[syahra712]

Summary

The multi_accuracy and f1-score metrics in multilabel_recognition.py mishandle calculate_average: False.

• f1-score builds its per-class values from MultiLabelPrecision and MultiLabelRecall, which receive the same config. When calculate_average is False, those two return only the per-class values. F1Score.evaluate still does precisions[:-1] / recalls[:-1], so the last class is silently dropped. With 3 labels, the metric returns 2 values while meta['names'] lists 3.
• multi_accuracy ignores the option and always appends the average. meta['names'] has no average entry in that case, so the values and names don't match.

The fix slices the precision/recall results to the number of labels in f1-score, and returns only the per-class values from multi_accuracy when the average is disabled. This is how multi_precision and multi_recall already behave.

Tests

New tests/test_multilabel_metrics.py covers both metrics with and without the average, on a 3-label example with hand-computed values. On master, the two calculate_average: False tests fail: f1-score returns [1.0, 0.667] and multi_accuracy returns [1.0, 0.5, 0.5, 0.667]. With this change all 4 pass. The changed module adds no new pylint findings (.pylintrc, max line 120).

[workflow-lab]

Can one of the admins verify this patch?