error while transferring weights of a trained CNN network to an empty CNN network

Hi,

I am trying to transfer the weights of layer 11 from 'original_net' to layer 11 of 'layers_final'. Both have same structure and 'layer_final' is just the empty, untrained version of 'original net'. i am using the following command:

Layers_final(11).Weights = net_1.Layers(11).Weights

I get the following error while doing so:

Error using nnet.cnn.layer.TransposedConvolution2DLayer/set.Weights (line 204)

Expected input to be of size 4x4x8x1, but it is of size 4x4x8x8.

code for layers_final:

imageLayer_final = imageInputLayer([32,32,1]);
encodingLayers_final = [ ...
    convolution2dLayer(3,16,'Padding','same'), ...
    reluLayer, ...
    maxPooling2dLayer(2,'Padding','same','Stride',2), ...
    convolution2dLayer(3,8,'Padding','same'), ...
    reluLayer, ...
    maxPooling2dLayer(2,'Padding','same','Stride',2), ...
    convolution2dLayer(3,8,'Padding','same'), ...
    reluLayer, ...
    maxPooling2dLayer(2,'Padding','same','Stride',2)];
decodingLayers_final = [ ...
    createUpsampleTransponseConvLayer(2,8), ...
    reluLayer, ...
    createUpsampleTransponseConvLayer(2,8), ...
    reluLayer, ...
    createUpsampleTransponseConvLayer(2,16), ...
    reluLayer, ...
    convolution2dLayer(3,1,'Padding','same'), ...
    clippedReluLayer(1.0), ...
    regressionLayer];
layers_final = [imageLayer,encodingLayers,decodingLayers];

net_original attached with the question.

Thanks

Answer 1

[Srivardhan Gadila]

If the function createUpsampleTransponseConvLayer is the helper function from the example Prepare Datastore for Image-to-Image Regression then change the 'NumChannels' Name-Value Pair Argument to 'auto' or don't mention it in the transposedConv2dLayer function.

% helper function from the example Prepare Datastore for Image-to-Image Regression
function out = createUpsampleTransponseConvLayer(factor,numFilters)

filterSize = 2*factor - mod(factor,2); 
cropping = (factor-mod(factor,2))/2;
numChannels = 1;

out = transposedConv2dLayer(filterSize,numFilters, ... 
    'NumChannels',numChannels,'Stride',factor,'Cropping',cropping);
end

Since the layer is defined with 'NumChannels' (number of channels of the input to this transposedConv2dLayer) as 1 hence it can accept wieghts of size "filterSize x filterSize x numFilters x numChannels" which is 4x4x8x1 in this case.

Change the function as follows:

function out = createUpsampleTransponseConvLayer(factor,numFilters)

filterSize = 2*factor - mod(factor,2); 
cropping = (factor-mod(factor,2))/2;

out = transposedConv2dLayer(filterSize,numFilters, ... 
    'Stride',factor,'Cropping',cropping);
end

and then define the layers.

[Radians]

why does it then show in the above picture the dimensions of the 11th layer of layers_final as 4x4x8x8, instead of 4x4x8x1, which is created by the same function as you mentioned?

[Srivardhan Gadila]

I have heard that this issue is known and the concerned parties might be working on it.

[Radians]

Hello

I know its been long now but if I try to perform the same operation with the transpose convolution function as follows:

O_dltconv1=dltranspconv(O_maxpool3,K_tconv_1,B_tconv_1,'Stride',2,'Cropping',1);

with:

K>> size(O_maxpool3)

ans =

4 4 8 500

K>> size(K_tconv_1)

ans =

4 4 8

K>> size(B_tconv_1)

ans =

1 1 8

I get the following error:

Number of channels to convolve (1, specified by the size of the 'U' dimension of the weights) must be equal

to the size of the 'C' dimension of the input data (8).

Could you please highlight what am I doing wrong, since I set the channels of the transpose convolution filter to 1 as you said?