GAN example: Only one backward() call?

[alainjungo]

In the PyTorch GAN tutorial https://pytorch.org/tutorials/beginner/dcgan_faces_tutorial.html there are two backward() calls for the discriminator. How do you ensure this with your structure, where backward() gets called after the training step?

Best, Alain

[williamFalcon]

good catch. we don't actually support this atm.

A solution here is to allow a dictionary of options for each optimizer which allows arbitrary number of calls. Something like:

def configure_optimizers(self,...):
    opt_G = {'optimizer': Adam(...), 'frequency': 2, 'lr_scheduler': LRScheduler(...)}
    opt_D = {'optimizer': Adam(...), 'frequency': 1, 'lr_scheduler': LRScheduler(...)}

    return opt_G, opt_D

Here G would be called twice back to back, and G once after

@jeffling @neggert

But not sure if this is a clean user experience

[jeffling]

@williamFalcon that API would work for us.

@alainjungo: Some workarounds, none of them ideal:

1. Skip every other generator step. You'll have to double your iterations
2. Double learning rate on D (not algorithmically the same but can have similar effect)

[kwanUm]

You can always return one loss at the training step that captures both losses. In the dcgan example, this would look like errD = errD_real + errD_fake; errD.backword();.

Not sure if I'm correct here, but this seems equivalent and matches PTL paradigms.

[asafmanor]

I have implemented an API that allows returning optimizer, lr_schedulers, optimizer_frequencies, and then based on the batch_idx will determine the current optimizer to use.

If there is an agreement on this API, I'll proceed to testing, documenting and submitting a PR.

Another option would be to allow returning a tuple of dictionaries as @williamFalcon suggested. that would be a minor change for me and I am willing to that if it is agreed upon.

[asafmanor]

I am adding here that the official implementation of WGAN-GP (for example) needs this feature in order to converge. This is a very fundamental feature in GAN training.

[Borda]

I like the @williamFalcon which seems clear to me... @asafmanor mind sens a PR or describe what API you have in mind? cc: @PyTorchLightning/core-contributors any comments on this?

[asafmanor]

I'll implement the @williamFalcon API and send a detailed PR over the weekend 👍