1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
|
import numpy as np
from .extmath import stable_cumsum
def _weighted_percentile(array, sample_weight, percentile=50):
"""
Compute the weighted ``percentile`` of ``array`` with ``sample_weight``.
"""
sorted_idx = np.argsort(array)
# Find index of median prediction for each sample
weight_cdf = stable_cumsum(sample_weight[sorted_idx])
percentile_idx = np.searchsorted(
weight_cdf, (percentile / 100.) * weight_cdf[-1])
# in rare cases, percentile_idx equals to len(sorted_idx)
percentile_idx = np.clip(percentile_idx, 0, len(sorted_idx)-1)
return array[sorted_idx[percentile_idx]]
|
