程序实现 softmax classifier, 含有三个隐含层的情况。activation function 是 ReLU :
function Out=Softmax_Classifier_3(train_x, train_y, opts)
% activation function RELU. y=max(0, x);
% setting learning parameters
step_size=opts.step_size;
reg=opts.reg;
batchsize = opts.batchsize;
numepochs = opts.numepochs;
K=opts.class;
h1=opts.hidden_1;
h2=opts.hidden_2;
h3=opts.hidden_3;
D=size(train_x, 2);
W1=0.01*randn(D, h1);
b1=zeros(1, h1);
W2=0.01*randn(h1, h2);
b2=zeros(1, h2);
W3=0.01*randn(h2, h3);
b3=zeros(1, h3);
W4=0.01*randn(h3, K);
b4=zeros(1, K);
loss(1 : numepochs)=0;
num_examples=size(train_x, 1);
numbatches = num_examples / batchsize;
for epoch=1:numepochs
kk = randperm(num_examples);
loss(epoch)=0;
tic;
sprintf('epoch %d:
' , epoch)
for bat=1:numbatches
batch_x = train_x(kk((bat - 1) * batchsize + 1 : bat * batchsize), :);
batch_y = train_y(kk((bat - 1) * batchsize + 1 : bat * batchsize), :);
%% forward
f1=batch_x*W1+repmat(b1, batchsize, 1);
hiddenval_1=max(0, f1);
f2=hiddenval_1*W2+repmat(b2, batchsize, 1);
hiddenval_2=max(0, f2);
f3=hiddenval_2*W3+repmat(b3, batchsize, 1);
hiddenval_3=max(0, f3);
scores=hiddenval_3*W4+repmat(b4, batchsize, 1);
%% the loss
exp_scores=exp(scores);
dd=repmat(sum(exp_scores, 2), 1, K);
probs=exp_scores./dd;
correct_logprobs=-log(sum(probs.*batch_y, 2));
data_loss=sum(correct_logprobs)/batchsize;
reg_loss=0.5*reg*sum(sum(W1.*W1))+0.5*reg*sum(sum(W2.*W2))+0.5*reg*sum(sum(W3.*W3))+0.5*reg*sum(sum(W4.*W4));
loss(epoch) =loss(epoch)+ data_loss + reg_loss;
%% back propagation
% output layer
dscores = probs-batch_y;
dscores=dscores/batchsize;
dW4=hiddenval_3'*dscores;
db4=sum(dscores);
% hidden layer 3
dhiddenval_3=dscores*W4';
mask=max(sign(hiddenval_3), 0);
df_3=dhiddenval_3.*mask;
dW3=hiddenval_2'*df_3;
db3=sum(df_3);
% hidden layer 2
dhiddenval_2=df_3*W3';
mask=max(sign(hiddenval_2), 0);
df_2=dhiddenval_2.*mask;
dW2=hiddenval_1'*df_2;
db2=sum(df_2);
% hidden layer 1
dhiddenval_1=df_2*W2';
mask=max(sign(hiddenval_1), 0);
df_1=dhiddenval_1.*mask;
dW1=batch_x'*df_1;
db1=sum(df_1);
%% update
dW4=dW4+reg*W4;
dW3=dW3+reg*W3;
dW2=dW2+reg*W2;
dW1=dW1+reg*W1;
W4=W4-step_size*dW4;
b4=b4-step_size*db4;
W3=W3-step_size*dW3;
b3=b3-step_size*db3;
W2=W2-step_size*dW2;
b2=b2-step_size*db2;
W1=W1-step_size*dW1;
b1=b1-step_size*db1;
end
loss(epoch)=loss(epoch)/numbatches;
sprintf('training loss is %f:
', loss(epoch))
toc;
end
Out.W1=W1;
Out.W2=W2;
Out.W3=W3;
Out.W4=W4;
Out.b1=b1;
Out.b2=b2;
Out.b3=b3;
Out.b4=b4;
Out.loss=loss;