balanced_topk_accuracy.py

import numpy as np
import evaluate
import datasets

_DESCRIPTION = """
Balanced (macro) Top-K Accuracy for multiclass classification.
For each class c, compute recall@k (fraction of samples of class c whose top-k predictions contain c),
then macro-average over classes. Accepts (N, K) score/prob arrays.

Supports sample_weight to compute weighted recalls per class.
"""

_KWARGS_DESCRIPTION = """
Args:
    predictions: 2D array-like of shape (N, K) with class scores/probabilities.
    references:  1D list/array of integer labels in [0, K-1].
    k:           int, top-k (default 1). If None, use k_list instead.
    k_list:      Optional[list[int]] to compute multiple ks at once (e.g., [1,5]).
    class_mask:  Optional[list[int]] — only average over these classes (e.g., tail classes).
    sample_weight: Optional[list[float]] — per-sample weights.
    zero_division: float, default 0.0.
    return_per_class: bool, default False. If True, also return per-class recalls@k.
"""

_CITATION = ""


def _div(a, b, zero_div=0.0):
    return (a / b) if b != 0 else float(zero_div)


class BalancedTopKAccuracy(evaluate.Metric):
    def _info(self):
        return evaluate.MetricInfo(
            description=_DESCRIPTION,
            citation=_CITATION,
            inputs_description=_KWARGS_DESCRIPTION,
            features=datasets.Features(
                {"predictions": datasets.Sequence(datasets.Value("float64")),
                 "references":  datasets.Value("int64")}
            ),
        )

    def _compute(
        self,
        predictions,
        references,
        k: int | None = 1,
        k_list: list[int] | None = None,
        class_mask: list[int] | None = None,
        sample_weight: list[float] | None = None,
        zero_division: float = 0.0,
        return_per_class: bool = False,
    ):
        y_true = np.asarray(references, dtype=int)
        scores = np.asarray(predictions, dtype=float)

        if scores.ndim != 2 or y_true.ndim != 1:
            raise ValueError("`predictions` must be (N, K) scores; `references` must be 1D labels.")
        N, K = scores.shape
        if y_true.shape[0] != N:
            raise ValueError(f"Length mismatch: references {y_true.shape[0]} vs predictions {N}.")
        if not np.all(np.isfinite(scores)):
            raise ValueError("`predictions` contains NaN/Inf.")
        if (y_true < 0).any() or (y_true >= K).any():
            raise ValueError(f"`references` must be within [0, {K-1}] for given predictions shape (N,{K}).")

        w = None
        if sample_weight is not None:
            w = np.asarray(sample_weight, dtype=float)
            if w.shape[0] != N:
                raise ValueError("`sample_weight` length must match number of samples.")

        ks_in = k_list if k_list is not None else [k if k is not None else 1]
        ks = sorted(set(int(x) for x in ks_in if x is not None and int(x) >= 1))
        if not ks:
            raise ValueError("Provide a positive integer `k` or a non-empty `k_list`.")
        ks = [min(kk, K) for kk in ks]

        if class_mask is None or len(class_mask) == 0:
            classes = list(range(K))
        else:
            classes = [c for c in class_mask if 0 <= c < K]
            if not classes:
                return {"balanced_topk_accuracy": float("nan"), "reason": "empty_class_mask_after_filtering"}

        sorted_idx = np.argsort(-scores, axis=1)

        results = {}
        per_class = {}

        for kk in ks:
            topk = sorted_idx[:, :kk]
            recalls = []
            for c in classes:
                mask_c = (y_true == c)
                denom = float(w[mask_c].sum()) if w is not None else float(mask_c.sum())
                if denom == 0:
                    recalls.append(float(zero_division))
                    continue
                hit_mask = np.any(topk[mask_c] == c, axis=1)
                hits = float(w[mask_c][hit_mask].sum()) if w is not None else float(hit_mask.sum())
                recalls.append(_div(hits, denom, zero_division))

            ba_k = float(np.mean(recalls)) if recalls else float("nan")
            results[kk] = ba_k
            if return_per_class:
                per_class[kk] = [float(x) for x in recalls]

        out = {"balanced_topk_accuracy": results[ks[0]] if len(ks) == 1 else results}
        if return_per_class:
            out["per_class_recall"] = per_class if len(ks) > 1 else per_class[ks[0]]
        return out