Using gradient_clip_val only for Discriminator in GAN

[Bajo1994]

How should I use gradient_clip_val if I only want to clip the discriminator in GAN?
Currently, I try to clip the discriminator as follows and I get an error:

   def optimizer_step(self, current_epoch, batch_idx, optimizer, optimizer_idx,*args,**kwargs):
     optimizer.step()
     optimizer.zero_grad()
     if optimizer_idx==0:
       for p in self.Discriminator.parameters():#weight clipping
         p.data.clamp_(-0.01, 0.01)

The error is:

TypeError: optimizer_step() missing 1 required positional argument: 'current_epoch'

I use pl==1.1.0 on colab

[rohitgr7]

should clip_gradient be a part of LightningModule? something like configure_gradient_clipping? cc @PyTorchLightning/core-contributors

[carmocca]

should clip_gradient be a part of LightningModule? something like configure_gradient_clipping`?

I'd say no if we can do it with the current API

[rohitgr7]

I'd say no if we can do it with the current API

Do you mean using manual optimization?

[carmocca]

After some discussion, I agree with @rohitgr7 that configuring clipping should be a responsibility of the LightningModule.

Relevant part:

https://github.com/PyTorchLightning/pytorch-lightning/blob/c00d5709c4fb40259ad05fe827e24ad421ce2da2/pytorch_lightning/accelerators/accelerator.py#L114-L130

Trainer flag (to deprecate)

https://github.com/PyTorchLightning/pytorch-lightning/blob/c00d5709c4fb40259ad05fe827e24ad421ce2da2/pytorch_lightning/trainer/trainer.py#L90