File size: 1,285 Bytes
fc329a3 | 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 | """Size-stratified coverage metrics."""
import numpy as np
def size_stratified_coverage(
covered: np.ndarray,
radius: np.ndarray,
n_bins: int = 5,
) -> dict[int, float]:
"""Coverage within radius/size bins.
Bins are quantile-based to avoid empty tails when set sizes are skewed.
"""
covered = np.asarray(covered, dtype=bool)
radius = np.asarray(radius, dtype=float)
if len(radius) == 0:
return {}
if np.allclose(radius, radius[0]):
return {0: float(covered.mean())}
quantiles = np.linspace(0.0, 1.0, n_bins + 1)
edges = np.unique(np.quantile(radius, quantiles))
if len(edges) <= 1:
return {0: float(covered.mean())}
bin_ids = np.digitize(radius, edges[1:-1], right=True)
result = {}
for b in np.unique(bin_ids):
mask = bin_ids == b
if mask.any():
result[int(b)] = float(covered[mask].mean())
return result
def size_stratified_coverage_violation(
covered: np.ndarray,
radius: np.ndarray,
alpha: float,
n_bins: int = 5,
) -> float:
"""Maximum coverage deviation across prediction-size bins."""
sc = size_stratified_coverage(covered, radius, n_bins=n_bins)
target = 1.0 - alpha
return max(abs(v - target) for v in sc.values())
|