Is there a bug about cyclegan?

[onepeachbiubiubiu]

I noticed that the training code does not set ‘requires_grad=False’ for the discriminator when training generators, which will conduct gradients to discriminators and force the discriminator to treat the fake image as a 'true image'.

[GaParmar]

I am a little confused by this. Could you point to the relevant line in the training code?

Here, the fake image is detached which should stop the flow of gradients:
https://github.com/GaParmar/img2img-turbo/blob/main/src/train_cyclegan_turbo.py#L232

-Gaurav

[onepeachbiubiubiu]

When training the generator, the discriminator's 'requires_grad' is Ture. So https://github.com/GaParmar/img2img-turbo/blob/main/src/train_cyclegan_turbo.py#L205 will produce gradients for the discriminator. The gradients will update to the discriminator's parameter in https://github.com/GaParmar/img2img-turbo/blob/main/src/train_cyclegan_turbo.py#L238.

[yahshibu]

I think that is the case for train_pix2pix_turbo.py.

Gradients for the discriminator will be produced at

img2img-turbo/src/train_pix2pix_turbo.py

Line 200 in 573dfdd

accelerator.backward(lossG)

.
And, the gradients won't be deleted until

img2img-turbo/src/train_pix2pix_turbo.py

Line 217 in 573dfdd

optimizer_disc.zero_grad(set_to_none=args.set_grads_to_none)

.

I think
accelerator.unwrap_model(net_disc).requires_grad_(False)
and
accelerator.unwrap_model(net_disc).requires_grad_(True)
should be called at appropriate places.

Thank you for finding this! @onepeachbiubiubiu