forked from BobWang21/cluster
/
k_init.py
108 lines (92 loc) · 4.16 KB
/
k_init.py
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
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
https://github.com/scikit-learn/scikit-learn/blob/ab93d65/sklearn/cluster/k_means_.py#L704
"""
import warnings
import numpy as np
import scipy.sparse as sp
from sklearn.metrics.pairwise import euclidean_distances
from sklearn.utils.extmath import row_norms
###############################################################################
# Initialization heuristic
def _k_init(X, n_clusters, random_state, n_local_trials=None):
"""Init n_clusters seeds according to k-means++
Parameters
-----------
X: array or sparse matrix, shape (n_samples, n_features)
The data to pick seeds for. To avoid memory copy, the input data
should be double precision (dtype=np.float64).
n_clusters: integer
The number of seeds to choose
x_squared_norms: array, shape (n_samples,)
Squared Euclidean norm of each data point.
random_state: numpy.RandomState
The generator used to initialize the centers.
n_local_trials: integer, optional
The number of seeding trials for each center (except the first),
of which the one reducing inertia the most is greedily chosen.
Set to None to make the number of trials depend logarithmically
on the number of seeds (2+log(k)); this is the default.
Notes
-----
Selects initial cluster centers for k-mean clustering in a smart way
to speed up convergence. see: Arthur, D. and Vassilvitskii, S.
"k-means++: the advantages of careful seeding". ACM-SIAM symposium
on Discrete algorithms. 2007
Version ported from http://www.stanford.edu/~darthur/kMeansppTest.zip,
which is the implementation used in the aforementioned paper.
"""
n_samples, n_features = X.shape
centers = np.empty((n_clusters, n_features), dtype=X.dtype)
# Modified !!!!
# assert x_squared_norms is not None, 'x_squared_norms None in _k_init'
x_squared_norms = row_norms(X, squared=True)
# Set the number of local seeding trials if none is given
if n_local_trials is None:
# This is what Arthur/Vassilvitskii tried, but did not report
# specific results for other than mentioning in the conclusion
# that it helped.
n_local_trials = 2 + int(np.log(n_clusters))
# Pick first center randomly
center_id = random_state.randint(n_samples)
if sp.issparse(X):
centers[0] = X[center_id].toarray()
else:
centers[0] = X[center_id]
# Initialize list of closest distances and calculate current potential
closest_dist_sq = euclidean_distances(
centers[0, np.newaxis], X, Y_norm_squared=x_squared_norms,
squared=True)
current_pot = closest_dist_sq.sum()
# Pick the remaining n_clusters-1 points
for c in range(1, n_clusters):
# Choose center candidates by sampling with probability proportional
# to the squared distance to the closest existing center
rand_vals = random_state.random_sample(n_local_trials) * current_pot
candidate_ids = np.searchsorted(closest_dist_sq.cumsum(), rand_vals)
# Compute distances to center candidates
distance_to_candidates = euclidean_distances(
X[candidate_ids], X, Y_norm_squared=x_squared_norms, squared=True)
# Decide which candidate is the best
best_candidate = None
best_pot = None
best_dist_sq = None
for trial in range(n_local_trials):
# Compute potential when including center candidate
new_dist_sq = np.minimum(closest_dist_sq,
distance_to_candidates[trial])
new_pot = new_dist_sq.sum()
# Store result if it is the best local trial so far
if (best_candidate is None) or (new_pot < best_pot):
best_candidate = candidate_ids[trial]
best_pot = new_pot
best_dist_sq = new_dist_sq
# Permanently add best center candidate found in local tries
if sp.issparse(X):
centers[c] = X[best_candidate].toarray()
else:
centers[c] = X[best_candidate]
current_pot = best_pot
closest_dist_sq = best_dist_sq
return centers