torch.optim.lr_scheduler.ReduceLROnPlateau not registered

[lgessler]

🐛 Describe the bug

For some reason, PyTorch's ReduceLROnPlateau learning rate scheduler is not registered as a subclass of LRScheduler. This causes Tango's registration to fail to register it, as Tango uses a subclass of LRScheduler check to see whether a given class ought to be registered.

Suggested fix: manually include ReduceLROnPlateau in Tango's registration code. I'd be happy to open a PR for this.

Versions

ai2tango==1.0.2
torch==1.13.0

[lgessler]

OK, I actually see that ReduceLROnPlateau.step requires a metrics argument, while other PyTorch LR schedulers do not. This which would complicate Tango's training implementation, so the fix is not as easy as I thought.