File size: 2,092 Bytes
dadf189
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
from __future__ import annotations

import numpy as np


def _safe_div(num: float, den: float) -> float:
    return num / den if den > 0 else 0.0


def _compute_threshold_for_fmr(
    impostor_scores: np.ndarray,
    fmr_threshold: float,
) -> float:
    # Accept if match_score >= threshold. We pick quantile so that expected FMR
    # is approximately fmr_threshold.
    q = float(np.clip(1.0 - fmr_threshold, 0.0, 1.0))
    return float(np.quantile(impostor_scores, q))


def compute_erc(
    quality_scores: np.ndarray,
    match_scores: np.ndarray,
    labels: np.ndarray,
    fmr_threshold: float = 1e-4,
    rejection_ratios: np.ndarray | None = None,
) -> tuple[np.ndarray, float]:
    """Compute FNMR under progressive rejection of lowest-quality samples."""

    if rejection_ratios is None:
        rejection_ratios = np.linspace(0.0, 0.5, 50)

    quality_scores = np.asarray(quality_scores)
    match_scores = np.asarray(match_scores)
    labels = np.asarray(labels)

    if not (quality_scores.shape == match_scores.shape == labels.shape):
        raise ValueError("quality_scores, match_scores, and labels must have same shape")

    sorted_idx = np.argsort(quality_scores)
    fnmr_curve = []
    for rr in rejection_ratios:
        n_reject = int(rr * len(sorted_idx))
        keep_mask = np.ones_like(labels, dtype=bool)
        keep_mask[sorted_idx[:n_reject]] = False

        kept_scores = match_scores[keep_mask]
        kept_labels = labels[keep_mask]
        impostor = kept_scores[kept_labels == 0]
        genuine = kept_scores[kept_labels == 1]
        if len(impostor) == 0 or len(genuine) == 0:
            fnmr_curve.append(1.0)
            continue

        threshold = _compute_threshold_for_fmr(impostor, fmr_threshold)
        fn = float((genuine < threshold).sum())
        fnmr = _safe_div(fn, float(len(genuine)))
        fnmr_curve.append(fnmr)

    fnmr_curve_arr = np.asarray(fnmr_curve, dtype=np.float64)
    auc = float(np.trapz(fnmr_curve_arr, rejection_ratios) / (rejection_ratios[-1] - rejection_ratios[0] + 1e-12))
    return fnmr_curve_arr, auc