openai/CLIP
1
0
Fork 0
mirror of https://github.com/openai/CLIP.git synced 2026-11-05 01:37:26 +00:00
Code Issues 274 Projects Packages Wiki Activity

[GH-ISSUE #183] Attention is not masked on padding tokens #157

New issue
Closed
opened 2026-10-31 21:11:49 +00:00 by bsakaguchi · 2 comments
bsakaguchi commented 2026-10-31 21:11:49 +00:00
Owner
Copy link

Originally created by @mesnico on GitHub (Nov 17, 2021).
Original GitHub issue: https://github.com/openai/CLIP/issues/183

I noticed that no padding masks are used in the attention computation of the input sentence.
The transformer in the encode_text function should be masked with the actual sentence length. However, in this implementation, it appears not to be the case:
github.com/openai/CLIP@573315e83f/clip/model.py (L344)

All the 77 tokens (padding tokens comprised) seem to participate in the attention computation.
Am I missing something?

Thank you in advance!

Originally created by @mesnico on GitHub (Nov 17, 2021). Original GitHub issue: https://github.com/openai/CLIP/issues/183 I noticed that no padding masks are used in the attention computation of the input sentence. The transformer in the `encode_text` function should be masked with the actual sentence length. However, in this implementation, it appears not to be the case: https://github.com/openai/CLIP/blob/573315e83f07b53a61ff5098757e8fc885f1703e/clip/model.py#L344 All the 77 tokens (padding tokens comprised) seem to participate in the attention computation. Am I missing something? Thank you in advance!
bsakaguchi commented 2026-10-31 21:11:50 +00:00
Author
Owner
Copy link

@jongwook commented on GitHub (Nov 17, 2021):

We use a causal attention mask (code, cf. #179) and it basically blocks any information from "future" tokens from flowing back in. So we can take the embedding at the EOT token (code), while safely ignoring any attention calculation of the padding tokens.

<!-- gh-comment-id:971752528 --> @jongwook commented on GitHub (Nov 17, 2021): We use a causal attention mask ([code](https://github.com/openai/CLIP/blob/573315e83f07b53a61ff5098757e8fc885f1703e/clip/model.py#L324-L330), cf. #179) and it basically blocks any information from "future" tokens from flowing back in. So we can take the embedding at the EOT token ([code](https://github.com/openai/CLIP/blob/573315e83f07b53a61ff5098757e8fc885f1703e/clip/model.py#L349-L350)), while safely ignoring any attention calculation of the padding tokens.
bsakaguchi commented 2026-10-31 21:11:50 +00:00
Author
Owner
Copy link

@mesnico commented on GitHub (Nov 17, 2021):

I was totally ignoring the pad masking "side effect" of causal attention.
Thank you!

<!-- gh-comment-id:971789130 --> @mesnico commented on GitHub (Nov 17, 2021): I was totally ignoring the pad masking "side effect" of causal attention. Thank you!
Sign in to join this conversation.
No Branch/Tag specified
Branches Tags
main
codex/pin-github-workflow-refs-20260325-132838
jongwook/issue-396
jongwook-patch-1
No results found.
Labels
Clear labels   
pull-request
Mirrored from GitHub Pull Request

No labels
pull-request
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
openai/CLIP#157
No description provided.