Remove config assumption in Trainer #7464
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
LysandreJik
approved these changes
Sep 30, 2020
| assert not getattr( | ||
| self.model.config, "output_hidden_states", False | ||
| ), "The prediction loop does not work with `output_hidden_states=True`." | ||
|
|
Member
There was a problem hiding this comment.
Wouldn’t we want to put these lines inside an if statement? The prediction loop still doesn’t work with these outputs right?
Collaborator
Author
There was a problem hiding this comment.
Nope it does now since the functions that detach/concat etc. all work on nested list/tuples of tensors :-)
TevenLeScao
approved these changes
Sep 30, 2020
Contributor
TevenLeScao
left a comment
TevenLeScao
left a comment
There was a problem hiding this comment.
LGTM ! Much cleaner this way, thanks!
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What does this PR do?
This PR tries to limit the access to
model.configinTrainerto the minimum so that it works with regular PyTorch modules (as long as they accept dict inputs and return loss first like our models). The most challenging part was the storing/restoring of thetotal_flos, which I moved to the newly createdTrainerState. It should work as before and be saved along the rest of the training state.