Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
Scribd
Upload
51 views

Tutorial Ex3

This document provides tutorials for implementing the oneVsAll() and predictOneVsAll() functions for multi-class classification and neural networks. It explains that oneVsAll() trains a linear regression classifier for each class and stores the theta vectors in the rows of all_theta. predictOneVsAll() calculates the sigmoid of X multiplied by all_theta to get class probabilities, and returns the class with the maximum probability.

Uploaded by

Kartik Chouhan
Copyright
© © All Rights Reserved
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
51 views

Tutorial Ex3

This document provides tutorials for implementing the oneVsAll() and predictOneVsAll() functions for multi-class classification and neural networks. It explains that oneVsAll() trains a linear regression classifier for each class and stores the theta vectors in the rows of all_theta. predictOneVsAll() calculates the sigmoid of X multiplied by all_theta to get class probabilities, and returns the class with the maximum probability.

Uploaded by

Kartik Chouhan
Copyright
© © All Rights Reserved
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
You are on page 1/ 2

Tutorial: ex3 oneVsAll()

Tom Mosher
MentorAssignment: Multi-class Classification and Neural Networks · 4 years ago · Edited
all_theta is a matrix, where there is a row for each of the trained thetas. In the exercise example,
there are 10 rows, of 401 elements each. You know this because that's how all_theta was
initialized in line 15 of the script template.

(note that the submit grader's test case doesn't have 401 elements or 10 rows - your function
must work for any size data set - so use the "num_labels" variable).

Each call to fmincg() returns a theta vector. Be sure you use the lambda value provided in the
function header.

You then need to copy that vector into a row of all_theta.

The oneVsAll.m script template contains several Hints and a code example to guide your work.

The "y == c" statement creates a vector of 0's and 1's for each value of 'c' as you iterate from 1 to
num_labels. Those are the effective 'y' values that are used for training to detect each label.

Type these commands in your workspace to see how to copy a vector into a matrix:

Q = zeros(5,3) % create a test matrix of all-zeros

v = [1 2 3]' % create a column vector

Q(2,:) = v % copy v into the 2nd row of Q

The syntax "(2,:)" means "use all columns of the 2nd row".

ex3: tutorial for predictOneVsAll()


Tom Mosher
MentorWeek 4 · 4 years ago · Edited

The code you add to predictOneVsAll.m can be as little as two lines:


1. one line to calculate the sigmoid() of the product of X and all_theta. X is (m x n), and all_theta is
(num_labels x n), so you'll need a transposition to get a result of (m x num_labels)
2. one line to return the classifier which has the max value. The size will be (m x 1). Use the "help
max" command in your workspace to learn how the max() function returns two values.

Note that your function must return the predictions as a column vector - size (m x 1). If you return
a row vector, the script will not compute the accuracy correctly.

ex3 tutorial for predict()


Tom Mosher
MentorWeek 4 · 4 years ago · Edited
Here is an outline for forward propagation using the vectorized method. This is an
implementation of the formula in Figure 2 on Page 11 of ex3.pdf.

1. Add a column of 1's to X (the first column), and it becomes 'a1'.


2. Multiply by Theta1 and you have 'z2'.
3. Compute the sigmoid() of 'z2', then add a column of 1's, and it becomes 'a2'
4. Multiply by Theta2, compute the sigmoid() and it becomes 'a3'.
5. Now use the max(a3, [], 2) function to return two vectors - one of the highest value for each
row, and one with its index. Ignore the highest values. Keep the vector of the indexes where the
highest values were found. These are your predictions.

Note: When you multiply by the Theta matrices, you'll have to use transposition to get a result
that is the correct size.

Note: The predictions must be returned as a column vector - size (m x 1). If you return a row
vector, the script will not compute the accuracy correctly.

Note: Not getting the correct results? In the hidden layer, be sure you use sigmoid() first, then
add the bias unit.

------ dimensions of the variables ---------

a1 is (m x n), where 'n' is the number of features including the bias unit

Theta1 is (h x n) where 'h' is the number of hidden units

a2 is (m x (h + 1))

Theta2 is (c x (h + 1)), where 'c' is the number of labels.

a3 is (m x c)

p is a vector of size (m x 1)

You might also like