Fix macro/weighted F1 and two sibling averaging bugs in deep.metric

[gaoflow]

Follow-up to #874, as requested there.

deep.metric.F1Score had the same F-of-means defect that #874 fixed in smile.validation.metric.FScore. While auditing the rest of the package against scikit-learn I found two more variants of the same mistake — the reduction of an averaging strategy applied to the wrong quantity, or without the guard the other strategies have — so this fixes the class rather than the one method.

1. Macro and weighted F1 averaged the wrong thing

F1Score.compute() called precision.compute() and recall.compute(), which already reduce the per-class vector to a scalar, then took their harmonic mean. F1 is not linear in (P, R), so F1(mean P, mean R) is not the mean of the per-class F1 scores.

Confusion matrix (rows = truth, columns = prediction):

        pred 0  pred 1  pred 2   support
true 0       4       6       0        10
true 1       0       2       0         2
true 2       0       0       3         3

per-class precision = [1, 1/4, 1], recall = [2/5, 1, 1], F1 = [4/7, 2/5, 1]

	before	correct
macro F1	0.77419355	23/35 = 0.65714286
weighted F1	0.72	111/175 = 0.63428571

2. Micro recall stopped equalling accuracy when a class is empty

Recall.compute() divided the micro score by where(size > 0, size, 1).sum(). That guard exists to keep the per-class recall of an empty class at 0 instead of NaN, but summing it makes every empty class add one phantom sample to the micro denominator. Averaging.Micro's javadoc states that micro precision and recall equal accuracy, and validation.metric.Recall divides by n.

With 3 output classes, no instance of class 2 in the batch and 8 of 10 samples correct, micro recall was 8/11 = 0.72727275 and micro F1 was 16/21 = 0.76190478, against an accuracy of 0.8. Any mini-batch that happens to miss a rare class hits this.

3. Weighted metrics returned NaN after reset()

The weighted reduction score.mul(size).sum().div(size.sum()) has no guard for the empty state, so Precision, Recall and F1Score all returned NaN after reset() — called at the end of every epoch, per the Metric javadoc. Macro and micro return 0 there, and there is already a test asserting that for macro F1.

Changes

Precision.score() and Recall.score() expose the per-class score with the guard each already had, so F1Score can average the per-class F1 without duplicating them. Averaging.weighted() performs the support-weighted reduction once for the three metrics. Recall.compute() now branches on the strategy rather than on tp.size(0).

Verification

Differential run against scikit-learn 1.8.0 over 10 label sets (perfect, imbalanced 3-class 900/80/20, imbalanced 4-class, binary, never-predicted class, empty class, all-wrong, multi-batch, random 5-class, single-class truth), covering every metric and averaging strategy — 109 comparisons. 18 divergences before, none after, largest remaining difference 9e-08 in float32. Accuracy, all three precisions, macro/weighted recall and the binary metrics already agreed, so the audit closes the class.

7 new tests in MetricTest, every one of them failing before the change. ./gradlew :deep:test goes from 391 tests / 0 failures to 398 / 0. The existing tests missed all three bugs because they only cover perfect predictions, binary classification, and micro averaging on batches where every class is present.

[haifengl]

Thanks a lot!