Question about the Matlab Wasserstein GAN example

The original Wasserstein gan paper suggest removing the Critic's last dense layer activation function(sigmoid) such that the output value is not limited to fake or real. The posted example still uses sigmoid layer, am I right?

Answer 1

[Malay Agarwal]

Hi @Mohammed,

The diagram of the Discriminator model in the example (https://www.mathworks.com/help/deeplearning/ug/trainwasserstein-gan-with-gradient-penalty-wgan-gp.html) shows that the model does have a “sigmoid” layer at the end:

This can also be confirmed by looking at how the Discriminator model is defined:

layersD = [
    imageInputLayer(inputSize,Normalization="none")
    convolution2dLayer(filterSize,numFilters,Stride=2,Padding="same")
    leakyReluLayer(scale)
    convolution2dLayer(filterSize,2*numFilters,Stride=2,Padding="same")
    layerNormalizationLayer
    leakyReluLayer(scale)
    convolution2dLayer(filterSize,4*numFilters,Stride=2,Padding="same")
    layerNormalizationLayer
    leakyReluLayer(scale)
    convolution2dLayer(filterSize,8*numFilters,Stride=2,Padding="same")
    layerNormalizationLayer
    leakyReluLayer(scale)
    convolution2dLayer(4,1)
    sigmoidLayer]; % Notice the sigmoid layer at the end

Hope this helps!