fix Bug in BertScore calculation: pred target misalignment #2347
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
gxy-gxy
requested review from
SkafteNicki,
Borda,
justusschock and
stancld
as code owners
|
Here is the test code: from torchmetrics.text.bert import BERTScore
score_model = BERTScore(model_name_or_path='roberta-large', batch_size=2)
text1 = [ "Claim A from machine", "Claim A from machine"]
text2 = ["Claim A from machine", "Claim B"]
similarities = score_model(text1, text2)
print(similarities) |
|
@stancld could you help here, pls? |
baskrahmer
reviewed
May 6, 2024
There was a problem hiding this comment.
Good catch! I can reproduce the bug using your test code snippet. Although the actual metric values seem to be correct, the ordering is not always valid.
It might be nice to somehow integrate this case with the current test suite. E.g. an assertation that reversing the targets/preds also reverses the scores.
| @@ -419,18 +419,12 @@ def bert_score( | |||
| preds_loader, preds_dataset.max_length, model, device, num_layers, all_layers, idf, verbose, user_forward_fn | |||
| ) | |||
|
|
|||
| preds_embeddings = preds_embeddings[preds_loader.dataset.sorting_indices] | |||
| target_embeddings = target_embeddings[target_loader.dataset.sorting_indices] | |||
There was a problem hiding this comment.
I'm pretty sure the IDF factors also need to be unsorted too
There was a problem hiding this comment.
I'm pretty sure the IDF factors also need to be unsorted too
Thanks for your suggestion, I will fix it soon.
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Outdated suggestions cannot be applied.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.
fix Bug in BertScore calculation: pred target misalignment
Fixes bug in BertScore cal.
This pull request addresses a bug identified in the BertScore calculation within the TextDataset class in
src/torchmetrics/functional/text/helper_embedding_metric.py.The class is designed with a preprocess function automatically sorts input text by length to optimize batch encoding efficiency. However, this behavior introduces an issue during the BertScore calculation process, as predictions (preds) and targets (targets) are initialized in separate datasets. This results in a mismatched ordering of text pairs, which is problematic given the pairwise nature of BertScore's calculation. To ensure accurate scoring, it is critical to re-align the datasets to their original order before computing the scores. The proposed fix involves ensuring that the datasets for predictions and targets are processed in a way that maintains their original pairing throughout the calculation process.
Here is the fixed code:
This change is essential for preserving the integrity of the BertScore evaluation, ensuring that each prediction is accurately compared against its corresponding target.
馃摎 Documentation preview 馃摎: https://torchmetrics--2347.org.readthedocs.build/en/2347/