About discriminator loss

[qibao77]

Why there are two types of discriminator loss as follows:
# pred_d_real = self.netD(self.var_ref)
# pred_d_fake = self.netD(self.fake_H.detach()) # detach to avoid BP to G
# l_d_real = self.cri_gan(pred_d_real - torch.mean(pred_d_fake), True)
# l_d_fake = self.cri_gan(pred_d_fake - torch.mean(pred_d_real), False)
# l_d_total = (l_d_real + l_d_fake) / 2
# l_d_total.backward()
pred_d_fake = self.netD(self.fake_H.detach()).detach()
pred_d_real = self.netD(self.var_ref)
l_d_real = self.cri_gan(pred_d_real - torch.mean(pred_d_fake), True) * 0.5
l_d_real.backward()
pred_d_fake = self.netD(self.fake_H.detach())
l_d_fake = self.cri_gan(pred_d_fake - torch.mean(pred_d_real.detach()), False) * 0.5
l_d_fake.backward()

[xinntao]

The discriminator has two losses for real and fake samples, respectively~