Visitar URL original
Wrong predictions read in ctd_tracking by Ayskin · Pull Request #3559 · DeepLabCut/DeepLabCut · GitHub
Skip to content

Wrong predictions read in ctd_tracking - #3559

Closed
Ayskin wants to merge 1 commit into
DeepLabCut:devfrom
Ayskin:yann/fix-ctd-predictions
Closed

Wrong predictions read in ctd_tracking#3559
Ayskin wants to merge 1 commit into
DeepLabCut:devfrom
Ayskin:yann/fix-ctd-predictions

Conversation

@Ayskin

@Ayskin Ayskin commented Oct 19, 2026

Copy link
Copy Markdown
Collaborator

The ctd_tracking branch of analyze_videos now reads the correct list ctd_predictions instead of the wrong list predictions.

Previous silent misread as both lists are the same but would fail if use_shelve=True as the list predictions is empty.

So currently ctd_predictions actually rebuilds the same data as in predictions from the pickle file (potential refactoring related to #3553).

@deruyter92
deruyter92 self-requested a review October 19, 2026 15:44
deruyter92
deruyter92 reviewed Oct 20, 2026
View reviewed changes

@deruyter92 deruyter92 left a comment •
edited
Loading

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Good fix! Minor suggestion: shall we add a tiny regression test that locks the correct behavior?

Just for reference:
Seems this bug has been there since the introduction of BUCTD (see #2952). It only affects users who ran analyze_videos from the Python API with a multi-animal CTD model and ctd_tracking turned on, then re-ran it with overwrite=False on a batch where some videos already had a _full.pickle. Is that correct, @Ayskin?

@Ayskin

Ayskin commented Oct 20, 2026

Copy link
Copy Markdown
Collaborator Author

It indeed only affects users running analyze_videos with multi_animal and ctd_tracking on, but would also only fail in the following cases :

  • use_shelve=True & overwrite=True : predictions == []
  • use_shelve=True & overwrite=False & no pkl exists : predictions == []

If use_shelve=False or (overwrite=False & pkl exists), the whole inference block is either skipped or predictions is filled.

@Ayskin

Ayskin commented Oct 20, 2026 •
edited
Loading

Copy link
Copy Markdown
Collaborator Author

Closing as this fix is actually already included in the merged PR #3488.

@Ayskin Ayskin closed this Oct 22, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@deruyter92 deruyter92 deruyter92 left review comments

Assignees

@Ayskin Ayskin

Labels

bug fix pytorch

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

3 participants

@Ayskin @deruyter92 @C-Achard