test(inductive): extend VAT/Noisy Student coverage and smoke readiness#20
test(inductive): extend VAT/Noisy Student coverage and smoke readiness#20melvinbarbaux merged 2 commits intomainfrom
Conversation
Codecov Report✅ All modified and coverable lines are covered by tests. 📢 Thoughts on this report? Let us know! |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 085e8a056e
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| else: | ||
| logits_l = extract_logits(student(x_lb)) | ||
| logits_us = extract_logits(student(x_us)) |
There was a problem hiding this comment.
When freeze_bn=True, only the teacher forward is wrapped with freeze_batchnorm later in this loop, but the student forward on x_us (and x_lb) happens outside any freeze context. For models with BatchNorm, this still updates running stats using unlabeled data even though freeze_bn is meant to prevent that, which can skew BN statistics and change evaluation behavior (other methods like MeanTeacher/PiModel freeze BN on student unlabeled passes). Consider wrapping the student forward(s) in freeze_batchnorm when the flag is enabled.
Useful? React with 👍 / 👎.
1 participant
Summary
What does this PR change?
Checklist
Notes
Anything reviewers should know?