🔀 Rename get_batch_sample and add num_items_in_batch to compute_loss
#2246
Conversation
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
qgallouedec
changed the title
get_batch_sample and add num_items_in_batch to compute_lossget_batch_sample and add num_items_in_batch to compute_loss
lewtun
left a comment
lewtun
left a comment
There was a problem hiding this comment.
Thanks for the fast fix! LGTM and I left a comment about some follow-up cleaning we can do with these generate methods in favour of extending the LogCompletions callback
| return SequentialSampler(self.train_dataset) | ||
|
|
||
| def get_batch_samples(self, model, batch: Dict[str, torch.LongTensor]) -> Tuple[str, str]: | ||
| def generate_from_model_and_ref(self, model, batch: Dict[str, torch.LongTensor]) -> Tuple[str, str]: |
There was a problem hiding this comment.
Note for future: now that we have LogCompletions callback, it might be possible to enable the generative aspects of this method directly as a callback. We'd probably have to extend the LogCompletions callback to check if a reference model exists and generate for that too, but that seems better than having this code duplicated all over our preference trainers
4 participants
What does this PR do?
follows huggingface/transformers#34198
Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.