You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I feel like I'm broadly happy with this PR and there are some minor kinks to work out now. That said, I think there's two main extensions to this PR for the future:
revamp the hyperparameter loading. The current implementation works and correctly stores the transforms in their respective modules, but it reintroduces duplicated memory that more efficient module checkpointing #444 attempted to address. In short, by saving the modules themselves as hyperparameters, their respective state dicts are stored twice in the checkpoint file: once with the scaler object itself and and once inside the entire model's state dict. Again, not a big deal considering these are O(24)−O(29), i.e, less than 1 KB each. There are only about five of them in the entire model, so this is practically a non-issue. However, it wouldn't be bad for someone to try and address this to maintain thematic consistency with the rest of the code, but it may be more trouble than it's worth...
The text was updated successfully, but these errors were encountered:
Notes
From @davidegraff 's comment from #726
The text was updated successfully, but these errors were encountered: