Hello Yang Liu
From what I understand, you are following the “Generate Synthetic Signals Using Conditional GAN” MATLAB example to train a conditional GAN work with a 2-Dimensional input.
The error you have encountered is occurring due to a mismatch in the output dimension of the generator network and the input dimension of the discriminator network.
The discriminator network has been adapted correctly for the use case and expects a 14x8x1 input. The last transposed convolutional layer of the generator network has ‘numFilters’ set as 8. This results in an output dimension of 14x8x8. Kindly make the following change to fix the network.
transposedConv2dLayer([3 3],1,'Stride',1,'Name','TransConv4') % 14*8*1
Additionally, you can either comment out or remove the following lines from the ‘trainGAN.m’ file since they are used to visualise the signal data specific to the example and for that reason it will throw an error for your specific data.
% if mod(ct,50) == 0 || ct == 1
% % Generate signals using held-out generator input
% dlXGeneratedValidation = predict(dlnetGenerator, dlZValidation, dlTValidation);
% dlXGeneratedValidation = squeeze(extractdata(gather(dlXGeneratedValidation)));
%
% % Display spectra of validation signals
% subplot(1,2,1);
% pspectrum(dlXGeneratedValidation);
% set(gca, 'XScale', 'log')
% legend('Healthy', 'Faulty')
% title("Spectra of Generated Signals")
% end
For further understanding, I suggest you refer to the following MathWorks Documentation and resources:
- Refer to the “Input Arguments” section to understand further about various parameters of ‘transposedConv2dlayer’: https://www.mathworks.com/help/releases/R2023b/deeplearning/ref/transposedconv2dlayer.html
I hope you find the above explanation and suggestions useful!
