After groupedCon​volution2d​Layer in network branches a corresponding depthConcatenationLayer needed

Hi, All!

I use groupedConvolution2dLayer in my network for processing in parallel 5 image fragments. So, I am using 5 groups of channels.

In my network I have three branches and I need to concatenate outputs of these branches. There is a depthConcatenationLayer for this purpose in Matlab:

However I need to concatenate layers in the right order: first channels of group #1 of all three branches, then channels of group #2 of all three branches, ..., and finally channels of group #5 of all three branches. How to do it? Matlab's depthConcatenationLayer does not allow to specify order of channels.

I tryed to create a custom grouped depth concatenation layer:

classdef myGroupedDepthConcatenationLayer < nnet.layer.Layer

properties

GroupNumber

end

methods

function layer = myGroupedDepthConcatenationLayer(groups, numinputs, name)

layer.Name = name;

layer.NumInputs = numinputs;

layer.Description = 'Custom grouped depth concatenation layer';

layer.GroupNumber = groups;

end

function Z = predict(layer,varargin)

X = varargin;

c = zeros(1, layer.NumInputs);

for i = 1:layer.NumInputs

s = size(X{i});

c(i) = s(3);

end

if length(s)<4

n=1;

else

n=s(4);

end

Z = X{1}; Z(s(1),s(2),sum(c),n)=0; % memory is allocated

ofset = 0;

for j=1:layer.GroupNumber

for i=1:layer.NumInputs

len = c(i)/layer.GroupNumber;

Z(:,:,ofset+1:ofset+len,:,:)=X{i}(:,:,(j-1)*len+1:j*len,:);

ofset=ofset+len;

end

end

end

end

end

Is it correct? How will the channels reordering in the custom layer affect backward propagation duiring network training?

How to distribute gradients between inputs correctly?