About loss

[joyyang1997]

Hi, it seems that you ignore the fake image condition loss: E[||D(G(I|y)) - y||^2], is this better for training?

# D(fake_I)
d_fake_desired_img_prob, **_** = self._D.forward(fake_imgs_masked.detach())
self._loss_d_fake = self._compute_loss_D(d_fake_desired_img_prob, False) * self._opt.lambda_D_prob

[vipermu]

In this step you're training the Discriminator in the case when you have a fake image in the input. You don't want to update the Action Units classification in this case.