src/training/metrics.py

"""src/training/metrics.py — Shared metric computation utilities."""
from __future__ import annotations

import numpy as np
from sklearn.metrics import (
    roc_auc_score, f1_score, accuracy_score,
    precision_score, recall_score, average_precision_score,
    roc_curve,
)


def compute_auc(y_true: list, y_scores: list) -> float:
    if len(set(y_true)) < 2:
        return 0.5
    return float(roc_auc_score(y_true, y_scores))


def compute_eer(y_true: list, y_scores: list) -> float:
    if len(set(y_true)) < 2:
        return 0.5
    fpr, tpr, _ = roc_curve(y_true, y_scores)
    idx = int(np.argmin(np.abs(fpr - (1 - tpr))))
    return float(fpr[idx])


def optimal_threshold(y_true: list, y_scores: list) -> float:
    """Threshold that maximises F1 on the provided set."""
    from sklearn.metrics import precision_recall_curve
    precision, recall, thresholds = precision_recall_curve(y_true, y_scores)
    f1 = 2 * precision * recall / (precision + recall + 1e-8)
    best = int(np.argmax(f1[:-1]))
    return float(thresholds[best])


def compute_all(
    y_true:   list,
    y_scores: list,
    threshold: float = 0.5,
) -> dict:
    y_pred = (np.array(y_scores) >= threshold).astype(int)
    return {
        "auc":       compute_auc(y_true, y_scores),
        "auc_pr":    float(average_precision_score(y_true, y_scores)) if len(set(y_true)) > 1 else 0.5,
        "eer":       compute_eer(y_true, y_scores),
        "f1":        float(f1_score(y_true, y_pred, zero_division=0)),
        "accuracy":  float(accuracy_score(y_true, y_pred)),
        "precision": float(precision_score(y_true, y_pred, zero_division=0)),
        "recall":    float(recall_score(y_true, y_pred, zero_division=0)),
        "threshold": threshold,
    }