# Logistic Regression Vectorization Example

### From Ufldl

m |
m (Added semicolons, deleted stray quote) |
||

Line 7: | Line 7: | ||

and <math>\textstyle \theta \in \Re^{n+1}</math>, and <math>\textstyle \theta_0</math> is our intercept term. We have a training set | and <math>\textstyle \theta \in \Re^{n+1}</math>, and <math>\textstyle \theta_0</math> is our intercept term. We have a training set | ||

<math>\textstyle \{(x^{(1)}, y^{(1)}), \ldots, (x^{(m)}, y^{(m)})\}</math> of <math>\textstyle m</math> examples, and the batch gradient | <math>\textstyle \{(x^{(1)}, y^{(1)}), \ldots, (x^{(m)}, y^{(m)})\}</math> of <math>\textstyle m</math> examples, and the batch gradient | ||

- | ascent update rule is | + | ascent update rule is <math>\textstyle \theta := \theta + \alpha \nabla_\theta \ell(\theta)</math>, where <math>\textstyle \ell(\theta)</math> |

is the log likelihood and <math>\textstyle \nabla_\theta \ell(\theta)</math> is its derivative. | is the log likelihood and <math>\textstyle \nabla_\theta \ell(\theta)</math> is its derivative. | ||

Line 56: | Line 56: | ||

% Fast implementation of matrix-vector multiple | % Fast implementation of matrix-vector multiple | ||

- | grad = A*b | + | grad = A*b; |

</syntaxhighlight> | </syntaxhighlight> | ||

Line 63: | Line 63: | ||

<syntaxhighlight lang="matlab"> | <syntaxhighlight lang="matlab"> | ||

% Implementation 3 | % Implementation 3 | ||

- | grad = x * (y- sigmoid(theta'*x))' | + | grad = x * (y- sigmoid(theta'*x))'; |

</syntaxhighlight> | </syntaxhighlight> | ||

Here, we assume that the Matlab/Octave <tt>sigmoid(z)</tt> takes as input a vector <tt>z</tt>, applies the sigmoid function component-wise to the input, and returns the result. The output of <tt>sigmoid(z)</tt> is therefore itself also a vector, of the same dimension as the input <tt>z</tt> | Here, we assume that the Matlab/Octave <tt>sigmoid(z)</tt> takes as input a vector <tt>z</tt>, applies the sigmoid function component-wise to the input, and returns the result. The output of <tt>sigmoid(z)</tt> is therefore itself also a vector, of the same dimension as the input <tt>z</tt> |

## Revision as of 00:53, 23 April 2011

Consider training a logistic regression model using batch gradient ascent. Suppose our hypothesis is

where (following CS229 notational convention) we let , so that and , and is our intercept term. We have a training set of examples, and the batch gradient ascent update rule is , where is the log likelihood and is its derivative.

[Note: Most of the notation below follows that defined in the class CS229: Machine Learning. Please see Lecture notes #1 from http://cs229.stanford.edu/ for details.]

We thus need to compute the gradient:

Suppose that the Matlab/Octave variable `x` is a matrix containing the training inputs, so that
`x(:,i)` is the -th training example , and `x(j,i)` is .
Further, suppose the Matlab/Octave variable `y` is a *row* vector of the labels in the
training set, so that the variable `y(i)` is . (Here we differ from the
CS229 notation. Specifically, in the matrix-valued `x` we stack the training inputs in columns rather than in rows;
and `y` is a row vector rather than a column vector.)

Here's truly horrible, extremely slow, implementation of the gradient computation:

% Implementation 1 grad = zeros(n+1,1); for i=1:m, h = sigmoid(theta'*x(:,i)); temp = y(i) - h; for j=1:n+1, grad(j) = grad(j) + temp * x(j,i); end; end;

The two nested for-loops makes this very slow. Here's a more typical implementation, that partially vectorizes the algorithm and gets better performance:

% Implementation 2 grad = zeros(n+1,1); for i=1:m, grad = grad + (y(i) - sigmoid(theta'*x(:,i)))* x(:,i); end;

However, it turns out to be possible to even further vectorize this. If we can get rid of the for-loop, we can significantly speed up the implementation. In particular, suppose `b` is a column vector, and `A` is a matrix. Consider the following ways of computing `A * b`:

% Slow implementation of matrix-vector multiply grad = zeros(n+1,1); for i=1:m, grad = grad + b(i) * A(:,i); % more commonly written A(:,i)*b(i) end; % Fast implementation of matrix-vector multiple grad = A*b;

We recognize that Implementation 2 of our gradient descent calculation above is using the slow version with a for-loop, with
`b(i)` playing the role of `(y(i) - sigmoid(theta'*x(:,i)))`, and `A` playing the role of `x`. We can derive a fast implementation as follows:

% Implementation 3 grad = x * (y- sigmoid(theta'*x))';

Here, we assume that the Matlab/Octave `sigmoid(z)` takes as input a vector `z`, applies the sigmoid function component-wise to the input, and returns the result. The output of `sigmoid(z)` is therefore itself also a vector, of the same dimension as the input `z`

When the training set is large, this final implementation takes the greatest advantage of Matlab/Octave's highly optimized numerical linear algebra libraries to carry out the matrix-vector operations, and so this is far more efficient than the earlier implementations.

Coming up with vectorized implementations isn't always easy, and sometimes requires careful thought. But as you gain familiarity with vectorized operations, you'll find that there are design patterns (i.e., a small number of ways of vectorizing) that apply to many different pieces of code.