I believe this is due to a bug in the R2022b version of the custom projectAndReshapeLayer attached to the example. In particular in the initialize method the layer.Weights and layer.Bias are replaced with their initial values even if they already have trained values. The initialize method is called when you load the saved generator network.
You can update the initialize method in the custom layer to the following:
function layer = initialize(layer,layout)
% layer = initialize(layer,layout) initializes the layer
% learnable parameters.
%
% Inputs:
% layer - Layer to initialize
% layout - Data layout, specified as a
% networkDataLayout object
%
% Outputs:
% layer - Initialized layer
% Layer output size.
outputSize = layer.OutputSize;
% Initialize fully connect weights.
if isempty(layer.Weights)
% Find number of channels.
idx = finddim(layout,"C");
numChannels = layout.Size(idx);
% Initialize using Glorot.
sz = [prod(outputSize) numChannels];
numOut = prod(outputSize);
numIn = numChannels;
layer.Weights = initializeGlorot(sz,numOut,numIn);
end
% Initialize fully connect bias.
if isempty(layer.Bias)
% Initialize with zeros.
layer.Bias = initializeZeros([prod(outputSize) 1]);
end
end
The if isempty(layer.Weights) and if isempty(layer.Bias) checks ensure that the trained projection is not lost on load.
