19 lines
631 B
Python
19 lines
631 B
Python
|
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]]
|