-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathclustering.py
201 lines (176 loc) · 6.55 KB
/
clustering.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
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
import numpy as np
import math
from collections import Counter
from sklearn.cluster import KMeans
#from matplotlib import pyplot as plt
from sklearn import manifold
import scipy as sp
import scipy.special
import random
def hierarchical_clustering(vector:np.ndarray, max_depth=3, weight=None):
n, d = vector.shape
if weight is None:
weight = np.ones(n)
depth_ = math.floor(np.log(n)) - 1
if max_depth > depth_:
max_depth = depth_
num_clusters = 2**(max_depth + 1)
cluster_center = np.zeros((num_clusters, d))
weight_sum_in_clusters = np.zeros(num_clusters)
label = np.zeros(n, dtype=np.int) - 1
def clustering_(items, root_id):
depth = int(np.log2(root_id))
if depth < max_depth:
left_root = 2 * root_id
right_root = 2 * root_id + 1
code, cluster_center[[left_root, right_root]] = balance(KMeans(2).fit(vector[items]), vector[items])
left = items[(code == 0).nonzero()[0]]
right = items[(code == 1).nonzero()[0]]
left_weight = clustering_(left, left_root)
right_weight = clustering_(right, right_root)
weight_sum_in_clusters[[left_root, right_root]] = left_weight, right_weight
return left_weight + right_weight
else:
weight_sum = 0.
for i in items:
label[i] = root_id
weight_sum += weight[i]
return weight_sum
clustering_(np.arange(n), 1)
return cluster_center, weight_sum_in_clusters, label
def balance(kmeans, vector):
code = kmeans.labels_
vector_transform = kmeans.transform(vector)
n = code.shape[0]
priority = np.abs(vector_transform[:, 1] - vector_transform[:, 0])
cnt = Counter(code)
left_num = int(np.ceil(n/2))
if cnt[0] > cnt[1]:
count_larger, count_smaller = cnt[0], cnt[1]
label_larger, label_smaller = 0, 1
else:
count_larger, count_smaller = cnt[1], cnt[0]
label_larger, label_smaller = 1, 0
idx_larger = (code == label_larger).nonzero()[0]
idx_smaller = (code == label_smaller).nonzero()[0]
priority_larger = priority[idx_larger]
priority_smaller = priority[idx_smaller]
if left_num < count_larger: # move larger to smaller
count = count_larger - left_num
idx_move_local = np.argpartition(priority_larger, count)[:count]
idx_move_global = idx_larger[idx_move_local]
code[idx_move_global] = label_smaller
elif left_num > count_larger: # move smaller to larger
count = left_num - count_larger
idx_move_local = np.argpartition(priority_smaller, count)[:count]
idx_move_global = idx_smaller[idx_move_local]
code[idx_move_global] = label_larger
code_mat = np.eye(2)[code]
cluster_sum = 1 / code_mat.sum(axis=0)
code_mat = np.diag(cluster_sum)[code]
center = code_mat.T @ vector
return code, center
def distribute_into_tree(exist, label, length, weight=None):
n = label.shape[0]
if weight is None:
weight = np.ones(n)
label2weight = np.zeros(length)
#label2weight = dict()
for e in exist:
l = label[e]
while l > 1:
label2weight[l] += weight[e]
# if l not in label2weight:
# label2weight[l] = weight[e]
# else:
# label2weight[l] += weight[e]
l = int(l/2)
return label2weight
def distribute_into_leaf(items, label, leaf_start, num_leaves):
items_in_leaves = [[] for _ in range(num_leaves)]
for i in items:
items_in_leaves[label[i] - leaf_start].append(i)
return items_in_leaves
def leaf_sampling(cluster_score):
max_depth = int(np.log2(cluster_score.shape[0] - 1))
curr_node = 1
prob = 1
while int(np.log2(curr_node)) < max_depth:
left = 2 * curr_node
#prob_l = cluster_score[left] / (cluster_score[left + 1] + cluster_score[left])
prob_l = cluster_score[left]
if random.random() < prob_l:
curr_node = left
prob = prob * prob_l
else:
curr_node = left + 1
prob = prob * (1 - prob_l)
return curr_node, prob
def leaf_sampling1(cluster_score, weight_sum_cluster, label2weight):
max_depth = int(np.log2(cluster_score.shape[0] - 1))
def sampling(root):
if int(np.log2(root)) == max_depth:
return root, 1
left = 2 * root
right = 2 * root + 1
score, weight = cluster_score[left], weight_sum_cluster[left]
score_l = score * (weight - label2weight[left])
score, weight = cluster_score[right], weight_sum_cluster[right]
score_r = score * (weight - label2weight[right])
prob_l = score_l / (score_l + score_r)
if random.random() < prob_l:
id, prob = sampling(left)
return id, prob * prob_l
else:
id, prob = sampling(right)
return id, prob * (1 - prob_l)
return sampling(1)
# pretrain_uv = 'uv.npz'
# data = np.load(pretrain_uv)
# U = data['U']
# V = data['V']
#
# cluster_center_, weight_sum_in_clusters_, label_ = hierarchical_clustering(V, 6)
#
# label2weight = distribute_into_tree(np.random.randint(0, V.shape[0], 100), label_)
# cnt = Counter()
# prob = 0
# cluster2prob = {}
# for _ in range(100000):
# c = leaf_sampling(U[0], cluster_center_, weight_sum_in_clusters_, label2weight)
# if c[0] not in cnt:
# prob += c[1]
# cluster2prob[c[0]] = c[1]
# cnt[c[0]] += 1
# print(prob)
# sum_ = sum(cnt.values())
# for e in sorted(cnt.items(), key=lambda x:-x[1]):
# print(e[0], e[1]/sum_, cluster2prob[e[0]])
# for node_id in range(2,2**4):
# v = np.zeros(32)
# node_depth = int(np.log2(node_id))
# count = 0
# for i in range(len(label_)):
# l = label_[i]
# depth = int(np.log2(l))
# while depth > node_depth:
# l = int(l/2)
# depth = int(np.log2(l))
# if l == node_id:
# v += V[i]
# count += 1
#
# print(np.mean(np.abs(v / num_items_in_clusters_[node_id] - cluster_center_[node_id])))
# x = np.zeros((5000,2))
# x[:3000,:] = np.random.randn(3000,2) + [0.5,1]
# x[3000:,:] = np.random.randn(2000,2) + [2,0.3]
# idx1 = np.logical_and(code1 == code, code == 0)
# idx2 = np.logical_and(code1 == code, code == 1)
# idx3 = code1 != code
#
# #vector = manifold.TSNE().fit_transform(vector)
# plt.plot(vector[idx1, 0], vector[idx1, 1], '.')
# plt.plot(vector[idx2, 0], vector[idx2, 1], '.')
# plt.plot(vector[idx3, 0], vector[idx3, 1], '.')
# plt.show()
#def hierarchical_clustering_balanced_binary(vector:np.ndarray):