Develop WGAN-GP for 3-D image

Hi, I am beginner in using MATLAB to develop generative adversarial networks (GANs). Based on the MATLAB WGAN-GP tutorial, I have developed a WGAN-GP model for 3-D images (H x W x D x C). The modified 'modelGradientsD' function is shown as below.

function [gradientsD, lossD, lossDUnregularized, D_X, D_G_Z1] = modelGradientsD(dlnetG, dlnetD, dlZ, dlX, lambda)
% Calculate the prediction for training images with D
dlYPred = forward(dlnetD, dlX);

% Calculate the prediction for G-generated images with D
dlXGenerated = forward(dlnetG, dlZ);
dlYPredGenerated = forward(dlnetD, dlXGenerated);

% Calculate D(X) and D(G(Z))
D_X = mean(dlYPred);
D_G_Z1 = mean(dlYPredGenerated);

% Calculate the unregularized loss
lossDUnregularized = D_G_Z1 - D_X;

% Get the interpolated image from the training and generated images
epsilon = rand([1 1 1 1 size(dlX,5)], 'like', dlX);
dlXInterpolated = epsilon.*dlX + (1-epsilon).*dlXGenerated;
dlYPredInterpolated = forward(dlnetD, dlXInterpolated);

% Calculate the gradient penalty
gradientsInterpolated = dlgradient(sum(dlYPredInterpolated), dlXInterpolated, 'EnableHigherDerivatives', true);
gradientsInterpolatedNorm = sqrt(sum(gradientsInterpolated.^2,1:4) + 1e-10);
gradientPenalty = lambda.*mean((gradientsInterpolatedNorm - 1).^2);

% Calculate the loss with gradient penalty
lossD = lossDUnregularized + gradientPenalty;

% Calculate the gradients of the loss with respect to learnable parameters
gradientsD = dlgradient(lossD, dlnetD.Learnables);
end

When running the program, however, an error pops up (as shown below). It seems that there is something wrong in calculating 'gradientsD'. After many debug attempts, I find that removing 'EnableHigherDerivatives' from calculating 'gradientsInterpolated' can make it. But the WGAN-GP perform not well, and I have several questions: (1) Does removing 'EnableHigherDerivatives' affect the model training significantly? (2) Is there robustness issue in the 'dlgradient' function? (3) Are there other solutions to this error?

I really appreciate it if you could offer any idea or suggestion. Thanks a lot!

Error using  + 
Arrays have incompatible sizes for this operation.

Error in gpuArray/internal_dlconv (line 57)
    stride, dilation, numGroups) + bias;

Error in deep.internal.recording.operations.DlconvBackwardOp/backward (line 88)
                ddZ2 = internal_dlconv(ddX,weights,zeroBias,op.Args{:});

Error in deep.internal.recording.RecordingArray/backwardPass (line 89)
    grad = backwardTape(tm,{y},{initialAdjoint},x,retainData,false,0);

Error in dlarray/dlgradient (line 132)
[grad,isTracedGrad] = backwardPass(y,xc,pvpairs{:});

Error in WGANGP_V1>modelGradientsD (line 442)
gradientsD = dlgradient(lossD, dlnetD.Learnables);

Error in deep.internal.dlfeval (line 17)
[varargout{1:nargout}] = fun(x{:});

Error in dlfeval (line 40)
    [varargout{1:nargout}] = deep.internal.dlfeval(fun,varargin{:});

Error in WGANGP_V1 (line 226)
        [gradientsD, lossD, lossDUnregularized, D_X, D_G_Z1] = dlfeval(@modelGradientsD, dlnetG, dlnetD, dlZ, dlX, lambda);

Answer 1

[Joss Knight]

You definitely need to use EnableHigherOrderDerivatives here because you are including computed gradients in the loss term. Without it your training will not work correctly.

It looks like this is a bug with higher order derivatives and 3-D data, which is fixed in R2022a. Can you get the latest version of MATLAB?

https://www.mathworks.com/support/bugreports/2661783

[Shuaibin WAN]

Hi! Joss Knight. Thanks a lot for your reply! Glad to hear that this bug has been fixed in R2022a. This crazy bug has bothered me for days. Sadly, my university MATLAB license remains R2021b (update 4), and I can't enjoy the latest version now. Hope it will be fixed as well soon.

[Joss Knight]

I have requested for this to be fixed in a future update of R2021b.

[Shuaibin WAN]

That sounds GREAT! Thank you so much, sir.