Repository navigation
fix(metrics): count true positives for bool inputs in confusion matrix and F-beta score - #9189
tritsystem wants to merge 2 commits into
Conversation
…atrix get_confusion_matrix finds true positives with (y_pred + y) == 2. For bool tensors that sum is a logical OR, which is never 2, so bool masks such as pred > 0.5 lost every true positive: a perfect prediction scored sensitivity 0 and F1 0. Cast bool inputs to float before counting; float and integer inputs are unchanged. Assisted-by: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: tritsystem <gbranaa4@gmail.com>
get_f_beta_score has the same (y_pred + y) == 2 true-positive test as get_confusion_matrix; both lost their float casts in Project-MONAI#7187. For bool masks the sum is a logical OR, so every true positive was dropped and a perfect prediction scored F-beta 0. Cast bool inputs to float before counting, as in the previous commit; float and integer inputs are unchanged. Assisted-by: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: tritsystem <gbranaa4@gmail.com>
|
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info
📝 Walkthrough
Merge Risk: ⚪ Minimal · up to
|
| Check name | Status | Explanation | Resolution |
|---|---|---|---|
| Docstring Coverage |  |
Docstring coverage is 50.00% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 6 functions across 4 files. | Write docstrings for the functions missing them to satisfy the coverage threshold. |
-
Check name Status Explanation Title check The title clearly and concisely describes the fix for true-positive counting with Boolean inputs in the confusion matrix and F-beta score metrics. Description check The description explains the defect, impact, root cause, scope, fix, and added tests. It omits the issue number and does not report integration or quick test results, but it is otherwise sufficiently … Linked Issues check Check skipped because no linked issues were found for this pull request. Out of Scope Changes check Check skipped because no linked issues were found for this pull request. Full details: Description check
-
Explanation
The description explains the defect, impact, root cause, scope, fix, and added tests. It omits the issue number and does not report integration or quick test results, but it is otherwise sufficiently complete.
-
- Fix all pre-merge checks with AI
✨ Finishing Touches
-
🧪 Generate unit tests (beta)
-
- Create a new PR
-
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.
Comment @coderabbitai help to get the list of available commands.
This branch has not been deployed
1 participant
Files:
monai/metrics/confusion_matrix.py(get_confusion_matrix) andmonai/metrics/f_beta_score.py(get_f_beta_score). Both count true positives with(y_pred + y) == 2, which works for float/int tensors. If bothy_predandyare bool tensors, PyTorch's+does logical OR, not addition, soTrue + TruegivesTrue(1), never 2 -- the true-positive count comes out as zero every time for bool inputs.Impact: a perfect bool prediction (identical bool masks) scores sensitivity 0, precision NaN, F1 0, F-beta 0 -- a perfect match graded as a total failure. The same perfect prediction in float form correctly scores 1.0. This only triggers when both inputs are bool; if either is float, PyTorch promotes the other and it works. Checked every other metric in MONAI (Dice, MeanIoU, GeneralizedDice, Hausdorff, surface distance) -- none have this pattern, so it's isolated to these two functions.
This is a regression: #7187 (merged Nov 2023) removed the
.float()casts from both functions as "unnecessary".Fix: cast bool inputs to float at the top of both functions; float/int inputs are unchanged.
Types of changes
Fixes # .