SpectralNormalizati​on Layer (for SNGAN)

spectral normalization layer input : layer , name

recommend using spectralNormalize

layers = [

imageInputLayer(inputSize)

convolution2dLayer(filterSize,fs,'Stride',s,'Padding','same','Name','conv1')

] ;

lg = layerGraph(layers);

net = dlnetwork(lg);

u = nan;

v = nan;

% in each iteration before foward(net,x)

[u,v,net] = spectralNormalize(u,v,net)

--------------------------------------------------------------------------------------------------------------------------------

input a conv2d layer or transposedconv2d layer and name for sn layer.

inputlayer need : filter size , num filters (output channel size), num channels(input channel size).

example :

SpectralNormalization(convolution2dLayer(filterSize,numFilters,"NumChannels", 3, 'Stride',2,'Padding','same','Name','conv1') , "sn1" )

SpectralNormalization( transposedConv2dLayer(filterSize, numFilters,"NumChannels", 64, 'Name','detc5',"Stride",2 ,'Cropping','same') , "sn19" )

or you can see how the SpectralNormalization work in code, then coding by self