I can`t solve out this problem, there is always Output argument "varargout{2}" (and possibly others) not assigned a value in the execution with "dlarray/dlgradient" function.

function [netG, stateG, lossG] = modelGStep(netG, wrappedImage, realImage, stateG, learningRate, beta1, beta2)
% insure GPU dlarray
if ~isa(wrappedImage, 'dlarray')
    wrappedImage = dlarray(gpuArray(wrappedImage), 'SSCB');
elseif ~strcmp(underlyingType(wrappedImage), 'gpuArray')
    wrappedImage = dlarray(gpuArray(extractdata(wrappedImage)), 'SSCB');
end
if ~isa(realImage, 'dlarray')
    realImage = dlarray(gpuArray(realImage), 'SSCB');
elseif ~strcmp(underlyingType(realImage), 'gpuArray')
    realImage = dlarray(gpuArray(extractdata(realImage)), 'SSCB');
end
wrappedImage = dlarray(gpuArray(wrappedImage), 'SSCB');
realImage = dlarray(gpuArray(realImage), 'SSCB');
% insure dlfeval use dlgradient
[gradG, lossG] = dlfeval(@dlgradient, lossG, netG.Learnables);
fakeImage = predict(netG, wrappedImage);
lossG = mean((fakeImage - realImage).^2, 'all');
[gradG, lossG] = dlgradient(lossG, netG.Learnables);
[netG, stateG] = adamupdate(netG, gradG, stateG, learningRate, beta1, beta2);
return
end

• this is my function.
• below is my code

for epoch = 1:epochs
    for i = 1:size(unwrapImages, 4)
        realImage = unwrapImages(:,:,:,i);
        wrappedImage = wrappedImages(:,:,:,i);
        [netG, stateG, lossG] = modelGStep(netG, wrappedImage, realImage, stateG, learningRate, beta1, beta2);
        [lossD, gradD] = modelDStep(netD, realImage, wrappedImage, netG);
        [netD, stateD] = adamupdate(netD, gradD, stateD, learningRate, beta1, beta2);
        gLosses(epoch) = gLosses(epoch) + double(gather(extractdata(lossG)));
        dLosses(epoch) = dLosses(epoch) + double(gather(extractdata(lossD)));
    end
    gLosses(epoch) = gLosses(epoch) / size(unwrapImages, 4);
    dLosses(epoch) = dLosses(epoch) / size(unwrapImages, 4);
    fprintf('Epoch %d, Generator Loss: %.4f, Discriminator Loss: %.4f\n', ...
        epoch, gLosses(epoch), dLosses(epoch));
end

what should i do to solve this,thanks!

Answer 1

[Joss Knight]

Your mistake is dlfeval(@dlgradient,...). You need to put your code that computes the loss and the loss gradients into a function and then pass that to dlfeval.

[CHEN]

oh i get.It`s a simple use-problem.I make it complex,thanks！