Incorrect discriminator update for opt.use_same_D

[SanderGielisse]

Hi, I believe there is a little mistake in the way the use_same_D option is processed. In case this option is enabled, in bicycle_gan_model.py, self.backward_D is called twice, both on self.netD. Since the backward call is done twice then, without retain_graph=True, the first backward pass is overwritten, thus D is only updated on self.real_data_random and self.fake_data_random, not on self.real_data_encoded and self.fake_data_encoded. Either retain_graph=True should be used, or the backward pass should incorporate both losses like, (self.loss_D + self.loss_D2).backward()

BicycleGAN/models/bicycle_gan_model.py

Line 175 in 40b9d52

if self.opt.use_same_D:

[SanderGielisse]

My bad, it seems that when backward() is called twice, it accumulates into .grad, so the implementation is correct. Closing this issue.