kplum
commented
3 years ago The third loss loss_G_GAN is the adversarial loss for the generator G. This loss models the generator's objective to make the discriminator "classify" fake images as real (thus the True instead of False), thus fooling him.
This loss is used alongside the feature matching loss and VGG loss to update only the generator's weights.
Whereas the first two losses loss_D_fake and loss_D_real constitute the adversarial loss(es) for the discriminator D, whose objective it is to classify real images as real and fake images as fake.
In other words, the second argument to self.criterionGAN, i.e. the __call__ function of the class GANLoss, states not whether the input image was real, but whether or not the discriminator should, for this loss function to be minimal, classify the input image as real.
Wangyulin-user
pred_fake_pool = self.discriminate(input_label, fake_image, use_pool=True) loss_D_fake = self.criterionGAN(pred_fake_pool, False)
pred_real = self.discriminate(input_label, real_image) loss_D_real = self.criterionGAN(pred_real, True)
pred_fake = self.netD.forward(torch.cat((input_label, fake_image), dim=1))
loss_G_GAN = self.criterionGAN(pred_fake, True)
In the last line, calculating loss_G_GAN, the input is the input label and the generated fake image, but in the criterionGAN(), the input is pred_fake and True. Shouldn't that be False?
I can see the the first line, pred_fake_pool, where the input is the input label and the generated fake image, the input for criterionGAN is pred_fake_pool and False. This is as expected. But what does the third loss do?