Channel-wise convolution deep learning layer for 1d and 3d data i.e. groupedCon​volution1d​Layer, groupedCon​volution3d​Layer

I am looking to implement a layer for channel-wise convolution similar to groupedConvolution2dLayer(filtSize, 1, "channel-wise"...). Unfortunately there are no 1D and 3D analogs of groupedConvolution2dLayer.

Any recommendations would be helpful?

Answer 1

[Matt J]

A 1D image is a special case of a 2D image, so you should be able to use groupedConvolution2dLayers for the 1D case.

[Matt J]

For the 3D case, you can use the deepnetwork designer to build an equivalent sub-network out of functionLayers and a depthConcatenationLayer, e.g.,

The functionLayers serve to extract a subset of channels from the input 3D image and can be built using the following

subsets={1:2,3:4};

subslayers=channelSubset(subsets);


function subslayers=channelSubset(subsets)

    N=numel(subsets);
     
    for i=1:N
     subslayers(i)=functionLayer(@(z) z(:,:,:,subsets{i}),'Name',"Group"+i); %#ok<AGROW> 
    end


end

[Artem Lensky]

Matt, thank you for the prompt response! Very much appreciated.

The input data for 1D case are 1D signals (in principal of different length, that is why I chose sequenceInputLayer as input layer). The tensor that is passed to the convolution layer is of "CBT" shape. To make it compatible with groupedConvolution2dLayer, I implemented two simple layers (cbt2sscbLayer and sscb2cbtLayer) that map "CBT" to "SSCB" and then "SSCB" to "CBT". So the network is organized in the following way

[...

cbt2sscbLayer()...

groupedConvolution2dLayer([recSize 1], 1, "channel-wise")...

sscb2cbtLayer()...

instanceNormalizationLayer()...

]

The mapping layers imlement method predict as follows

% cbt2sscbLayer

function Z = predict(layer, X)

idx = finddim(X, "T");

numSamples = X.size(idx(1));

idx = finddim(X, "C");

numChannels = X.size(idx);

idx = finddim(X, "B");

if(isnan(X.size(idx)))

numBatches = 1;

else

numBatches = X.size(idx);

end

X = dlarray(X,'UUU');

Z = ones(numSamples, 1, numChannels, numBatches);

Z(:,1,:,:) = permute(X, [3, 1, 2]);

Z = dlarray(Z,'SSCB');

end

and

% sscb2cbtLayer

function Z = predict(layer, X)

idx = finddim(X, "S");

numSamples = X.size(idx(1));

idx = finddim(X, "C");

numChannels = X.size(idx);

idx = finddim(X, "B");

if(isnan(X.size(idx)))

numBatches = 1;

else

numBatches = X.size(idx);

end

X = X.reshape([numSamples, numChannels, numBatches]);

X = dlarray(X,'UUU');

Z = X.permute([2 3 1]);

Z = dlarray(Z,'CBT');

end

Unfortunately the model is not able to learn.

Do I need to implement function backward? Any recommendations would be apprecitated.

[Artem Lensky]

UPDATE: The solution above seemed to work for 1D signals. It didn't work inititally because of some hyperparameter that is not relevant to the implementation above.

I will do a bit of testing and then will try implementing/testing the idea for the 3D case.