[Feature] Reload experiment without re-passing config #179
Conversation
facebook-github-bot
added
the
CLA Signed
benchmarl/evaluate.py
Outdated
| args = parser.parse_args() | ||
| checkpoint_file = str(Path(args.checkpoint_file).resolve()) | ||
| experiment = reload_experiment_from_file(checkpoint_file) | ||
| try: |
There was a problem hiding this comment.
nit: for readability i'd suggest adapting the hydra_config util such that you can do:
if hydra_config.checkpoint_from_hydra_run(checkpoint_file)
experiment = hydra_config.reload_experiment_from_file()
else:
# Assume experiment can be loaded directly.
experiment = Experiment.reload_from_file(checkpoint_file)
There was a problem hiding this comment.
Does it make sense in the medium term to maintain both hydra-based and non-hydra based saving/loading?
There was a problem hiding this comment.
Does it make sense in the medium term to maintain both hydra-based and non-hydra based saving/loading?
I would think so, the hydra one is more interpretable so it should be the first choice if you can
|
Thanks a lot for the review!! This should be ready now if you have any last comments |
benchmarl/hydra_config.py
Outdated
| try: | ||
| hydra_folder = _find_hydra_folder(restore_file) | ||
| except ValueError as e: | ||
| if ( |
There was a problem hiding this comment.
feel free to ignore style comments, but the best way to do this would be to raise a more specific error within _find_hydra_folder that you can catch here. I.e. MissingHydraMetadataError(FileNotFoundError). If someone changes the exception string in _find_hydra_folder, they may have no idea they are changing the functionality here.
There was a problem hiding this comment.
thanksss for pushing me out of my lazyness ahahah
fixes #178
cc @itwasabhi