Add MiniLM cross-encoder reranker #200
Merged
Conversation
This file contains 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
ronakice
reviewed
Jul 10, 2021
pygaggle/rerank/transformer.py
Outdated
| @@ -247,3 +249,29 @@ def rescore(self, query: Query, texts: List[Text]) -> List[Text]: | |||
| text.score = max(smax_val.item(), emax_val.item()) | |||
|
|
|||
| return texts | |||
|
|
|||
|
|
|||
| class CrossEncoderReranker(Reranker): | |||
There was a problem hiding this comment.
I think calling this a CrossEncoderReranker while the others not is not a good naming convention since all of them are, how about change it to be SentenceTransformersReranker
There was a problem hiding this comment.
Sure, that'd be better!
pygaggle/rerank/transformer.py
Outdated
| device=None, | ||
| use_amp=None): | ||
| device = device or ('cuda' if torch.cuda.is_available() else 'cpu') | ||
| self.use_amp = use_amp or (device == 'cuda') |
There was a problem hiding this comment.
Why would we want to default to amp if there is a gpu it could bring performance drop, more explicit a flag would be better.
There was a problem hiding this comment.
You're right, I'm changing it to disabled by default.
rodrigonogueira4
approved these changes
Jul 10, 2021
There was a problem hiding this comment.
LGTM! Thanks a lot for doing this
@ronakice, I will let you merge when you think it is ready.
ronakice
reviewed
Jul 12, 2021
|
Thanks a lot @hugoabonizio, I'm merging! |
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Outdated suggestions cannot be applied.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.
This PR adds a new reranker based on a MiniLM cross-encoder pretrained on MS-MARCO provided by the SentenceTransformers package. MiniLM-based models are much faster than MonoT5/MonoBERT while maintaining a similar result when compared to MonoT5.
Although it adds a new dependency, SentenceTransformers handles smart batching and other optimizations that otherwise would need to be implemented in PyGaggle. I also added AMP integration for FP16 inference, which should make inferences even faster (I can add this to other models if that's desirable).
Besides MiniLM-based models, there is a list of supported models.