AttributeError: 'LightningDistributedModule' object has no attribute 'require_backward_grad_sync'

[sunshuofeng]

🐛 Bug

When I use manual_backward ，this bug appear 。
I know the new version may fix this, but for some reasons I can only use version 1.2.4, what should I do?

Environment

• CUDA:
  - GPU:
  - available: True
  - version: 11.1
• Packages:
  - numpy: 1.21.2
  - pyTorch_debug: False
  - pyTorch_version: 1.8.1+cu111
  - pytorch-lightning: 1.2.4
  - tqdm: 4.62.3
• System:
  - OS: Linux
  - architecture:
  - 64bit
  -
  - processor: x86_64
  - python: 3.7.11
  - version: ModuleNotFoundError: No module named 'demo' #4 SMP Mon Mar 30 12:42:28 HKT 2020

Thanks for your help!!

[kaushikb11]

Hey @sunshuofeng, could you update PyTorch Lightning to the latest?

[sunshuofeng]

I wish I could do this, but due to some version conflict issues, I can only use version 1.2

What I want to do is optimize center Loss, so there will be two optimizers,I've tried many other solutions

Optimization of Center Loss requires a coefficient multiplication of the gradient, and an error-free approach is：

def optimizer_step（.......）：
       ## optimize model
       if optimizer_idx == 0:
            optimizer.step(closure=optimizer_closure)

       ## optimize center loss
       else:
            optimizer_closure()
            for param in center_loss.parameters():
                     param.grad.data*=weight
           optimizer.step()

But this will propagate forward twice, which is obviously not necessary,So I tried the following:

def optimizer_step（.......）：
       if optimizer_idx == 0:
            optimizer,optimizer_center=self.optimizers()
            optimizer_closure()
            for param in center_loss.parameters():
                     param.grad.data*=weight
             optimizer.step()
             optimizer_center.step()

But the error will show that param.grad is None

So I guess I'll just have to use manual optimization

[awaelchli]

Perhaps you could share more about what the version conflict issues are and we could try to help with that?

[stale]

This issue has been automatically marked as stale because it hasn't had any recent activity. This issue will be closed in 7 days if no further activity occurs. Thank you for your contributions, Pytorch Lightning Team!