Problem with vl_nnconv while training a FCN about matconvnet-fcn

[omrysendik]

I was able to find the problem and fix it by changing lines 49 to 60 in fcnInitializeModel to:

% Modify the last fully-connected layer to have 21 output classes
% Initialize the new filters to zero
Biases = 0;
for i = net.getParamIndex(net.layers(end-1).params) ;
sz = size(net.params(i).value) ;
if(~Biases)
sz(end) = 21 ;
else
sz(end-1) = 21 ;
end
net.params(i).value = zeros(sz, 'single') ;
Biases = 1;
end

[codesteller86]

@omrysendik: Thank you. it works