Angrew Machine Learning ex3

lrCostFunction.m

h = sigmoid(X * theta);
theta(1,:) = 0; %This is the same with the file 'costFunctionReg.m' of ex2 except this line.
								%For adding the multi-classification,we should take the matrix instead of vector.Therefore, we set the first row to 0 in order to get them away from the computation of regularization.
grad = (X' * (h - y) + lambda * theta) / m;
J = (y' * log(h) + (1 .- y') * log(1 .- h) - lambda / 2 * sum(theta .^ 2)) / m;
J = 0 - J;

predictOneVsAll.m

p = X * all_theta';
[_, p] = max(p, [], 2);   
%max(p, [], 2) return two elements. The first of them is the max element of each row
%The second of them is the index of the max element.

 

oneVsAll.m

 

options = optimset('GradObj', 'on', 'MaxIter', 50);
for c = 1:num_labels
  initial_theta = zeros((n + 1) , 1);
  [theta, J, exitflag] = ...
    fmincg (@(t)(lrCostFunction(t, X, (y == c), lambda)), ...  %Train on the set in batch size. Each batch consisted of the pictures whhose num is c.
                initial_theta, options);
   all_theta(i,:) = theta';
end

predict.m

 

X = [ones(m, 1) X];
a = sigmoid(Theta1 * X');
a = [ones(1, size(a, 2)) ;  a];
[_, p] = max(a' * Theta2', [] ,2);

 

这是笔者自己想到一些对吴恩达机器学习课程的编程作业的实现方式，如果你有更好的实现方式，欢迎在评论区讨论。

这里只是部分代码，全部代码在 

https://download.csdn.net/download/ti_an_di/10590380