ cd advanced/interfacing_with_c/cython
functions and annotate variables and class attributes with c types. In this sense one could also call it a Python $ ls
with types. cos_module.pyx setup.py
In addition to the basic use case of wrapping native code, Cython supports an additional use-case, namely $ python setup.py build_ext --inplace
running build_ext
interactive optimization. Basically, one starts out with a pure-Python script and incrementally adds Cython
cythoning cos_module.pyx to cos_module.c
types to the bottleneck code to optimize only those code paths that really matter.
building 'cos_module' extension
In this sense it is quite similar to SWIG, since the code can be autogenerated but in a sense it also quite similar creating build
to ctypes since the wrapping code can (almost) be written in Python. creating build/temp.linux-x86_64-2.7
gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -
While others solutions that autogenerate code can be quite difficult to debug (for example SWIG) Cython ,→fPIC -I/home/esc/anaconda/include/python2.7 -c cos_module.c -o build/temp.linux-x86_64-2.7/
comes with an extension to the GNU debugger that helps debug Python, Cython and C code. ,→cos_module.o
gcc -pthread -shared build/temp.linux-x86_64-2.7/cos_module.o -L/home/esc/anaconda/lib -
,→lpython2.7 -o /home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/cython/
Note: The autogenerated C code uses the Python-C-Api. ,→cos_module.so
$ ls
build/ cos_module.c cos_module.pyx cos_module.so* setup.py
Advantages
• Python like language for writing C-extensions And running it:
• Autogenerated code In [1]: import cos_module
• Supports incremental optimization
In [2]: cos_module?
• Includes a GNU debugger extension Type: module
String Form:<module 'cos_module' from 'cos_module.so'>
• Support for C++ (Since version 0.13)
File: /home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/cython/cos_
Disadvantages ,→module.so
Docstring: <no docstring>
• Must be compiled
• Requires an additional library ( but only at build time, at this problem can be overcome by shipping the In [3]: dir(cos_module)
Out[3]:
generated C files)
['__builtins__',
'__doc__',
'__file__',
14.5.1 Example '__name__',
'__package__',
The main Cython code for our cos_module is contained in the file cos_module.pyx: '__test__',
'cos_func']
""" Example of wrapping cos function from math.h using Cython. """
In [4]: cos_module.cos_func(1.0)
cdef extern from "math.h": Out[4]: 0.5403023058681398
double cos(double arg)
In [5]: cos_module.cos_func(0.0)
def cos_func(arg): Out[5]: 1.0
return cos(arg)
In [6]: cos_module.cos_func(3.14159265359)
Note the additional keywords such as cdef and extern. Also the cos_func is then pure Python. Out[6]: -1.0
Again we can use the standard distutils module, but this time we need some additional pieces from the
Cython.Distutils: And, testing a little for robustness, we can see that we get good error messages:

from distutils.core import setup, Extension In [7]: cos_module.cos_func('foo')

from Cython.Distutils import build_ext ---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
setup( <ipython-input-7-11bee483665d> in <module>()
cmdclass={'build_ext': build_ext}, ----> 1 cos_module.cos_func('foo')
ext_modules=[Extension("cos_module", ["cos_module.pyx"])]
) /home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/cython/cos_module.so in␣
,→cos_module.cos_func (cos_module.c:506)()

Compiling this: TypeError: a float is required

Additionally, it is worth noting that Cython ships with complete declarations for the C math library, which

14.5. Cython 454 14.5. Cython 455

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

simplifies the code above to become:

from Cython.Distutils import build_ext
""" Simpler example of wrapping cos function from math.h using Cython. """
setup(
from libc.math cimport cos cmdclass={'build_ext': build_ext},
ext_modules=[Extension("cos_doubles",
def cos_func(arg): sources=["_cos_doubles.pyx", "cos_doubles.c"],
return cos(arg) include_dirs=[numpy.get_include()])],
)
In this case the cimport statement is used to import the cos function.
• As with the previous compiled Numpy examples, we need the include_dirs option.

$ ls
14.5.2 Numpy Support cos_doubles.c cos_doubles.h _cos_doubles.pyx setup.py test_cos_doubles.py
$ python setup.py build_ext -i
Cython has support for Numpy via the numpy.pyx file which allows you to add the Numpy array type to your running build_ext
Cython code. I.e. like specifying that variable i is of type int, you can specify that variable a is of type numpy. cythoning _cos_doubles.pyx to _cos_doubles.c
ndarray with a given dtype. Also, certain optimizations such as bounds checking are supported. Look at the building 'cos_doubles' extension
corresponding section in the Cython documentation. In case you want to pass Numpy arrays as C arrays to creating build
your Cython wrapped C functions, there is a section about this in the Cython wiki. creating build/temp.linux-x86_64-2.7
gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -
In the following example, we will show how to wrap the familiar cos_doubles function using Cython. ,→fPIC -I/home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include -I/home/esc/
,→anaconda/include/python2.7 -c _cos_doubles.c -o build/temp.linux-x86_64-2.7/_cos_doubles.o
void cos_doubles(double * in_array, double * out_array, int size); In file included from /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/
,→ndarraytypes.h:1722,

# include <math.h> from /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/

,→ndarrayobject.h:17,

/* Compute the cosine of each element in in_array, storing the result in from /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/
* out_array. */ ,→arrayobject.h:15,

void cos_doubles(double * in_array, double * out_array, int size){ from _cos_doubles.c:253:

int i; /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/npy_deprecated_api.
for(i=0;i<size;i++){ ,→h:11:2: warning: #warning "Using deprecated NumPy API, disable it by #defining NPY_NO_

out_array[i] = cos(in_array[i]); ,→DEPRECATED_API NPY_1_7_API_VERSION"

} /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/__ufunc_api.h:236:␣
} ,→warning: ‘_import_umath’ defined but not used
gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -
,→fPIC -I/home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include -I/home/esc/
This is wrapped as cos_doubles_func using the following Cython code:
,→anaconda/include/python2.7 -c cos_doubles.c -o build/temp.linux-x86_64-2.7/cos_doubles.o

""" Example of wrapping a C function that takes C double arrays as input using gcc -pthread -shared build/temp.linux-x86_64-2.7/_cos_doubles.o build/temp.linux-x86_64-2.7/
the Numpy declarations from Cython """ ,→cos_doubles.o -L/home/esc/anaconda/lib -lpython2.7 -o /home/esc/git-working/scipy-lecture-
,→notes/advanced/interfacing_with_c/cython_numpy/cos_doubles.so

# cimport the Cython declarations for numpy $ ls

cimport numpy as np build/ _cos_doubles.c cos_doubles.c cos_doubles.h _cos_doubles.pyx cos_doubles.so* setup.
,→py test_cos_doubles.py
# if you want to use the Numpy-C-API from Cython
# (not strictly necessary for this example, but good practice) And, as before, we convince ourselves that it worked:
np.import_array()
import numpy as np
# cdefine the signature of our c function import pylab
cdef extern from "cos_doubles.h": import cos_doubles
void cos_doubles (double * in_array, double * out_array, int size)
x = np.arange(0, 2 * np.pi, 0.1)
# create the wrapper code, with numpy type annotations y = np.empty_like(x)
def cos_doubles_func(np.ndarray[double, ndim=1, mode="c"] in_array not None,
np.ndarray[double, ndim=1, mode="c"] out_array not None): cos_doubles.cos_doubles_func(x, y)
cos_doubles(<double*> np.PyArray_DATA(in_array), pylab.plot(x, y)
<double*> np.PyArray_DATA(out_array), pylab.show()
in_array.shape[0])

And can be compiled using distutils:

from distutils.core import setup, Extension

import numpy

14.5. Cython 456 14.5. Cython 457

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

14.8.1 Python-C-API

1. Modify the Numpy example such that the function takes two input arguments, where the second is the
preallocated output array, making it similar to the other Numpy examples.
2. Modify the example such that the function only takes a single input array and modifies this in place.
3. Try to fix the example to use the new Numpy iterator protocol. If you manage to obtain a working solu-
tion, please submit a pull-request on github.
4. You may have noticed, that the Numpy-C-API example is the only Numpy example that does not wrap
cos_doubles but instead applies the cos function directly to the elements of the Numpy array. Does
this have any advantages over the other techniques.
5. Can you wrap cos_doubles using only the Numpy-C-API. You may need to ensure that the arrays have

14.6 Summary
the correct type, are one dimensional and contiguous in memory.

In this section four different techniques for interfacing with native code have been presented. The table below 14.8.2 Ctypes
roughly summarizes some of the aspects of the techniques.
1. Modify the Numpy example such that cos_doubles_func handles the preallocation for you, thus mak-
ing it more like the Numpy-C-API example.
x Part of CPython Compiled Autogenerated Numpy Support
Python-C-API True True False True
Ctypes True False False True 14.8.3 SWIG
Swig False True True True
Cython False True True True 1. Look at the code that SWIG autogenerates, how much of it do you understand?
2. Modify the Numpy example such that cos_doubles_func handles the preallocation for you, thus mak-
Of all three presented techniques, Cython is the most modern and advanced. In particular, the ability to opti-
ing it more like the Numpy-C-API example.
mize code incrementally by adding types to your Python code is unique.
3. Modify the cos_doubles C function so that it returns an allocated array. Can you wrap this using SWIG
typemaps? If not, why not? Is there a workaround for this specific situation? (Hint: you know the size of
14.7 Further Reading and References the output array, so it may be possible to construct a Numpy array from the returned double *.)

• Gaël Varoquaux’s blog post about avoiding data copies provides some insight on how to handle memory 14.8.4 Cython
management cleverly. If you ever run into issues with large datasets, this is a reference to come back to
for some inspiration. 1. Look at the code that Cython autogenerates. Take a closer look at some of the comments that Cython
inserts. What do you see?

14.8 Exercises 2. Look at the section Working with Numpy from the Cython documentation to learn how to incrementally
optimize a pure python script that uses Numpy.
3. Modify the Numpy example such that cos_doubles_func handles the preallocation for you, thus mak-
Since this is a brand new section, the exercises are considered more as pointers as to what to look at next, so
ing it more like the Numpy-C-API example.
pick the ones that you find more interesting. If you have good ideas for exercises, please let us know!
1. Download the source code for each example and compile and run them on your machine.
2. Make trivial changes to each example and convince yourself that this works. ( E.g. change cos for sin.)
3. Most of the examples, especially the ones involving Numpy may still be fragile and respond badly to
input errors. Look for ways to crash the examples, figure what the problem is and devise a potential
solution. Here are some ideas:
(a) Numerical overflow.
(b) Input and output arrays that have different lengths.
(c) Multidimensional array.
(d) Empty array
(e) Arrays with non-double types
4. Use the %timeit IPython magic to measure the execution time of the various solutions

14.6. Summary 458 14.8. Exercises 459

 Scipy lecture notes, Edition 2017.1

This part of the Scipy lecture notes is dedicated to various scientific packages useful for extended needs.

Part III

Packages and applications

460 461
 Scipy lecture notes, Edition 2017.1

R is a language dedicated to statistics. Python is a general-purpose language with statistics modules. R has
more statistical analysis features than Python, and specialized syntaxes. However, when it comes to building
complex analysis pipelines that mix statistics with e.g. image analysis, text mining, or control of a physical
experiment, the richness of Python is an invaluable asset.

15
Contents

• Data representation and interaction

– Data as a table
CHAPTER – The pandas data-frame
• Hypothesis testing: comparing two groups
– Student’s t-test: the simplest statistical test
– Paired tests: repeated measurements on the same indivuals
• Linear models, multiple factors, and analysis of variance
– “formulas” to specify statistical models in Python

Statistics in Python – Multiple Regression: including multiple factors

– Post-hoc hypothesis testing: analysis of variance (ANOVA)
• More visualization: seaborn for statistical exploration
– Pairplot: scatter matrices
– lmplot: plotting a univariate regression
• Testing for interactions
Author: Gaël Varoquaux
• Full code for the figures
• Solutions to this chapter’s exercises
Requirements

• Standard scientific Python environment (numpy, scipy, matplotlib)

Tip: In this document, the Python inputs are represented with the sign “>>>”.
• Pandas
• Statsmodels
• Seaborn
To install Python and these dependencies, we recommend that you download Anaconda Python or En-
thought Canopy, or preferably use the package manager if you are under Ubuntu or other linux. Disclaimer: Gender questions
Some of the examples of this tutorial are chosen around gender questions. The reason is that on such questions
See also: controlling the truth of a claim actually matters to many people.
• Bayesian statistics in Python: This chapter does not cover tools for Bayesian statistics. Of particular
interest for Bayesian modelling is PyMC, which implements a probabilistic programming language in
Python.
15.1 Data representation and interaction
• Read a statistics book: The Think stats book is available as free PDF or in print and is a great introduction
to statistics.
15.1.1 Data as a table

The setting that we consider for statistical analysis is that of multiple observations or samples described by a
set of different attributes or features. The data can than be seen as a 2D table, or matrix, with columns giving
the different attributes of the data, and rows the observations. For instance, the data contained in examples/
brain_size.csv:
Tip: Why Python for statistics?

462 15.1. Data representation and interaction 463

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

"";"Gender";"FSIQ";"VIQ";"PIQ";"Weight";"Height";"MRI_Count" >>> pandas.DataFrame({'t': t, 'sin': sin_t, 'cos': cos_t})

"1";"Female";133;132;124;"118";"64.5";816932 cos sin t
"2";"Male";140;150;124;".";"72.5";1001121 0 0.960170 0.279415 -6.000000
"3";"Male";139;123;150;"143";"73.3";1038437 1 0.609977 0.792419 -5.368421
"4";"Male";133;129;128;"172";"68.8";965353 2 0.024451 0.999701 -4.736842
"5";"Female";137;132;134;"147";"65.0";951545 3 -0.570509 0.821291 -4.105263
4 -0.945363 0.326021 -3.473684
5 -0.955488 -0.295030 -2.842105
6 -0.596979 -0.802257 -2.210526
15.1.2 The pandas data-frame 7 -0.008151 -0.999967 -1.578947
8 0.583822 -0.811882 -0.947368
...
Tip: We will store and manipulate this data in a pandas.DataFrame, from the pandas module. It is the Python
equivalent of the spreadsheet table. It is different from a 2D numpy array as it has named columns, can contain
a mixture of different data types by column, and has elaborate selection and pivotal mechanisms.

Creating dataframes: reading data files or converting arrays Other inputs: pandas can input data from SQL, excel files, or other formats. See the pandas documentation.

Separator

It is a CSV file, but the separator is “;”

Manipulating data
Reading from a CSV file: Using the above CSV file that gives observations of brain size and weight and IQ
data is a pandas.DataFrame, that resembles R’s dataframe:
(Willerman et al. 1991), the data are a mixture of numerical and categorical values:
>>> data.shape # 40 rows and 8 columns
>>> import pandas
(40, 8)
>>> data = pandas.read_csv('examples/brain_size.csv', sep=';', na_values=".")
>>> data
>>> data.columns # It has columns
Unnamed: 0 Gender FSIQ VIQ PIQ Weight Height MRI_Count
Index([u'Unnamed: 0', u'Gender', u'FSIQ', u'VIQ', u'PIQ', u'Weight', u'Height', u'MRI_Count'],␣
0 1 Female 133 132 124 118.0 64.5 816932
,→dtype='object')
1 2 Male 140 150 124 NaN 72.5 1001121
2 3 Male 139 123 150 143.0 73.3 1038437
>>> print(data['Gender']) # Columns can be addressed by name
3 4 Male 133 129 128 172.0 68.8 965353
0 Female
4 5 Female 137 132 134 147.0 65.0 951545
1 Male
...
2 Male
3 Male
4 Female
Warning: Missing values ...

The weight of the second individual is missing in the CSV file. If we don’t specify the missing value (NA = >>> # Simpler selector
not available) marker, we will not be able to do statistical analysis. >>> data[data['Gender'] == 'Female']['VIQ'].mean()
109.45

Note: For a quick view on a large dataframe, use its describe method: pandas.DataFrame.describe().

Creating from arrays: A pandas.DataFrame can also be seen as a dictionary of 1D ‘series’, eg arrays or lists. If
we have 3 numpy arrays:

>>> import numpy as np

>>> t = np.linspace(-6, 6, 20)
>>> sin_t = np.sin(t) groupby: splitting a dataframe on values of categorical variables:
>>> cos_t = np.cos(t)
>>> groupby_gender = data.groupby('Gender')
>>> for gender, value in groupby_gender['VIQ']:
We can expose them as a pandas.DataFrame:
... print((gender, value.mean()))

15.1. Data representation and interaction 464 15.1. Data representation and interaction 465
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Plotting data
('Female', 109.45)
('Male', 115.25)
Pandas comes with some plotting tools (pandas.tools.plotting, using matplotlib behind the scene) to dis-
groupby_gender is a powerful object that exposes many operations on the resulting group of dataframes: play statistics of the data in dataframes:
Scatter matrices:
>>> groupby_gender.mean()
Unnamed: 0 FSIQ VIQ PIQ Weight Height MRI_Count >>> from pandas.tools import plotting
Gender >>> plotting.scatter_matrix(data[['Weight', 'Height', 'MRI_Count']])
Female 19.65 111.9 109.45 110.45 137.200000 65.765000 862654.6
Male 21.35 115.0 115.25 111.60 166.444444 71.431579 954855.4

Tip: Use tab-completion on groupby_gender to find more. Other common grouping functions are median,
count (useful for checking to see the amount of missing values in different subsets) or sum. Groupby evalua-
tion is lazy, no work is done until an aggregation function is applied.

>>> plotting.scatter_matrix(data[['PIQ', 'VIQ', 'FSIQ']])

Exercise
Two populations
• What is the mean value for VIQ for the full population?
The IQ metrics are bimodal, as if there are 2 sub-populations.
• How many males/females were included in this study?
Hint use ‘tab completion’ to find out the methods that can be called, instead of ‘mean’ in the above
example.
• What is the average value of MRI counts expressed in log units, for males and females?

Note: groupby_gender.boxplot is used for the plots above (see this example).

15.1. Data representation and interaction 466 15.1. Data representation and interaction 467
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

15.2.1 Student’s t-test: the simplest statistical test

1-sample t-test: testing the value of a population mean

scipy.stats.ttest_1samp() tests if the population mean of data is likely to be equal to a given value (tech-
nically if observations are drawn from a Gaussian distributions of given population mean). It returns the T
statistic, and the p-value (see the function’s help):
Exercise
>>> stats.ttest_1samp(data['VIQ'], 0)
Ttest_1sampResult(statistic=30.088099970..., pvalue=1.32891964...e-28)
Plot the scatter matrix for males only, and for females only. Do you think that the 2 sub-populations corre-
spond to gender?
Tip: With a p-value of 10^-28 we can claim that the population mean for the IQ (VIQ measure) is not 0.

15.2 Hypothesis testing: comparing two groups

2-sample t-test: testing for difference across populations
For simple statistical tests, we will use the scipy.stats sub-module of scipy:
We have seen above that the mean VIQ in the male and female populations were different. To test if this is
>>> from scipy import stats significant, we do a 2-sample t-test with scipy.stats.ttest_ind():

>>> female_viq = data[data['Gender'] == 'Female']['VIQ']

See also:
>>> male_viq = data[data['Gender'] == 'Male']['VIQ']
Scipy is a vast library. For a quick summary to the whole library, see the scipy chapter. >>> stats.ttest_ind(female_viq, male_viq)
Ttest_indResult(statistic=-0.77261617232..., pvalue=0.4445287677858...)

15.2. Hypothesis testing: comparing two groups 468 15.2. Hypothesis testing: comparing two groups 469
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

15.2.2 Paired tests: repeated measurements on the same indivuals Exercise

• Test the difference between weights in males and females.

• Use non parametric statistics to test the difference between VIQ in males and females.
Conclusion: we find that the data does not support the hypothesis that males and females have different
VIQ.

PIQ, VIQ, and FSIQ give 3 measures of IQ. Let us test if FISQ
15.3 Linear models, multiple factors, and analysis of variance
and PIQ are significantly different. We can use a 2 sample test:
15.3.1 “formulas” to specify statistical models in Python
>>> stats.ttest_ind(data['FSIQ'], data['PIQ'])
Ttest_indResult(statistic=0.46563759638..., pvalue=0.64277250...)
A simple linear regression
The problem with this approach is that it forgets that there are links between observations: FSIQ and PIQ are
measured on the same individuals. Thus the variance due to inter-subject variability is confounding, and can
be removed, using a “paired test”, or “repeated measures test”:

>>> stats.ttest_rel(data['FSIQ'], data['PIQ'])

Ttest_relResult(statistic=1.784201940..., pvalue=0.082172638183...)

Given two set of observations, x and y, we want to test the

hypothesis that y is a linear function of x. In other terms:
This is equivalent to a 1-sample test on the difference:
y = x ∗ coe f + i nt er cept + e
>>> stats.ttest_1samp(data['FSIQ'] - data['PIQ'], 0) where e is observation noise. We will use the statsmodels module to:
Ttest_1sampResult(statistic=1.784201940..., pvalue=0.082172638...)
1. Fit a linear model. We will use the simplest strategy, ordinary least squares (OLS).
2. Test that coef is non zero.

T-tests assume Gaussian errors. We can use a Wilcoxon signed-rank test, that relaxes this assumption:

>>> stats.wilcoxon(data['FSIQ'], data['PIQ']) First, we generate simulated data according to the model:
WilcoxonResult(statistic=274.5, pvalue=0.106594927...)
>>> import numpy as np
>>> x = np.linspace(-5, 5, 20)
Note: The corresponding test in the non paired case is the Mann–Whitney U test, scipy.stats. >>> np.random.seed(1)
>>> # normal distributed noise
mannwhitneyu().
>>> y = -5 + 3*x + 4 * np.random.normal(size=x.shape)
>>> # Create a data frame containing all the relevant variables
>>> data = pandas.DataFrame({'x': x, 'y': y})

15.2. Hypothesis testing: comparing two groups 470 15.3. Linear models, multiple factors, and analysis of variance 471
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

“formulas” for statistics in Python Categorical variables: comparing groups or multiple categories

See the statsmodels documentation Let us go back the data on brain size:

>>> data = pandas.read_csv('examples/brain_size.csv', sep=';', na_values=".")

We can write a comparison between IQ of male and female using a linear model:

>>> model = ols("VIQ ~ Gender + 1", data).fit()

>>> print(model.summary())
Then we specify an OLS model and fit it: OLS Regression Results
==========================...
>>> from statsmodels.formula.api import ols
Dep. Variable: VIQ R-squared: 0.015
>>> model = ols("y ~ x", data).fit()
Model: OLS Adj. R-squared: -0.010
Method: Least Squares F-statistic: 0.5969
We can inspect the various statistics derived from the fit: Date: ... Prob (F-statistic): 0.445
Time: ... Log-Likelihood: -182.42
>>> print(model.summary())
No. Observations: 40 AIC: 368.8
OLS Regression Results
Df Residuals: 38 BIC: 372.2
==========================...
Df Model: 1
Dep. Variable: y R-squared: 0.804
Covariance Type: nonrobust
Model: OLS Adj. R-squared: 0.794
==========================...
Method: Least Squares F-statistic: 74.03
coef std err t P>|t| [95.0% Conf. Int.]
Date: ... Prob (F-statistic): 8.56e-08
-----------------------------------------------------------------------...
Time: ... Log-Likelihood: -57.988
Intercept 109.4500 5.308 20.619 0.000 98.704 120.196
No. Observations: 20 AIC: 120.0
Gender[T.Male] 5.8000 7.507 0.773 0.445 -9.397 20.997
Df Residuals: 18 BIC: 122.0
==========================...
Df Model: 1
Omnibus: 26.188 Durbin-Watson: 1.709
Covariance Type: nonrobust
Prob(Omnibus): 0.000 Jarque-Bera (JB): 3.703
==========================...
Skew: 0.010 Prob(JB): 0.157
coef std err t P>|t| [95.0% Conf. Int.]
Kurtosis: 1.510 Cond. No. 2.62
------------------------------------------...
==========================...
Intercept -5.5335 1.036 -5.342 0.000 -7.710 -3.357
x 2.9369 0.341 8.604 0.000 2.220 3.654
Warnings:
==========================...
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.
Omnibus: 0.100 Durbin-Watson: 2.956
Prob(Omnibus): 0.951 Jarque-Bera (JB): 0.322
Skew: -0.058 Prob(JB): 0.851
Kurtosis: 2.390 Cond. No. 3.03 Tips on specifying model
==========================...
Forcing categorical: the ‘Gender’ is automatically detected as a categorical variable, and thus each of its
Warnings: different values are treated as different entities.
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.
An integer column can be forced to be treated as categorical using:
>>> model = ols('VIQ ~ C(Gender)', data).fit()
Terminology:
Intercept: We can remove the intercept using - 1 in the formula, or force the use of an intercept using + 1.
Statsmodels uses a statistical terminology: the y variable in statsmodels is called ‘endogenous’ while the x
variable is called exogenous. This is discussed in more detail here. Tip: By default, statsmodels treats a categorical variable with K possible values as K-1 ‘dummy’ boolean
To simplify, y (endogenous) is the value you are trying to predict, while x (exogenous) represents the features variables (the last level being absorbed into the intercept term). This is almost always a good default
you are using to make the prediction. choice - however, it is possible to specify different encodings for categorical variables (http://statsmodels.
sourceforge.net/devel/contrasts.html).

Exercise

Retrieve the estimated parameters from the model above. Hint: use tab-completion to find the relevent
attribute.

15.3. Linear models, multiple factors, and analysis of variance 472 15.3. Linear models, multiple factors, and analysis of variance 473
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Link to t-tests between different FSIQ and PIQ Such a model can be seen in 3D as fitting a plane to a cloud of (x, y, z) points.

To compare different types of IQ, we need to create a “long-form” table, listing IQs, where the type of IQ is
indicated by a categorical variable:
>>> data_fisq = pandas.DataFrame({'iq': data['FSIQ'], 'type': 'fsiq'})
>>> data_piq = pandas.DataFrame({'iq': data['PIQ'], 'type': 'piq'})
>>> data_long = pandas.concat((data_fisq, data_piq))
>>> print(data_long) Example: the iris data (examples/iris.csv)
iq type
0 133 fsiq
Tip: Sepal and petal size tend to be related: bigger flowers are bigger! But is there in addition a systematic
1 140 fsiq
effect of species?
2 139 fsiq
...
31 137 piq
32 110 piq
33 86 piq
...

>>> model = ols("iq ~ type", data_long).fit()

>>> print(model.summary())
OLS Regression Results
...
==========================...
coef std err t P>|t| [95.0% Conf. Int.]
------------------------------------------...
Intercept 113.4500 3.683 30.807 0.000 106.119 120.781
type[T.piq] -2.4250 5.208 -0.466 0.643 -12.793 7.943
...

We can see that we retrieve the same values for t-test and corresponding p-values for the effect of the type
of iq than the previous t-test:
>>> stats.ttest_ind(data['FSIQ'], data['PIQ'])
Ttest_indResult(statistic=0.46563759638..., pvalue=0.64277250...)

15.3.2 Multiple Regression: including multiple factors

>>> data = pandas.read_csv('examples/iris.csv')

>>> model = ols('sepal_width ~ name + petal_length', data).fit()
>>> print(model.summary())
OLS Regression Results
==========================...
Dep. Variable: sepal_width R-squared: 0.478
Model: OLS Adj. R-squared: 0.468
Method: Least Squares F-statistic: 44.63
Date: ... Prob (F-statistic): 1.58e-20
Time: ... Log-Likelihood: -38.185
No. Observations: 150 AIC: 84.37
Df Residuals: 146 BIC: 96.41
Df Model: 3
Covariance Type: nonrobust
==========================...
Consider a linear model explaining a variable z (the dependent variable) with 2 variables x and y: coef std err t P>|t| [95.0% Conf. Int.]
z = x c1 + y c2 + i + e ------------------------------------------...

15.3. Linear models, multiple factors, and analysis of variance 474 15.3. Linear models, multiple factors, and analysis of variance 475
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Intercept 2.9813 0.099 29.989 0.000 2.785 3.178 0 8 0 1 21 0 0.707570 35 2

name[T.versicolor] -1.4821 0.181 -8.190 0.000 -1.840 -1.124 1 9 0 1 42 0 0.694605 57 3
name[T.virginica] -1.6635 0.256 -6.502 0.000 -2.169 -1.158 2 12 0 0 1 0 0.824126 19 3
petal_length 0.2983 0.061 4.920 0.000 0.178 0.418 3 12 0 0 4 0 0.602060 22 3
==========================... ...
Omnibus: 2.868 Durbin-Watson: 1.753
Prob(Omnibus): 0.238 Jarque-Bera (JB): 2.885
Skew: -0.082 Prob(JB): 0.236
Kurtosis: 3.659 Cond. No. 54.0 15.4.1 Pairplot: scatter matrices
==========================...
We can easily have an intuition on the interactions between continuous variables using seaborn.pairplot()
Warnings: to display a scatter matrix:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.
>>> import seaborn
>>> seaborn.pairplot(data, vars=['WAGE', 'AGE', 'EDUCATION'],
... kind='reg')

15.3.3 Post-hoc hypothesis testing: analysis of variance (ANOVA)

In the above iris example, we wish to test if the petal length is different between versicolor and virginica, after
removing the effect of sepal width. This can be formulated as testing the difference between the coefficient as-
sociated to versicolor and virginica in the linear model estimated above (it is an Analysis of Variance, ANOVA).
For this, we write a vector of ‘contrast’ on the parameters estimated: we want to test "name[T.versicolor]
- name[T.virginica]", with an F-test:

>>> print(model.f_test([0, 1, -1, 0]))

<F test: F=array([[ 3.24533535]]), p=0.07369..., df_denom=146, df_num=1>

Is this difference significant?

Exercise

Going back to the brain size + IQ data, test if the VIQ of male and female are different after removing the
effect of brain size, height and weight.

Categorical variables can be

plotted as the hue:

>>> seaborn.pairplot(data, vars=['WAGE', 'AGE', 'EDUCATION'],

15.4 More visualization: seaborn for statistical exploration ... kind='reg', hue='SEX')

Seaborn combines simple statistical fits with plotting on pandas dataframes.

Let us consider a data giving wages and many other personal information on 500 individuals (Berndt, ER. The
Practice of Econometrics. 1991. NY: Addison-Wesley).

Tip: The full code loading and plotting of the wages data is found in corresponding example.

>>> print data

EDUCATION SOUTH SEX EXPERIENCE UNION WAGE AGE RACE \

15.4. More visualization: seaborn for statistical exploration 476 15.4. More visualization: seaborn for statistical exploration 477
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

15.4.2 lmplot: plotting a univariate regression

A regression capturing the relation between one variable

and another, eg wage and eduction, can be plotted using seaborn.lmplot():

>>> seaborn.lmplot(y='WAGE', x='EDUCATION', data=data)

Robust regression

Tip: Given that, in the above plot, there seems to be a couple of data points that are outside of the main
Look and feel and matplotlib settings
cloud to the right, they might be outliers, not representative of the population, but driving the regression.
Seaborn changes the default of matplotlib figures to achieve a more “modern”, “excel-like” look. It does that
upon import. You can reset the default using: To compute a regression that is less sentive to outliers, one must use a robust model. This is done in seaborn
using robust=True in the plotting functions, or in statsmodels by replacing the use of the OLS by a “Robust
>>> from matplotlib import pyplot as plt
>>> plt.rcdefaults() Linear Model”, statsmodels.formula.api.rlm().

Tip: To switch back to seaborn settings, or understand better styling in seaborn, see the relevent section of
the seaborn documentation.

15.4. More visualization: seaborn for statistical exploration 478 15.4. More visualization: seaborn for statistical exploration 479
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

15.5 Testing for interactions

15.6 Full code for the figures

Code examples for the statistics chapter.

15.6.1 Boxplots and paired differences

Plot boxplots for FSIQ, PIQ, and the paired difference between the two: while the spread (error bars) for FSIQ
and PIQ are very large, there is a systematic (common) effect due to the subjects. This effect is cancelled out
in the difference and the spread of the difference (“paired” by subject) is much smaller than the spread of the
individual measures.

Do wages increase more with educa- •

tion for males than females?

Tip: The plot above is made of two different fits. We need to formulate a single model that tests for a variance
of slope across the to population. This is done via an “interaction”.

>>> result = sm.ols(formula='wage ~ education + gender + education * gender',

... data=data).fit()
>>> print(result.summary()) •
...
coef std err t P>|t| [95.0% Conf. Int.] import pandas
------------------------------------------------------------------------------
Intercept 0.2998 0.072 4.173 0.000 0.159 0.441 import matplotlib.pyplot as plt
gender[T.male] 0.2750 0.093 2.972 0.003 0.093 0.457
education 0.0415 0.005 7.647 0.000 0.031 0.052 data = pandas.read_csv('brain_size.csv', sep=';', na_values='.')
education:gender[T.male] -0.0134 0.007 -1.919 0.056 -0.027 0.000
==========================... # Box plot of FSIQ and PIQ (different measures od IQ)
... plt.figure(figsize=(4, 3))
data.boxplot(column=['FSIQ', 'PIQ'])

Can we conclude that education benefits males more than females? # Boxplot of the difference
plt.figure(figsize=(4, 3))
plt.boxplot(data['FSIQ'] - data['PIQ'])
plt.xticks((1, ), ('FSIQ - PIQ', ))

plt.show()

Take home messages

Total running time of the script: ( 0 minutes 0.091 seconds)

• Hypothesis testing and p-value give you the significance of an effect / difference Download Python source code: plot_paired_boxplots.py
• Formulas (with categorical variables) enable you to express rich links in your data Download Jupyter notebook: plot_paired_boxplots.ipynb
• Visualizing your data and simple model fits matters! Generated by Sphinx-Gallery

• Conditionning (adding factors that can explain all or part of the variation) is important modeling
aspect that changes the interpretation.

15.5. Testing for interactions 480 15.6. Full code for the figures 481
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

15.6.2 Plotting simple quantities of a pandas dataframe

This example loads from a CSV file data with mixed numerical and categorical entries, and plots a few quan-
tities, separately for females and males, thanks to the pandas integrated plotting tool (that uses matplotlib
behind the scene).
See http://pandas.pydata.org/pandas-docs/stable/visualization.html

import pandas

data = pandas.read_csv('brain_size.csv', sep=';', na_values='.')

# Box plots of different columns for each gender

groupby_gender = data.groupby('Gender')
• groupby_gender.boxplot(column=['FSIQ', 'VIQ', 'PIQ'])

from pandas.tools import plotting

# Scatter matrices for different columns

plotting.scatter_matrix(data[['Weight', 'Height', 'MRI_Count']])
plotting.scatter_matrix(data[['PIQ', 'VIQ', 'FSIQ']])

import matplotlib.pyplot as plt

plt.show()

Total running time of the script: ( 0 minutes 1.288 seconds)

Download Python source code: plot_pandas.py
Download Jupyter notebook: plot_pandas.ipynb
Generated by Sphinx-Gallery

• 15.6.3 Analysis of Iris petal and sepal sizes

Ilustrate an analysis on a real dataset:

• Visualizing the data to formulate intuitions
• Fitting of a linear model
• Hypothesis test of the effect of a categorical variable in the presence of a continuous confound

import matplotlib.pyplot as plt

import pandas
from pandas.tools import plotting

from statsmodels.formula.api import ols

# Load the data

data = pandas.read_csv('iris.csv')

15.6. Full code for the figures 482 15.6. Full code for the figures 483
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Plot a scatter matrix

Method: Least Squares F-statistic: 44.63
# Express the names as categories Date: Tue, 03 Oct 2017 Prob (F-statistic): 1.58e-20
categories = pandas.Categorical(data['name']) Time: 16:00:29 Log-Likelihood: -38.185
No. Observations: 150 AIC: 84.37
# The parameter 'c' is passed to plt.scatter and will control the color Df Residuals: 146 BIC: 96.41
plotting.scatter_matrix(data, c=categories.labels, marker='o') Df Model: 3
Covariance Type: nonrobust
fig = plt.gcf() ======================================================================================
fig.suptitle("blue: setosa, green: versicolor, red: virginica", size=13) coef std err t P>|t| [95.0% Conf. Int.]
--------------------------------------------------------------------------------------
Intercept 2.9813 0.099 29.989 0.000 2.785 3.178
name[T.versicolor] -1.4821 0.181 -8.190 0.000 -1.840 -1.124
name[T.virginica] -1.6635 0.256 -6.502 0.000 -2.169 -1.158
petal_length 0.2983 0.061 4.920 0.000 0.178 0.418
==============================================================================
Omnibus: 2.868 Durbin-Watson: 1.753
Prob(Omnibus): 0.238 Jarque-Bera (JB): 2.885
Skew: -0.082 Prob(JB): 0.236
Kurtosis: 3.659 Cond. No. 54.0
==============================================================================

Warnings:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.
Testing the difference between effect of versicolor and virginica
<F test: F=array([[ 3.24533535]]), p=0.07369058781700738, df_denom=146, df_num=1>

Total running time of the script: ( 0 minutes 0.944 seconds)

Download Python source code: plot_iris_analysis.py
Download Jupyter notebook: plot_iris_analysis.ipynb
Generated by Sphinx-Gallery

15.6.4 Simple Regression

Fit a simple linear regression using ‘statsmodels’, compute corresponding p-values.

# Original author: Thomas Haslwanter

import numpy as np
import matplotlib.pyplot as plt
Statistical analysis import pandas
# Let us try to explain the sepal length as a function of the petal
# For statistics. Requires statsmodels 5.0 or more
# width and the category of iris
from statsmodels.formula.api import ols
# Analysis of Variance (ANOVA) on linear models
model = ols('sepal_width ~ name + petal_length', data).fit()
from statsmodels.stats.anova import anova_lm
print(model.summary())

# Now formulate a "contrast", to test if the offset for versicolor and Generate and show the data
# virginica are identical
x = np.linspace(-5, 5, 20)
print('Testing the difference between effect of versicolor and virginica')
# To get reproducable values, provide a seed value
print(model.f_test([0, 1, -1, 0]))
np.random.seed(1)
plt.show()
y = -5 + 3*x + 4 * np.random.normal(size=x.shape)
Out:
# Plot the data
OLS Regression Results
plt.figure(figsize=(5, 4))
==============================================================================
plt.plot(x, y, 'o')
Dep. Variable: sepal_width R-squared: 0.478
Model: OLS Adj. R-squared: 0.468

15.6. Full code for the figures 484 15.6. Full code for the figures 485
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

==============================================================================
Omnibus: 0.100 Durbin-Watson: 2.956
Prob(Omnibus): 0.951 Jarque-Bera (JB): 0.322
Skew: -0.058 Prob(JB): 0.851
Kurtosis: 2.390 Cond. No. 3.03
==============================================================================

Warnings:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.

ANOVA results
df sum_sq mean_sq F PR(>F)
x 1.0 1588.873443 1588.873443 74.029383 8.560649e-08
Residual 18.0 386.329330 21.462741 NaN NaN

Plot the fitted model

# Retrieve the parameter estimates

offset, coef = model._results.params
plt.plot(x, x*coef + offset)
plt.xlabel('x')
plt.ylabel('y')

plt.show()

Multilinear regression model, calculating fit, P-values, confidence intervals etc.

# Convert the data into a Pandas DataFrame to use the formulas framework
# in statsmodels
data = pandas.DataFrame({'x': x, 'y': y})

# Fit the model

model = ols("y ~ x", data).fit()

# Print the summary

print(model.summary())

# Peform analysis of variance on fitted linear model

anova_results = anova_lm(model)

print('\nANOVA results')
print(anova_results)

Out:

OLS Regression Results

==============================================================================
Dep. Variable: y R-squared: 0.804
Model: OLS Adj. R-squared: 0.794
Method: Least Squares F-statistic: 74.03
Date: Tue, 03 Oct 2017 Prob (F-statistic): 8.56e-08
Time: 16:00:31 Log-Likelihood: -57.988
No. Observations: 20 AIC: 120.0
Df Residuals: 18 BIC: 122.0
Df Model: 1
Covariance Type: nonrobust Total running time of the script: ( 0 minutes 0.081 seconds)
==============================================================================
coef std err t P>|t| [95.0% Conf. Int.] Download Python source code: plot_regression.py
------------------------------------------------------------------------------
Intercept -5.5335 1.036 -5.342 0.000 -7.710 -3.357 Download Jupyter notebook: plot_regression.ipynb
x 2.9369 0.341 8.604 0.000 2.220 3.654 Generated by Sphinx-Gallery

15.6. Full code for the figures 486 15.6. Full code for the figures 487
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

15.6.5 Multiple Regression

Calculate using ‘statsmodels’ just the best fit, or all the corresponding statistical parameters.
Also shows how to make 3d plots.

# Original author: Thomas Haslwanter

import numpy as np
import matplotlib.pyplot as plt
import pandas

# For 3d plots. This import is necessary to have 3D plotting below

from mpl_toolkits.mplot3d import Axes3D

# For statistics. Requires statsmodels 5.0 or more

from statsmodels.formula.api import ols
# Analysis of Variance (ANOVA) on linear models
from statsmodels.stats.anova import anova_lm

Generate and show the data

x = np.linspace(-5, 5, 21)
# We generate a 2D grid
X, Y = np.meshgrid(x, x)

# To get reproducable values, provide a seed value

np.random.seed(1)

# Z is the elevation of this 2D grid

Z = -5 + 3*X - 0.5*Y + 8 * np.random.normal(size=X.shape)

# Plot the data Multilinear regression model, calculating fit, P-values, confidence intervals etc.
fig = plt.figure()
ax = fig.gca(projection='3d') # Convert the data into a Pandas DataFrame to use the formulas framework
surf = ax.plot_surface(X, Y, Z, cmap=plt.cm.coolwarm, # in statsmodels
rstride=1, cstride=1)
ax.view_init(20, -120) # First we need to flatten the data: it's 2D layout is not relevent.
ax.set_xlabel('X') X = X.flatten()
ax.set_ylabel('Y') Y = Y.flatten()
ax.set_zlabel('Z') Z = Z.flatten()

data = pandas.DataFrame({'x': X, 'y': Y, 'z': Z})

# Fit the model

model = ols("z ~ x + y", data).fit()

# Print the summary

print(model.summary())

print("\nRetrieving manually the parameter estimates:")

print(model._results.params)
# should be array([-4.99754526, 3.00250049, -0.50514907])

# Peform analysis of variance on fitted linear model

anova_results = anova_lm(model)

print('\nANOVA results')
print(anova_results)

plt.show()

Out:

15.6. Full code for the figures 488 15.6. Full code for the figures 489
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

OLS Regression Results 'wages.txt')

==============================================================================
Dep. Variable: z R-squared: 0.594 # EDUCATION: Number of years of education
Model: OLS Adj. R-squared: 0.592 # SEX: 1=Female, 0=Male
Method: Least Squares F-statistic: 320.4 # WAGE: Wage (dollars per hour)
Date: Tue, 03 Oct 2017 Prob (F-statistic): 1.89e-86 data = pandas.read_csv('wages.txt', skiprows=27, skipfooter=6, sep=None,
Time: 16:00:32 Log-Likelihood: -1537.7 header=None, names=['education', 'gender', 'wage'],
No. Observations: 441 AIC: 3081. usecols=[0, 2, 5],
Df Residuals: 438 BIC: 3094. )
Df Model: 2
Covariance Type: nonrobust # Convert genders to strings (this is particulary useful so that the
============================================================================== # statsmodels formulas detects that gender is a categorical variable)
coef std err t P>|t| [95.0% Conf. Int.] import numpy as np
------------------------------------------------------------------------------ data['gender'] = np.choose(data.gender, ['male', 'female'])
Intercept -4.5052 0.378 -11.924 0.000 -5.248 -3.763
x 3.1173 0.125 24.979 0.000 2.872 3.363 # Log-transform the wages, because they typically are increased with
y -0.5109 0.125 -4.094 0.000 -0.756 -0.266 # multiplicative factors
============================================================================== data['wage'] = np.log10(data['wage'])
Omnibus: 0.260 Durbin-Watson: 2.057
Prob(Omnibus): 0.878 Jarque-Bera (JB): 0.204 simple plotting
Skew: -0.052 Prob(JB): 0.903
Kurtosis: 3.015 Cond. No. 3.03 import seaborn
==============================================================================
# Plot 2 linear fits for male and female.
Warnings: seaborn.lmplot(y='wage', x='education', hue='gender', data=data)
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.

Retrieving manually the parameter estimates:

[-4.50523303 3.11734237 -0.51091248]

ANOVA results
df sum_sq mean_sq F PR(>F)
x 1.0 39284.301219 39284.301219 623.962799 2.888238e-86
y 1.0 1055.220089 1055.220089 16.760336 5.050899e-05
Residual 438.0 27576.201607 62.959364 NaN NaN

Total running time of the script: ( 0 minutes 0.047 seconds)

Download Python source code: plot_regression_3d.py
Download Jupyter notebook: plot_regression_3d.ipynb
Generated by Sphinx-Gallery

15.6.6 Test for an education/gender interaction in wages

Wages depend mostly on education. Here we investigate how this dependence is related to gender: not only
does gender create an offset in wages, it also seems that wages increase more with education for males than
females.
Does our data support this last hypothesis? We will test this using statsmodels’ formulas (http://statsmodels.
sourceforge.net/stable/example_formulas.html).
Load and massage the data

import pandas

import urllib
import os

if not os.path.exists('wages.txt'): statistical analysis

# Download the file if it is not present
urllib.urlretrieve('http://lib.stat.cmu.edu/datasets/CPS_85_Wages',

15.6. Full code for the figures 490 15.6. Full code for the figures 491
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

import statsmodels.formula.api as sm education 0.0415 0.005 7.647 0.000 0.031 0.052

education:gender[T.male] -0.0134 0.007 -1.919 0.056 -0.027 0.000
# Note that this model is not the plot displayed above: it is one ==============================================================================
# joined model for male and female, not separate models for male and Omnibus: 4.838 Durbin-Watson: 1.825
# female. The reason is that a single model enables statistical testing Prob(Omnibus): 0.089 Jarque-Bera (JB): 5.000
result = sm.ols(formula='wage ~ education + gender', data=data).fit() Skew: -0.156 Prob(JB): 0.0821
print(result.summary()) Kurtosis: 3.356 Cond. No. 194.
==============================================================================
Out:
Warnings:
OLS Regression Results [1] Standard Errors assume that the covariance matrix of the errors is correctly specified.
==============================================================================
Dep. Variable: wage R-squared: 0.193 Looking at the p-value of the interaction of gender and education, the data does not support the hypothesis
Model: OLS Adj. R-squared: 0.190 that education benefits males more than female (p-value > 0.05).
Method: Least Squares F-statistic: 63.42
Date: Tue, 03 Oct 2017 Prob (F-statistic): 2.01e-25 import matplotlib.pyplot as plt
Time: 16:00:42 Log-Likelihood: 86.654 plt.show()
No. Observations: 534 AIC: -167.3
Df Residuals: 531 BIC: -154.5 Total running time of the script: ( 0 minutes 0.393 seconds)
Df Model: 2
Covariance Type: nonrobust Download Python source code: plot_wage_education_gender.py
==================================================================================
coef std err t P>|t| [95.0% Conf. Int.]
Download Jupyter notebook: plot_wage_education_gender.ipynb
---------------------------------------------------------------------------------- Generated by Sphinx-Gallery
Intercept 0.4053 0.046 8.732 0.000 0.314 0.496
gender[T.male] 0.1008 0.018 5.625 0.000 0.066 0.136
education 0.0334 0.003 9.768 0.000 0.027 0.040 15.6.7 Visualizing factors influencing wages
==============================================================================
Omnibus: 4.675 Durbin-Watson: 1.792
This example uses seaborn to quickly plot various factors relating wages, experience and eduction.
Prob(Omnibus): 0.097 Jarque-Bera (JB): 4.876
Skew: -0.147 Prob(JB): 0.0873 Seaborn (http://stanford.edu/~mwaskom/software/seaborn/) is a library that combines visualization and sta-
Kurtosis: 3.365 Cond. No. 69.7 tistical fits to show trends in data.
==============================================================================
Note that importing seaborn changes the matplotlib style to have an “excel-like” feeling. This changes affect
Warnings: other matplotlib figures. To restore defaults once this example is run, we would need to call plt.rcdefaults().
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.
# Standard library imports
import urllib
The plots above highlight that there is not only a different offset in wage but also a different slope
import os
We need to model this using an interaction
import matplotlib.pyplot as plt
result = sm.ols(formula='wage ~ education + gender + education * gender',
data=data).fit() Load the data
print(result.summary())
import pandas
Out:
if not os.path.exists('wages.txt'):
OLS Regression Results # Download the file if it is not present
============================================================================== urllib.urlretrieve('http://lib.stat.cmu.edu/datasets/CPS_85_Wages',
Dep. Variable: wage R-squared: 0.198 'wages.txt')
Model: OLS Adj. R-squared: 0.194
Method: Least Squares F-statistic: 43.72 # Give names to the columns
Date: Tue, 03 Oct 2017 Prob (F-statistic): 2.94e-25 names = [
Time: 16:00:42 Log-Likelihood: 88.503 'EDUCATION: Number of years of education',
No. Observations: 534 AIC: -169.0 'SOUTH: 1=Person lives in South, 0=Person lives elsewhere',
Df Residuals: 530 BIC: -151.9 'SEX: 1=Female, 0=Male',
Df Model: 3 'EXPERIENCE: Number of years of work experience',
Covariance Type: nonrobust 'UNION: 1=Union member, 0=Not union member',
============================================================================================ 'WAGE: Wage (dollars per hour)',
coef std err t P>|t| [95.0% Conf. Int.] 'AGE: years',
-------------------------------------------------------------------------------------------- 'RACE: 1=Other, 2=Hispanic, 3=White',
Intercept 0.2998 0.072 4.173 0.000 0.159 0.441 'OCCUPATION: 1=Management, 2=Sales, 3=Clerical, 4=Service, 5=Professional, 6=Other',
gender[T.male] 0.2750 0.093 2.972 0.003 0.093 0.457 'SECTOR: 0=Other, 1=Manufacturing, 2=Construction',

15.6. Full code for the figures 492 15.6. Full code for the figures 493
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

'MARR: 0=Unmarried, 1=Married',

]

short_names = [n.split(':')[0] for n in names]

data = pandas.read_csv('wages.txt', skiprows=27, skipfooter=6, sep=None,

header=None)
data.columns = short_names

# Log-transform the wages, because they typically are increased with

# multiplicative factors
import numpy as np
data['WAGE'] = np.log10(data['WAGE'])

Plot scatter matrices highlighting different aspects

import seaborn
seaborn.pairplot(data, vars=['WAGE', 'AGE', 'EDUCATION'],
kind='reg')

seaborn.pairplot(data, vars=['WAGE', 'AGE', 'EDUCATION'],

kind='reg', hue='SEX')
plt.suptitle('Effect of gender: 1=Female, 0=Male') •

seaborn.pairplot(data, vars=['WAGE', 'AGE', 'EDUCATION'],

kind='reg', hue='RACE')
plt.suptitle('Effect of race: 1=Other, 2=Hispanic, 3=White')

seaborn.pairplot(data, vars=['WAGE', 'AGE', 'EDUCATION'],

kind='reg', hue='UNION')
plt.suptitle('Effect of union: 1=Union member, 0=Not union member')

15.6. Full code for the figures 494 15.6. Full code for the figures 495
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Total running time of the script: ( 0 minutes 9.437 seconds)

Download Python source code: plot_wage_data.py
Download Jupyter notebook: plot_wage_data.ipynb
Generated by Sphinx-Gallery

15.6.8 Air fares before and after 9/11

This is a business-intelligence (BI) like application.

What is interesting here is that we may want to study fares as a function of the year, paired accordingly to the
trips, or forgetting the year, only as a function of the trip endpoints.
Using statsmodels’ linear models, we find that both with an OLS (ordinary least square) and a robust fit, the
intercept and the slope are significantly non-zero: the air fares have decreased between 2000 and 2001, and
their dependence on distance travelled has also decreased

# Standard library imports

import urllib
import os

•
Load the data
Plot a simple regression
import pandas
seaborn.lmplot(y='WAGE', x='EDUCATION', data=data)
if not os.path.exists('airfares.txt'):
plt.show() # Download the file if it is not present
urllib.urlretrieve(
'http://www.stat.ufl.edu/~winner/data/airq4.dat',
'airfares.txt')

# As a seperator, ' +' is a regular expression that means 'one of more

# space'
data = pandas.read_csv('airfares.txt', sep=' +', header=0,
names=['city1', 'city2', 'pop1', 'pop2',
'dist', 'fare_2000', 'nb_passengers_2000',
'fare_2001', 'nb_passengers_2001'])

# we log-transform the number of passengers

import numpy as np
data['nb_passengers_2000'] = np.log10(data['nb_passengers_2000'])
data['nb_passengers_2001'] = np.log10(data['nb_passengers_2001'])

Make a dataframe whith the year as an attribute, instead of separate columns

# This involves a small danse in which we separate the dataframes in 2,

# one for year 2000, and one for 2001, before concatenating again.

# Make an index of each flight

data_flat = data.reset_index()

data_2000 = data_flat[['city1', 'city2', 'pop1', 'pop2',

'dist', 'fare_2000', 'nb_passengers_2000']]
# Rename the columns
data_2000.columns = ['city1', 'city2', 'pop1', 'pop2', 'dist', 'fare',
'nb_passengers']
# Add a column with the year
data_2000['year'] = 2000

data_2001 = data_flat[['city1', 'city2', 'pop1', 'pop2',

'dist', 'fare_2001', 'nb_passengers_2001']]

15.6. Full code for the figures 496 15.6. Full code for the figures 497
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

# Rename the columns

data_2001.columns = ['city1', 'city2', 'pop1', 'pop2', 'dist', 'fare',
'nb_passengers']
# Add a column with the year
data_2001['year'] = 2001

data_flat = pandas.concat([data_2000, data_2001])

Plot scatter matrices highlighting different aspects

import seaborn
seaborn.pairplot(data_flat, vars=['fare', 'dist', 'nb_passengers'],
kind='reg', markers='.')

# A second plot, to show the effect of the year (ie the 9/11 effect)
seaborn.pairplot(data_flat, vars=['fare', 'dist', 'nb_passengers'],
kind='reg', hue='year', markers='.')

•
Plot the difference in fare

import matplotlib.pyplot as plt

plt.figure(figsize=(5, 2))
seaborn.boxplot(data.fare_2001 - data.fare_2000)
plt.title('Fare: 2001 - 2000')
plt.subplots_adjust()

plt.figure(figsize=(5, 2))
seaborn.boxplot(data.nb_passengers_2001 - data.nb_passengers_2000)
plt.title('NB passengers: 2001 - 2000')
plt.subplots_adjust()

•
Statistical testing: dependence of fare on distance and number of passengers

import statsmodels.formula.api as sm

result = sm.ols(formula='fare ~ 1 + dist + nb_passengers', data=data_flat).fit()

print(result.summary())

# Using a robust fit

result = sm.rlm(formula='fare ~ 1 + dist + nb_passengers', data=data_flat).fit()
print(result.summary())

Out:

15.6. Full code for the figures 498 15.6. Full code for the figures 499
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

OLS Regression Results

==============================================================================
Dep. Variable: fare R-squared: 0.275
Model: OLS Adj. R-squared: 0.275
Method: Least Squares F-statistic: 1585.
Date: Tue, 03 Oct 2017 Prob (F-statistic): 0.00
Time: 16:00:28 Log-Likelihood: -45532.
No. Observations: 8352 AIC: 9.107e+04
Df Residuals: 8349 BIC: 9.109e+04
Df Model: 2
Covariance Type: nonrobust
=================================================================================
coef std err t P>|t| [95.0% Conf. Int.]
---------------------------------------------------------------------------------
Intercept 211.2428 2.466 85.669 0.000 206.409 216.076
dist 0.0484 0.001 48.149 0.000 0.046 0.050
nb_passengers -32.8925 1.127 -29.191 0.000 -35.101 -30.684
==============================================================================
Omnibus: 604.051 Durbin-Watson: 1.446
Prob(Omnibus): 0.000 Jarque-Bera (JB): 740.733
Skew: 0.710 Prob(JB): 1.42e-161
Kurtosis: 3.338 Cond. No. 5.23e+03
==============================================================================

Warnings:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.
[2] The condition number is large, 5.23e+03. This might indicate that there are
strong multicollinearity or other numerical problems.
Robust linear Model Regression Results
==============================================================================
Dep. Variable: fare No. Observations: 8352
Model: RLM Df Residuals: 8349
Method: IRLS Df Model: 2
Norm: HuberT Out:
Scale Est.: mad
Cov Type: H1 OLS Regression Results
Date: Tue, 03 Oct 2017 ==============================================================================
Time: 16:00:28 Dep. Variable: fare_2001 R-squared: 0.159
No. Iterations: 12 Model: OLS Adj. R-squared: 0.159
================================================================================= Method: Least Squares F-statistic: 791.7
coef std err z P>|z| [95.0% Conf. Int.] Date: Tue, 03 Oct 2017 Prob (F-statistic): 1.20e-159
--------------------------------------------------------------------------------- Time: 16:00:28 Log-Likelihood: -22640.
Intercept 215.0848 2.448 87.856 0.000 210.287 219.883 No. Observations: 4176 AIC: 4.528e+04
dist 0.0460 0.001 46.166 0.000 0.044 0.048 Df Residuals: 4174 BIC: 4.530e+04
nb_passengers -35.2686 1.119 -31.526 0.000 -37.461 -33.076 Df Model: 1
================================================================================= Covariance Type: nonrobust
==============================================================================
If the model instance has been used for another fit with different fit coef std err t P>|t| [95.0% Conf. Int.]
parameters, then the fit options might not be the correct ones anymore . ------------------------------------------------------------------------------
Intercept 148.0279 1.673 88.480 0.000 144.748 151.308
Statistical testing: regression of fare on distance: 2001/2000 difference dist 0.0388 0.001 28.136 0.000 0.036 0.041
==============================================================================
result = sm.ols(formula='fare_2001 - fare_2000 ~ 1 + dist', data=data).fit() Omnibus: 136.558 Durbin-Watson: 1.544
print(result.summary()) Prob(Omnibus): 0.000 Jarque-Bera (JB): 149.624
Skew: 0.462 Prob(JB): 3.23e-33
# Plot the corresponding regression Kurtosis: 2.920 Cond. No. 2.40e+03
data['fare_difference'] = data['fare_2001'] - data['fare_2000'] ==============================================================================
seaborn.lmplot(x='dist', y='fare_difference', data=data)
Warnings:
plt.show() [1] Standard Errors assume that the covariance matrix of the errors is correctly specified.
[2] The condition number is large, 2.4e+03. This might indicate that there are
strong multicollinearity or other numerical problems.

15.6. Full code for the figures 500 15.6. Full code for the figures 501
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Total running time of the script: ( 0 minutes 8.958 seconds)

strong multicollinearity or other numerical problems.
Download Python source code: plot_airfare.py <F test: F=array([[ 0.68319608]]), p=0.4140878441244694, df_denom=35, df_num=1>

Download Jupyter notebook: plot_airfare.ipynb Here we plot a scatter matrix to get intuitions on our results. This goes beyond what was asked in the exercise
Generated by Sphinx-Gallery
# This plotting is useful to get an intuitions on the relationships between
# our different variables

15.7 Solutions to this chapter’s exercises from pandas.tools import plotting

import matplotlib.pyplot as plt

15.7.1 Relating Gender and IQ # Fill in the missing values for Height for plotting
data['Height'].fillna(method='pad', inplace=True)
Going back to the brain size + IQ data, test if the VIQ of male and female are different after removing the effect
of brain size, height and weight. # The parameter 'c' is passed to plt.scatter and will control the color
# The same holds for parameters 'marker', 'alpha' and 'cmap', that
Notice that here ‘Gender’ is a categorical value. As it is a non-float data type, statsmodels is able to automati- # control respectively the type of marker used, their transparency and
cally infer this. # the colormap
plotting.scatter_matrix(data[['VIQ', 'MRI_Count', 'Height']],
import pandas c=(data['Gender'] == 'Female'), marker='o',
from statsmodels.formula.api import ols alpha=1, cmap='winter')
data = pandas.read_csv('../brain_size.csv', sep=';', na_values='.') fig = plt.gcf()
fig.suptitle("blue: male, green: female", size=13)
model = ols('VIQ ~ Gender + MRI_Count + Height', data).fit()
print(model.summary()) plt.show()
# Here, we don't need to define a contrast, as we are testing a single
# coefficient of our model, and not a combination of coefficients.
# However, defining a contrast, which would then be a 'unit contrast',
# will give us the same results
print(model.f_test([0, 1, 0, 0]))

Out:

OLS Regression Results

==============================================================================
Dep. Variable: VIQ R-squared: 0.246
Model: OLS Adj. R-squared: 0.181
Method: Least Squares F-statistic: 3.809
Date: Tue, 03 Oct 2017 Prob (F-statistic): 0.0184
Time: 16:00:42 Log-Likelihood: -172.34
No. Observations: 39 AIC: 352.7
Df Residuals: 35 BIC: 359.3
Df Model: 3
Covariance Type: nonrobust
==================================================================================
coef std err t P>|t| [95.0% Conf. Int.]
----------------------------------------------------------------------------------
Intercept 166.6258 88.824 1.876 0.069 -13.696 346.948
Gender[T.Male] 8.8524 10.710 0.827 0.414 -12.890 30.595
MRI_Count 0.0002 6.46e-05 2.615 0.013 3.78e-05 0.000
Height -3.0837 1.276 -2.417 0.021 -5.674 -0.494
==============================================================================
Omnibus: 7.373 Durbin-Watson: 2.109
Prob(Omnibus): 0.025 Jarque-Bera (JB): 2.252
Skew: 0.005 Prob(JB): 0.324
Kurtosis: 1.823 Cond. No. 2.40e+07
==============================================================================
Total running time of the script: ( 0 minutes 0.537 seconds)
Warnings:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified. Download Python source code: plot_brain_size.py
[2] The condition number is large, 2.4e+07. This might indicate that there are
Download Jupyter notebook: plot_brain_size.ipynb

15.7. Solutions to this chapter’s exercises 502 15.7. Solutions to this chapter’s exercises 503
 Scipy lecture notes, Edition 2017.1

Generated by Sphinx-Gallery
Download all examples in Python source code: auto_examples_python.zip
Download all examples in Jupyter notebooks: auto_examples_jupyter.zip
Generated by Sphinx-Gallery

CHAPTER 16
Sympy : Symbolic Mathematics in
Python

Author: Fabian Pedregosa

Objectives

1. Evaluate expressions with arbitrary precision.

2. Perform algebraic manipulations on symbolic expressions.
3. Perform basic calculus tasks (limits, differentiation and integration) with symbolic expressions.
4. Solve polynomial and transcendental equations.
5. Solve some differential equations.

What is SymPy? SymPy is a Python library for symbolic mathematics. It aims to be an alternative to systems
such as Mathematica or Maple while keeping the code as simple as possible and easily extensible. SymPy is
written entirely in Python and does not require any external libraries.
Sympy documentation and packages for installation can be found on http://www.sympy.org/

Chapters contents

• First Steps with SymPy

– Using SymPy as a calculator
– Symbols

15.7. Solutions to this chapter’s exercises 504 505

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

• Algebraic manipulations
Exercises
– Expand
p
– Simplify 1. Calculate 2 with 100 decimals.

• Calculus 2. Calculate 1/2 + 1/3 in rational arithmetic.

– Limits
– Differentiation
16.1.2 Symbols
– Series expansion
In contrast to other Computer Algebra Systems, in SymPy you have to declare symbolic variables explicitly:
– Integration
• Equation solving >>> x = sym.Symbol('x')
>>> y = sym.Symbol('y')
• Linear Algebra
– Matrices Then you can manipulate them:

– Differential Equations >>> x + y + x - y

2*x

>>> (x + y) ** 2
(x + y)**2
16.1 First Steps with SymPy
Symbols can now be manipulated using some of python operators: +, -`, ``*, ** (arithmetic), &, |, ~ , >>, <<
16.1.1 Using SymPy as a calculator (boolean).

SymPy defines three numerical types: Real, Rational and Integer. Printing
The Rational class represents a rational number as a pair of two Integers: the numerator and the denominator,
Sympy allows for control of the display of the output. From here we use the following setting for printing:
so Rational(1, 2) represents 1/2, Rational(5, 2) 5/2 and so on:
>>> sym.init_printing(use_unicode=False, wrap_line=True)
>>> import sympy as sym
>>> a = sym.Rational(1, 2)

>>> a
1/2 16.2 Algebraic manipulations
>>> a*2
1 SymPy is capable of performing powerful algebraic manipulations. We’ll take a look into some of the most
frequently used: expand and simplify.
SymPy uses mpmath in the background, which makes it possible to perform computations using arbitrary-
precision arithmetic. That way, some special constants, like e, pi , oo (Infinity), are treated as symbols and can
be evaluated with arbitrary precision: 16.2.1 Expand
>>> sym.pi**2 Use this to expand an algebraic expression. It will try to denest powers and multiplications:
pi**2
>>> sym.expand((x + y) ** 3)
>>> sym.pi.evalf() 3 2 2 3
3.14159265358979 x + 3*x *y + 3*x*y + y
>>> 3 * x * y ** 2 + 3 * y * x ** 2 + x ** 3 + y ** 3
>>> (sym.pi + sym.exp(1)).evalf() 3 2 2 3
5.85987448204884 x + 3*x *y + 3*x*y + y

as you see, evalf evaluates the expression to a floating-point number. Further options can be given in form on keywords:

There is also a class representing mathematical infinity, called oo: >>> sym.expand(x + y, complex=True)
re(x) + re(y) + I*im(x) + I*im(y)
>>> sym.oo > 99999 >>> sym.I * sym.im(x) + sym.I * sym.im(y) + sym.re(x) + sym.re(y)
True re(x) + re(y) + I*im(x) + I*im(y)
>>> sym.oo + 1
oo >>> sym.expand(sym.cos(x + y), trig=True)

16.1. First Steps with SymPy 506 16.2. Algebraic manipulations 507
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

You can check, that it is correct by:

-sin(x)*sin(y) + cos(x)*cos(y)
>>> sym.cos(x) * sym.cos(y) - sym.sin(x) * sym.sin(y) >>> sym.limit((sym.tan(x + y) - sym.tan(x)) / y, y, 0)
-sin(x)*sin(y) + cos(x)*cos(y) 2
tan (x) + 1

16.2.2 Simplify Higher derivatives can be calculated using the diff(func, var, n) method:

>>> sym.diff(sym.sin(2 * x), x, 1)

Use simplify if you would like to transform an expression into a simpler form:
2*cos(2*x)
>>> sym.simplify((x + x * y) / x)
y + 1 >>> sym.diff(sym.sin(2 * x), x, 2)
-4*sin(2*x)
Simplification is a somewhat vague term, and more precises alternatives to simplify exists: powsimp (simplifi-
>>> sym.diff(sym.sin(2 * x), x, 3)
cation of exponents), trigsimp (for trigonometric expressions) , logcombine, radsimp, together. -8*cos(2*x)

Exercises
16.3.3 Series expansion
1. Calculate the expanded form of (x + y)6 .
2. Simplify the trigonometric expression sin(x)/ cos(x) SymPy also knows how to compute the Taylor series of an expression at a point. Use series(expr, var):

>>> sym.series(sym.cos(x), x)
2 4
x x / 6\
16.3 Calculus 1 - -- + -- + O\x /
2 24
>>> sym.series(1/sym.cos(x), x)
16.3.1 Limits 2 4
x 5*x / 6\
Limits are easy to use in SymPy, they follow the syntax limit(function, variable, point), so to compute 1 + -- + ---- + O\x /
the limit of f (x) as x → 0, you would issue limit(f, x, 0): 2 24

>>> sym.limit(sym.sin(x) / x, x, 0)
1 Exercises

you can also calculate the limit at infinity: 1. Calculate limx→0 sin(x)/x
>>> sym.limit(x, x, sym.oo) 2. Calculate the derivative of l og (x) for x.
oo

>>> sym.limit(1 / x, x, sym.oo)

0 16.3.4 Integration
>>> sym.limit(x ** x, x, 0) SymPy has support for indefinite and definite integration of transcendental elementary and special functions
1
via integrate() facility, which uses the powerful extended Risch-Norman algorithm and some heuristics and
pattern matching. You can integrate elementary functions:

16.3.2 Differentiation >>> sym.integrate(6 * x ** 5, x)

6
x
You can differentiate any SymPy expression using diff(func, var). Examples:
>>> sym.integrate(sym.sin(x), x)
>>> sym.diff(sym.sin(x), x) -cos(x)
cos(x) >>> sym.integrate(sym.log(x), x)
>>> sym.diff(sym.sin(2 * x), x) x*log(x) - x
2*cos(2*x) >>> sym.integrate(2 * x + sym.sinh(x), x)
2
>>> sym.diff(sym.tan(x), x) x + cosh(x)
2
tan (x) + 1 Also special functions are handled easily:

16.3. Calculus 508 16.3. Calculus 509

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

SymPy is also able to solve boolean equations, that is, to decide if a certain boolean expression is satisfiable or
>>> sym.integrate(sym.exp(-x ** 2) * sym.erf(x), x)
____ 2 not. For this, we use the function satisfiable:
\/ pi *erf (x)
>>> sym.satisfiable(x & y)
--------------
{x: True, y: True}
4

This tells us that (x & y) is True whenever x and y are both True. If an expression cannot be true, i.e. no
It is possible to compute definite integral:
values of its arguments can make the expression True, it will return False:
>>> sym.integrate(x**3, (x, -1, 1))
>>> sym.satisfiable(x & ~x)
0
False
>>> sym.integrate(sym.sin(x), (x, 0, sym.pi / 2))
1
>>> sym.integrate(sym.cos(x), (x, -sym.pi / 2, sym.pi / 2))
2 Exercises

Also improper integrals are supported as well: 1. Solve the system of equations x + y = 2, 2 · x + y = 0

>>> sym.integrate(sym.exp(-x), (x, 0, sym.oo)) 2. Are there boolean values x, y that make (~x | y) & (~y | x) true?
1
>>> sym.integrate(sym.exp(-x ** 2), (x, -sym.oo, sym.oo))
____
\/ pi 16.5 Linear Algebra

16.5.1 Matrices
16.4 Equation solving
Matrices are created as instances from the Matrix class:
SymPy is able to solve algebraic equations, in one and several variables using solveset(): >>> sym.Matrix([[1, 0], [0, 1]])
[1 0]
>>> sym.solveset(x ** 4 - 1, x)
[ ]
{-1, 1, -I, I}
[0 1]

As you can see it takes as first argument an expression that is supposed to be equaled to 0. It also has (limited)
unlike a NumPy array, you can also put Symbols in it:
support for transcendental equations:
>>> x, y = sym.symbols('x, y')
>>> sym.solveset(sym.exp(x) + 1, x)
>>> A = sym.Matrix([[1, x], [y, 1]])
{I*(2*n*pi + pi) | n in Integers()}
>>> A
[1 x]
[ ]
Systems of linear equations [y 1]

Sympy is able to solve a large part of polynomial equations, and is also capable of solving multiple equa- >>> A**2
tions with respect to multiple variables giving a tuple as second argument. To do this you use the solve() [x*y + 1 2*x ]
command: [ ]
[ 2*y x*y + 1]
>>> solution = sym.solve((x + 5 * y - 2, -3 * x + 6 * y - 15), (x, y))
>>> solution[x], solution[y]

(-3, 1) 16.5.2 Differential Equations

Another alternative in the case of polynomial equations is factor. factor returns the polynomial factorized into SymPy is capable of solving (some) Ordinary Differential. To solve differential equations, use dsolve. First,
irreducible terms, and is capable of computing the factorization over various domains: create an undefined function by passing cls=Function to the symbols function:

>>> f = x ** 4 - 3 * x ** 2 + 1 >>> f, g = sym.symbols('f g', cls=sym.Function)

>>> sym.factor(f)
/ 2 \ / 2 \ f and g are now undefined functions. We can call f(x), and it will represent an unknown function:
\x - x - 1/*\x + x - 1/
>>> f(x)
>>> sym.factor(f, modulus=5) f(x)
2 2
(x - 2) *(x + 2) >>> f(x).diff(x, x) + f(x)

16.4. Equation solving 510 16.5. Linear Algebra 511

 Scipy lecture notes, Edition 2017.1

2
d
f(x) + ---(f(x))
2
dx

17
>>> sym.dsolve(f(x).diff(x, x) + f(x), f(x))
f(x) = C1*sin(x) + C2*cos(x)

Keyword arguments can be given to this function in order to help if find the best possible resolution system.
For example, if you know that it is a separable equations, you can use keyword hint='separable' to force
dsolve to resolve it as a separable equation:
CHAPTER
>>> sym.dsolve(sym.sin(x) * sym.cos(f(x)) + sym.cos(x) * sym.sin(f(x)) * f(x).diff(x), f(x),␣
,→hint='separable')
/ _________________\ / ________________
| / C1 | | / C1
[f(x) = - asin| / ----------- + 1 | + pi, f(x) = asin| / ----------- + 1
| / 2 | | / 2
\\/ sin (x) - 1 / \\/ sin (x) - 1

Scikit-image: image processing

_\ / _________________\ / ______________
| | / C1 | | / C1
| + pi, f(x) = -asin| / ----------- + 1 |, f(x) = asin| / ----------- +
| | / 2 | | / 2
/ \\/ sin (x) - 1 / \\/ sin (x) - 1

___\
|
1 |]
|
/ Author: Emmanuelle Gouillart
scikit-image is a Python package dedicated to image processing, and using natively NumPy arrays as image
objects. This chapter describes how to use scikit-image on various image processing tasks, and insists on
Exercises the link with other scientific Python modules such as NumPy and SciPy.

1. Solve the Bernoulli differential equation See also:

For basic image manipulation, such as image cropping or simple filtering, a large number of simple operations
can be realized with NumPy and SciPy only. See Image manipulation and processing using Numpy and Scipy.
d f (x)
x + f (x) − f (x)2 = 0 Note that you should be familiar with the content of the previous chapter before reading the current one, as
x
basic operations such as masking and labeling are a prerequisite.
2. Solve the same equation using hint='Bernoulli'. What do you observe ?

Chapters contents

• Introduction and concepts

– scikit-image and the SciPy ecosystem
– What’s to be found in scikit-image
• Input/output, data types and colorspaces
– Data types
– Colorspaces
• Image preprocessing / enhancement
– Local filters
– Non-local filters

16.5. Linear Algebra 512 513

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

– Mathematical morphology 17.1.1 scikit-image and the SciPy ecosystem

• Image segmentation Recent versions of scikit-image is packaged in most Scientific Python distributions, such as Anaconda or
– Binary segmentation: foreground + background Enthought Canopy. It is also packaged for Ubuntu/Debian.

– Marker based methods >>> import skimage

>>> from skimage import data # most functions are in subpackages
• Measuring regions’ properties
• Data visualization and interaction Most scikit-image functions take NumPy ndarrays as arguments

• Feature extraction for computer vision >>> camera = data.camera()

>>> camera.dtype
• Full code examples dtype('uint8')
• Examples for the scikit-image chapter >>> camera.shape
(512, 512)
>>> from skimage import filters
>>> filtered_camera = filters.gaussian(camera, 1)
>>> type(filtered_camera)
17.1 Introduction and concepts <type 'numpy.ndarray'>

Images are NumPy’s arrays np.ndarray Other Python packages are available for image processing and work with NumPy arrays:

image np.ndarray • scipy.ndimage : for nd-arrays. Basic filtering, mathematical morphology, regions properties

pixels array values: a[2, 3] • Mahotas

channels array dimensions Also, powerful image processing libraries have Python bindings:

image encoding dtype (np.uint8, np.uint16, np.float) • OpenCV (computer vision)

filters functions (numpy, skimage, scipy) • ITK (3D images and registration)
• and many others
>>> import numpy as np
>>> check = np.zeros((9, 9)) (but they are less Pythonic and NumPy friendly, to a variable extent).
>>> check[::2, 1::2] = 1
>>> check[1::2, ::2] = 1
>>> import matplotlib.pyplot as plt 17.1.2 What’s to be found in scikit-image
>>> plt.imshow(check, cmap='gray', interpolation='nearest')
• Website: http://scikit-image.org/
• Gallery of examples: http://scikit-image.org/docs/stable/auto_examples/
Different kinds of functions, from boilerplate utility functions to high-level recent algorithms.
• Filters: functions transforming images into other images.
– NumPy machinery
– Common filtering algorithms
• Data reduction functions: computation of image histogram, position of local maxima, of corners, etc.
• Other actions: I/O, visualization, etc.

17.2 Input/output, data types and colorspaces

I/O: skimage.io

>>> from skimage import io

Reading from files: skimage.io.imread()

>>> import os
>>> filename = os.path.join(skimage.data_dir, 'camera.png')
>>> camera = io.imread(filename)

17.1. Introduction and concepts 514 17.2. Input/output, data types and colorspaces 515
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

>>> from skimage import img_as_float

>>> camera_float = img_as_float(camera)
>>> camera.max(), camera_float.max()
(255, 1.0)

Some image processing routines need to work with float arrays, and may hence output an array with a different
type and the data range from the input array

>>> from skimage import filters

>>> camera_sobel = filters.sobel(camera)
>>> camera_sobel.max()
0.591502...

Utility functions are provided in skimage to convert both the dtype and the data range, following skimage’s
conventions: util.img_as_float, util.img_as_ubyte, etc.
See the user guide for more details.

Works with all data formats supported by the Python 17.2.2 Colorspaces
Imaging Library (or any other I/O plugin provided to imread with the plugin keyword argument).
Also works with URL image paths: Color images are of shape (N, M, 3) or (N, M, 4) (when an alpha channel encodes transparency)

>>> logo = io.imread('http://scikit-image.org/_static/img/logo.png') >>> face = scipy.misc.face()

>>> face.shape
(768, 1024, 3)
Saving to files:

>>> io.imsave('local_logo.png', logo) Routines converting between different colorspaces (RGB, HSV, LAB etc.) are available in skimage.color :
color.rgb2hsv, color.lab2rgb, etc. Check the docstring for the expected dtype (and data range) of input
(imsave also uses an external plugin such as PIL) images.

3D images
17.2.1 Data types
Most functions of skimage can take 3D images as input arguments. Check the docstring to know if a func-
tion can be used on 3D images (for example MRI or CT images).

Exercise

Open a color image on your disk as a NumPy array.

Find a skimage function computing the histogram of an image and plot the histogram of each
color channel
Image ndarrays can be represented either by integers
(signed or unsigned) or floats. Convert the image to grayscale and plot its histogram.
Careful with overflows with integer data types

>>> camera = data.camera()

>>> camera.dtype 17.3 Image preprocessing / enhancement
dtype('uint8')
>>> camera_multiply = 3 * camera
Goals: denoising, feature (edges) extraction, . . .
Different integer sizes are possible: 8-, 16- or 32-bytes, signed or unsigned.
17.3.1 Local filters
Warning: An important (if questionable) skimage convention: float images are supposed to lie in [-1, 1]
(in order to have comparable contrast for all float images) Local filters replace the value of pixels by a function of the values of neighboring pixels. The function can be
linear or non-linear.

17.2. Input/output, data types and colorspaces 516 17.3. Image preprocessing / enhancement 517
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Neighbourhood: square (choose size), disk, or more complicated structuring element. Default structuring element: 4-connectivity of a pixel

>>> from skimage import morphology

>>> morphology.diamond(1)
array([[0, 1, 0],
[1, 1, 1],
[0, 1, 0]], dtype=uint8)

Example : horizontal Sobel filter

>>> text = data.text()

>>> hsobel_text = filters.sobel_h(text)

Uses the following linear kernel for computing horizontal gradients:

1 2 1
0 0 0
-1 -2 -1 Erosion = minimum filter. Replace the value of a pixel by the minimal value covered by the structuring ele-
ment.:

>>> a = np.zeros((7,7), dtype=np.uint8)

>>> a[1:6, 2:5] = 1
>>> a
array([[0, 0, 0, 0, 0, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
17.3.2 Non-local filters [0, 0, 0, 0, 0, 0, 0]], dtype=uint8)
>>> morphology.binary_erosion(a, morphology.diamond(1)).astype(np.uint8)
Non-local filters use a large region of the image (or all the image) to transform the value of one pixel: array([[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
>>> from skimage import exposure [0, 0, 0, 1, 0, 0, 0],
>>> camera = data.camera() [0, 0, 0, 1, 0, 0, 0],
>>> camera_equalized = exposure.equalize_hist(camera) [0, 0, 0, 1, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0]], dtype=uint8)
>>> #Erosion removes objects smaller than the structure
>>> morphology.binary_erosion(a, morphology.diamond(2)).astype(np.uint8)
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, 0, 0],
[0, 0, 0, 0, 0, 0, 0]], dtype=uint8)

Enhances contrast in large almost uniform regions. Dilation: maximum filter:

>>> a = np.zeros((5, 5))

>>> a[2, 2] = 1
17.3.3 Mathematical morphology >>> a
array([[ 0., 0., 0., 0., 0.],
See wikipedia for an introduction on mathematical morphology. [ 0., 0., 0., 0., 0.],
[ 0., 0., 1., 0., 0.],
Probe an image with a simple shape (a structuring element), and modify this image according to how the
[ 0., 0., 0., 0., 0.],
shape locally fits or misses the image.
[ 0., 0., 0., 0., 0.]])

17.3. Image preprocessing / enhancement 518 17.3. Image preprocessing / enhancement 519
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

>>> morphology.binary_dilation(a, morphology.diamond(1)).astype(np.uint8) 17.4 Image segmentation

array([[0, 0, 0, 0, 0],
[0, 0, 1, 0, 0],
Image segmentation is the attribution of different labels to different regions of the image, for example in order
[0, 1, 1, 1, 0],
to extract the pixels of an object of interest.
[0, 0, 1, 0, 0],
[0, 0, 0, 0, 0]], dtype=uint8)
17.4.1 Binary segmentation: foreground + background
Opening: erosion + dilation:

>>> a = np.zeros((5,5), dtype=np.int) Histogram-based method: Otsu thresholding

>>> a[1:4, 1:4] = 1; a[4, 4] = 1
>>> a
array([[0, 0, 0, 0, 0], Tip: The Otsu method is a simple heuristic to find a threshold to separate the foreground from the background.
[0, 1, 1, 1, 0],
[0, 1, 1, 1, 0],
[0, 1, 1, 1, 0],
[0, 0, 0, 0, 1]]) Earlier scikit-image versions
>>> morphology.binary_opening(a, morphology.diamond(1)).astype(np.uint8)
array([[0, 0, 0, 0, 0], skimage.filters is called skimage.filter in earlier versions of scikit-image
[0, 0, 1, 0, 0],
[0, 1, 1, 1, 0],
[0, 0, 1, 0, 0], from skimage import data
[0, 0, 0, 0, 0]], dtype=uint8) from skimage import filters
camera = data.camera()
Opening removes small objects and smoothes corners. val = filters.threshold_otsu(camera)
mask = camera < val

Grayscale mathematical morphology

Mathematical morphology operations are also available for (non-binary) grayscale images (int or float type).
Erosion and dilation correspond to minimum (resp. maximum) filters.

Higher-level mathematical morphology are available: tophat, skeletonization, etc.

See also:
Basic mathematical morphology is also implemented in scipy.ndimage.morphology. The scipy.ndimage
implementation works on arbitrary-dimensional arrays.

Example of filters comparison: image denoising

Labeling connected components of a discrete image
>>> from skimage.morphology import disk
>>> coins = data.coins()
>>> coins_zoom = coins[10:80, 300:370] Tip: Once you have separated foreground objects, it is use to separate them from each other. For this, we can
>>> median_coins = filters.median(coins_zoom, disk(1)) assign a different integer labels to each one.
>>> from skimage import restoration
>>> tv_coins = restoration.denoise_tv_chambolle(coins_zoom, weight=0.1)
>>> gaussian_coins = filters.gaussian(coins, sigma=2) Synthetic data:

>>> n = 20
>>> l = 256
>>> im = np.zeros((l, l))
>>> points = l * np.random.random((2, n ** 2))
>>> im[(points[0]).astype(np.int), (points[1]).astype(np.int)] = 1
>>> im = filters.gaussian(im, sigma=l / (4. * n))
>>> blobs = im > im.mean()

Label all connected components:

17.3. Image preprocessing / enhancement 520 17.4. Image segmentation 521

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

>>> from skimage import measure >>> from skimage import segmentation
>>> all_labels = measure.label(blobs) >>> # Transform markers image so that 0-valued pixels are to
>>> # be labelled, and -1-valued pixels represent background
Label only foreground connected components: >>> markers[~image] = -1
>>> labels_rw = segmentation.random_walker(image, markers)
>>> blobs_labels = measure.label(blobs, background=0)

Postprocessing label images

skimage provides several utility functions that can be used on label images (ie images where
See also: different discrete values identify different regions). Functions names are often self-explaining:
skimage.segmentation.clear_border(), skimage.segmentation.relabel_from_one(), skimage.
scipy.ndimage.find_objects() is useful to return slices on object in an image. morphology.remove_small_objects(), etc.

17.4.2 Marker based methods

Exercise
If you have markers inside a set of regions, you can use these to segment the regions.
• Load the coins image from the data submodule.

Watershed segmentation • Separate the coins from the background by testing several segmentation methods: Otsu thresholding,
adaptive thresholding, and watershed or random walker segmentation.
The Watershed (skimage.morphology.watershed()) is a region-growing approach that fills “basins” in the • If necessary, use a postprocessing function to improve the coins / background segmentation.
image

>>> from skimage.morphology import watershed

>>> from skimage.feature import peak_local_max
>>> 17.5 Measuring regions’ properties
>>> # Generate an initial image with two overlapping circles
>>> x, y = np.indices((80, 80))
>>> x1, y1, x2, y2 = 28, 28, 44, 52 >>> from skimage import measure
>>> r1, r2 = 16, 20
>>> mask_circle1 = (x - x1) ** 2 + (y - y1) ** 2 < r1 ** 2 Example: compute the size and perimeter of the two segmented regions:
>>> mask_circle2 = (x - x2) ** 2 + (y - y2) ** 2 < r2 ** 2
>>> image = np.logical_or(mask_circle1, mask_circle2) >>> properties = measure.regionprops(labels_rw)
>>> # Now we want to separate the two objects in image >>> [prop.area for prop in properties]
>>> # Generate the markers as local maxima of the distance [770, 1168]
>>> # to the background >>> [prop.perimeter for prop in properties]
>>> from scipy import ndimage [100.91..., 126.81...]
>>> distance = ndimage.distance_transform_edt(image)
>>> local_maxi = peak_local_max(distance, indices=False, footprint=np.ones((3, 3)),␣ See also:
,→labels=image)
for some properties, functions are available as well in scipy.ndimage.measurements with a different API (a
>>> markers = morphology.label(local_maxi)
>>> labels_ws = watershed(-distance, markers, mask=image) list is returned).

Exercise (continued)
Random walker segmentation
• Use the binary image of the coins and background from the previous exercise.
The random walker algorithm (skimage.segmentation.random_walker()) is similar to the Watershed, but
• Compute an image of labels for the different coins.
with a more “probabilistic” approach. It is based on the idea of the diffusion of labels in the image:

17.4. Image segmentation 522 17.5. Measuring regions’ properties 523

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

• Compute the size and eccentricity of all coins. >>> new_viewer = viewer.ImageViewer(coins)
>>> from skimage.viewer.plugins import lineprofile
>>> new_viewer += lineprofile.LineProfile()
>>> new_viewer.show()
17.6 Data visualization and interaction

Meaningful visualizations are useful when testing a given processing pipeline.

Some image processing operations:

>>> coins = data.coins()

>>> mask = coins > filters.threshold_otsu(coins)
>>> clean_border = segmentation.clear_border(mask)

Visualize binary result:

>>> plt.figure()
<matplotlib.figure.Figure object at 0x...>
>>> plt.imshow(clean_border, cmap='gray')
<matplotlib.image.AxesImage object at 0x...>

Visualize contour

>>> plt.figure()
<matplotlib.figure.Figure object at 0x...>
>>> plt.imshow(coins, cmap='gray')
<matplotlib.image.AxesImage object at 0x...>
>>> plt.contour(clean_border, [0.5])
<matplotlib.contour.QuadContourSet ...>

Use skimage dedicated utility function:

>>> coins_edges = segmentation.mark_boundaries(coins, clean_border.astype(np.int))

The (ex-
perimental) scikit-image viewer
skimage.viewer = matplotlib-based canvas for displaying images + experimental Qt-based GUI-toolkit

>>> from skimage import viewer

>>> new_viewer = viewer.ImageViewer(coins)
>>> new_viewer.show()

Useful for displaying pixel values. 17.7 Feature extraction for computer vision
For more interaction, plugins can be added to the viewer:
Geometric or textural descriptor can be extracted from images in order to

17.6. Data visualization and interaction 524 17.7. Feature extraction for computer vision 525
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

• classify parts of the image (e.g. sky vs. buildings) This examples show how to create a simple checkerboard.
• match parts of different images (e.g. for object detection)
• and many other applications of Computer Vision

>>> from skimage import feature

Example: detecting corners using Harris detector

from skimage.feature import corner_harris, corner_subpix, corner_peaks

from skimage.transform import warp, AffineTransform

tform = AffineTransform(scale=(1.3, 1.1), rotation=1, shear=0.7,

translation=(210, 50))
image = warp(data.checkerboard(), tform.inverse, output_shape=(350, 350))

coords = corner_peaks(corner_harris(image), min_distance=5)

coords_subpix = corner_subpix(image, coords, window_size=13)

import numpy as np
import matplotlib.pyplot as plt

check = np.zeros((9, 9))

check[::2, 1::2] = 1
(this ex- check[1::2, ::2] = 1
ample is taken from the plot_corner example in scikit-image) plt.matshow(check, cmap='gray')
plt.show()
Points of interest such as corners can then be used to match objects in different images, as described in the
plot_matching example of scikit-image. Total running time of the script: ( 0 minutes 0.033 seconds)
Download Python source code: plot_check.py
17.8 Full code examples Download Jupyter notebook: plot_check.ipynb
Generated by Sphinx-Gallery
17.9 Examples for the scikit-image chapter
17.9.2 Displaying a simple image
17.9.1 Creating an image
Load and display an image
How to create an image with basic NumPy commands : np.zeros, slicing. . .

17.8. Full code examples 526 17.9. Examples for the scikit-image chapter 527
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

import matplotlib.pyplot as plt

from skimage import data

camera = data.camera()
camera_multiply = 3 * camera

import matplotlib.pyplot as plt plt.figure(figsize=(8, 4))

from skimage import data plt.subplot(121)
plt.imshow(camera, cmap='gray', interpolation='nearest')
camera = data.camera() plt.axis('off')
plt.subplot(122)
plt.imshow(camera_multiply, cmap='gray', interpolation='nearest')
plt.figure(figsize=(4, 4)) plt.axis('off')
plt.imshow(camera, cmap='gray', interpolation='nearest')
plt.axis('off') plt.tight_layout()
plt.show()
plt.tight_layout()
plt.show()
Total running time of the script: ( 0 minutes 0.118 seconds)

Total running time of the script: ( 0 minutes 0.062 seconds) Download Python source code: plot_camera_uint.py
Download Python source code: plot_camera.py Download Jupyter notebook: plot_camera_uint.ipynb
Download Jupyter notebook: plot_camera.ipynb Generated by Sphinx-Gallery

Generated by Sphinx-Gallery
17.9.4 Computing horizontal gradients with the Sobel filter
17.9.3 Integers can overflow This example illustrates the use of the horizontal Sobel filter, to compute horizontal gradients.

An illustration of overflow problem arising when working with integers

17.9. Examples for the scikit-image chapter 528 17.9. Examples for the scikit-image chapter 529
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

from skimage import data plt.tight_layout()

from skimage import filters plt.show()
import matplotlib.pyplot as plt
Total running time of the script: ( 0 minutes 0.123 seconds)
text = data.text()
hsobel_text = filters.sobel_h(text) Download Python source code: plot_equalize_hist.py
Download Jupyter notebook: plot_equalize_hist.ipynb
plt.figure(figsize=(12, 3))
Generated by Sphinx-Gallery
plt.subplot(121)
plt.imshow(text, cmap='gray', interpolation='nearest')
plt.axis('off') 17.9.6 Segmentation contours
plt.subplot(122)
plt.imshow(hsobel_text, cmap='spectral', interpolation='nearest')
Visualize segmentation contours on original grayscale image.
plt.axis('off')
plt.tight_layout()
plt.show()

Total running time of the script: ( 0 minutes 0.104 seconds)

Download Python source code: plot_sobel.py
Download Jupyter notebook: plot_sobel.ipynb
Generated by Sphinx-Gallery

17.9.5 Equalizing the histogram of an image

Histogram equalizing makes images have a uniform histogram.

from skimage import data, segmentation

from skimage import filters
import matplotlib.pyplot as plt
import numpy as np

coins = data.coins()
mask = coins > filters.threshold_otsu(coins)
clean_border = segmentation.clear_border(mask).astype(np.int)

coins_edges = segmentation.mark_boundaries(coins, clean_border)

plt.figure(figsize=(8, 3.5))
plt.subplot(121)
plt.imshow(clean_border, cmap='gray')
plt.axis('off')
from skimage import data, exposure
plt.subplot(122)
import matplotlib.pyplot as plt
plt.imshow(coins_edges)
plt.axis('off')
camera = data.camera()
camera_equalized = exposure.equalize_hist(camera)
plt.tight_layout()
plt.show()
plt.figure(figsize=(7, 3))

plt.subplot(121) Total running time of the script: ( 0 minutes 0.146 seconds)

plt.imshow(camera, cmap='gray', interpolation='nearest') Download Python source code: plot_boundaries.py
plt.axis('off')
plt.subplot(122) Download Jupyter notebook: plot_boundaries.ipynb
plt.imshow(camera_equalized, cmap='gray', interpolation='nearest')
plt.axis('off') Generated by Sphinx-Gallery

17.9. Examples for the scikit-image chapter 530 17.9. Examples for the scikit-image chapter 531
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

17.9.7 Otsu thresholding

This example illustrates automatic Otsu thresholding.

from skimage import measure

from skimage import filters
import matplotlib.pyplot as plt
import numpy as np

n = 12
import matplotlib.pyplot as plt l = 256
from skimage import data np.random.seed(1)
from skimage import filters im = np.zeros((l, l))
from skimage import exposure points = l * np.random.random((2, n ** 2))
im[(points[0]).astype(np.int), (points[1]).astype(np.int)] = 1
camera = data.camera() im = filters.gaussian_filter(im, sigma= l / (4. * n))
val = filters.threshold_otsu(camera) blobs = im > 0.7 * im.mean()

hist, bins_center = exposure.histogram(camera) all_labels = measure.label(blobs)

blobs_labels = measure.label(blobs, background=0)
plt.figure(figsize=(9, 4))
plt.subplot(131) plt.figure(figsize=(9, 3.5))
plt.imshow(camera, cmap='gray', interpolation='nearest') plt.subplot(131)
plt.axis('off') plt.imshow(blobs, cmap='gray')
plt.subplot(132) plt.axis('off')
plt.imshow(camera < val, cmap='gray', interpolation='nearest') plt.subplot(132)
plt.axis('off') plt.imshow(all_labels, cmap='spectral')
plt.subplot(133) plt.axis('off')
plt.plot(bins_center, hist, lw=2) plt.subplot(133)
plt.axvline(val, color='k', ls='--') plt.imshow(blobs_labels, cmap='spectral')
plt.axis('off')
plt.tight_layout()
plt.show() plt.tight_layout()
plt.show()
Total running time of the script: ( 0 minutes 0.240 seconds)
Total running time of the script: ( 0 minutes 0.169 seconds)
Download Python source code: plot_threshold.py
Download Python source code: plot_labels.py
Download Jupyter notebook: plot_threshold.ipynb
Download Jupyter notebook: plot_labels.ipynb
Generated by Sphinx-Gallery
Generated by Sphinx-Gallery

17.9.8 Labelling connected components of an image

17.9.9 Affine transform
This example shows how to label connected components of a binary image, using the dedicated skim-
age.measure.label function. Warping and affine transforms of images.

17.9. Examples for the scikit-image chapter 532 17.9. Examples for the scikit-image chapter 533
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

17.9.10 Various denoising filters

This example compares several denoising filters available in scikit-image: a Gaussian filter, a median filter, and
total variation denoising.

import numpy as np
import matplotlib.pyplot as plt
from skimage import data
from skimage import filters
from skimage import restoration

coins = data.coins()
gaussian_filter_coins = filters.gaussian(coins, sigma=2)
med_filter_coins = filters.median(coins, np.ones((3, 3)))
tv_filter_coins = restoration.denoise_tv_chambolle(coins, weight=0.1)

plt.figure(figsize=(16, 4))
plt.subplot(141)
plt.imshow(coins[10:80, 300:370], cmap='gray', interpolation='nearest')
plt.axis('off')
plt.title('Image')
from matplotlib import pyplot as plt
plt.subplot(142)
plt.imshow(gaussian_filter_coins[10:80, 300:370], cmap='gray',
from skimage import data
interpolation='nearest')
from skimage.feature import corner_harris, corner_subpix, corner_peaks
plt.axis('off')
from skimage.transform import warp, AffineTransform
plt.title('Gaussian filter')
plt.subplot(143)
plt.imshow(med_filter_coins[10:80, 300:370], cmap='gray',
tform = AffineTransform(scale=(1.3, 1.1), rotation=1, shear=0.7,
interpolation='nearest')
translation=(210, 50))
plt.axis('off')
image = warp(data.checkerboard(), tform.inverse, output_shape=(350, 350))
plt.title('Median filter')
plt.subplot(144)
coords = corner_peaks(corner_harris(image), min_distance=5)
plt.imshow(tv_filter_coins[10:80, 300:370], cmap='gray',
coords_subpix = corner_subpix(image, coords, window_size=13)
interpolation='nearest')
plt.axis('off')
plt.gray()
plt.title('TV filter')
plt.imshow(image, interpolation='nearest')
plt.show()
plt.plot(coords_subpix[:, 1], coords_subpix[:, 0], '+r', markersize=15, mew=5)
plt.plot(coords[:, 1], coords[:, 0], '.b', markersize=7)
plt.axis('off') Total running time of the script: ( 0 minutes 0.285 seconds)
plt.show() Download Python source code: plot_filter_coins.py

Total running time of the script: ( 0 minutes 0.072 seconds) Download Jupyter notebook: plot_filter_coins.ipynb

Download Python source code: plot_features.py Generated by Sphinx-Gallery

Download Jupyter notebook: plot_features.ipynb

Generated by Sphinx-Gallery
17.9.11 Watershed and random walker for segmentation

This example compares two segmentation methods in order to separate two connected disks: the watershed
algorithm, and the random walker algorithm.

17.9. Examples for the scikit-image chapter 534 17.9. Examples for the scikit-image chapter 535
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Both segmentation methods require seeds, that are pixels belonging unambigusouly to a reagion. Here, local Total running time of the script: ( 0 minutes 0.250 seconds)
maxima of the distance map to the background are used as seeds.
Download Python source code: plot_segmentations.py
Download Jupyter notebook: plot_segmentations.ipynb
Generated by Sphinx-Gallery
Download all examples in Python source code: auto_examples_python.zip
Download all examples in Jupyter notebooks: auto_examples_jupyter.zip
Generated by Sphinx-Gallery

import numpy as np
from skimage.morphology import watershed
from skimage.feature import peak_local_max
from skimage import measure
from skimage.segmentation import random_walker
import matplotlib.pyplot as plt
from scipy import ndimage

# Generate an initial image with two overlapping circles

x, y = np.indices((80, 80))
x1, y1, x2, y2 = 28, 28, 44, 52
r1, r2 = 16, 20
mask_circle1 = (x - x1) ** 2 + (y - y1) ** 2 < r1 ** 2
mask_circle2 = (x - x2) ** 2 + (y - y2) ** 2 < r2 ** 2
image = np.logical_or(mask_circle1, mask_circle2)
# Now we want to separate the two objects in image
# Generate the markers as local maxima of the distance
# to the background
distance = ndimage.distance_transform_edt(image)
local_maxi = peak_local_max(
distance, indices=False, footprint=np.ones((3, 3)), labels=image)
markers = measure.label(local_maxi)
labels_ws = watershed(-distance, markers, mask=image)

markers[~image] = -1
labels_rw = random_walker(image, markers)

plt.figure(figsize=(12, 3.5))
plt.subplot(141)
plt.imshow(image, cmap='gray', interpolation='nearest')
plt.axis('off')
plt.title('image')
plt.subplot(142)
plt.imshow(-distance, interpolation='nearest')
plt.axis('off')
plt.title('distance map')
plt.subplot(143)
plt.imshow(labels_ws, cmap='spectral', interpolation='nearest')
plt.axis('off')
plt.title('watershed segmentation')
plt.subplot(144)
plt.imshow(labels_rw, cmap='spectral', interpolation='nearest')
plt.axis('off')
plt.title('random walker segmentation')

plt.tight_layout()
plt.show()

17.9. Examples for the scikit-image chapter 536 17.9. Examples for the scikit-image chapter 537
 Scipy lecture notes, Edition 2017.1

Tutorial content

• Introduction
• Example

18
• What are Traits
– Initialisation
– Validation
– Documentation

CHAPTER – Visualization: opening a dialog

– Deferral
– Notification
– Some more advanced traits

18.1 Introduction
Traits: building interactive dialogs
Tip: The Enthought Tool Suite enable the construction of sophisticated application frameworks for data anal-
ysis, 2D plotting and 3D visualization. These powerful, reusable components are released under liberal BSD-
style licenses.

Author: Didrik Pinte

The Traits project allows you to simply add validation, initialization, delegation, notification and a graphical
user interface to Python object attributes.

Tip: In this tutorial we will explore the Traits toolset and learn how to dramatically reduce the amount of
boilerplate code you write, do rapid GUI application development, and understand the ideas which underly
other parts of the Enthought Tool Suite.
Traits and the Enthought Tool Suite are open source projects licensed under a BSD-style license.

Intended Audience

Intermediate to advanced Python programmers

Requirements

• Either wxPython, PyQt or PySide

• Numpy and Scipy
• Enthought Tool Suite
• All required software can be obtained by installing the EPD Free
The main packages of the Enthought Tool Suite are:
• Traits - component based approach to build our applications.
• Kiva - 2D primitives supporting path based rendering, affine transforms, alpha blending and more.

538 18.1. Introduction 539

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

• Enable - object based 2D drawing canvas. – Initialization

• Chaco - plotting toolkit for building complex interactive 2D plots. – Validation
• Mayavi - 3D visualization of scientific data based on VTK. – Deferral
• Envisage - application plugin framework for building scriptable and extensible applications • Notification
In this tutorial, we will focus on Traits. • Visualization
• Documentation

18.2 Example A class can freely mix trait-based attributes with normal Python attributes, or can opt to allow the use of only
a fixed or open set of trait attributes within the class. Trait attributes defined by a class are automatically
inherited by any subclass derived from the class.
Throughout this tutorial, we will use an example based on a water resource management simple case. We will
try to model a dam and reservoir system. The reservoir and the dams do have a set of parameters : The common way of creating a traits class is by extending from the HasTraits base class and defining class
traits :
• Name
from traits.api import HasTraits, Str, Float
• Minimal and maximal capacity of the reservoir [hm3]
• Height and length of the dam [m] class Reservoir(HasTraits):

• Catchment area [km2] name = Str

max_storage = Float
• Hydraulic head [m]
• Power of the turbines [MW]
• Minimal and maximal release [m3/s] Warning: For Traits 3.x users

• Efficiency of the turbines If using Traits 3.x, you need to adapt the namespace of the traits packages:

The reservoir has a known behaviour. One part is related to the energy production based on the water released. • traits.api should be enthought.traits.api
A simple formula for approximating electric power production at a hydroelectric plant is P = ρhr g k, where: • traitsui.api should be enthought.traits.ui.api
• P is Power in watts,
• ρ is the density of water (~1000 kg/m3), Using a traits class like that is as simple as any other Python class. Note that the trait value are passed using
keyword arguments:
• h is height in meters,
reservoir = Reservoir(name='Lac de Vouglans', max_storage=605)
• r is flow rate in cubic meters per second,
• g is acceleration due to gravity of 9.8 m/s2,
• k is a coefficient of efficiency ranging from 0 to 1. 18.3.1 Initialisation

All the traits do have a default value that initialise the variables. For example, the basic python types do have
Tip: Annual electric energy production depends on the available water supply. In some installations the water
the following trait equivalents:
flow rate can vary by a factor of 10:1 over the course of a year.

Trait Python Type Built-in Default Value

The second part of the behaviour is the state of the storage that depends on controlled and uncontrolled pa- Bool Boolean False
rameters : Complex Complex number 0+0j
st or ag e t +1 = st or ag e t + i n f l ow s − r el ease − spi l l ag e − i r r i g at i on Float Floating point number 0.0
Int Plain integer 0
Long Long integer 0L
Warning: The data used in this tutorial are not real and might even not have sense in the reality. Str String ‘’
Unicode Unicode u’‘

18.3 What are Traits A number of other predefined trait type do exist : Array, Enum, Range, Event, Dict, List, Color, Set, Expression,
Code, Callable, Type, Tuple, etc.

A trait is a type definition that can be used for normal Python object attributes, giving the attributes some Custom default values can be defined in the code:
additional characteristics: from traits.api import HasTraits, Str, Float
• Standardization:
class Reservoir(HasTraits):

18.2. Example 540 18.3. What are Traits 541

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

name = Str
name = Str max_storage = Float(100, desc='Maximal storage [hm3]')
max_storage = Float(100)
Let’s now define the complete reservoir class:
reservoir = Reservoir(name='Lac de Vouglans')
from traits.api import HasTraits, Str, Float, Range

Complex initialisation class Reservoir(HasTraits):

name = Str
max_storage = Float(1e6, desc='Maximal storage [hm3]')
When a complex initialisation is required for a trait, a _XXX_default magic method can be implemented. It
max_release = Float(10, desc='Maximal release [m3/s]')
will be lazily called when trying to access the XXX trait. For example:
head = Float(10, desc='Hydraulic head [m]')
def _name_default(self): efficiency = Range(0, 1.)
""" Complex initialisation of the reservoir name. """
def energy_production(self, release):
return 'Undefined' ''' Returns the energy production [Wh] for the given release [m3/s]
'''
power = 1000 * 9.81 * self.head * release * self.efficiency
return power * 3600

18.3.2 Validation
if __name__ == '__main__':
Every trait does validation when the user tries to set its content: reservoir = Reservoir(
name = 'Project A',
reservoir = Reservoir(name='Lac de Vouglans', max_storage=605) max_storage = 30,
max_release = 100.0,
reservoir.max_storage = '230' head = 60,
--------------------------------------------------------------------------- efficiency = 0.8
TraitError Traceback (most recent call last) )
.../scipy-lecture-notes/advanced/traits/<ipython-input-7-979bdff9974a> in <module>()
----> 1 reservoir.max_storage = '230' release = 80
print 'Releasing {} m3/s produces {} kWh'.format(
.../traits/trait_handlers.pyc in error(self, object, name, value) release, reservoir.energy_production(release)
166 """ )
167 raise TraitError( object, name, self.full_info( object, name, value ),
--> 168 value )
169
170 def arg_error ( self, method, arg_num, object, name, value ): 18.3.4 Visualization: opening a dialog
TraitError: The 'max_storage' trait of a Reservoir instance must be a float, but a value of '23 The Traits library is also aware of user interfaces and can pop up a default view for the Reservoir class:
,→' <type 'str'> was specified.
reservoir1 = Reservoir()
reservoir1.edit_traits()

18.3.3 Documentation

By essence, all the traits do provide documentation about the model itself. The declarative approach to the
creation of classes makes it self-descriptive:

from traits.api import HasTraits, Str, Float

class Reservoir(HasTraits):

name = Str
max_storage = Float(100)

The desc metadata of the traits can be used to provide a more descriptive information about the trait :

from traits.api import HasTraits, Str, Float

class Reservoir(HasTraits):

18.3. What are Traits 542 18.3. What are Traits 543
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

TraitsUI simplifies the way user interfaces are created. Every trait on a HasTraits class has a default editor that
class ReservoirState(HasTraits):
will manage the way the trait is rendered to the screen (e.g. the Range trait is displayed as a slider, etc.). """Keeps track of the reservoir state given the initial storage.
In the very same vein as the Traits declarative way of creating classes, TraitsUI provides a declarative interface """
to build user interfaces code: reservoir = Instance(Reservoir, ())
min_storage = Float
from traits.api import HasTraits, Str, Float, Range max_storage = DelegatesTo('reservoir')
from traitsui.api import View min_release = Float
max_release = DelegatesTo('reservoir')
class Reservoir(HasTraits):
name = Str # state attributes
max_storage = Float(1e6, desc='Maximal storage [hm3]') storage = Range(low='min_storage', high='max_storage')
max_release = Float(10, desc='Maximal release [m3/s]')
head = Float(10, desc='Hydraulic head [m]') # control attributes
efficiency = Range(0, 1.) inflows = Float(desc='Inflows [hm3]')
release = Range(low='min_release', high='max_release')
traits_view = View( spillage = Float(desc='Spillage [hm3]')
'name', 'max_storage', 'max_release', 'head', 'efficiency',
title = 'Reservoir', def print_state(self):
resizable = True, print 'Storage\tRelease\tInflows\tSpillage'
) str_format = '\t'.join(['{:7.2f} 'for i in range(4)])
print str_format.format(self.storage, self.release, self.inflows,
def energy_production(self, release): self.spillage)
''' Returns the energy production [Wh] for the given release [m3/s] print '-' * 79
'''
power = 1000 * 9.81 * self.head * release * self.efficiency
return power * 3600 if __name__ == '__main__':
projectA = Reservoir(
name = 'Project A',
if __name__ == '__main__': max_storage = 30,
reservoir = Reservoir( max_release = 100.0,
name = 'Project A', hydraulic_head = 60,
max_storage = 30, efficiency = 0.8
max_release = 100.0, )
head = 60,
efficiency = 0.8 state = ReservoirState(reservoir=projectA, storage=10)
) state.release = 90
state.inflows = 0
reservoir.configure_traits() state.print_state()

print 'How do we update the current storage ?'

A special trait allows to manage events and trigger function calls using the magic _xxxx_fired method:

from traits.api import HasTraits, Instance, DelegatesTo, Float, Range, Event

from reservoir import Reservoir

class ReservoirState(HasTraits):
"""Keeps track of the reservoir state given the initial storage.

For the simplicity of the example, the release is considered in

hm3/timestep and not in m3/s.
"""
reservoir = Instance(Reservoir, ())
min_storage = Float
18.3.5 Deferral max_storage = DelegatesTo('reservoir')
min_release = Float
Being able to defer the definition of a trait and its value to another object is a powerful feature of Traits. max_release = DelegatesTo('reservoir')

from traits.api import HasTraits, Instance, DelegatesTo, Float, Range # state attributes
storage = Range(low='min_storage', high='max_storage')
from reservoir import Reservoir
# control attributes

18.3. What are Traits 544 18.3. What are Traits 545
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

inflows = Float(desc='Inflows [hm3]') min_release = Float

release = Range(low='min_release', high='max_release') max_release = DelegatesTo('reservoir')
spillage = Float(desc='Spillage [hm3]')
# state attributes
update_storage = Event(desc='Updates the storage to the next time step') storage = Property(depends_on='inflows, release')

def _update_storage_fired(self): # control attributes

# update storage state inflows = Float(desc='Inflows [hm3]')
new_storage = self.storage - self.release + self.inflows release = Range(low='min_release', high='max_release')
self.storage = min(new_storage, self.max_storage) spillage = Property(
overflow = new_storage - self.max_storage desc='Spillage [hm3]', depends_on=['storage', 'inflows', 'release']
self.spillage = max(overflow, 0) )

def print_state(self): ### Private traits.

print 'Storage\tRelease\tInflows\tSpillage' _storage = Float
str_format = '\t'.join(['{:7.2f} 'for i in range(4)])
print str_format.format(self.storage, self.release, self.inflows, ### Traits property implementation.
self.spillage) def _get_storage(self):
print '-' * 79 new_storage = self._storage - self.release + self.inflows
return min(new_storage, self.max_storage)

if __name__ == '__main__': def _set_storage(self, storage_value):

projectA = Reservoir( self._storage = storage_value
name = 'Project A',
max_storage = 30, def _get_spillage(self):
max_release = 5.0, new_storage = self._storage - self.release + self.inflows
hydraulic_head = 60, overflow = new_storage - self.max_storage
efficiency = 0.8 return max(overflow, 0)
)
def print_state(self):
state = ReservoirState(reservoir=projectA, storage=15) print 'Storage\tRelease\tInflows\tSpillage'
state.release = 5 str_format = '\t'.join(['{:7.2f} 'for i in range(4)])
state.inflows = 0 print str_format.format(self.storage, self.release, self.inflows,
self.spillage)
# release the maximum amount of water during 3 time steps print '-' * 79
state.update_storage = True
state.print_state() if __name__ == '__main__':
state.update_storage = True projectA = Reservoir(
state.print_state() name = 'Project A',
state.update_storage = True max_storage = 30,
state.print_state() max_release = 5,
hydraulic_head = 60,
Dependency between objects can be made automatic using the trait Property. The depends_on attribute efficiency = 0.8
expresses the dependency between the property and other traits. When the other traits gets changed, the )
property is invalidated. Again, Traits uses magic method names for the property :
state = ReservoirState(reservoir=projectA, storage=25)
• _get_XXX for the getter of the XXX Property trait state.release = 4
state.inflows = 0
• _set_XXX for the setter of the XXX Property trait
state.print_state()
from traits.api import HasTraits, Instance, DelegatesTo, Float, Range
from traits.api import Property

from reservoir import Reservoir Note: Caching property

Heavy computation or long running computation might be a problem when accessing a property where the
class ReservoirState(HasTraits):
inputs have not changed. The @cached_property decorator can be used to cache the value and only recompute
"""Keeps track of the reservoir state given the initial storage.
them once invalidated.
For the simplicity of the example, the release is considered in
hm3/timestep and not in m3/s. Let’s extend the TraitsUI introduction with the ReservoirState example:
"""
reservoir = Instance(Reservoir, ()) from traits.api import HasTraits, Instance, DelegatesTo, Float, Range, Property
max_storage = DelegatesTo('reservoir') from traitsui.api import View, Item, Group, VGroup

18.3. What are Traits 546 18.3. What are Traits 547
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

hydraulic_head = 60,
from reservoir import Reservoir efficiency = 0.8
)
class ReservoirState(HasTraits):
"""Keeps track of the reservoir state given the initial storage. state = ReservoirState(reservoir=projectA, storage=25)
state.release = 4
For the simplicity of the example, the release is considered in state.inflows = 0
hm3/timestep and not in m3/s.
""" state.print_state()
reservoir = Instance(Reservoir, ()) state.configure_traits()
name = DelegatesTo('reservoir')
max_storage = DelegatesTo('reservoir')
max_release = DelegatesTo('reservoir')
min_release = Float

# state attributes
storage = Property(depends_on='inflows, release')

# control attributes
inflows = Float(desc='Inflows [hm3]')
release = Range(low='min_release', high='max_release')
spillage = Property(
desc='Spillage [hm3]', depends_on=['storage', 'inflows', 'release']
)

### Traits view

traits_view = View(
Group(
VGroup(Item('name'), Item('storage'), Item('spillage'),
label = 'State', style = 'readonly'
),
Some use cases need the delegation mechanism to be broken by the user when setting the value of the trait.
VGroup(Item('inflows'), Item('release'), label='Control'),
)
The PrototypeFrom trait implements this behaviour.
)
from traits.api import HasTraits, Str, Float, Range, PrototypedFrom, Instance
### Private traits.
class Turbine(HasTraits):
_storage = Float
turbine_type = Str
power = Float(1.0, desc='Maximal power delivered by the turbine [Mw]')
### Traits property implementation.
def _get_storage(self):
new_storage = self._storage - self.release + self.inflows
class Reservoir(HasTraits):
return min(new_storage, self.max_storage)
name = Str
max_storage = Float(1e6, desc='Maximal storage [hm3]')
def _set_storage(self, storage_value):
max_release = Float(10, desc='Maximal release [m3/s]')
self._storage = storage_value
head = Float(10, desc='Hydraulic head [m]')
efficiency = Range(0, 1.)
def _get_spillage(self):
new_storage = self._storage - self.release + self.inflows
turbine = Instance(Turbine)
overflow = new_storage - self.max_storage
installed_capacity = PrototypedFrom('turbine', 'power')
return max(overflow, 0)

def print_state(self):
if __name__ == '__main__':
print 'Storage\tRelease\tInflows\tSpillage'
turbine = Turbine(turbine_type='type1', power=5.0)
str_format = '\t'.join(['{:7.2f} 'for i in range(4)])
print str_format.format(self.storage, self.release, self.inflows,
reservoir = Reservoir(
self.spillage)
name = 'Project A',
print '-' * 79
max_storage = 30,
max_release = 100.0,
if __name__ == '__main__':
head = 60,
projectA = Reservoir(
efficiency = 0.8,
name = 'Project A',
turbine = turbine,
max_storage = 30,
)
max_release = 5,

18.3. What are Traits 548 18.3. What are Traits 549
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

if __name__ == '__main__':
print 'installed capacity is initialised with turbine.power' projectA = Reservoir(
print reservoir.installed_capacity name = 'Project A',
max_storage = 30,
print '-' * 15 max_release = 100.0,
print 'updating the turbine power updates the installed capacity' hydraulic_head = 60,
turbine.power = 10 efficiency = 0.8
print reservoir.installed_capacity )

print '-' * 15 state = ReservoirState(reservoir=projectA, storage=10)

print 'setting the installed capacity breaks the link between turbine.power' state.release = 90
print 'and the installed_capacity trait' state.inflows = 0
state.print_state()
reservoir.installed_capacity = 8
print turbine.power, reservoir.installed_capacity The static trait notification signatures can be:
• def _release_changed(self ): pass

18.3.6 Notification • def _release_changed(self, new): pass

• def _release_changed(self, old, new): pass
Traits implements a Listener pattern. For each trait a list of static and dynamic listeners can be fed with call-
backs. When the trait does change, all the listeners are called. • def _release_changed(self, name, old, new pass

Static listeners are defined using the _XXX_changed magic methods:

Listening to all the changes
from traits.api import HasTraits, Instance, DelegatesTo, Float, Range
To listen to all the changes on a HasTraits class, the magic _any_trait_changed method can be implemented.
from reservoir import Reservoir

class ReservoirState(HasTraits): In many situations, you do not know in advance what type of listeners need to be activated. Traits offers the
"""Keeps track of the reservoir state given the initial storage. ability to register listeners on the fly with the dynamic listeners
"""
reservoir = Instance(Reservoir, ()) from reservoir import Reservoir
min_storage = Float from reservoir_state_property import ReservoirState
max_storage = DelegatesTo('reservoir')
min_release = Float def wake_up_watchman_if_spillage(new_value):
max_release = DelegatesTo('reservoir') if new_value > 0:
print 'Wake up watchman! Spilling {} hm3'.format(new_value)
# state attributes
storage = Range(low='min_storage', high='max_storage') if __name__ == '__main__':
projectA = Reservoir(
# control attributes name = 'Project A',
inflows = Float(desc='Inflows [hm3]') max_storage = 30,
release = Range(low='min_release', high='max_release') max_release = 100.0,
spillage = Float(desc='Spillage [hm3]') hydraulic_head = 60,
efficiency = 0.8
def print_state(self): )
print 'Storage\tRelease\tInflows\tSpillage'
str_format = '\t'.join(['{:7.2f} 'for i in range(4)]) state = ReservoirState(reservoir=projectA, storage=10)
print str_format.format(self.storage, self.release, self.inflows,
self.spillage) #register the dynamic listener
print '-' * 79 state.on_trait_change(wake_up_watchman_if_spillage, name='spillage')

### Traits listeners ########### state.release = 90

def _release_changed(self, new): state.inflows = 0
"""When the release is higher than zero, warn all the inhabitants of state.print_state()
the valley.
""" print 'Forcing spillage'
state.inflows = 100
if new > 0: state.release = 0
print 'Warning, we are releasing {} hm3 of water'.format(new)
print 'Why do we have two executions of the callback ?'

18.3. What are Traits 550 18.3. What are Traits 551
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

The dynamic trait notification signatures are not the same as the static ones :
print 'Storage\tRelease\tInflows\tSpillage'
• def wake_up_watchman(): pass str_format = '\t'.join(['{:7.2f} 'for i in range(4)])
print str_format.format(self.storage, self.release, self.inflows,
• def wake_up_watchman(new): pass self.spillage)
print '-' * 79
• def wake_up_watchman(name, new): pass
• def wake_up_watchman(object, name, new): pass if __name__ == '__main__':
projectA = Reservoir(
• def wake_up_watchman(object, name, old, new): pass name = 'Project A',
Removing a dynamic listener can be done by: max_storage = 30,
max_release = 5,
• calling the remove_trait_listener method on the trait with the listener method as argument, hydraulic_head = 60,
efficiency = 0.8
• calling the on_trait_change method with listener method and the keyword remove=True, )
• deleting the instance that holds the listener.
state = ReservoirState(reservoir=projectA, storage=25)
Listeners can also be added to classes using the on_trait_change decorator: state.release = 4
state.inflows = 0
from traits.api import HasTraits, Instance, DelegatesTo, Float, Range
from traits.api import Property, on_trait_change
The patterns supported by the on_trait_change method and decorator are powerful. The reader should look at
from reservoir import Reservoir the docstring of HasTraits.on_trait_change for the details.

class ReservoirState(HasTraits):
"""Keeps track of the reservoir state given the initial storage. 18.3.7 Some more advanced traits

For the simplicity of the example, the release is considered in The following example demonstrate the usage of the Enum and List traits :
hm3/timestep and not in m3/s.
""" from traits.api import HasTraits, Str, Float, Range, Enum, List
reservoir = Instance(Reservoir, ()) from traitsui.api import View, Item
max_storage = DelegatesTo('reservoir')
min_release = Float class IrrigationArea(HasTraits):
max_release = DelegatesTo('reservoir') name = Str
surface = Float(desc='Surface [ha]')
# state attributes crop = Enum('Alfalfa', 'Wheat', 'Cotton')
storage = Property(depends_on='inflows, release')

# control attributes class Reservoir(HasTraits):

inflows = Float(desc='Inflows [hm3]') name = Str
release = Range(low='min_release', high='max_release') max_storage = Float(1e6, desc='Maximal storage [hm3]')
spillage = Property( max_release = Float(10, desc='Maximal release [m3/s]')
desc='Spillage [hm3]', depends_on=['storage', 'inflows', 'release'] head = Float(10, desc='Hydraulic head [m]')
) efficiency = Range(0, 1.)
irrigated_areas = List(IrrigationArea)

### Private traits. ########## def energy_production(self, release):

_storage = Float ''' Returns the energy production [Wh] for the given release [m3/s]
'''
### Traits property implementation. power = 1000 * 9.81 * self.head * release * self.efficiency
def _get_storage(self): return power * 3600
new_storage = self._storage - self.release + self.inflows
return min(new_storage, self.max_storage) traits_view = View(
Item('name'),
def _set_storage(self, storage_value): Item('max_storage'),
self._storage = storage_value Item('max_release'),
Item('head'),
def _get_spillage(self): Item('efficiency'),
new_storage = self._storage - self.release + self.inflows Item('irrigated_areas'),
overflow = new_storage - self.max_storage resizable = True
return max(overflow, 0) )

@on_trait_change('storage') if __name__ == '__main__':

def print_state(self): upper_block = IrrigationArea(name='Section C', surface=2000, crop='Wheat')

18.3. What are Traits 552 18.3. What are Traits 553
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

reservoir = Reservoir(
reservoir = Reservoir( name='Project A',
name='Project A', max_storage=30,
max_storage=30, max_release=100.0,
max_release=100.0, head=60,
head=60, efficiency=0.8,
efficiency=0.8, irrigated_areas=[upper_block],
irrigated_areas=[upper_block] )
)
release = 80
release = 80 print 'Releasing {} m3/s produces {} kWh'.format(
print 'Releasing {} m3/s produces {} kWh'.format( release, reservoir.energy_production(release)
release, reservoir.energy_production(release) )
)
The next example shows how the Array trait can be used to feed a specialised TraitsUI Item, the ChacoPlotItem:
Trait listeners can be used to listen to changes in the content of the list to e.g. keep track of the total crop
surface on linked to a given reservoir. import numpy as np

from traits.api import HasTraits, Str, Float, Range, Enum, List, Property from traits.api import HasTraits, Array, Instance, Float, Property
from traitsui.api import View, Item from traits.api import DelegatesTo
from traitsui.api import View, Item, Group
class IrrigationArea(HasTraits): from chaco.chaco_plot_editor import ChacoPlotItem
name = Str
surface = Float(desc='Surface [ha]') from reservoir import Reservoir
crop = Enum('Alfalfa', 'Wheat', 'Cotton')

class ReservoirEvolution(HasTraits):
class Reservoir(HasTraits): reservoir = Instance(Reservoir)
name = Str
max_storage = Float(1e6, desc='Maximal storage [hm3]') name = DelegatesTo('reservoir')
max_release = Float(10, desc='Maximal release [m3/s]')
head = Float(10, desc='Hydraulic head [m]') inflows = Array(dtype=np.float64, shape=(None))
efficiency = Range(0, 1.) releass = Array(dtype=np.float64, shape=(None))

irrigated_areas = List(IrrigationArea) initial_stock = Float

stock = Property(depends_on='inflows, releases, initial_stock')

total_crop_surface = Property(depends_on='irrigated_areas.surface') month = Property(depends_on='stock')

def _get_total_crop_surface(self): ### Traits view ###########

return sum([iarea.surface for iarea in self.irrigated_areas]) traits_view = View(
Item('name'),
def energy_production(self, release): Group(
''' Returns the energy production [Wh] for the given release [m3/s] ChacoPlotItem('month', 'stock', show_label=False),
''' ),
power = 1000 * 9.81 * self.head * release * self.efficiency width = 500,
return power * 3600 resizable = True
)
traits_view = View(
Item('name'), ### Traits properties #####
Item('max_storage'), def _get_stock(self):
Item('max_release'), """
Item('head'), fixme: should handle cases where we go over the max storage
Item('efficiency'), """
Item('irrigated_areas'), return self.initial_stock + (self.inflows - self.releases).cumsum()
Item('total_crop_surface'),
resizable = True def _get_month(self):
) return np.arange(self.stock.size)

if __name__ == '__main__': if __name__ == '__main__':

upper_block = IrrigationArea(name='Section C', surface=2000, crop='Wheat') reservoir = Reservoir(
name = 'Project A',
max_storage = 30,

18.3. What are Traits 554 18.3. What are Traits 555
 Scipy lecture notes, Edition 2017.1

max_release = 100.0,
head = 60,
efficiency = 0.8
)

19
initial_stock = 10.
inflows_ts = np.array([6., 6, 4, 4, 1, 2, 0, 0, 3, 1, 5, 3])
releases_ts = np.array([4., 5, 3, 5, 3, 5, 5, 3, 2, 1, 3, 3])

view = ReservoirEvolution(
reservoir = reservoir,
inflows = inflows_ts,
releases = releases_ts CHAPTER
)
view.configure_traits()

3D plotting with Mayavi

Author: Gaël Varoquaux

Tip: Mayavi is an interactive 3D plotting package. matplotlib can also do simple 3D plotting, but Mayavi relies
on a more powerful engine ( VTK ) and is more suited to displaying large or complex data.

See also:
References
Chapters contents
• ETS repositories
• Mlab: the scripting interface
• Traits manual
– 3D plotting functions
• Traits UI manual
* Points
• Mailing list : enthought-dev@enthought.com
* Lines
* Elevation surface
* Arbitrary regular mesh
* Volumetric data
– Figures and decorations

* Figure management
* Changing plot properties
* Decorations
• Interactive work
– The “pipeline dialog”

18.3. What are Traits 556 557

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

– The script recording button Lines

• Slicing and dicing data: sources, modules and filters

– An example: inspecting magnetic fields
– Different views on data: sources and modules

* Different sources: scatters and fields

* Transforming data: filters
* mlab.pipeline: the scripting layer
• Animating the data
• Making interactive dialogs
– A simple dialog
– Making it interactive Hint: A line connecting points in 3D, with optional thickness and varying color.
• Putting it together
mlab.clf() # Clear the figure
t = np.linspace(0, 20, 200)
mlab.plot3d(np.sin(t), np.cos(t), 0.1*t, t)
19.1 Mlab: the scripting interface

The mayavi.mlab module provides simple plotting functions to apply to numpy arrays, similar to matplotlib
Elevation surface
or matlab’s plotting interface. Try using them in IPython, by starting IPython with the switch --gui=wx.

19.1.1 3D plotting functions

Points

Hint: A surface given by its elevation, coded as a 2D array

mlab.clf()
x, y = np.mgrid[-10:10:100j, -10:10:100j]
r = np.sqrt(x**2 + y**2)
z = np.sin(r)/r
Hint: Points in 3D, represented with markers (or “glyphs”) and optionaly different sizes. mlab.surf(z, warp_scale='auto')

x, y, z, value = np.random.random((4, 40))

mlab.points3d(x, y, z, value)

19.1. Mlab: the scripting interface 558 19.1. Mlab: the scripting interface 559
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Arbitrary regular mesh

This function works with a regular orthogonal grid: the value array is a 3D array that gives the shape of the
Hint: A surface mesh given by x, y, z positions of its node points grid.

mlab.clf() 19.1.2 Figures and decorations

phi, theta = np.mgrid[0:np.pi:11j, 0:2*np.pi:11j]
x = np.sin(phi) * np.cos(theta) Figure management
y = np.sin(phi) * np.sin(theta)
z = np.cos(phi)
mlab.mesh(x, y, z)
mlab.mesh(x, y, z, representation='wireframe', color=(0, 0, 0)) Tip: Here is a list of functions useful to control the current figure

Note: A surface is defined by points connected to form triangles or polygones. In mayavi.mlab.surf() and Get the current figure: mlab.gcf()
mayavi.mlab.mesh(), the connectivity is implicity given by the layout of the arrays. See also mayavi.mlab. Clear the current figure: mlab.clf()
triangular_mesh(). Set the current figure: mlab.figure(1, bgcolor=(1, 1, 1), fgcolor=(0.5, 0.5, 0.5)
Save figure to image file: mlab.savefig(‘foo.png’, size=(300, 300))
Change the view: mlab.view(azimuth=45, elevation=54, distance=1.)
Our data is often more than points and values: it needs some connectivity information

Volumetric data Changing plot properties

Tip: In general, many properties of the various objects on the figure can be changed. If these visualization are
created via mlab functions, the easiest way to change them is to use the keyword arguments of these functions,
as described in the docstrings.

Example docstring: mlab.mesh

Plots a surface using grid-spaced data supplied as 2D arrays.

Function signatures:
mesh(x, y, z, ...)

Hint: If your data is dense in 3D, it is more difficult to display. One option is to take iso-contours of the data. x, y, z are 2D arrays, all of the same shape, giving the positions of the vertices of the surface. The connectivity
between these points is implied by the connectivity on the arrays.

mlab.clf() For simple structures (such as orthogonal grids) prefer the surf function, as it will create more efficient data
x, y, z = np.mgrid[-5:5:64j, -5:5:64j, -5:5:64j] structures.
values = x*x*0.5 + y*y + z*z*2.0 Keyword arguments:
mlab.contour3d(values)
color the color of the vtk object. Overides the colormap, if any, when specified. This
is specified as a triplet of float ranging from 0 to 1, eg (1, 1, 1) for white.

19.1. Mlab: the scripting interface 560 19.1. Mlab: the scripting interface 561
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

colormap type of colormap to use.

extent [xmin, xmax, ymin, ymax, zmin, zmax] Default is the x, y, z arrays extents.
Use this to change the extent of the object created.
figure Figure to populate.
line_width The with of the lines, if any used. Must be a float. Default: 2.0
mask boolean mask array to suppress some data points.
mask_points If supplied, only one out of ‘mask_points’ data point is displayed. This
option is usefull to reduce the number of points displayed on large datasets Must
be an integer or None.
mode the mode of the glyphs. Must be ‘2darrow’ or ‘2dcircle’ or ‘2dcross’ or
‘2ddash’ or ‘2ddiamond’ or ‘2dhooked_arrow’ or ‘2dsquare’ or ‘2dthick_arrow’
or ‘2dthick_cross’ or ‘2dtriangle’ or ‘2dvertex’ or ‘arrow’ or ‘cone’ or ‘cube’ or
‘cylinder’ or ‘point’ or ‘sphere’. Default: sphere
name the name of the vtk object created.
representation the representation type used for the surface. Must be ‘surface’ or
‘wireframe’ or ‘points’ or ‘mesh’ or ‘fancymesh’. Default: surface
resolution The resolution of the glyph created. For spheres, for instance, this is the
number of divisions along theta and phi. Must be an integer. Default: 8 Example:
scalars optional scalar data. In [1]: import numpy as np
scale_factor scale factor of the glyphs used to represent the vertices, in fancy_mesh
In [2]: r, theta = np.mgrid[0:10, -np.pi:np.pi:10j]
mode. Must be a float. Default: 0.05
scale_mode the scaling mode for the glyphs (‘vector’, ‘scalar’, or ‘none’). In [3]: x = r * np.cos(theta)

transparent make the opacity of the actor depend on the scalar. In [4]: y = r * np.sin(theta)
tube_radius radius of the tubes used to represent the lines, in mesh mode. If None,
In [5]: z = np.sin(r)/r
simple lines are used.
tube_sides number of sides of the tubes used to represent the lines. Must be an In [6]: from mayavi import mlab
integer. Default: 6
In [7]: mlab.mesh(x, y, z, colormap='gist_earth', extent=[0, 1, 0, 1, 0, 1])
vmax vmax is used to scale the colormap If None, the max of the data will be used Out[7]: <mayavi.modules.surface.Surface object at 0xde6f08c>
vmin vmin is used to scale the colormap If None, the min of the data will be used
In [8]: mlab.mesh(x, y, z, extent=[0, 1, 0, 1, 0, 1],
...: representation='wireframe', line_width=1, color=(0.5, 0.5, 0.5))
Out[8]: <mayavi.modules.surface.Surface object at 0xdd6a71c>

Decorations

Tip: Different items can be added to the figure to carry extra information, such as a colorbar or a title.

In [9]: mlab.colorbar(Out[7], orientation='vertical')

Out[9]: <tvtk_classes.scalar_bar_actor.ScalarBarActor object at 0xd897f8c>

In [10]: mlab.title('polar mesh')

Out[10]: <enthought.mayavi.modules.text.Text object at 0xd8ed38c>

In [11]: mlab.outline(Out[7])
Out[11]: <enthought.mayavi.modules.outline.Outline object at 0xdd21b6c>

In [12]: mlab.axes(Out[7])
Out[12]: <enthought.mayavi.modules.axes.Axes object at 0xd2e4bcc>

19.1. Mlab: the scripting interface 562 19.1. Mlab: the scripting interface 563
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

• Set the background of the figure in the Mayavi Scene node

• Set the colormap in the Colors and legends node
• Right click on the node to add modules or filters

19.2.2 The script recording button

To find out what code can be used to program these changes, click on the red button as you modify those
Warning: extent: If we specified extents for a plotting object, mlab.outline’ and ‘mlab.axes don’t get them properties, and it will generate the corresponding lines of code.
by default.

19.3 Slicing and dicing data: sources, modules and filters

19.2 Interactive work 19.3.1 An example: inspecting magnetic fields

Suppose we are simulating the magnetic field generated by Helmholtz coils. The examples/compute_field.
Tip: The quickest way to create beautiful visualization with Mayavi is probably to interactively tweak the py script does this computation and gives you a B array, that is (3 x n), where the first axis is the direction of
various settings. the field (Bx, By, Bz), and the second axis the index number of the point. Arrays X, Y and Z give the positions
of these data points.

19.2.1 The “pipeline dialog” Excercise

Click on the ‘Mayavi’ button in the scene, and you can control properties of objects with dialogs. Visualize this field. Your goal is to make sure that the simulation code is correct.

19.2. Interactive work 564 19.3. Slicing and dicing data: sources, modules and filters 565
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

• A 3D block of regularly-spaced value is structured: it is easy to know how one measurement is related to
another neighboring and how to continuously interpolate between these. We can call such data a field,
borrowing from terminology used in physics, as it is continuously defined in space.
• A set of data points measured at random positions in a random order gives rise to much more difficult
and ill-posed interpolation problems: the data structure itself does not tell us what are the neighbors of
a data point. We call such data a scatter.

Unstructured and unconnected data: a scatter Structured and connected data: a field
mlab.points3d, mlab.quiver3d mlab.contour3d

Data sources corresponding to scatters can be created with mayavi.mlab.pipeline.scalar_scatter()

or mayavi.mlab.pipeline.vector_scatter(); field data sources can be created with mlab.pipeline.
Suggestions
scalar_field() or mlab.pipeline.vector_field().
• If you compute the norm of the vector field, you can apply an isosurface to it.
Exercice:
• using mayavi.mlab.quiver3d() you can plot vectors. You can also use the ‘masking’ options (in the
GUI) to make the plot a bit less dense.
1. Create a contour (for instance of the magnetic field norm) by using one of those functions and adding
the right module by clicking on the GUI dialog.
2. Create the right source to apply a ‘vector_cut_plane’ and reproduce the picture of the magnetic field
19.3.2 Different views on data: sources and modules shown previously.
Note that one of the difficulties is providing the data in the right form (number of arrays, shape) to the
Tip: As we see above, it may be desirable to look at the same data in different ways. functions. This is often the case with real-life data.

See also:
Mayavi visualization are created by loading the data in a data source and then displayed on the screen using
modules. Sources are described in details in the Mayavi manual.

This can be seen by looking at the “pipeline” view. By right-clicking on the nodes of the pipeline, you can add
new modules. Transforming data: filters

Quiz If you create a vector field, you may want to visualize the iso-contours of its magnitude. But the isosurface
module can only be applied to scalar data, and not vector data. We can use a filter, ExtractVectorNorm to
Why is it not possible to add a VectorCutPlane to the vectors created by mayavi.mlab.quiver3d()? add this scalar value to the vector field.
Filters apply a transformation to data, and can be added between sources and modules

Different sources: scatters and fields Excercice

Using the GUI, add the ExtractVectorNorm filter to display iso-contours of the field magnitude.
Tip: Data comes in different descriptions.

19.3. Slicing and dicing data: sources, modules and filters 566 19.3. Slicing and dicing data: sources, modules and filters 567
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

mlab.pipeline: the scripting layer

iso = mlab.contour3d(scalars, transparent=True, contours=[0.5])
for i in range(1, 20):
The mlab scripting layer builds pipelines for you. You can reproduce these pipelines programmatically with scalars = np.sin(i * x * y * z) /(x * y * z)
the mlab.pipeline interface: each step has a corresponding mlab.pipeline function (simply convert the iso.mlab_source.scalars = scalars
name of the step to lower-case underscore-separated: ExtractVectorNorm gives extract_vector_norm). This
function takes as an argument the node that it applies to, as well as optional parameters, and returns the new
node. See also:
For example, iso-contours of the magnitude are coded as: More details in the Mayavi documentation
mlab.pipeline.iso_surface(mlab.pipeline.extract_vector_norm(field),
contours=[0.1*Bmax, 0.4*Bmax], Event loops
opacity=0.5)
For the interaction with the user (for instance changing the view with the mouse), Mayavi needs some time
to process these events. The for loop above prevents this. The Mayavi documentation details a workaround

19.5 Making interactive dialogs

It is very simple to make interactive dialogs with Mayavi using the Traits library (see the dedicated chapter
Traits: building interactive dialogs).

19.5.1 A simple dialog

from traits.api import HasTraits, Instance

from traitsui.api import View, Item, HGroup
from mayavi.core.ui.api import SceneEditor, MlabSceneModel

def curve(n_turns):
"The function creating the x, y, z coordinates needed to plot"
phi = np.linspace(0, 2*np.pi, 2000)
return [np.cos(phi) * (1 + 0.5*np.cos(n_turns*phi)),
np.sin(phi) * (1 + 0.5*np.cos(n_turns*phi)),
0.5*np.sin(n_turns*phi)]

Excercice class Visualization(HasTraits):

"The class that contains the dialog"
scene = Instance(MlabSceneModel, ())
Using the mlab.pipeline interface, generate a complete visualization, with iso-contours of the field magni-
tude, and a vector cut plane. def __init__(self):
(click on the figure for a solution) HasTraits.__init__(self)
x, y, z = curve(n_turns=2)
# Populating our plot
self.plot = self.scene.mlab.plot3d(x, y, z)

19.4 Animating the data # Describe the dialog

view = View(Item('scene', height=300, show_label=False,
editor=SceneEditor()),
HGroup('n_turns'), resizable=True)
Tip: To make movies, or interactive application, you may want to change the data represented on a given
visualization.
# Fire up the dialog
Visualization().configure_traits()
If you have built a visualization, using the mlab plotting functions, or the mlab.pipeline function, we can
update the data by assigning new values to the mlab_source attributes
Tip: Let us read a bit the code above (examples/mlab_dialog.py).
x , y , z = np.ogrid[-5:5:100j ,-5:5:100j, -5:5:100j]
scalars = np.sin(x * y * z) / (x * y * z) First, the curve function is used to compute the coordinate of the curve we want to plot.

19.4. Animating the data 568 19.5. Making interactive dialogs 569
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Second, the dialog is defined by an object inheriting from HasTraits, as it is done with Traits. The important
view = View(Item('scene', height=300, show_label=False,
point here is that a Mayavi scene is added as a specific Traits attribute (Instance). This is important for editor=SceneEditor()),
embedding it in the dialog. The view of this dialog is defined by the view attribute of the object. In the init of HGroup('n_turns'), resizable=True)
this object, we populate the 3D scene with a curve.
# Fire up the dialog
Finally, the configure_traits method creates the dialog and starts the event loop.
Visualization().configure_traits()

See also:
Tip: Full code of the example: examples/mlab_dialog.py.
There are a few things to be aware of when doing dialogs with Mayavi. Please read the Mayavi documentation

19.5.2 Making it interactive

19.6 Putting it together
We can combine the Traits events handler with the mlab_source to modify the visualization with the dialog.
We will enable the user to vary the n_turns parameter in the definition of the curve. For this, we need:
Exercise
• to define an n_turns attribute on our visualization object, so that it can appear in the dialog. We use a
Range type. Using the code from the magnetic field simulation, create a dialog that enable to move the 2 coils: change
• to wire modification of this attribute to a recomputation of the curve. For this, we use the their parameters.
on_traits_change decorator. Hint: to define a dialog entry for a vector of dimension 3
direction = Array(float, value=(0, 0, 1), cols=3, shape=(3,))

You can look at the example_coil_application.py to see a full-blown application for coil design in 270 lines of
code.

from traits.api import Range, on_trait_change

class Visualization(HasTraits):
n_turns = Range(0, 30, 11)
scene = Instance(MlabSceneModel, ())

def __init__(self):
HasTraits.__init__(self)
x, y, z = curve(self.n_turns)
self.plot = self.scene.mlab.plot3d(x, y, z)

@on_trait_change('n_turns')
def update_plot(self):
x, y, z = curve(self.n_turns)
self.plot.mlab_source.set(x=x, y=y, z=z)

19.5. Making interactive dialogs 570 19.6. Putting it together 571

 Scipy lecture notes, Edition 2017.1

Chapters contents

• Introduction: problem settings

• Basic principles of machine learning with scikit-learn

20
• Supervised Learning: Classification of Handwritten Digits
• Supervised Learning: Regression of Housing Data
• Measuring prediction performance
• Unsupervised Learning: Dimensionality Reduction and Visualization
CHAPTER
• The eigenfaces example: chaining PCA and SVMs
• Parameter selection, Validation, and Testing
• Examples for the scikit-learn chapter

20.1 Introduction: problem settings

scikit-learn: machine learning in Python 20.1.1 What is machine learning?

Tip: Machine Learning is about building programs with tunable parameters that are adjusted automatically
so as to improve their behavior by adapting to previously seen data.
Machine Learning can be considered a subfield of Artificial Intelligence since those algorithms can be seen
as building blocks to make computers learn to behave more intelligently by somehow generalizing rather that
Authors: Gael Varoquaux
just storing and retrieving data items like a database system would do.

Prerequisites

• numpy
• scipy
• matplotlib (optional)
• ipython (the enhancements come handy)

Acknowledgements

This chapter is adapted from a tutorial given by Gaël Varoquaux, Jake Vanderplas, Olivier Grisel. Fig. 20.1: A classification problem

See also:
We’ll take a look at two very simple machine learning tasks here. The first is a classification task: the figure
Data science in Python shows a collection of two-dimensional data, colored according to two different class labels. A classification
algorithm may be used to draw a dividing boundary between the two clusters of points:
• The Statistics in Python chapter may also be of interest for readers looking into machine learning.
By drawing this separating line, we have learned a model which can generalize to new data: if you were to drop
• The documentation of scikit-learn is very complete and didactic.
another point onto the plane which is unlabeled, this algorithm could now predict whether it’s a blue or a red

572 20.1. Introduction: problem settings 573

 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

point.

Setosa Iris Versicolor Iris Virginica Iris

Quick Question:

If we want to design an algorithm to recognize iris species, what might the data be?
Remember: we need a 2D array of size [n_samples x n_features].
Fig. 20.2: A regression problem • What would the n_samples refer to?
• What might the n_features refer to?

Remember that there must be a fixed number of features for each sample, and feature number i must be a
similar kind of quantity for each sample.
The next simple task we’ll look at is a regression task: a simple best-fit line to a set of data.
Again, this is an example of fitting a model to data, but our focus here is that the model can make general- Loading the Iris Data with Scikit-learn
izations about new data. The model has been learned from the training data, and can be used to predict the
result of test data: here, we might be given an x-value, and the model would allow us to predict the y value. Scikit-learn has a very straightforward set of data on these iris species. The data consist of the following:
• Features in the Iris dataset:
20.1.2 Data in scikit-learn
– sepal length (cm)

The data matrix – sepal width (cm)

– petal length (cm)
Machine learning algorithms implemented in scikit-learn expect data to be stored in a two-dimensional array
or matrix. The arrays can be either numpy arrays, or in some cases scipy.sparse matrices. The size of the – petal width (cm)
array is expected to be [n_samples, n_features] • Target classes to predict:
• n_samples: The number of samples: each sample is an item to process (e.g. classify). A sample can be a – Setosa
document, a picture, a sound, a video, an astronomical object, a row in database or CSV file, or whatever
you can describe with a fixed set of quantitative traits. – Versicolour

• n_features: The number of features or distinct traits that can be used to describe each item in a quanti- – Virginica
tative manner. Features are generally real-valued, but may be boolean or discrete-valued in some cases. scikit-learn embeds a copy of the iris CSV file along with a function to load it into numpy arrays:
>>> from sklearn.datasets import load_iris
Tip: The number of features must be fixed in advance. However it can be very high dimensional (e.g. millions
>>> iris = load_iris()
of features) with most of them being zeros for a given sample. This is a case where scipy.sparse matrices
can be useful, in that they are much more memory-efficient than numpy arrays.
Note: Import sklearn Note that scikit-learn is imported as sklearn

A Simple Example: the Iris Dataset The features of each sample flower are stored in the data attribute of the dataset:

The application problem >>> print(iris.data.shape)

(150, 4)
>>> n_samples, n_features = iris.data.shape
As an example of a simple dataset, let us a look at the iris data stored by scikit-learn. Suppose we want to
>>> print(n_samples)
recognize species of irises. The data consists of measurements of three different species of irises:

20.1. Introduction: problem settings 574 20.1. Introduction: problem settings 575
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

150 20.2 Basic principles of machine learning with scikit-learn

>>> print(n_features)
4
>>> print(iris.data[0]) 20.2.1 Introducing the scikit-learn estimator object
[ 5.1 3.5 1.4 0.2]
Every algorithm is exposed in scikit-learn via an ‘’Estimator” object. For instance a linear regression is:
The information about the class of each sample is stored in the target attribute of the dataset: sklearn.linear_model.LinearRegression

>>> print(iris.target.shape) >>> from sklearn.linear_model import LinearRegression

(150,)
>>> print(iris.target) Estimator parameters: All the parameters of an estimator can be set when it is instantiated:
[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 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 >>> model = LinearRegression(normalize=True)
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 >>> print(model.normalize)
2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 True
2 2] >>> print(model)
LinearRegression(copy_X=True, fit_intercept=True, n_jobs=1, normalize=True)
The names of the classes are stored in the last attribute, namely target_names:

>>> print(iris.target_names) Fitting on data

['setosa' 'versicolor' 'virginica']
Let’s create some simple data with numpy:
This data is four-dimensional, but we can visualize two of the dimensions at a time using a scatter plot:
>>> import numpy as np
>>> x = np.array([0, 1, 2])
>>> y = np.array([0, 1, 2])

>>> X = x[:, np.newaxis] # The input data for sklearn is 2D: (samples == 3 x features == 1)
>>> X
array([[0],
[1],
[2]])

>>> model.fit(X, y)
LinearRegression(copy_X=True, fit_intercept=True, n_jobs=1, normalize=True)

Estimated parameters: When data is fitted with an estimator, parameters are estimated from the data at hand.
All the estimated parameters are attributes of the estimator object ending by an underscore:

>>> model.coef_
array([ 1.])

20.2.2 Supervised Learning: Classification and regression

In Supervised Learning, we have a dataset consisting of both features and labels. The task is to construct an
estimator which is able to predict the label of an object given the set of features. A relatively simple example
is predicting the species of iris given a set of measurements of its flower. This is a relatively simple task. Some
more complicated examples are:

Excercise: • given a multicolor image of an object through a telescope, determine whether that object is a star, a
quasar, or a galaxy.
Can you choose 2 features to find a plot where it is easier to seperate the different classes of irises? • given a photograph of a person, identify the person in the photo.
Hint: click on the figure above to see the code that generates it, and modify this code. • given a list of movies a person has watched and their personal rating of the movie, recommend a list of
movies they would like (So-called recommender systems: a famous example is the Netflix Prize).

Tip: What these tasks have in common is that there is one or more unknown quantities associated with the
object which needs to be determined from other observed quantities.

20.1. Introduction: problem settings 576 20.2. Basic principles of machine learning with scikit-learn 577
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Supervised learning is further broken down into two categories, classification and regression. In classifica-
tion, the label is discrete, while in regression, the label is continuous. For example, in astronomy, the task
of determining whether an object is a star, a galaxy, or a quasar is a classification problem: the label is from
three distinct categories. On the other hand, we might wish to estimate the age of an object based on such
observations: this would be a regression problem, because the label (age) is a continuous quantity.
Classification: K nearest neighbors (kNN) is one of the simplest learning strategies: given a new, unknown ob-
servation, look up in your reference database which ones have the closest features and assign the predominant
class. Let’s try it out on our iris classification problem:

from sklearn import neighbors, datasets

iris = datasets.load_iris()
X, y = iris.data, iris.target
knn = neighbors.KNeighborsClassifier(n_neighbors=1)
knn.fit(X, y)
# What kind of iris has 3cm x 5cm sepal and 4cm x 2cm petal?
print(iris.target_names[knn.predict([[3, 5, 4, 2]])])

Fig. 20.4: A plot of a simple linear regression.

• model.fit() : fit training data. For supervised learning applications, this accepts two
arguments: the data X and the labels y (e.g. model.fit(X, y)). For unsupervised learn-
ing applications, this accepts only a single argument, the data X (e.g. model.fit(X)).
In supervised estimators
• model.predict() : given a trained model, predict the label of a new set of data. This
method accepts one argument, the new data X_new (e.g. model.predict(X_new)), and
returns the learned label for each object in the array.
• model.predict_proba() : For classification problems, some estimators also provide
this method, which returns the probability that a new observation has each categor-
ical label. In this case, the label with the highest probability is returned by model.
predict().
• model.score() : for classification or regression problems, most (all?) estimators imple-
ment a score method. Scores are between 0 and 1, with a larger score indicating a better
fit.
In unsupervised estimators
• model.transform() : given an unsupervised model, transform new data into the new
basis. This also accepts one argument X_new, and returns the new representation of the
data based on the unsupervised model.
• model.fit_transform() : some estimators implement this method, which more effi-
ciently performs a fit and a transform on the same input data.

Fig. 20.3: A plot of the sepal space and the prediction of the KNN
20.2.4 Regularization: what it is and why it is necessary
Regression: The simplest possible regression setting is the linear regression one:
Prefering simpler models

20.2.3 A recap on Scikit-learn’s estimator interface Train errors Suppose you are using a 1-nearest neighbor estimator. How many errors do you expect on your
train set?
Scikit-learn strives to have a uniform interface across all methods, and we’ll see examples of these below. Given • Train set error is not a good measurement of prediction performance. You need to leave out a test set.
a scikit-learn estimator object named model, the following methods are available:
• In general, we should accept errors on the train set.
In all Estimators

20.2. Basic principles of machine learning with scikit-learn 578 20.2. Basic principles of machine learning with scikit-learn 579
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

An example of regularization The core idea behind regularization is that we are going to prefer models that
are simpler, for a certain definition of ‘’simpler’‘, even if they lead to more errors on the train set.
As an example, let’s generate with a 9th order polynomial, with noise:

the decision. k=1 amounts to no regularization: 0 error on the training set, whereas large k will push toward
smoother decision boundaries in the feature space.

And now, let’s fit a 4th order and a 9th order polynomial to the data.
Simple versus complex models for classification

A linear separation A non-linear separation

Tip: For classification models, the decision boundary, that separates the class expresses the complexity of the
With your naked eyes, which model do you prefer, the 4th order one, or the 9th order one? model. For instance, a linear model, that makes a decision based on a linear combination of features, is more
Let’s look at the ground truth: complex than a non-linear one.

Tip: Regularization is ubiquitous in machine learning. Most scikit-learn estimators have a parameter to tune
the amount of regularization. For instance, with k-NN, it is ‘k’, the number of nearest neighbors used to make

20.2. Basic principles of machine learning with scikit-learn 580 20.2. Basic principles of machine learning with scikit-learn 581
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

20.3 Supervised Learning: Classification of Handwritten Digits ax.imshow(digits.images[i], cmap=plt.cm.binary, interpolation='nearest')

20.3.1 The nature of the data

20.3.2 Visualizing the Data on its principal components

Code and notebook A good first-step for many problems is to visualize the data using a Dimensionality Reduction technique. We’ll
start with the most straightforward one, Principal Component Analysis (PCA).
Python code and Jupyter notebook for this section are found here
PCA seeks orthogonal linear combinations of the features which show the greatest variance, and as such, can
help give you a good idea of the structure of the data set.
In this section we’ll apply scikit-learn to the classification of handwritten digits. This will go a bit beyond
>>> from sklearn.decomposition import PCA
the iris classification we saw before: we’ll discuss some of the metrics which can be used in evaluating the
>>> pca = PCA(n_components=2)
effectiveness of a classification model.
>>> proj = pca.fit_transform(digits.data)
>>> from sklearn.datasets import load_digits >>> plt.scatter(proj[:, 0], proj[:, 1], c=digits.target)
>>> digits = load_digits() <matplotlib.collections.PathCollection object at ...>
>>> plt.colorbar()
<matplotlib.colorbar.Colorbar object at ...>

Question

Given these projections of the data, which numbers do you think a classifier might have trouble distinguish-
ing?
Let
us visualize the data and remind us what we’re looking at (click on the figure for the full code):

# plot the digits: each image is 8x8 pixels

for i in range(64):
ax = fig.add_subplot(8, 8, i + 1, xticks=[], yticks=[])

20.3. Supervised Learning: Classification of Handwritten Digits 582 20.3. Supervised Learning: Classification of Handwritten Digits 583
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

20.3.3 Gaussian Naive Bayes Classification

For most classification problems, it’s nice to have a simple, fast method to provide a quick baseline classifi-
cation. If the simple and fast method is sufficient, then we don’t have to waste CPU cycles on more complex
models. If not, we can use the results of the simple method to give us clues about our data.
One good method to keep in mind is Gaussian Naive Bayes (sklearn.naive_bayes.GaussianNB).

Old scikit-learn versions

train_test_split() is imported from sklearn.cross_validation

Tip: Gaussian Naive Bayes fits a Gaussian distribution to each training label independantly on each feature,
and uses this to quickly give a rough classification. It is generally not sufficiently accurate for real-world data,
but can perform surprisingly well, for instance on text data.

>>> from sklearn.naive_bayes import GaussianNB

>>> from sklearn.model_selection import train_test_split

>>> # split the data into training and validation sets

>>> X_train, X_test, y_train, y_test = train_test_split(digits.data, digits.target)

>>> # train the model

>>> clf = GaussianNB()
>>> clf.fit(X_train, y_train)
GaussianNB(priors=None)

>>> # use the model to predict the labels of the test data
>>> predicted = clf.predict(X_test)
>>> expected = y_test
>>> print(predicted)
[1 7 7 7 8 2 8 0 4 8 7 7 0 8 2 3 5 8 5 3 7 9 6 2 8 2 2 7 3 5...]
>>> print(expected)
[1 0 4 7 8 2 2 0 4 3 7 7 0 8 2 3 4 8 5 3 7 9 6 3 8 2 2 9 3 5...]

As above, we plot the digits with the predicted labels to get an idea of how well the classification is working.

Question

Why did we split the data into training and validation sets?

20.3.4 Quantitative Measurement of Performance

We’d like to measure the performance of our estimator without having to resort to plotting examples. A simple
method might be to simply compare the number of matches:

>>> matches = (predicted == expected)

>>> print(matches.sum())
367
>>> print(len(matches))
450
>>> matches.sum() / float(len(matches))
0.81555555555555559

We see that more than 80% of the 450 predictions match the input. But there are other more sophisticated
metrics that can be used to judge the performance of a classifier: several are available in the sklearn.metrics

20.3. Supervised Learning: Classification of Handwritten Digits 584 20.3. Supervised Learning: Classification of Handwritten Digits 585
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

submodule.
>>> print(data.target.shape)
One of the most useful metrics is the classification_report, which combines several measures and prints (506,)
a table with the results:
We can see that there are just over 500 data points.
>>> from sklearn import metrics
>>> print(metrics.classification_report(expected, predicted)) The DESCR variable has a long description of the dataset:
precision recall f1-score support
>>> print(data.DESCR)
Boston House Prices dataset
0 1.00 0.91 0.95 46
===========================
1 0.76 0.64 0.69 44
2 0.85 0.62 0.72 47
Notes
3 0.98 0.82 0.89 49
------
4 0.89 0.86 0.88 37
Data Set Characteristics:
5 0.97 0.93 0.95 41
6 1.00 0.98 0.99 44
:Number of Instances: 506
7 0.73 1.00 0.84 45
8 0.50 0.90 0.64 49
:Number of Attributes: 13 numeric/categorical predictive
9 0.93 0.54 0.68 48
:Median Value (attribute 14) is usually the target
avg / total 0.86 0.82 0.82 450
:Attribute Information (in order):
Another enlightening metric for this sort of multi-label classification is a confusion matrix: it helps us visualize - CRIM per capita crime rate by town
which labels are being interchanged in the classification errors: - ZN proportion of residential land zoned for lots over 25,000 sq.ft.
- INDUS proportion of non-retail business acres per town
>>> print(metrics.confusion_matrix(expected, predicted))
- CHAS Charles River dummy variable (= 1 if tract bounds river; 0 otherwise)
[[42 0 0 0 3 0 0 1 0 0]
- NOX nitric oxides concentration (parts per 10 million)
[ 0 28 0 0 0 0 0 1 13 2]
- RM average number of rooms per dwelling
[ 0 3 29 0 0 0 0 0 15 0]
- AGE proportion of owner-occupied units built prior to 1940
[ 0 0 2 40 0 0 0 2 5 0]
- DIS weighted distances to five Boston employment centres
[ 0 0 1 0 32 1 0 3 0 0]
- RAD index of accessibility to radial highways
[ 0 0 0 0 0 38 0 2 1 0]
- TAX full-value property-tax rate per $10,000
[ 0 0 1 0 0 0 43 0 0 0]
- PTRATIO pupil-teacher ratio by town
[ 0 0 0 0 0 0 0 45 0 0]
- B 1000(Bk - 0.63)^2 where Bk is the proportion of blacks by town
[ 0 3 1 0 0 0 0 1 44 0]
- LSTAT % lower status of the population
[ 0 3 0 1 1 0 0 7 10 26]]
- MEDV Median value of owner-occupied homes in $1000's
...
We see here that in particular, the numbers 1, 2, 3, and 9 are often being labeled 8.
It often helps to quickly visualize pieces of the data using histograms, scatter plots, or other plot types. With
pylab, let us show a histogram of the target values: the median price in each neighborhood:
20.4 Supervised Learning: Regression of Housing Data
>>> plt.hist(data.target)
(array([...
Here we’ll do a short example of a regression problem: learning a continuous value from a set of features.

20.4.1 A quick look at the data

Code and notebook

Python code and Jupyter notebook for this section are found here

We’ll use the simple Boston house prices set, available in scikit-learn. This records measurements of 13 at-
tributes of housing markets around Boston, as well as the median price. The question is: can you predict the
price of a new market given its attributes?:
Let’s have a quick look to see if some features are more rele-
>>> from sklearn.datasets import load_boston vant than others for our problem:
>>> data = load_boston()
>>> print(data.data.shape) >>> for index, feature_name in enumerate(data.feature_names):
(506, 13) ... plt.figure()
... plt.scatter(data.data[:, index], data.target)

20.4. Supervised Learning: Regression of Housing Data 586 20.4. Supervised Learning: Regression of Housing Data 587
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

<matplotlib.figure.Figure object...
Tip: The prediction at least correlates with the true price, though there are clearly some biases. We could
imagine evaluating the performance of the regressor by, say, computing the RMS residuals between the true
and predicted price. There are some subtleties in this, however, which we’ll cover in a later section.

Exercise: Gradient Boosting Tree Regression

There are many other types of regressors available in scikit-learn: we’ll try a more powerful one here.
Use the GradientBoostingRegressor class to fit the housing data.

This is a manual version of a technique called feature selection. hint You can copy and paste some of the above code, replacing LinearRegression with
GradientBoostingRegressor:
Tip: Sometimes, in Machine Learning it is useful to use feature selection to decide which features are the most from sklearn.ensemble import GradientBoostingRegressor
useful for a particular problem. Automated methods exist which quantify this sort of exercise of choosing the # Instantiate the model, fit the results, and scatter in vs. out
most informative features.
Solution The solution is found in the code of this chapter

20.4.2 Predicting Home Prices: a Simple Linear Regression

20.5 Measuring prediction performance
Now we’ll use scikit-learn to perform a simple linear regression on the housing data. There are many
possibilities of regressors to use. A particularly simple one is LinearRegression: this is basically a wrapper
around an ordinary least squares calculation.
20.5.1 A quick test on the K-neighbors classifier

>>> from sklearn.model_selection import train_test_split Here we’ll continue to look at the digits data, but we’ll switch to the K-Neighbors classifier. The K-neighbors
>>> X_train, X_test, y_train, y_test = train_test_split(data.data, data.target) classifier is an instance-based classifier. The K-neighbors classifier predicts the label of an unknown point
>>> from sklearn.linear_model import LinearRegression based on the labels of the K nearest points in the parameter space.
>>> clf = LinearRegression()
>>> clf.fit(X_train, y_train) >>> # Get the data
LinearRegression(copy_X=True, fit_intercept=True, n_jobs=1, normalize=False) >>> from sklearn.datasets import load_digits
>>> predicted = clf.predict(X_test) >>> digits = load_digits()
>>> expected = y_test >>> X = digits.data
>>> print("RMS: %s " % np.sqrt(np.mean((predicted - expected) ** 2))) >>> y = digits.target
RMS: 5.0059...
>>> # Instantiate and train the classifier
>>> from sklearn.neighbors import KNeighborsClassifier
>>> clf = KNeighborsClassifier(n_neighbors=1)
>>> clf.fit(X, y)
KNeighborsClassifier(...)

>>> # Check the results using metrics

>>> from sklearn import metrics
>>> y_pred = clf.predict(X)

>>> print(metrics.confusion_matrix(y_pred, y))

[[178 0 0 0 0 0 0 0 0 0]
[ 0 182 0 0 0 0 0 0 0 0]
[ 0 0 177 0 0 0 0 0 0 0]
[ 0 0 0 183 0 0 0 0 0 0]
[ 0 0 0 0 181 0 0 0 0 0]
[ 0 0 0 0 0 182 0 0 0 0]
[ 0 0 0 0 0 0 181 0 0 0]
[ 0 0 0 0 0 0 0 179 0 0]
[ 0 0 0 0 0 0 0 0 174 0]
We can plot the error: expected as a [ 0 0 0 0 0 0 0 0 0 180]]
function of predicted:
Apparently, we’ve found a perfect classifier! But this is misleading for the reasons we saw before: the classifier
>>> plt.scatter(expected, predicted) essentially “memorizes” all the samples it has already seen. To really test how well this algorithm does, we need
<matplotlib.collections.PathCollection object at ...>
to try some samples it hasn’t yet seen.

20.4. Supervised Learning: Regression of Housing Data 588 20.5. Measuring prediction performance 589
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

This problem also occurs with regression models. In the following we fit an other instance-based model named
“decision tree” to the Boston Housing price dataset we introduced previously: >>> print("%r , %r , %r " % (X.shape, X_train.shape, X_test.shape))
>>> from sklearn.datasets import load_boston (1797, 64), (1347, 64), (450, 64)
>>> from sklearn.tree import DecisionTreeRegressor
Now we train on the training data, and test on the testing data:
>>> data = load_boston()
>>> clf = DecisionTreeRegressor().fit(data.data, data.target) >>> clf = KNeighborsClassifier(n_neighbors=1).fit(X_train, y_train)
>>> predicted = clf.predict(data.data) >>> y_pred = clf.predict(X_test)
>>> expected = data.target
>>> print(metrics.confusion_matrix(y_test, y_pred))
>>> plt.scatter(expected, predicted) [[37 0 0 0 0 0 0 0 0 0]
<matplotlib.collections.PathCollection object at ...> [ 0 43 0 0 0 0 0 0 0 0]
>>> plt.plot([0, 50], [0, 50], '--k') [ 0 0 43 1 0 0 0 0 0 0]
[<matplotlib.lines.Line2D object at ...] [ 0 0 0 45 0 0 0 0 0 0]
[ 0 0 0 0 38 0 0 0 0 0]
[ 0 0 0 0 0 47 0 0 0 1]
[ 0 0 0 0 0 0 52 0 0 0]
[ 0 0 0 0 0 0 0 48 0 0]
[ 0 0 0 0 0 0 0 0 48 0]
[ 0 0 0 1 0 1 0 0 0 45]]
>>> print(metrics.classification_report(y_test, y_pred))
precision recall f1-score support

0 1.00 1.00 1.00 37

1 1.00 1.00 1.00 43
2 1.00 0.98 0.99 44
3 0.96 1.00 0.98 45
4 1.00 1.00 1.00 38
5 0.98 0.98 0.98 48
6 1.00 1.00 1.00 52
7 1.00 1.00 1.00 48
8 1.00 1.00 1.00 48
9 0.98 0.96 0.97 47

avg / total 0.99 0.99 0.99 450

Here again the predictions are seemingly perfect as the model was able to perfectly memorize the training set.
The averaged f1-score is often used as a convenient measure of the overall performance of an algorithm. It
appears in the bottom row of the classification report; it can also be accessed directly:
Warning: Performance on test set
>>> metrics.f1_score(y_test, y_pred, average="macro")
Performance on test set does not measure overfit (as described above) 0.991367...

The over-fitting we saw previously can be quantified by computing the f1-score on the training data itself:
20.5.2 A correct approach: Using a validation set >>> metrics.f1_score(y_train, clf.predict(X_train), average="macro")
1.0
Learning the parameters of a prediction function and testing it on the same data is a methodological mistake:
a model that would just repeat the labels of the samples that it has just seen would have a perfect score but
would fail to predict anything useful on yet-unseen data. Note: Regression metrics In the case of regression models, we need to use different metrics, such as explained
To avoid over-fitting, we have to define two different sets: variance.

• a training set X_train, y_train which is used for learning the parameters of a predictive model
• a testing set X_test, y_test which is used for evaluating the fitted predictive model
20.5.3 Model Selection via Validation
In scikit-learn such a random split can be quickly computed with the train_test_split() function:

>>> from sklearn import model_selection

Tip: We have applied Gaussian Naives, support vectors machines, and K-nearest neighbors classifiers to the
>>> X = digits.data
>>> y = digits.target digits dataset. Now that we have these validation tools in place, we can ask quantitatively which of the three
estimators works best for this dataset.
>>> X_train, X_test, y_train, y_test = model_selection.train_test_split(X, y,
... test_size=0.25, random_state=0)

20.5. Measuring prediction performance 590 20.5. Measuring prediction performance 591
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

• With the default hyper-parameters for each estimator, which gives the best f1 score on the validation account non iid datasets.
set? Recall that hyperparameters are the parameters set when you instantiate the classifier: for example,
the n_neighbors in clf = KNeighborsClassifier(n_neighbors=1)

>>> from sklearn.naive_bayes import GaussianNB 20.5.5 Hyperparameter optimization with cross-validation
>>> from sklearn.neighbors import KNeighborsClassifier
>>> from sklearn.svm import LinearSVC Consider regularized linear models, such as Ridge Regression, which uses l2 regularlization, and Lasso Regres-
sion, which uses l1 regularization. Choosing their regularization parameter is important.
>>> X = digits.data
>>> y = digits.target Let us set these parameters on the Diabetes dataset, a simple regression problem. The diabetes data consists
>>> X_train, X_test, y_train, y_test = model_selection.train_test_split(X, y, of 10 physiological variables (age, sex, weight, blood pressure) measure on 442 patients, and an indication of
... test_size=0.25, random_state=0) disease progression after one year:

>>> for Model in [GaussianNB, KNeighborsClassifier, LinearSVC]: >>> from sklearn.datasets import load_diabetes
... clf = Model().fit(X_train, y_train) >>> data = load_diabetes()
... y_pred = clf.predict(X_test) >>> X, y = data.data, data.target
... print('%s : %s ' % >>> print(X.shape)
... (Model.__name__, metrics.f1_score(y_test, y_pred, average="macro"))) (442, 10)
GaussianNB: 0.8332741681...
KNeighborsClassifier: 0.9804562804... With the default hyper-parameters: we compute the cross-validation score:
LinearSVC: 0.93...
>>> from sklearn.linear_model import Ridge, Lasso
• For each classifier, which value for the hyperparameters gives the best results for the digits data? For
LinearSVC, use loss='l2' and loss='l1'. For KNeighborsClassifier we use n_neighbors be- >>> for Model in [Ridge, Lasso]:
tween 1 and 10. Note that GaussianNB does not have any adjustable hyperparameters. ... model = Model()
... print('%s : %s ' % (Model.__name__, cross_val_score(model, X, y).mean()))
LinearSVC(loss='l1'): 0.930570687535 Ridge: 0.409427438303
LinearSVC(loss='l2'): 0.933068826918 Lasso: 0.353800083299
-------------------
KNeighbors(n_neighbors=1): 0.991367521884
KNeighbors(n_neighbors=2): 0.984844206884 Basic Hyperparameter Optimization
KNeighbors(n_neighbors=3): 0.986775344954
KNeighbors(n_neighbors=4): 0.980371905382
KNeighbors(n_neighbors=5): 0.980456280495 We compute the cross-validation score as a function of alpha, the strength of the regularization for Lasso and
KNeighbors(n_neighbors=6): 0.975792419414 Ridge. We choose 20 values of alpha between 0.0001 and 1:
KNeighbors(n_neighbors=7): 0.978064579214
>>> alphas = np.logspace(-3, -1, 30)
KNeighbors(n_neighbors=8): 0.978064579214
KNeighbors(n_neighbors=9): 0.978064579214
>>> for Model in [Lasso, Ridge]:
KNeighbors(n_neighbors=10): 0.975555089773
... scores = [cross_val_score(Model(alpha), X, y, cv=3).mean()
... for alpha in alphas]
Solution: code source ... plt.plot(alphas, scores, label=Model.__name__)
[<matplotlib.lines.Line2D object at ...

20.5.4 Cross-validation

Cross-validation consists in repetively splitting the data in pairs of train and test sets, called ‘folds’. Scikit-learn
comes with a function to automatically compute score on all these folds. Here we do KFold with k=5.

>>> clf = KNeighborsClassifier()

>>> from sklearn.model_selection import cross_val_score
>>> cross_val_score(clf, X, y, cv=5)
array([ 0.9478022 , 0.9558011 , 0.96657382, 0.98039216, 0.96338028])

We can use different splitting strategies, such as random splitting:

>>> from sklearn.model_selection import ShuffleSplit

>>> cv = ShuffleSplit(n_splits=5)
>>> cross_val_score(clf, X, y, cv=cv)
array([...])
Question

Can we trust our results to be actually useful?

Tip: There exists many different cross-validation strategies in scikit-learn. They are often useful to take in

20.5. Measuring prediction performance 592 20.5. Measuring prediction performance 593
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Automatically Performing Grid Search

>>> X = iris.data
>>> y = iris.target
sklearn.grid_search.GridSearchCV is constructed with an estimator, as well as a dictionary of parameter
values to be searched. We can find the optimal parameters this way:

>>> from sklearn.grid_search import GridSearchCV Tip: PCA computes linear combinations of the original features using a truncated Singular Value Decomposi-
>>> for Model in [Ridge, Lasso]: tion of the matrix X, to project the data onto a base of the top singular vectors.
... gscv = GridSearchCV(Model(), dict(alpha=alphas), cv=3).fit(X, y)
... print('%s : %s ' % (Model.__name__, gscv.best_params_))
Ridge: {'alpha': 0.062101694189156162} >>> from sklearn.decomposition import PCA
Lasso: {'alpha': 0.01268961003167922} >>> pca = PCA(n_components=2, whiten=True)
>>> pca.fit(X)
PCA(..., n_components=2, ...)

Built-in Hyperparameter Search Once fitted, PCA exposes the singular vectors in the components_ attribute:

For some models within scikit-learn, cross-validation can be performed more efficiently on large datasets. In >>> pca.components_
this case, a cross-validated version of the particular model is included. The cross-validated versions of Ridge array([[ 0.36158..., -0.08226..., 0.85657..., 0.35884...],
and Lasso are RidgeCV and LassoCV, respectively. Parameter search on these estimators can be performed as [ 0.65653..., 0.72971..., -0.17576..., -0.07470...]])
follows:
Other attributes are available as well:
>>> from sklearn.linear_model import RidgeCV, LassoCV
>>> for Model in [RidgeCV, LassoCV]: >>> pca.explained_variance_ratio_
... model = Model(alphas=alphas, cv=3).fit(X, y) array([ 0.92461..., 0.05301...])
... print('%s : %s ' % (Model.__name__, model.alpha_))
RidgeCV: 0.0621016941892 Let us project the iris dataset along those first two dimensions::
LassoCV: 0.0126896100317
>>> X_pca = pca.transform(X)
We see that the results match those returned by GridSearchCV >>> X_pca.shape
(150, 2)

Nested cross-validation PCA normalizes and whitens the data, which means that the data is now centered on both components with
unit variance:
How do we measure the performance of these estimators? We have used data to set the hyperparameters, so
we need to test on actually new data. We can do this by running cross_val_score() on our CV objects. Here >>> X_pca.mean(axis=0)
there are 2 cross-validation loops going on, this is called ‘nested cross validation’: array([ ...e-15, ...e-15])
>>> X_pca.std(axis=0)
for Model in [RidgeCV, LassoCV]: array([ 1., 1.])
scores = cross_val_score(Model(alphas=alphas, cv=3), X, y, cv=3)
print(Model.__name__, np.mean(scores)) Furthermore, the samples components do no longer carry any linear correlation:

>>> np.corrcoef(X_pca.T)
Note: Note that these results do not match the best results of our curves above, and LassoCV seems to under- array([[ 1.00000000e+00, ...],
[ ..., 1.00000000e+00]])
perform RidgeCV. The reason is that setting the hyper-parameter is harder for Lasso, thus the estimation error
on this hyper-parameter is larger.
With a number of retained components 2 or 3, PCA is useful to visualize the dataset:

>>> target_ids = range(len(iris.target_names))

>>> for i, c, label in zip(target_ids, 'rgbcmykw', iris.target_names):
20.6 Unsupervised Learning: Dimensionality Reduction and Visual- ... plt.scatter(X_pca[y == i, 0], X_pca[y == i, 1],
... c=c, label=label)
ization <matplotlib.collections.PathCollection ...

Unsupervised learning is applied on X without y: data without labels. A typical use case is to find hidden
structure in the data.

20.6.1 Dimensionality Reduction: PCA

Dimensionality reduction derives a set of new artificial features smaller than the original feature set. Here we’ll
use Principal Component Analysis (PCA), a dimensionality reduction that strives to retain most of the variance
of the original data. We’ll use sklearn.decomposition.PCA on the iris dataset:

20.6. Unsupervised Learning: Dimensionality Reduction and Visualization 594 20.6. Unsupervised Learning: Dimensionality Reduction and Visualization 595
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Tip: Note that this projection was determined without any information about the labels (represented by the fit_transform
colors): this is the sense in which the learning is unsupervised. Nevertheless, we see that the projection gives
us insight into the distribution of the different flowers in parameter space: notably, iris setosa is much more As TSNE cannot be applied to new data, we need to use its fit_transform method.
distinct than the other two species.

20.6.2 Visualization with a non-linear embedding: tSNE

For visualization, more complex embeddings can be useful (for statistical analysis, they are harder to control). sklearn.manifold.TSNE separates quite well the different classes of digits eventhough it had no access to
sklearn.manifold.TSNE is such a powerful manifold learning method. We apply it to the digits dataset, as the class information.
the digits are vectors of dimension 8*8 = 64. Embedding them in 2D enables visualization:

>>> # Take the first 500 data points: it's hard to see 1500 points Exercise: Other dimension reduction of digits
>>> X = digits.data[:500]
>>> y = digits.target[:500] sklearn.manifold has many other non-linear embeddings. Try them out on the digits dataset. Could you
judge their quality without knowing the labels y?
>>> # Fit and transform with a TSNE
>>> from sklearn.manifold import TSNE >>> from sklearn.datasets import load_digits
>>> tsne = TSNE(n_components=2, random_state=0) >>> digits = load_digits()
>>> X_2d = tsne.fit_transform(X) >>> # ...

>>> # Visualize the data

>>> plt.scatter(X_2d[:, 0], X_2d[:, 1], c=y)
<matplotlib.collections.PathCollection object at ...>
20.7 The eigenfaces example: chaining PCA and SVMs

Code and notebook

Python code and Jupyter notebook for this section are found here

The goal of this example is to show how an unsupervised method and a supervised one can be chained for
better prediction. It starts with a didactic but lengthy way of doing things, and finishes with the idiomatic
approach to pipelining in scikit-learn.

20.6. Unsupervised Learning: Dimensionality Reduction and Visualization 596 20.7. The eigenfaces example: chaining PCA and SVMs 597
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Here we’ll take a look at a simple facial recognition example. Ideally, we would use a dataset consisting of a
subset of the Labeled Faces in the Wild data that is available with sklearn.datasets.fetch_lfw_people(). print(X_train.shape, X_test.shape)
However, this is a relatively large download (~200MB) so we will do the tutorial on a simpler, less rich dataset.
Feel free to explore the LFW dataset.
Out:
from sklearn import datasets
(300, 4096) (100, 4096)
faces = datasets.fetch_olivetti_faces()
faces.data.shape

Let’s visualize these faces to see what we’re working with 20.7.1 Preprocessing: Principal Component Analysis
from matplotlib import pyplot as plt 1850 dimensions is a lot for SVM. We can use PCA to reduce these 1850 features to a manageable size, while
fig = plt.figure(figsize=(8, 6)) maintaining most of the information in the dataset.
# plot several images
for i in range(15): from sklearn import decomposition
ax = fig.add_subplot(3, 5, i + 1, xticks=[], yticks=[]) pca = decomposition.PCA(n_components=150, whiten=True)
ax.imshow(faces.images[i], cmap=plt.cm.bone) pca.fit(X_train)

One interesting part of PCA is that it computes the “mean” face, which can be interesting to examine:

plt.imshow(pca.mean_.reshape(faces.images[0].shape),
cmap=plt.cm.bone)

Tip: Note is that these faces have already been localized and scaled to a common size. This is an important
preprocessing piece for facial recognition, and is a process that can require a large collection of training data.
This can be done in scikit-learn, but the challenge is gathering a sufficient amount of training data for the
algorithm to work. Fortunately, this piece is common enough that it has been done. One good resource is
OpenCV, the Open Computer Vision Library. The principal components measure deviations about this mean along orthogonal axes.

print(pca.components_.shape)
We’ll perform a Support Vector classification of the images. We’ll do a typical train-test split on the images:
Out:
from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(faces.data, (150, 4096)
faces.target, random_state=0)

20.7. The eigenfaces example: chaining PCA and SVMs 598 20.7. The eigenfaces example: chaining PCA and SVMs 599
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

It is also interesting to visualize these principal components:

ax = fig.add_subplot(3, 5, i + 1, xticks=[], yticks=[])
fig = plt.figure(figsize=(16, 6)) ax.imshow(X_test[i].reshape(faces.images[0].shape),
for i in range(30): cmap=plt.cm.bone)
ax = fig.add_subplot(3, 10, i + 1, xticks=[], yticks=[]) y_pred = clf.predict(X_test_pca[i, np.newaxis])[0]
ax.imshow(pca.components_[i].reshape(faces.images[0].shape), color = ('black' if y_pred == y_test[i] else 'red')
cmap=plt.cm.bone) ax.set_title(faces.target[y_pred],
fontsize='small', color=color)

The components (“eigenfaces”) are ordered by their importance from top-left to bottom-right. We see that the
first few components seem to primarily take care of lighting conditions; the remaining components pull out
certain identifying features: the nose, eyes, eyebrows, etc.
With this projection computed, we can now project our original training and test data onto the PCA basis:

X_train_pca = pca.transform(X_train)
X_test_pca = pca.transform(X_test)
print(X_train_pca.shape)

Out:

(300, 150)

print(X_test_pca.shape)

Out: The classifier is correct on an impressive number of images given the simplicity of its learning model! Using
a linear classifier on 150 features derived from the pixel-level data, the algorithm correctly identifies a large
(100, 150) number of the people in the images.
Again, we can quantify this effectiveness using one of several measures from sklearn.metrics. First we can
These projected components correspond to factors in a linear combination of component images such that
do the classification report, which shows the precision, recall and other measures of the “goodness” of the
the combination approaches the original face.
classification:

from sklearn import metrics

20.7.2 Doing the Learning: Support Vector Machines y_pred = clf.predict(X_test_pca)
print(metrics.classification_report(y_test, y_pred))
Now we’ll perform support-vector-machine classification on this reduced dataset:
Out:
from sklearn import svm
clf = svm.SVC(C=5., gamma=0.001) precision recall f1-score support
clf.fit(X_train_pca, y_train)
0 1.00 0.67 0.80 6
Finally, we can evaluate how well this classification did. First, we might plot a few of the test-cases with the 1 1.00 1.00 1.00 4
labels learned from the training set: 2 0.50 1.00 0.67 2
3 1.00 1.00 1.00 1
import numpy as np 4 1.00 1.00 1.00 1
fig = plt.figure(figsize=(8, 6)) 5 1.00 1.00 1.00 5
for i in range(15): 6 1.00 1.00 1.00 4

20.7. The eigenfaces example: chaining PCA and SVMs 600 20.7. The eigenfaces example: chaining PCA and SVMs 601
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

7 1.00 0.67 0.80 3 clf.fit(X_train, y_train)

9 1.00 1.00 1.00 1
10 1.00 1.00 1.00 4 y_pred = clf.predict(X_test)
11 1.00 1.00 1.00 1 print(metrics.confusion_matrix(y_pred, y_test))
12 0.67 1.00 0.80 2
13 1.00 1.00 1.00 3
14 1.00 1.00 1.00 5
15 1.00 1.00 1.00 3 20.8 Parameter selection, Validation, and Testing
17 1.00 1.00 1.00 6
19 1.00 1.00 1.00 4
20 1.00 1.00 1.00 1 20.8.1 Hyperparameters, Over-fitting, and Under-fitting
21 1.00 1.00 1.00 1
22 1.00 1.00 1.00 2 See also:
23 0.50 1.00 0.67 1
24 1.00 1.00 1.00 2 This section is adapted from Andrew Ng’s excellent Coursera course
25 1.00 0.50 0.67 2
The issues associated with validation and cross-validation are some of the most important aspects of the prac-
26 1.00 0.75 0.86 4
27 1.00 1.00 1.00 1 tice of machine learning. Selecting the optimal model for your data is vital, and is a piece of the problem that
28 0.67 1.00 0.80 2 is not often appreciated by machine learning practitioners.
29 1.00 1.00 1.00 3 The central question is: If our estimator is underperforming, how should we move forward?
30 1.00 1.00 1.00 4
31 1.00 1.00 1.00 3 • Use simpler or more complicated model?
32 1.00 1.00 1.00 3
33 1.00 1.00 1.00 2 • Add more features to each observed data point?
34 1.00 1.00 1.00 3 • Add more training samples?
35 1.00 1.00 1.00 1
36 1.00 1.00 1.00 3 The answer is often counter-intuitive. In particular, Sometimes using a more complicated model will give
37 1.00 1.00 1.00 3 worse results. Also, Sometimes adding training data will not improve your results. The ability to determine
38 1.00 1.00 1.00 1 what steps will improve your model is what separates the successful machine learning practitioners from the
39 1.00 1.00 1.00 3 unsuccessful.

avg / total 0.97 0.95 0.95 100

Bias-variance trade-off: illustration on a simple regression problem
Another interesting metric is the confusion matrix, which indicates how often any two items are mixed-up.
The confusion matrix of a perfect classifier would only have nonzero entries on the diagonal, with zeros on the
off-diagonal: Code and notebook

print(metrics.confusion_matrix(y_test, y_pred)) Python code and Jupyter notebook for this section are found here

Out:
Let us start with a simple 1D regression problem. This will help us to easily visualize the data and the model,
[[4 0 0 ..., 0 0 0] and the results generalize easily to higher-dimensional datasets. We’ll explore a simple linear regression prob-
[0 4 0 ..., 0 0 0] lem, with sklearn.linear_model.
[0 0 2 ..., 0 0 0]
..., X = np.c_[ .5, 1].T
[0 0 0 ..., 3 0 0] y = [.5, 1]
[0 0 0 ..., 0 1 0] X_test = np.c_[ 0, 2].T
[0 0 0 ..., 0 0 3]]
Without noise, as linear regression fits the data perfectly

from sklearn import linear_model

20.7.3 Pipelining regr = linear_model.LinearRegression()
regr.fit(X, y)
Above we used PCA as a pre-processing step before applying our support vector machine classifier. Plugging plt.plot(X, y, 'o')
the output of one estimator directly into the input of a second estimator is a commonly used pattern; for this plt.plot(X_test, regr.predict(X_test))
reason scikit-learn provides a Pipeline object which automates this process. The above problem can be re-
expressed as a pipeline as follows:

from sklearn.pipeline import Pipeline

clf = Pipeline([('pca', decomposition.PCA(n_components=150, whiten=True)),
('svm', svm.LinearSVC(C=1.0))])

20.7. The eigenfaces example: chaining PCA and SVMs 602 20.8. Parameter selection, Validation, and Testing 603
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

In real life situation, we have noise (e.g. measurement noise) in our data: As we can see, the estimator displays much less variance. However it systematically under-estimates the coef-
ficient. It displays a biased behavior.
np.random.seed(0)
for _ in range(6): This is a typical example of bias/variance tradeof: non-regularized estimator are not biased, but they can
noisy_X = X + np.random.normal(loc=0, scale=.1, size=X.shape) display a lot of bias. Highly-regularized models have little variance, but high bias. This bias is not necessarily
plt.plot(noisy_X, y, 'o') a bad thing: what matters is choosing the tradeoff between bias and variance that leads to the best prediction
regr.fit(noisy_X, y) performance. For a specific dataset there is a sweet spot corresponding to the highest complexity that the data
plt.plot(X_test, regr.predict(X_test)) can support, depending on the amount of noise and of observations available.

20.8.2 Visualizing the Bias/Variance Tradeoff

Tip: Given a particular dataset and a model (e.g. a polynomial), we’d like to understand whether bias (underfit)
or variance limits prediction, and how to tune the hyperparameter (here d, the degree of the polynomial) to give
the best fit.

On a given data, let us fit a simple polynomial regression model with varying degrees:

As we can see, our linear model captures and amplifies the noise in the data. It displays a lot of variance.
We can use another linear estimator that uses regularization, the Ridge estimator. This estimator regularizes
the coefficients by shrinking them to zero, under the assumption that very high correlations are often spurious.
The alpha parameter controls the amount of shrinkage used.

regr = linear_model.Ridge(alpha=.1)
np.random.seed(0)
for _ in range(6):
noisy_X = X + np.random.normal(loc=0, scale=.1, size=X.shape)
plt.plot(noisy_X, y, 'o')
regr.fit(noisy_X, y)
plt.plot(X_test, regr.predict(X_test))

plt.show()
Tip: In the above figure, we see fits for three different values of d. For d = 1, the data is under-fit. This means
that the model is too simplistic: no straight line will ever be a good fit to this data. In this case, we say that the
model suffers from high bias. The model itself is biased, and this will be reflected in the fact that the data is
poorly fit. At the other extreme, for d = 6 the data is over-fit. This means that the model has too many free
parameters (6 in this case) which can be adjusted to perfectly fit the training data. If we add a new point to this
plot, though, chances are it will be very far from the curve representing the degree-6 fit. In this case, we say
that the model suffers from high variance. The reason for the term “high variance” is that if any of the input
points are varied slightly, it could result in a very different model.
In the middle, for d = 2, we have found a good mid-point. It fits the data fairly well, and does not suffer from
the bias and variance problems seen in the figures on either side. What we would like is a way to quantitatively

20.8. Parameter selection, Validation, and Testing 604 20.8. Parameter selection, Validation, and Testing 605
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

identify bias and variance, and optimize the metaparameters (in this case, the polynomial degree d) in order
... model, x[:, np.newaxis], y,
to determine the best algorithm. ... param_name='polynomialfeatures__degree',
... param_range=degrees)

>>> # Plot the mean train score and validation score across folds
Polynomial regression with scikit-learn
>>> plt.plot(degrees, validation_scores.mean(axis=1), label='cross-validation')
[<matplotlib.lines.Line2D object at ...>]
A polynomial regression is built by pipelining PolynomialFeatures and a LinearRegression: >>> plt.plot(degrees, train_scores.mean(axis=1), label='training')
>>> from sklearn.pipeline import make_pipeline [<matplotlib.lines.Line2D object at ...>]
>>> from sklearn.preprocessing import PolynomialFeatures >>> plt.legend(loc='best')
>>> from sklearn.linear_model import LinearRegression <matplotlib.legend.Legend object at ...>
>>> model = make_pipeline(PolynomialFeatures(degree=2), LinearRegression())

Validation Curves

Let us create a dataset like in the example above:

>>> def generating_func(x, err=0.5):

... return np.random.normal(10 - 1. / (x + 0.1), err)

>>> # randomly sample more data

>>> np.random.seed(1)
>>> x = np.random.random(size=200)
>>> y = generating_func(x, err=1.)
This figure shows why validation is important.
On the left side of the plot, we have very low-degree polynomial, which under-fit the data. This leads to a low
explained variance for both the training set and the validation set. On the far right side of the plot, we have a
very high degree polynomial, which over-fits the data. This can be seen in the fact that the training explained
variance is very high, while on the validation set, it is low. Choosing d around 4 or 5 gets us the best tradeoff.

Tip: The astute reader will realize that something is amiss here: in the above plot, d = 4 gives the best results.
But in the previous plot, we found that d = 6 vastly over-fits the data. What’s going on here? The difference
is the number of training points used. In the previous example, there were only eight training points. In this
example, we have 100. As a general rule of thumb, the more training points used, the more complicated model
can be used. But how can you determine for a given model whether more training points will be helpful? A
useful diagnostic for this are learning curves.

Central to quantify bias and variance of a

model is to apply it on test data, sampled from the same distribution as the train, but that will capture in- Learning Curves
dependent noise:
A learning curve shows the training and validation score as a function of the number of training points. Note
>>> xtrain, xtest, ytrain, ytest = train_test_split(x, y, test_size=0.4) that when we train on a subset of the training data, the training score is computed using this subset, not the
full training set. This curve gives a quantitative view into how beneficial it will be to add training samples.
Validation curve A validation curve consists in varying a model parameter that controls its complexity (here
the degree of the polynomial) and measures both error of the model on training data, and on test data (eg with
Questions:
cross-validation). The model parameter is then adjusted so that the test error is minimized:
We use sklearn.model_selection.validation_curve() to compute train and test error, and plot it: • As the number of training samples are increased, what do you expect to see for the training score? For
the validation score?
>>> from sklearn.model_selection import validation_curve
• Would you expect the training score to be higher or lower than the validation score? Would you ever
>>> degrees = np.arange(1, 21) expect this to change?

>>> model = make_pipeline(PolynomialFeatures(), LinearRegression())

scikit-learn provides sklearn.model_selection.learning_curve():
>>> # Vary the "degrees" on the pipeline step "polynomialfeatures"
>>> train_scores, validation_scores = validation_curve(

20.8. Parameter selection, Validation, and Testing 606 20.8. Parameter selection, Validation, and Testing 607
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

>>> from sklearn.model_selection import learning_curve

>>> train_sizes, train_scores, validation_scores = learning_curve(
... model, x[:, np.newaxis], y, train_sizes=np.logspace(-1, 0, 20))

>>> # Plot the mean train score and validation score across folds
>>> plt.plot(train_sizes, validation_scores.mean(axis=1), label='cross-validation')
[<matplotlib.lines.Line2D object at ...>]
>>> plt.plot(train_sizes, train_scores.mean(axis=1), label='training')
[<matplotlib.lines.Line2D object at ...>]

Fig. 20.6: For a degree=15 model

Here we show the learning curve for d = 15. From the above discussion, we know that d = 15 is a high-
variance estimator which over-fits the data. This is indicated by the fact that the training score is much higher
than the validation score. As we add more samples to this training set, the training score will continue to
decrease, while the cross-validation error will continue to increase, until they meet in the middle.
Learning curves that have not yet converged with the full training set indicate a high-variance, over-fit
model.
A high-variance model can be improved by:
Fig. 20.5: For a degree=1 model
• Gathering more training samples.

Note that the validation score generally increases with a growing training set, while the training score generally • Using a less-sophisticated model (i.e. in this case, make d smaller)
decreases with a growing training set. As the training size increases, they will converge to a single value. • Increasing regularization.
From the above discussion, we know that d = 1 is a high-bias estimator which under-fits the data. This is In particular, gathering more features for each sample will not help the results.
indicated by the fact that both the training and validation scores are low. When confronted with this type of
learning curve, we can expect that adding more training data will not help: both lines converge to a relatively
low score. 20.8.3 Summary on model selection
When the learning curves have converged to a low score, we have a high bias model.
We’ve seen above that an under-performing algorithm can be due to two possible situations: high bias (under-
A high-bias model can be improved by: fitting) and high variance (over-fitting). In order to evaluate our algorithm, we set aside a portion of our training
• Using a more sophisticated model (i.e. in this case, increase d) data for cross-validation. Using the technique of learning curves, we can train on progressively larger subsets
of the data, evaluating the training error and cross-validation error to determine whether our algorithm has
• Gather more features for each sample. high variance or high bias. But what do we do with this information?
• Decrease regularization in a regularized model.
Increasing the number of samples, however, does not improve a high-bias model. High Bias

Now let’s look at a high-variance (i.e. over-fit) model: If a model shows high bias, the following actions might help:
• Add more features. In our example of predicting home prices, it may be helpful to make use of infor-
mation such as the neighborhood the house is in, the year the house was built, the size of the lot, etc.
Adding these features to the training and test sets can improve a high-bias estimator
• Use a more sophisticated model. Adding complexity to the model can help improve on bias. For a
polynomial fit, this can be accomplished by increasing the degree d. Each learning technique has its
own methods of adding complexity.
• Use fewer samples. Though this will not improve the classification, a high-bias algorithm can attain
nearly the same error with a smaller training sample. For algorithms which are computationally expen-
sive, reducing the training sample size can lead to very large improvements in speed.

20.8. Parameter selection, Validation, and Testing 608 20.8. Parameter selection, Validation, and Testing 609
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

• Decrease regularization. Regularization is a technique used to impose simplicity in some machine

from sklearn.tree import DecisionTreeRegressor
learning models, by adding a penalty term that depends on the characteristics of the parameters. If a clf = DecisionTreeRegressor().fit(data.data, data.target)
model has high bias, decreasing the effect of regularization can lead to better results.
predicted = clf.predict(data.data)
expected = data.target
High Variance
Plot predicted as a function of expected
If a model shows high variance, the following actions might help:
from matplotlib import pyplot as plt
• Use fewer features. Using a feature selection technique may be useful, and decrease the over-fitting of
plt.figure(figsize=(4, 3))
the estimator.
plt.scatter(expected, predicted)
• Use a simpler model. Model complexity and over-fitting go hand-in-hand. plt.plot([0, 50], [0, 50], '--k')
plt.axis('tight')
• Use more training samples. Adding training samples can reduce the effect of over-fitting, and lead to plt.xlabel('True price ($1000s)')
improvements in a high variance estimator. plt.ylabel('Predicted price ($1000s)')
plt.tight_layout()
• Increase Regularization. Regularization is designed to prevent over-fitting. In a high-variance model,
increasing regularization can lead to better results.
These choices become very important in real-world situations. For example, due to limited telescope time,
astronomers must seek a balance between observing a large number of objects, and observing a large number
of features for each object. Determining which is more important for a particular learning task can inform the
observing strategy that the astronomer employs.

20.8.4 A last word of caution: separate validation and test set

Using validation schemes to determine hyper-parameters means that we are fitting the hyper-parameters to
the particular validation set. In the same way that parameters can be over-fit to the training set, hyperpa-
rameters can be over-fit to the validation set. Because of this, the validation error tends to under-predict the
classification error of new data.
For this reason, it is recommended to split the data into three sets:
• The training set, used to train the model (usually ~60% of the data)
• The validation set, used to validate the model (usually ~20% of the data)
• The test set, used to evaluate the expected error of the validated model (usually ~20% of the data)
Pretty much no errors!
Many machine learning practitioners do not separate test set and validation set. But if your goal is to gauge
the error of a model on unknown data, using an independent test set is vital. This is too good to be true: we are testing the model on the train data, which is not a mesure of generalization.
The results are not valid
Total running time of the script: ( 0 minutes 0.069 seconds)
Download Python source code: plot_measuring_performance.py
Download Jupyter notebook: plot_measuring_performance.ipynb
20.9 Examples for the scikit-learn chapter
Generated by Sphinx-Gallery

20.9.1 Measuring Decision Tree performance

20.9.2 Demo PCA in 2D
Demonstrates overfit when testing on train set.
Load the iris data
Get the data
from sklearn import datasets
from sklearn.datasets import load_boston
iris = datasets.load_iris()
data = load_boston()
X = iris.data
y = iris.target
Train and test a model
Fit a PCA

20.9. Examples for the scikit-learn chapter 610 20.9. Examples for the scikit-learn chapter 611
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

from sklearn.decomposition import PCA

20.9.3 A simple linear regression
pca = PCA(n_components=2, whiten=True)
pca.fit(X)

Project the data in 2D

X_pca = pca.transform(X)

Visualize the data

target_ids = range(len(iris.target_names))

from matplotlib import pyplot as plt

plt.figure(figsize=(6, 5))
for i, c, label in zip(target_ids, 'rgbcmykw', iris.target_names):
plt.scatter(X_pca[y == i, 0], X_pca[y == i, 1],
c=c, label=label)
plt.legend()
plt.show()

import numpy as np
import matplotlib.pyplot as plt
from sklearn.linear_model import LinearRegression

# x from 0 to 30
x = 30 * np.random.random((20, 1))

# y = a*x + b with noise

y = 0.5 * x + 1.0 + np.random.normal(size=x.shape)

# create a linear regression model

model = LinearRegression()
model.fit(x, y)

# predict y from the data

x_new = np.linspace(0, 30, 100)
y_new = model.predict(x_new[:, np.newaxis])

# plot the results

plt.figure(figsize=(4, 3))
ax = plt.axes()
ax.scatter(x, y)
ax.plot(x_new, y_new)

ax.set_xlabel('x')
ax.set_ylabel('y')

ax.axis('tight')

plt.show()
Total running time of the script: ( 0 minutes 0.043 seconds)
Total running time of the script: ( 0 minutes 0.122 seconds)
Download Python source code: plot_pca.py
Download Python source code: plot_linear_regression.py
Download Jupyter notebook: plot_pca.ipynb
Download Jupyter notebook: plot_linear_regression.ipynb
Generated by Sphinx-Gallery
Generated by Sphinx-Gallery

20.9. Examples for the scikit-learn chapter 612 20.9. Examples for the scikit-learn chapter 613
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

20.9.4 Plot 2D views of the iris dataset 20.9.5 tSNE to visualize digits

Plot a simple scatter plot of 2 features of the iris dataset. Here we use sklearn.manifold.TSNE to visualize the digits datasets. Indeed, the digits are vectors in a 8*8
= 64 dimensional space. We want to project them in 2D for visualization. tSNE is often a good solution, as it
Note that more elaborate visualization of this dataset is detailed in the Statistics in Python chapter.
groups and separates data points based on their local relationship.
Load the iris data

from sklearn import datasets

digits = datasets.load_digits()
# Take the first 500 data points: it's hard to see 1500 points
X = digits.data[:500]
y = digits.target[:500]

Fit and transform with a TSNE

from sklearn.manifold import TSNE

tsne = TSNE(n_components=2, random_state=0)

Project the data in 2D

X_2d = tsne.fit_transform(X)

Visualize the data

target_ids = range(len(digits.target_names))

from matplotlib import pyplot as plt

plt.figure(figsize=(6, 5))
colors = 'r', 'g', 'b', 'c', 'm', 'y', 'k', 'w', 'orange', 'purple'
for i, c, label in zip(target_ids, colors, digits.target_names):
plt.scatter(X_2d[y == i, 0], X_2d[y == i, 1], c=c, label=label)
plt.legend()
# Load the data plt.show()
from sklearn.datasets import load_iris
iris = load_iris()

from matplotlib import pyplot as plt

# The indices of the features that we are plotting

x_index = 0
y_index = 1

# this formatter will label the colorbar with the correct target names
formatter = plt.FuncFormatter(lambda i, *args: iris.target_names[int(i)])

plt.figure(figsize=(5, 4))
plt.scatter(iris.data[:, x_index], iris.data[:, y_index], c=iris.target)
plt.colorbar(ticks=[0, 1, 2], format=formatter)
plt.xlabel(iris.feature_names[x_index])
plt.ylabel(iris.feature_names[y_index])

plt.tight_layout()
plt.show()

Total running time of the script: ( 0 minutes 0.105 seconds)

Download Python source code: plot_iris_scatter.py
Download Jupyter notebook: plot_iris_scatter.ipynb
Generated by Sphinx-Gallery

20.9. Examples for the scikit-learn chapter 614 20.9. Examples for the scikit-learn chapter 615
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

print('%s : %s ' % (Model.__name__,

cross_val_score(model, X, y).mean()))

Out:

Ridge: 0.409427438303
Lasso: 0.353800083299

We compute the cross-validation score as a function of alpha, the strength of the regularization for Lasso and
Ridge

import numpy as np
from matplotlib import pyplot as plt

alphas = np.logspace(-3, -1, 30)

plt.figure(figsize=(5, 3))

for Model in [Lasso, Ridge]:

scores = [cross_val_score(Model(alpha), X, y, cv=3).mean()
for alpha in alphas]
plt.plot(alphas, scores, label=Model.__name__)

plt.legend(loc='lower left')
plt.xlabel('alpha')
plt.ylabel('cross validation score')
plt.tight_layout()
plt.show()

Total running time of the script: ( 0 minutes 1.713 seconds)

Download Python source code: plot_tsne.py
Download Jupyter notebook: plot_tsne.ipynb
Generated by Sphinx-Gallery

20.9.6 Use the RidgeCV and LassoCV to set the regularization parameter

Load the diabetes dataset

from sklearn.datasets import load_diabetes

data = load_diabetes()
X, y = data.data, data.target
print(X.shape)

Total running time of the script: ( 0 minutes 0.316 seconds)

Out:
Download Python source code: plot_linear_model_cv.py
(442, 10)
Download Jupyter notebook: plot_linear_model_cv.ipynb
Compute the cross-validation score with the default hyper-parameters Generated by Sphinx-Gallery
from sklearn.model_selection import cross_val_score
from sklearn.linear_model import Ridge, Lasso
20.9.7 Plot variance and regularization in linear models
for Model in [Ridge, Lasso]:
model = Model() import numpy as np

20.9. Examples for the scikit-learn chapter 616 20.9. Examples for the scikit-learn chapter 617
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

# Smaller figures for _ in range(6):

from matplotlib import pyplot as plt noisy_X = X + np.random.normal(loc=0, scale=.1, size=X.shape)
plt.rcParams['figure.figsize'] = (3, 2) plt.plot(noisy_X, y, 'o')
regr.fit(noisy_X, y)
We consider the situation where we have only 2 data point plt.plot(X_test, regr.predict(X_test))

X = np.c_[ .5, 1].T plt.show()

y = [.5, 1]
X_test = np.c_[ 0, 2].T

Without noise, as linear regression fits the data perfectly

from sklearn import linear_model

regr = linear_model.LinearRegression()
regr.fit(X, y)
plt.plot(X, y, 'o')
plt.plot(X_test, regr.predict(X_test))

Total running time of the script: ( 0 minutes 0.125 seconds)

Download Python source code: plot_variance_linear_regr.py
Download Jupyter notebook: plot_variance_linear_regr.ipynb
Generated by Sphinx-Gallery

20.9.8 Simple picture of the formal problem of machine learning

In real life situation, we have noise (e.g. measurement noise) in our data: This example generates simple synthetic data ploints and shows a separating hyperplane on them.

np.random.seed(0)
for _ in range(6):
noisy_X = X + np.random.normal(loc=0, scale=.1, size=X.shape)
plt.plot(noisy_X, y, 'o')
regr.fit(noisy_X, y)
plt.plot(X_test, regr.predict(X_test))

As we can see, our linear model captures and amplifies the noise in the data. It displays a lot of variance. import numpy as np
import matplotlib.pyplot as plt
We can use another linear estimator that uses regularization, the Ridge estimator. This estimator regularizes from sklearn.linear_model import SGDClassifier
the coefficients by shrinking them to zero, under the assumption that very high correlations are often spurious. from sklearn.datasets.samples_generator import make_blobs
The alpha parameter controls the amount of shrinkage used.
# we create 50 separable synthetic points
regr = linear_model.Ridge(alpha=.1) X, Y = make_blobs(n_samples=50, centers=2,
np.random.seed(0) random_state=0, cluster_std=0.60)

20.9. Examples for the scikit-learn chapter 618 20.9. Examples for the scikit-learn chapter 619
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

# fit the model

clf = SGDClassifier(loss="hinge", alpha=0.01,
n_iter=200, fit_intercept=True) from sklearn import model_selection, datasets, metrics
clf.fit(X, Y) from sklearn.svm import LinearSVC
from sklearn.naive_bayes import GaussianNB
# plot the line, the points, and the nearest vectors to the plane from sklearn.neighbors import KNeighborsClassifier
xx = np.linspace(-1, 5, 10)
yy = np.linspace(-1, 5, 10) digits = datasets.load_digits()
X = digits.data
X1, X2 = np.meshgrid(xx, yy) y = digits.target
Z = np.empty(X1.shape) X_train, X_test, y_train, y_test = model_selection.train_test_split(X, y,
for (i, j), val in np.ndenumerate(X1): test_size=0.25, random_state=0)
x1 = val
x2 = X2[i, j] for Model in [LinearSVC, GaussianNB, KNeighborsClassifier]:
p = clf.decision_function([[x1, x2]]) clf = Model().fit(X_train, y_train)
Z[i, j] = p[0] y_pred = clf.predict(X_test)
print('%s : %s ' %
plt.figure(figsize=(4, 3)) (Model.__name__, metrics.f1_score(y_test, y_pred, average="macro")))
ax = plt.axes()
ax.contour(X1, X2, Z, [-1.0, 0.0, 1.0], colors='k', print('------------------')
linestyles=['dashed', 'solid', 'dashed'])
ax.scatter(X[:, 0], X[:, 1], c=Y, cmap=plt.cm.Paired) # test SVC loss
for loss in ['l1', 'l2']:
ax.axis('tight') clf = LinearSVC(loss=loss).fit(X_train, y_train)
y_pred = clf.predict(X_test)
print("LinearSVC(loss='{0}'): {1}".format(loss,
plt.show() metrics.f1_score(y_test, y_pred, average="macro")))

Total running time of the script: ( 0 minutes 0.046 seconds) print('-------------------')

Download Python source code: plot_separator.py # test the number of neighbors

for n_neighbors in range(1, 11):
Download Jupyter notebook: plot_separator.ipynb
clf = KNeighborsClassifier(n_neighbors=n_neighbors).fit(X_train, y_train)
Generated by Sphinx-Gallery y_pred = clf.predict(X_test)
print("KNeighbors(n_neighbors={0}): {1}".format(n_neighbors,
metrics.f1_score(y_test, y_pred, average="macro")))
20.9.9 Compare classifiers on the digits data
Total running time of the script: ( 0 minutes 1.016 seconds)
Compare the performance of a variety of classifiers on a test set for the digits data. Download Python source code: plot_compare_classifiers.py
Out: Download Jupyter notebook: plot_compare_classifiers.ipynb
LinearSVC: 0.927190026916 Generated by Sphinx-Gallery
GaussianNB: 0.833274168101
KNeighborsClassifier: 0.980456280495
------------------ 20.9.10 Plot fitting a 9th order polynomial
LinearSVC(loss='l1'): 0.9325901988044822
LinearSVC(loss='l2'): 0.9165874249889226
Fits data generated from a 9th order polynomial with model of 4th order and 9th order polynomials, to demon-
-------------------
KNeighbors(n_neighbors=1): 0.9913675218842191 strate that often simpler models are to be prefered
KNeighbors(n_neighbors=2): 0.9848442068835102
import numpy as np
KNeighbors(n_neighbors=3): 0.9867753449543099
from matplotlib import pyplot as plt
KNeighbors(n_neighbors=4): 0.9803719053818863
from matplotlib.colors import ListedColormap
KNeighbors(n_neighbors=5): 0.9804562804949924
KNeighbors(n_neighbors=6): 0.9757924194139573
from sklearn import linear_model
KNeighbors(n_neighbors=7): 0.9780645792142071
KNeighbors(n_neighbors=8): 0.9780645792142071
# Create color maps for 3-class classification problem, as with iris
KNeighbors(n_neighbors=9): 0.9780645792142071
cmap_light = ListedColormap(['#FFAAAA', '#AAFFAA', '#AAAAFF'])
KNeighbors(n_neighbors=10): 0.9755550897728812
cmap_bold = ListedColormap(['#FF0000', '#00FF00', '#0000FF'])

20.9. Examples for the scikit-learn chapter 620 20.9. Examples for the scikit-learn chapter 621
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

rng = np.random.RandomState(0)
x = 2*rng.rand(100) - 1

f = lambda t: 1.2 * t**2 + .1 * t**3 - .4 * t **5 - .5 * t ** 9

y = f(x) + .4 * rng.normal(size=100)

x_test = np.linspace(-1, 1, 100)

The data

plt.figure(figsize=(6, 4))
plt.scatter(x, y, s=4)

Ground truth

plt.figure(figsize=(6, 4))
plt.scatter(x, y, s=4)
plt.plot(x_test, f(x_test), label="truth")
plt.axis('tight')
plt.title('Ground truth (9th order polynomial)')

plt.show()

Fitting 4th and 9th order polynomials

For this we need to engineer features: the n_th powers of x:

plt.figure(figsize=(6, 4))
plt.scatter(x, y, s=4)

X = np.array([x**i for i in range(5)]).T

X_test = np.array([x_test**i for i in range(5)]).T
regr = linear_model.LinearRegression()
regr.fit(X, y)
plt.plot(x_test, regr.predict(X_test), label='4th order')

X = np.array([x**i for i in range(10)]).T

X_test = np.array([x_test**i for i in range(10)]).T
regr = linear_model.LinearRegression()
regr.fit(X, y)
plt.plot(x_test, regr.predict(X_test), label='9th order')

plt.legend(loc='best')
plt.axis('tight')
plt.title('Fitting a 4th and a 9th order polynomial')

20.9. Examples for the scikit-learn chapter 622 20.9. Examples for the scikit-learn chapter 623
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Total running time of the script: ( 0 minutes 0.115 seconds)

Download Python source code: plot_polynomial_regression.py
Download Jupyter notebook: plot_polynomial_regression.ipynb
Generated by Sphinx-Gallery

20.9.11 A simple regression analysis on the Boston housing data •

Here we perform a simple regression analysis on the Boston housing data, exploring two types of regressors.

from sklearn.datasets import load_boston

data = load_boston()

Print a histogram of the quantity to predict: price

import matplotlib.pyplot as plt

plt.figure(figsize=(4, 3)) •
plt.hist(data.target)
plt.xlabel('price ($1000s)')
plt.ylabel('count')
plt.tight_layout()

Print the join histogram for each feature

for index, feature_name in enumerate(data.feature_names):

plt.figure(figsize=(4, 3))
•
plt.scatter(data.data[:, index], data.target)
plt.ylabel('Price', size=15)
plt.xlabel(feature_name, size=15)
plt.tight_layout()

20.9. Examples for the scikit-learn chapter 624 20.9. Examples for the scikit-learn chapter 625
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

from sklearn.linear_model import LinearRegression

clf = LinearRegression()
clf.fit(X_train, y_train)
predicted = clf.predict(X_test)
expected = y_test

plt.figure(figsize=(4, 3))
• plt.scatter(expected, predicted)
plt.plot([0, 50], [0, 50], '--k')
plt.axis('tight')
plt.xlabel('True price ($1000s)')
plt.ylabel('Predicted price ($1000s)')
plt.tight_layout()

Prediction with gradient boosted tree

• from sklearn.ensemble import GradientBoostingRegressor

clf = GradientBoostingRegressor()
clf.fit(X_train, y_train)

predicted = clf.predict(X_test)
expected = y_test

plt.figure(figsize=(4, 3))
plt.scatter(expected, predicted)
• plt.plot([0, 50], [0, 50], '--k')
plt.axis('tight')
plt.xlabel('True price ($1000s)')
plt.ylabel('Predicted price ($1000s)')
plt.tight_layout()

•
Simple prediction

from sklearn.model_selection import train_test_split

X_train, X_test, y_train, y_test = train_test_split(data.data, data.target)

20.9. Examples for the scikit-learn chapter 626 20.9. Examples for the scikit-learn chapter 627
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

y_min, y_max = X[:, 1].min() - .1, X[:, 1].max() + .1

xx, yy = np.meshgrid(np.linspace(x_min, x_max, 100),
np.linspace(y_min, y_max, 100))
Z = knn.predict(np.c_[xx.ravel(), yy.ravel()])

Put the result into a color plot

Z = Z.reshape(xx.shape)
plt.figure()
plt.pcolormesh(xx, yy, Z, cmap=cmap_light)

# Plot also the training points

plt.scatter(X[:, 0], X[:, 1], c=y, cmap=cmap_bold)
plt.xlabel('sepal length (cm)')
plt.ylabel('sepal width (cm)')
plt.axis('tight')

Print the error rate

import numpy as np
print("RMS: %r " % np.sqrt(np.mean((predicted - expected) ** 2)))

plt.show()

Out:

RMS: 3.3940265988986305

Total running time of the script: ( 0 minutes 1.082 seconds)

Download Python source code: plot_boston_prediction.py
Download Jupyter notebook: plot_boston_prediction.ipynb
Generated by Sphinx-Gallery

20.9.12 Nearest-neighbor prediction on iris

Plot the decision boundary of nearest neighbor decision on iris, first with a single nearest neighbor, and then
using 3 nearest neighbors.

import numpy as np
from matplotlib import pyplot as plt
from sklearn import neighbors, datasets
from matplotlib.colors import ListedColormap
And now, redo the analysis with 3 neighbors
# Create color maps for 3-class classification problem, as with iris
cmap_light = ListedColormap(['#FFAAAA', '#AAFFAA', '#AAAAFF']) knn = neighbors.KNeighborsClassifier(n_neighbors=3)
cmap_bold = ListedColormap(['#FF0000', '#00FF00', '#0000FF']) knn.fit(X, y)
iris = datasets.load_iris() Z = knn.predict(np.c_[xx.ravel(), yy.ravel()])
X = iris.data[:, :2] # we only take the first two features. We could
# avoid this ugly slicing by using a two-dim dataset # Put the result into a color plot
y = iris.target Z = Z.reshape(xx.shape)
plt.figure()
knn = neighbors.KNeighborsClassifier(n_neighbors=1) plt.pcolormesh(xx, yy, Z, cmap=cmap_light)
knn.fit(X, y)
# Plot also the training points
x_min, x_max = X[:, 0].min() - .1, X[:, 0].max() + .1 plt.scatter(X[:, 0], X[:, 1], c=y, cmap=cmap_bold)

20.9. Examples for the scikit-learn chapter 628 20.9. Examples for the scikit-learn chapter 629
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

plt.xlabel('sepal length (cm)')

plt.ylabel('sepal width (cm)') for i in range(64):
plt.axis('tight') ax = fig.add_subplot(8, 8, i + 1, xticks=[], yticks=[])
ax.imshow(digits.images[i], cmap=plt.cm.binary, interpolation='nearest')
plt.show() # label the image with the target value
ax.text(0, 7, str(digits.target[i]))

Total running time of the script: ( 0 minutes 0.091 seconds)

Download Python source code: plot_iris_knn.py
Download Jupyter notebook: plot_iris_knn.ipynb
Generated by Sphinx-Gallery

20.9.13 Simple visualization and classification of the digits dataset

Plot the first few samples of the digits dataset and a 2D representation built using PCA, then do a simple clas-
Plot a projection on the 2 first principal axis
sification

from sklearn.datasets import load_digits plt.figure()

digits = load_digits()
from sklearn.decomposition import PCA
pca = PCA(n_components=2)
proj = pca.fit_transform(digits.data)
Plot the data: images of digits
plt.scatter(proj[:, 0], proj[:, 1], c=digits.target)
plt.colorbar()
Each data in a 8x8 image

from matplotlib import pyplot as plt

fig = plt.figure(figsize=(6, 6)) # figure size in inches
fig.subplots_adjust(left=0, right=1, bottom=0, top=1, hspace=0.05, wspace=0.05)

20.9. Examples for the scikit-learn chapter 630 20.9. Examples for the scikit-learn chapter 631
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

ax.text(0, 7, str(predicted[i]), color='red')

Classify with Gaussian naive Bayes

from sklearn.naive_bayes import GaussianNB

from sklearn.model_selection import train_test_split

# split the data into training and validation sets

X_train, X_test, y_train, y_test = train_test_split(digits.data, digits.target)

# train the model

clf = GaussianNB()
clf.fit(X_train, y_train)
Quantify the performance
# use the model to predict the labels of the test data
predicted = clf.predict(X_test) First print the number of correct matches
expected = y_test
matches = (predicted == expected)
# Plot the prediction print(matches.sum())
fig = plt.figure(figsize=(6, 6)) # figure size in inches
fig.subplots_adjust(left=0, right=1, bottom=0, top=1, hspace=0.05, wspace=0.05) Out:

# plot the digits: each image is 8x8 pixels 384

for i in range(64):
ax = fig.add_subplot(8, 8, i + 1, xticks=[], yticks=[]) The total number of data points
ax.imshow(X_test.reshape(-1, 8, 8)[i], cmap=plt.cm.binary,
interpolation='nearest') print(len(matches))

# label the image with the target value Out:

if predicted[i] == expected[i]:
ax.text(0, 7, str(predicted[i]), color='green') 450
else:

20.9. Examples for the scikit-learn chapter 632 20.9. Examples for the scikit-learn chapter 633
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

And now, the ration of correct predictions

from sklearn import datasets
matches.sum() / float(len(matches)) faces = datasets.fetch_olivetti_faces()
faces.data.shape
Print the classification report
Let’s visualize these faces to see what we’re working with
from sklearn import metrics
print(metrics.classification_report(expected, predicted)) from matplotlib import pyplot as plt
fig = plt.figure(figsize=(8, 6))
# plot several images
Out:
for i in range(15):
precision recall f1-score support ax = fig.add_subplot(3, 5, i + 1, xticks=[], yticks=[])
ax.imshow(faces.images[i], cmap=plt.cm.bone)
0 1.00 0.93 0.96 40
1 0.82 0.85 0.84 48
2 0.91 0.69 0.78 45
3 0.96 0.83 0.89 58
4 0.91 0.89 0.90 36
5 0.91 0.85 0.88 48
6 0.93 0.98 0.95 43
7 0.75 0.98 0.85 47
8 0.59 0.81 0.68 42
9 0.91 0.74 0.82 43

avg / total 0.87 0.85 0.86 450

Print the confusion matrix

print(metrics.confusion_matrix(expected, predicted))

plt.show()

Out:

[[37 0 0 0 2 0 0 1 0 0]
[ 0 41 0 0 0 0 1 0 4 2]
[ 0 2 31 0 0 0 1 0 11 0]
[ 0 0 2 48 0 1 0 2 4 1]
[ 0 1 1 0 32 0 0 1 1 0]
[ 0 1 0 1 1 41 0 4 0 0]
[ 0 0 0 0 0 1 42 0 0 0]
[ 0 0 0 0 0 1 0 46 0 0]
[ 0 4 0 0 0 1 0 3 34 0]
[ 0 1 0 1 0 0 1 4 4 32]]

Total running time of the script: ( 0 minutes 5.208 seconds)

Download Python source code: plot_digits_simple_classif.py Tip: Note is that these faces have already been localized and scaled to a common size. This is an important
preprocessing piece for facial recognition, and is a process that can require a large collection of training data.
Download Jupyter notebook: plot_digits_simple_classif.ipynb This can be done in scikit-learn, but the challenge is gathering a sufficient amount of training data for the
Generated by Sphinx-Gallery algorithm to work. Fortunately, this piece is common enough that it has been done. One good resource is
OpenCV, the Open Computer Vision Library.

20.9.14 The eigenfaces example: chaining PCA and SVMs We’ll perform a Support Vector classification of the images. We’ll do a typical train-test split on the images:

The goal of this example is to show how an unsupervised method and a supervised one can be chained for from sklearn.model_selection import train_test_split
better prediction. It starts with a didactic but lengthy way of doing things, and finishes with the idiomatic X_train, X_test, y_train, y_test = train_test_split(faces.data,
approach to pipelining in scikit-learn. faces.target, random_state=0)

Here we’ll take a look at a simple facial recognition example. Ideally, we would use a dataset consisting of a print(X_train.shape, X_test.shape)
subset of the Labeled Faces in the Wild data that is available with sklearn.datasets.fetch_lfw_people().
However, this is a relatively large download (~200MB) so we will do the tutorial on a simpler, less rich dataset. Out:
Feel free to explore the LFW dataset.

20.9. Examples for the scikit-learn chapter 634 20.9. Examples for the scikit-learn chapter 635
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

(300, 4096) (100, 4096) ax.imshow(pca.components_[i].reshape(faces.images[0].shape),

cmap=plt.cm.bone)

Preprocessing: Principal Component Analysis

1850 dimensions is a lot for SVM. We can use PCA to reduce these 1850 features to a manageable size, while
maintaining most of the information in the dataset.

from sklearn import decomposition

pca = decomposition.PCA(n_components=150, whiten=True)
pca.fit(X_train)

One interesting part of PCA is that it computes the “mean” face, which can be interesting to examine:

plt.imshow(pca.mean_.reshape(faces.images[0].shape),
cmap=plt.cm.bone)

The components (“eigenfaces”) are ordered by their importance from top-left to bottom-right. We see that the
first few components seem to primarily take care of lighting conditions; the remaining components pull out
certain identifying features: the nose, eyes, eyebrows, etc.
With this projection computed, we can now project our original training and test data onto the PCA basis:

X_train_pca = pca.transform(X_train)
X_test_pca = pca.transform(X_test)
print(X_train_pca.shape)

Out:

(300, 150)

print(X_test_pca.shape)

Out:

(100, 150)

These projected components correspond to factors in a linear combination of component images such that
the combination approaches the original face.

Doing the Learning: Support Vector Machines

Now we’ll perform support-vector-machine classification on this reduced dataset:

from sklearn import svm

clf = svm.SVC(C=5., gamma=0.001)
The principal components measure deviations about this mean along orthogonal axes. clf.fit(X_train_pca, y_train)

print(pca.components_.shape)
Finally, we can evaluate how well this classification did. First, we might plot a few of the test-cases with the
labels learned from the training set:
Out:
import numpy as np
(150, 4096) fig = plt.figure(figsize=(8, 6))
for i in range(15):
It is also interesting to visualize these principal components: ax = fig.add_subplot(3, 5, i + 1, xticks=[], yticks=[])
ax.imshow(X_test[i].reshape(faces.images[0].shape),
fig = plt.figure(figsize=(16, 6)) cmap=plt.cm.bone)
for i in range(30): y_pred = clf.predict(X_test_pca[i, np.newaxis])[0]
ax = fig.add_subplot(3, 10, i + 1, xticks=[], yticks=[]) color = ('black' if y_pred == y_test[i] else 'red')

20.9. Examples for the scikit-learn chapter 636 20.9. Examples for the scikit-learn chapter 637
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

ax.set_title(faces.target[y_pred], 13 1.00 1.00 1.00 3

fontsize='small', color=color) 14 1.00 1.00 1.00 5
15 1.00 1.00 1.00 3
17 1.00 1.00 1.00 6
19 1.00 1.00 1.00 4
20 1.00 1.00 1.00 1
21 1.00 1.00 1.00 1
22 1.00 1.00 1.00 2
23 0.50 1.00 0.67 1
24 1.00 1.00 1.00 2
25 1.00 0.50 0.67 2
26 1.00 0.75 0.86 4
27 1.00 1.00 1.00 1
28 0.67 1.00 0.80 2
29 1.00 1.00 1.00 3
30 1.00 1.00 1.00 4
31 1.00 1.00 1.00 3
32 1.00 1.00 1.00 3
33 1.00 1.00 1.00 2
34 1.00 1.00 1.00 3
35 1.00 1.00 1.00 1
36 1.00 1.00 1.00 3
37 1.00 1.00 1.00 3
38 1.00 1.00 1.00 1
39 1.00 1.00 1.00 3

avg / total 0.97 0.95 0.95 100

Another interesting metric is the confusion matrix, which indicates how often any two items are mixed-up.
The confusion matrix of a perfect classifier would only have nonzero entries on the diagonal, with zeros on the
off-diagonal:

print(metrics.confusion_matrix(y_test, y_pred))

The classifier is correct on an impressive number of images given the simplicity of its learning model! Using Out:
a linear classifier on 150 features derived from the pixel-level data, the algorithm correctly identifies a large
[[4 0 0 ..., 0 0 0]
number of the people in the images. [0 4 0 ..., 0 0 0]
Again, we can quantify this effectiveness using one of several measures from sklearn.metrics. First we can [0 0 2 ..., 0 0 0]
do the classification report, which shows the precision, recall and other measures of the “goodness” of the ...,
[0 0 0 ..., 3 0 0]
classification:
[0 0 0 ..., 0 1 0]
from sklearn import metrics [0 0 0 ..., 0 0 3]]
y_pred = clf.predict(X_test_pca)
print(metrics.classification_report(y_test, y_pred))
Pipelining
Out:
Above we used PCA as a pre-processing step before applying our support vector machine classifier. Plugging
precision recall f1-score support
the output of one estimator directly into the input of a second estimator is a commonly used pattern; for this
0 1.00 0.67 0.80 6
reason scikit-learn provides a Pipeline object which automates this process. The above problem can be re-
1 1.00 1.00 1.00 4 expressed as a pipeline as follows:
2 0.50 1.00 0.67 2
from sklearn.pipeline import Pipeline
3 1.00 1.00 1.00 1
clf = Pipeline([('pca', decomposition.PCA(n_components=150, whiten=True)),
4 1.00 1.00 1.00 1
('svm', svm.LinearSVC(C=1.0))])
5 1.00 1.00 1.00 5
6 1.00 1.00 1.00 4
clf.fit(X_train, y_train)
7 1.00 0.67 0.80 3
9 1.00 1.00 1.00 1
y_pred = clf.predict(X_test)
10 1.00 1.00 1.00 4
print(metrics.confusion_matrix(y_pred, y_test))
11 1.00 1.00 1.00 1
plt.show()
12 0.67 1.00 0.80 2

20.9. Examples for the scikit-learn chapter 638 20.9. Examples for the scikit-learn chapter 639
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Out:
s=80, edgecolors="k", facecolors="none")
[[5 0 0 ..., 0 0 0]
[0 4 0 ..., 0 0 0] delta = 1
[0 0 1 ..., 0 0 0] y_min, y_max = -50, 50
..., x_min, x_max = -50, 50
[0 0 0 ..., 3 0 0] x = np.arange(x_min, x_max + delta, delta)
[0 0 0 ..., 0 1 0] y = np.arange(y_min, y_max + delta, delta)
[0 0 0 ..., 0 0 3]] X1, X2 = np.meshgrid(x, y)
Z = clf.decision_function(np.c_[X1.ravel(), X2.ravel()])
Z = Z.reshape(X1.shape)

A Note on Facial Recognition ax.contour(X1, X2, Z, [-1.0, 0.0, 1.0], colors='k',

linestyles=['dashed', 'solid', 'dashed'])
Here we have used PCA “eigenfaces” as a pre-processing step for facial recognition. The reason we chose this
is because PCA is a broadly-applicable technique, which can be useful for a wide array of data types. Research
in the field of facial recognition in particular, however, has shown that other more specific feature extraction
methods are can be much more effective.
Total running time of the script: ( 0 minutes 3.886 seconds)
Download Python source code: plot_eigenfaces.py
Download Jupyter notebook: plot_eigenfaces.ipynb
Generated by Sphinx-Gallery

20.9.15 Example of linear and non-linear models

This is an example plot from the tutorial which accompanies an explanation of the support vector machine
GUI.

import numpy as np
from matplotlib import pyplot as plt

from sklearn import svm

data that is linearly separable

def linear_model(rseed=42, n_samples=30):

" Generate data according to a linear model"
np.random.seed(rseed)
data with a non-linear separation
data = np.random.normal(0, 10, (n_samples, 2))
data[:n_samples // 2] -= 15 def nonlinear_model(rseed=42, n_samples=30):
data[n_samples // 2:] += 15 radius = 40 * np.random.random(n_samples)
far_pts = radius > 20
labels = np.ones(n_samples) radius[far_pts] *= 1.2
labels[:n_samples // 2] = -1 radius[~far_pts] *= 1.1

return data, labels theta = np.random.random(n_samples) * np.pi * 2

data = np.empty((n_samples, 2))

X, y = linear_model() data[:, 0] = radius * np.cos(theta)
clf = svm.SVC(kernel='linear') data[:, 1] = radius * np.sin(theta)
clf.fit(X, y)
labels = np.ones(n_samples)
plt.figure(figsize=(6, 4)) labels[far_pts] = -1
ax = plt.subplot(111, xticks=[], yticks=[])
ax.scatter(X[:, 0], X[:, 1], c=y, cmap=plt.cm.bone) return data, labels

ax.scatter(clf.support_vectors_[:, 0],
clf.support_vectors_[:, 1], X, y = nonlinear_model()
clf = svm.SVC(kernel='rbf', gamma=0.001, coef0=0, degree=3)

20.9. Examples for the scikit-learn chapter 640 20.9. Examples for the scikit-learn chapter 641
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Fit polynomes of different degrees to a dataset: for too small a degree, the model underfits, while for too large
clf.fit(X, y)
a degree, it overfits.
plt.figure(figsize=(6, 4))
import numpy as np
ax = plt.subplot(1, 1, 1, xticks=[], yticks=[])
import matplotlib.pyplot as plt
ax.scatter(X[:, 0], X[:, 1], c=y, cmap=plt.cm.bone, zorder=2)

ax.scatter(clf.support_vectors_[:, 0], clf.support_vectors_[:, 1],

def generating_func(x, err=0.5):
s=80, edgecolors="k", facecolors="none")
return np.random.normal(10 - 1. / (x + 0.1), err)
delta = 1
y_min, y_max = -50, 50 A polynomial regression
x_min, x_max = -50, 50
from sklearn.pipeline import make_pipeline
x = np.arange(x_min, x_max + delta, delta)
from sklearn.linear_model import LinearRegression
y = np.arange(y_min, y_max + delta, delta)
from sklearn.preprocessing import PolynomialFeatures
X1, X2 = np.meshgrid(x, y)
Z = clf.decision_function(np.c_[X1.ravel(), X2.ravel()])
Z = Z.reshape(X1.shape) A simple figure to illustrate the problem

ax.contour(X1, X2, Z, [-1.0, 0.0, 1.0], colors='k', n_samples = 8

linestyles=['dashed', 'solid', 'dashed'], zorder=1)
np.random.seed(0)
plt.show() x = 10 ** np.linspace(-2, 0, n_samples)
y = generating_func(x)

x_test = np.linspace(-0.2, 1.2, 1000)

titles = ['d = 1 (under-fit; high bias)',

'd = 2',
'd = 6 (over-fit; high variance)']
degrees = [1, 2, 6]

fig = plt.figure(figsize=(9, 3.5))

fig.subplots_adjust(left=0.06, right=0.98, bottom=0.15, top=0.85, wspace=0.05)

for i, d in enumerate(degrees):
ax = fig.add_subplot(131 + i, xticks=[], yticks=[])
ax.scatter(x, y, marker='x', c='k', s=50)

model = make_pipeline(PolynomialFeatures(d), LinearRegression())

model.fit(x[:, np.newaxis], y)
ax.plot(x_test, model.predict(x_test[:, np.newaxis]), '-b')

ax.set_xlim(-0.2, 1.2)
ax.set_ylim(0, 12)
ax.set_xlabel('house size')
if i == 0:
ax.set_ylabel('price')

ax.set_title(titles[i])

Total running time of the script: ( 0 minutes 0.147 seconds)

Download Python source code: plot_svm_non_linear.py
Download Jupyter notebook: plot_svm_non_linear.ipynb
Generated by Sphinx-Gallery

20.9.16 Bias and variance of polynomial fit

Demo overfitting, underfitting, and validation and learning curves with polynomial regression.

20.9. Examples for the scikit-learn chapter 642 20.9. Examples for the scikit-learn chapter 643
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Generate a larger dataset

from sklearn.model_selection import train_test_split

n_samples = 200
test_size = 0.4
error = 1.0

# randomly sample the data

np.random.seed(1)
x = np.random.random(n_samples)
Plot a validation curve
y = generating_func(x, error)
from sklearn.model_selection import validation_curve
# split into training, validation, and testing sets.
x_train, x_test, y_train, y_test = train_test_split(x, y, test_size=test_size) degrees = np.arange(1, 21)

# show the training and validation sets model = make_pipeline(PolynomialFeatures(), LinearRegression())

plt.figure(figsize=(6, 4))
plt.scatter(x_train, y_train, color='red', label='Training set') # The parameter to vary is the "degrees" on the pipeline step
plt.scatter(x_test, y_test, color='blue', label='Test set') # "polynomialfeatures"
plt.title('The data') train_scores, validation_scores = validation_curve(
plt.legend(loc='best') model, x[:, np.newaxis], y,
param_name='polynomialfeatures__degree',
param_range=degrees)

# Plot the mean train error and validation error across folds
plt.figure(figsize=(6, 4))
plt.plot(degrees, validation_scores.mean(axis=1), lw=2,
label='cross-validation')
plt.plot(degrees, train_scores.mean(axis=1), lw=2, label='training')

plt.legend(loc='best')
plt.xlabel('degree of fit')
plt.ylabel('explained variance')
plt.title('Validation curve')
plt.tight_layout()

20.9. Examples for the scikit-learn chapter 644 20.9. Examples for the scikit-learn chapter 645
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Learning curves

Plot train and test error with an increasing number of samples

# A learning curve for d=1, 5, 15

for d in [1, 5, 15]:
model = make_pipeline(PolynomialFeatures(degree=d), LinearRegression())

from sklearn.model_selection import learning_curve

train_sizes, train_scores, validation_scores = learning_curve( •
model, x[:, np.newaxis], y,
train_sizes=np.logspace(-1, 0, 20)) Total running time of the script: ( 0 minutes 0.935 seconds)
Download Python source code: plot_bias_variance.py
# Plot the mean train error and validation error across folds
plt.figure(figsize=(6, 4)) Download Jupyter notebook: plot_bias_variance.ipynb
plt.plot(train_sizes, validation_scores.mean(axis=1),
lw=2, label='cross-validation') Generated by Sphinx-Gallery
plt.plot(train_sizes, train_scores.mean(axis=1),
lw=2, label='training')
plt.ylim(ymin=-.1, ymax=1) 20.9.17 Tutorial Diagrams

plt.legend(loc='best') This script plots the flow-charts used in the scikit-learn tutorials.
plt.xlabel('number of train samples')
plt.ylabel('explained variance')
plt.title('Learning curve (degree=%i )' % d)
plt.tight_layout()

plt.show()

20.9. Examples for the scikit-learn chapter 646 20.9. Examples for the scikit-learn chapter 647
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

arrow1 = '#88CCFF',
arrow2 = '#88FF88',
supervised=True):
fig = pl.figure(figsize=(9, 6), facecolor='w')
ax = pl.axes((0, 0, 1, 1),
xticks=[], yticks=[], frameon=False)
ax.set_xlim(0, 9)
ax.set_ylim(0, 6)

patches = [Rectangle((0.3, 3.6), 1.5, 1.8, zorder=1, fc=box_bg),

Rectangle((0.5, 3.8), 1.5, 1.8, zorder=2, fc=box_bg),
Rectangle((0.7, 4.0), 1.5, 1.8, zorder=3, fc=box_bg),

Rectangle((2.9, 3.6), 0.2, 1.8, fc=box_bg),

Rectangle((3.1, 3.8), 0.2, 1.8, fc=box_bg),
Rectangle((3.3, 4.0), 0.2, 1.8, fc=box_bg),

• Rectangle((0.3, 0.2), 1.5, 1.8, fc=box_bg),

Rectangle((2.9, 0.2), 0.2, 1.8, fc=box_bg),

Circle((5.5, 3.5), 1.0, fc=box_bg),

Polygon([[5.5, 1.7],
[6.1, 1.1],
[5.5, 0.5],
[4.9, 1.1]], fc=box_bg),

FancyArrow(2.3, 4.6, 0.35, 0, fc=arrow1,

width=0.25, head_width=0.5, head_length=0.2),

FancyArrow(3.75, 4.2, 0.5, -0.2, fc=arrow1,

width=0.25, head_width=0.5, head_length=0.2),

FancyArrow(5.5, 2.4, 0, -0.4, fc=arrow1,

width=0.25, head_width=0.5, head_length=0.2),
•
FancyArrow(2.0, 1.1, 0.5, 0, fc=arrow2,
width=0.25, head_width=0.5, head_length=0.2),

FancyArrow(3.3, 1.1, 1.3, 0, fc=arrow2,

width=0.25, head_width=0.5, head_length=0.2),

FancyArrow(6.2, 1.1, 0.8, 0, fc=arrow2,

width=0.25, head_width=0.5, head_length=0.2)]

if supervised:
patches += [Rectangle((0.3, 2.4), 1.5, 0.5, zorder=1, fc=box_bg),
Rectangle((0.5, 2.6), 1.5, 0.5, zorder=2, fc=box_bg),
Rectangle((0.7, 2.8), 1.5, 0.5, zorder=3, fc=box_bg),
FancyArrow(2.3, 2.9, 2.0, 0, fc=arrow1,
width=0.25, head_width=0.5, head_length=0.2),
Rectangle((7.3, 0.85), 1.5, 0.5, fc=box_bg)]
else:
patches += [Rectangle((7.3, 0.2), 1.5, 1.8, fc=box_bg)]
•
for p in patches:
import numpy as np ax.add_patch(p)
import pylab as pl
from matplotlib.patches import Circle, Rectangle, Polygon, Arrow, FancyArrow pl.text(1.45, 4.9, "Training\nText,\nDocuments,\nImages,\netc.",
ha='center', va='center', fontsize=14)
def create_base(box_bg = '#CCCCCC',
pl.text(3.6, 4.9, "Feature\nVectors",

20.9. Examples for the scikit-learn chapter 648 20.9. Examples for the scikit-learn chapter 649
 Scipy lecture notes, Edition 2017.1 Scipy lecture notes, Edition 2017.1

Download Jupyter notebook: plot_ML_flow_chart.ipynb

ha='left', va='center', fontsize=14)
Generated by Sphinx-Gallery
pl.text(5.5, 3.5, "Machine\nLearning\nAlgorithm",
ha='center', va='center', fontsize=14) Download all examples in Python source code: auto_examples_python.zip
Download all examples in Jupyter notebooks: auto_examples_jupyter.zip
pl.text(1.05, 1.1, "New Text,\nDocument,\nImage,\netc.",
ha='center', va='center', fontsize=14) Generated by Sphinx-Gallery

pl.text(3.3, 1.7, "Feature\nVector", See also:

ha='left', va='center', fontsize=14) Going further
pl.text(5.5, 1.1, "Predictive\nModel", • The documentation of scikit-learn is very complete and didactic.
ha='center', va='center', fontsize=12)
• Introduction to Machine Learning with Python, by Sarah Guido, Andreas Müller
if supervised: (notebooks available here).
pl.text(1.45, 3.05, "Labels",
ha='center', va='center', fontsize=14)

pl.text(8.05, 1.1, "Expected\nLabel",

ha='center', va='center', fontsize=14)
pl.text(8.8, 5.8, "Supervised Learning Model",
ha='right', va='top', fontsize=18)

else:
pl.text(8.05, 1.1,
"Likelihood\nor Cluster ID\nor Better\nRepresentation",
ha='center', va='center', fontsize=12)
pl.text(8.8, 5.8, "Unsupervised Learning Model",
ha='right', va='top', fontsize=18)

def plot_supervised_chart(annotate=False):
create_base(supervised=True)
if annotate:
fontdict = dict(color='r', weight='bold', size=14)
pl.text(1.9, 4.55, 'X = vec.fit_transform(input)',
fontdict=fontdict,
rotation=20, ha='left', va='bottom')
pl.text(3.7, 3.2, 'clf.fit(X, y)',
fontdict=fontdict,
rotation=20, ha='left', va='bottom')
pl.text(1.7, 1.5, 'X_new = vec.transform(input)',
fontdict=fontdict,
rotation=20, ha='left', va='bottom')
pl.text(6.1, 1.5, 'y_new = clf.predict(X_new)',
fontdict=fontdict,
rotation=20, ha='left', va='bottom')

def plot_unsupervised_chart():
create_base(supervised=False)

if __name__ == '__main__':
plot_supervised_chart(False)
plot_supervised_chart(True)
plot_unsupervised_chart()
pl.show()

Total running time of the script: ( 0 minutes 0.137 seconds)

Download Python source code: plot_ML_flow_chart.py

20.9. Examples for the scikit-learn chapter 650 20.9. Examples for the scikit-learn chapter 651
 Index

D
diff, 508, 511
differentiation, 508
dsolve, 511

E
equations
algebraic, 510
differential, 511

I
integration, 509

M
Matrix, 511

P
Python Enhancement Proposals
PEP 255, 268
PEP 3118, 304
PEP 3129, 277
PEP 318, 270, 277
PEP 342, 268
PEP 343, 278
PEP 380, 269
PEP 380#id13, 269
PEP 8, 272

S
solve, 510

652