Fix TypeError in macro f1 when zero_division=0.0

[JayeshSuryavanshi]

Problem

f1(y_true, y_pred, average="macro", zero_division=0.0) raises:

TypeError: _f_score_micro() missing 1 required positional argument: 'zero_division'

0.0 is a documented, valid value for zero_division (the docstring lists {"warn", 0.0, 1.0, np.nan}), so this should work the same as the other options.

Root cause

In _compute_macro, the per-sample dispatch guards with a truthiness check:

for ground_truth, prediction in zip(y_true, y_pred):
    if zero_division:
        sample_score = _micro_function(np.array([ground_truth]), np.array([prediction]), zero_division)
    else:
        sample_score = _micro_function(np.array([ground_truth]), np.array([prediction]))

The intent of that branch is "if the micro function accepts a zero_division argument, forward it." precision/recall macro leave the argument at its None default and correctly take the 2-argument path. But the f-score path always passes a value, and 0.0 (and 0) is falsy, so it wrongly falls through to the 2-argument call, and _f_score_micro requires the third argument. The other documented values ("warn", 1.0, np.nan) are all truthy, which is why this went unnoticed.

Fix

Guard on is not None instead of truthiness:

if zero_division is not None:

precision/recall macro still take the None default path; the f-score path now forwards 0.0 correctly. No other behavior changes.

Tests

The existing macro tests exercised zero_division via the default "warn", plus 1.0 and np.nan, but never the explicit 0.0, which is exactly the gap that let this ship. Added assert 0.0 == f1(y_true, y_pred, "macro", 0.0) to all five macro f1 tests (1d/2d/3d, list and array).

Verified: reproduced the TypeError before the change, confirmed it is gone after, confirmed explicit 0.0 returns 0.0 with no spurious warning, and confirmed precision/recall macro are unaffected. Full tests/test_metrics.py passes (63/63).

Closes #161

[JayeshSuryavanshi]

Note on the failing docs/readthedocs.org check: it's the docs build, not the test suite, and it appears pre-existing rather than caused by this PR. This change only touches metrics.py (one line) and the metrics tests, nothing docs-related. I ran the Sphinx gallery examples locally on this branch and they execute fine except the ray-based plot_parallel_training.py (which downloads a dataset), which looks like the flaky cause; the same check also fails on some unrelated PRs. Separately, while checking this I found and fixed a genuine numpy 2.5 incompatibility (np.cross on 2-D vectors in the Venn-Abers calibrator) in #170.

[JayeshSuryavanshi]

Hi @mirand863, I have three small fixes open here, would you have time to look?

• Fix TypeError in macro f1 when zero_division=0.0 #168 fixes Exception when computing f1 metrics with average='macro' and zero_division=0.0 #161, a TypeError in macro f1 when zero_division=0.0
• Fix logger verbose level not restored after unpickling #169 fixes the logger verbose level is not maintained after unpickling it #146, logger verbose level lost after unpickling
• Fix numpy 2.5 incompatibility in Venn-Abers calibration #170 fixes numpy 2.5 breaking Venn-Abers calibration (14 tests fail on a clean main)

The red readthedocs check on all three is the docs build rather than the test suite, and it fails the same way on PRs that aren't mine.