Pin PTL, bump omegaconf

[titu1994]

Signed-off-by: smajumdar [email protected]

[titu1994]

@blisc The patch for the DictConfig being cast down to ordinary dict inside save_hyperparameters() has been patched in this PR - Lightning-AI/pytorch-lightning#2519. We can revert it safely once we bump PTL version.

[titu1994]

@blisc We can also enable multi data loader testing by removing the explicit checks inside model_utils once we bump PTL to 0.9 - Lightning-AI/pytorch-lightning#2581

trainer.test() stalling has been patched via - Lightning-AI/pytorch-lightning#2997, will revert the DDP checks inside ModelPT.prepare_test(Trainer) to allow pass through.

Experiment manager might need some updates due to this PR - Lightning-AI/pytorch-lightning#2719, but I'll defer to your call.

Documenting it here so I can apply those patches once we bump PTL ver.

[blisc]

@blisc We can also enable multi data loader testing by removing the explicit checks inside model_utils once we bump PTL to 0.9 - PyTorchLightning/pytorch-lightning#2581

trainer.test() stalling has been patched via - PyTorchLightning/pytorch-lightning#2997, will revert the DDP checks inside ModelPT.prepare_test(Trainer) to allow pass through.

Experiment manager might need some updates due to this PR - PyTorchLightning/pytorch-lightning#2719, but I'll defer to your call.

Documenting it here so I can apply those patches once we bump PTL ver.

Right Tomasz alerted me to those changes. IMO, those changes are still insufficient to make DDP work with Lightning. I also don't think it will make a change with exp_manager, since we set hydra.run.dir=. explicitly. Hence hydra.utils.get_original_cwd()==Path.cwd(), and that patch will not change any of exp_manager's behaviour.

[titu1994]

@blisc Yep I reached the same conclusion after going through the PR. It's fine, we can continue using the hydra decorator for the time being, we probably need a cleanup of the older configs to remove the redundant hydra setup (but that can be done much later).

[omry]

Use open_dict:

with open_dict(self._cfg.decoder):
  self._cfg.decoder = new_decoder_config