急求！！！！！！！！

cnn进行MNIST手写数字库识别

程序已调通

但查看中间层数据却提示Cannot display summaries of variables with more than 524288 elements.

后来查到可以保存为excel

但保存为excel 时提示Index exceeds matrix dimensions.

不知道该怎么办

求解

代码如下

clear all

% Hyper-parameters

imageSize = 28;

filterNum1 = 10;

filterSize1 = 5;

poolDim1 = 2;

filterNum2 = 10;

filterSize2 = 5;

poolDim2 = 2;

convDim1 = imageSize - filterSize1 + 1;

outputDim1 = convDim1/poolDim1;

convDim2 = outputDim1 - filterSize2 + 1;

outputDim2 = convDim2/poolDim2;

classNum = 10;

sampleSize = 60000;

mbSize = 50; % mini-batch sample size

r = 0.05;    % learning rate

wdr = 0.00; % weight decay rate

mom = 0.5;   % momentum

epochNum = 1;

% Load data

addpath('E:/YUI/common');

images = loadMNISTImages('E:/YUI/common/train-images-idx3-ubyte');

images = reshape(images,imageSize,imageSize,[]);

labels = loadMNISTLabels('E:/YUI/common/train-labels-idx1-ubyte');

labels(labels==0) = 10;

labelMat = full(sparse(labels,1:length(labels),1));

% Initialize parameters

Wc1 = 0.1*(randn(filterSize1,filterSize1,filterNum1));

bc1 = zeros(filterNum1,1);

Wc2 = 0.1*(randn(filterSize2,filterSize2,filterNum1,filterNum2));

bc2 = zeros(filterNum2,1);

Wd = 0.01*(rand(classNum,filterNum2*outputDim2^2) - 0.5);

bd = zeros(classNum,1);

% Velocity of parameters

vel_Wc1 = zeros(size(Wc1));

vel_bc1 = zeros(size(bc1));

vel_Wc2 = zeros(size(Wc2));

vel_bc2 = zeros(size(bc2));

vel_Wd = zeros(size(Wd));

vel_bd = zeros(size(bd));

%% Training

for e = 1:epochNum

err = 0;

rp = randperm(sampleSize);

for s = 1:mbSize:(sampleSize-mbSize+1)

mbImages = images(:,:,rp(s:s+mbSize-1));

% Feedforward

o1 = zeros(convDim1,convDim1,filterNum1,mbSize);

for i = 1:filterNum1

o1(:,:,i,:) = convn(mbImages,rot90(Wc1(:,:,i),2),'valid') + bc1(i);

end

o1Pooled = zeros(outputDim1,outputDim1,filterNum1,mbSize);

o1PoolIdx = zeros(outputDim1^2,filterNum1,mbSize);

for i = 1:mbSize

for j = 1:filterNum1

[o1Pooled(:,:,j,i) o1PoolIdx(:,j,i)] = MaxPooling(o1(:,:,j,i),[poolDim1 poolDim1]);

end

end

o2 = zeros(convDim2,convDim2,filterNum2,mbSize);

for i = 1:filterNum2

for j = 1:filterNum1

o2(:,:,i,:) = o2(:,:,i,:) + convn(o1Pooled(:,:,j,:),rot90(Wc2(:,:,j,i),2),'valid');

end

o2(:,:,i,:) = o2(:,:,i,:) + bc2(i);

end

o2Pooled = zeros(outputDim2,outputDim2,filterNum2,mbSize);

o2PoolIdx = zeros(outputDim2^2,filterNum2,mbSize);

for i = 1:mbSize

for j = 1:filterNum2

[o2Pooled(:,:,j,i) o2PoolIdx(:,j,i)] = MaxPooling(o2(:,:,j,i),[poolDim2 poolDim2]);

end

end

o2PooledVec = reshape(o2Pooled,[],mbSize);

o3 = Wd*o2PooledVec + repmat(bd,[1,mbSize]);

% Back Propagation

y = labelMat(:,rp(s:s+mbSize-1));

delta_d = o3 - y;

delta_s2 = Wd' * delta_d;

delta_s2 = reshape(delta_s2,outputDim2,outputDim2,filterNum2,mbSize);

delta_c2 = zeros(convDim2,convDim2,filterNum2,mbSize);

for i = 1:mbSize

for j = 1:filterNum2

delta_c2(:,:,j,i) = upsampleMax(delta_s2(:,:,j,i),o2PoolIdx(:,j,i),poolDim2);

end

end

delta_s1 = zeros(outputDim1,outputDim1,filterNum1,mbSize);

for i = 1:filterNum1

for j = 1:filterNum2

delta_s1(:,:,i,:) = delta_s1(:,:,i,:) + convn(delta_c2(:,:,j,:),Wc2(:,:,i,j),'full');

end

end

delta_c1 = zeros(convDim1,convDim1,filterNum1,mbSize);

for i = 1:mbSize

for j = 1:filterNum1

delta_c1(:,:,j,i) = upsampleMax(delta_s1(:,:,j,i),o1PoolIdx(:,j,i),poolDim1);

end

end

grad_Wd = (1/mbSize)*delta_d*o2PooledVec';

grad_bd = zeros(size(bd));

for i = 1:classNum

grad_bd(i) = (1/mbSize)*sum(delta_d(i,:));

end

grad_Wc2 = zeros(size(Wc2));

grad_bc2 = zeros(size(bc2));

for i = 1:filterNum2

for j = 1:filterNum1

for k = 1:mbSize

grad_Wc2(:,:,j,i) = grad_Wc2(:,:,j,i) + conv2(o1Pooled(:,:,j,k),rot90(delta_c2(:,:,i,k),2),'valid');

end

grad_Wc2(:,:,j,i) = (1/mbSize)*grad_Wc2(:,:,j,i);

end

tmp_grad_bc2 = delta_c2(:,:,i,:);

grad_bc2(i) = (1/mbSize)*sum(tmp_grad_bc2(:));

end

grad_Wc1 = zeros(size(Wc1));

grad_bc1 = zeros(size(bc1));

for i = 1:filterNum1

for j = 1:mbSize

grad_Wc1(:,:,i) = grad_Wc1(:,:,i) + conv2(mbImages(:,:,j),rot90(delta_c1(:,:,i,j),2),'valid');

end

grad_Wc1(:,:,i) = (1/mbSize)*grad_Wc1(:,:,i);

tmp_grad_bc1 = delta_c1(:,:,i,:);

grad_bc1(i) = (1/mbSize)*sum(tmp_grad_bc1(:));

end

vel_Wd = mom*vel_Wd - r*grad_Wd - wdr*Wd;

vel_bd = mom*vel_bd - r*grad_bd - wdr*bd;

vel_Wc2 = mom*vel_Wc2 - r*grad_Wc2 - wdr*Wc2;

vel_bc2 = mom*vel_bc2 - r*grad_bc2 - wdr*bc2;

vel_Wc1 = mom*vel_Wc1 - r*grad_Wc1 - wdr*Wc1;

vel_bc1 = mom*vel_bc1 - r*grad_bc1 - wdr*bc1;

Wd = Wd + vel_Wd;

bd = bd + vel_bd;

Wc2 = Wc2 + vel_Wc2;

bc2 = bc2 + vel_bc2;

Wc1 = Wc1 + vel_Wc1;

bc1 = bc1 + vel_bc1;

mbErr = mean(delta_d(:).^2);

err = err + mbErr;

end

fprintf('Epoch %d Training error: %f\n',e,err);

r = 0.98*r;

end

cnnTest;

源代码为 github.com/yaolubrain/cnn_linear_max