How to match output size in cnn

I am trying to train a cnn to take as input a grayscale image (25x25) and output also an image (25x25). I have created the training as follows: (I1 is the input and I2 is the response)

[I1, I2] = generateImage();
        X(:,:,:,i) = I1;
        Y(i,:,:,:) = I2;

The I set the network and try to train it:

%create network layers
layers = [...
    imageInputLayer([25 25 1])
    convolution2dLayer([4 3],12)
    reluLayer
    crossChannelNormalizationLayer(4)
    maxPooling2dLayer(2,'Stride',2)
    convolution2dLayer(5,16)
    reluLayer
    crossChannelNormalizationLayer(4)
    maxPooling2dLayer(2,'Stride',2)
    fullyConnectedLayer(256)
    reluLayer
    fullyConnectedLayer(10)
    regressionLayer];

%create training option
options = trainingOptions('sgdm','InitialLearnRate',0.001, ...
    'MaxEpochs',15);

%create network
net = trainNetwork(X,Y,layers,options)

The message I get is:

Error using trainNetwork (line 92)
The output size [1 1 10] of the last layer doesn't match the response size [1 25 25].

Any ideias on to fix this?

[Kanushka Gajjar]

Hi,

Did you find a solution to this problem?

I am having the same problem.

Thanks,

Kanushka

Answer 1

[Abel Babu]

This error is due to the ' fullyConnectedLayer'. If you see the following documentation it is clear that the output of this layer is a single dimension vector

https://www.mathworks.com/help/nnet/ref/fullyconnectedlayer-class.html?searchHighlight=outputsize&s_tid=doc_srchtitle#bu6spu9-3

The workaround is to unroll your input image matrix to make it single dimension vector and then training it. I have modified your code by taking random matrices as inputs and response.

   layers = [...
    imageInputLayer([25 25 1])
    convolution2dLayer([4 3],12)
    reluLayer
    crossChannelNormalizationLayer(4)
    maxPooling2dLayer(2,'Stride',2)
    convolution2dLayer(5,16)
    reluLayer
    crossChannelNormalizationLayer(4)
    maxPooling2dLayer(2,'Stride',2)
    fullyConnectedLayer(256)
    reluLayer
    fullyConnectedLayer(25*25)%Change the dimensions to match the outputs.
    regressionLayer;
    ];
%create training option
options = trainingOptions('sgdm','InitialLearnRate',0.001, ...
    'MaxEpochs',15);
%create network
        X(:,:,:,1) = rand(25);
        Y=randn(1,1,25*25,1);
net = trainNetwork(X,Y,layers,options)

I hope the above code will help you.

Regards,

Darshan Bhat

[Osama Tabbakh]

I wounder why u put two fullyConnectedLayers. And can I somehow filter the output also?