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)
dlYPred = forward(dlnetD, dlX);
dlXGenerated = forward(dlnetG, dlZ);
dlYPredGenerated = forward(dlnetD, dlXGenerated);
D_G_Z1 = mean(dlYPredGenerated);
lossDUnregularized = D_G_Z1 - D_X;
epsilon = rand([1 1 1 1 size(dlX,5)], 'like', dlX);
dlXInterpolated = epsilon.*dlX + (1-epsilon).*dlXGenerated;
dlYPredInterpolated = forward(dlnetD, dlXInterpolated);
gradientsInterpolated = dlgradient(sum(dlYPredInterpolated), dlXInterpolated, 'EnableHigherDerivatives', true);
gradientsInterpolatedNorm = sqrt(sum(gradientsInterpolated.^2,1:4) + 1e-10);
gradientPenalty = lambda.*mean((gradientsInterpolatedNorm - 1).^2);
lossD = lossDUnregularized + gradientPenalty;
gradientsD = dlgradient(lossD, dlnetD.Learnables);
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!
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);
