UFLDL练习-稀疏自编码实现

这个练习是要我们编码实现稀疏自编码的神经网络，并将其可视化出来，需要我们编码的有三个部分：sampleIMAGES.m，sparseAutoencoderCost.m，computeNumericalGradient.m，已下是我实现的代码。

sampleIMAGES.m：

rands = zeros(3,numpatches);
rands(1,:) = unidrnd(10,1,numpatches);
rands(2:3,:) = unidrnd(512-patchsize+1,2,numpatches);
for i=1:numpatches
    patch = IMAGES(rands(2,i):rands(2,i)+patchsize-1,rands(3,i):rands(3,i)+patchsize-1,rands(1,i));
    patches(:,i) = patch(:);
end

sparseAutoencoderCost.m：

num_example = size(data,2); %样本总数
a1 = sigmoid(W1 * data + repmat(b1,1,num_example));    %第二层激活值
a2 = sigmoid(W2 * a1 + repmat(b2,1,num_example));      %第三层（输出层）激活值

diff = (a2 - data).^2;
data_loss = 0.5 * sum(diff(:))/num_example;   %经验损失
reg_loss = 0.5 * lambda * (sum((sum(W1.^2))') + sum((sum(W2.^2))'));  %正则项

rou = sum(a1,2)/num_example;    %激活度
sparse_loss = beta * sum(sparsityParam * log(sparsityParam./rou) + (1 - sparsityParam) * log((1 - sparsityParam)./(1 - rou)));    %稀疏惩罚项

cost = data_loss + reg_loss + sparse_loss;  %损失函数

error2 = (a2 - data) .* (a2 .* (1 - a2));   %输出层残差
error1 = (W2' * error2 + repmat(beta * (-sparsityParam./rou + (1- sparsityParam)./(1- rou)),1,num_example)) .* (a1 .* (1 - a1));    %第二层残差（加稀疏性限制）

W2grad = error2 * a1' / num_example;
W2grad = W2grad + lambda * W2; 
b2grad = sum(error2,2) / num_example;
W1grad = error1 * data' / num_example;
W1grad = W1grad + lambda * W1;
b1grad = sum(error1,2) / num_example;

computeNumericalGradient.m：

epsilon = 10^-4;
e = epsilon * eye(size(theta,1));
for i=1:size(theta,1)
    numgrad(i) = (J(theta + e(:,i)) - J(theta - e(:,i)))/(2 * epsilon);
end

输出结果：