/
kmeans.py
173 lines (150 loc) · 6.05 KB
/
kmeans.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
import numpy as np
import torch
import random
import sys
import utils
import pdb
'''
Class for running Lloyd iterations for k-means and associated utility functions.
'''
chunk_size = 8192
num_iterations = 60
k = 10
device = utils.device
device_cpu = torch.device('cpu')
class FastKMeans:
def __init__(self, dataset, n_clusters, opt):
if isinstance(dataset, np.ndarray):
dataset = torch.from_numpy(dataset).to(utils.device)
self.centers, self.codes = self.build_kmeans(dataset, n_clusters)
self.centers_norm = torch.sum(self.centers**2, dim=0).view(1,-1).to(utils.device)
self.opt = opt
#self.k = opt.k
'''
Creates kmeans
'''
def build_kmeans(self, dataset, num_centers):
return build_kmeans(dataset, num_centers)
'''
Input: query. tensor, batched query.
Returns:
-indices of nearest centers
'''
def predict(self, query, k):
#query = query.to(utils.device)
if isinstance(query, np.ndarray):
query = torch.from_numpy(query).to(utils.device)
#self centers have dimension 1, torch.Size([100, 1024])
if hasattr(self, 'opt') and (self.opt.glove or self.opt.sift) and self.centers.size(1) > 512:
centers = self.centers.t()
idx = utils.dist_rank(query, k, data_y=centers, largest=False)
else:
q_norm = torch.sum(query ** 2, dim=1).view(-1, 1)
dist = q_norm + self.centers_norm - 2*torch.mm(query, self.centers)
if k > dist.size(1):
k = dist.size(1)
_, idx = torch.topk(dist, k=k, dim=1, largest=False)
#move predict to numpy
idx = idx.cpu().numpy()
return idx
def eval_kmeans(queries, centers, codes):
centers_norms = torch.sum(centers ** 2, dim=0).view(1, -1)
queries_norms = torch.sum(queries ** 2, dim=1).view(-1, 1)
distances = torch.mm(queries, centers)
distances *= -2.0
distances += queries_norms
distances += centers_norms
codes = codes.to(device_cpu)
#counts of points per center. To compute # of candidates.
cnt = torch.zeros(num_centers, dtype=torch.long)
bins = [[]] * num_centers
for i in range(num_points):
cnt[codes[i]] += 1 #don't recompute!!
bins[codes[i]].append(i)
num_queries = answers.size()[0]
for num_probes in range(1, num_centers + 1):
#ranking of indices to nearest centers
_, probes = torch.topk(distances, num_probes, dim=1, largest=False)
probes = probes.to(device_cpu)
total_score = 0
total_candidates = 0
for i in range(num_queries):
candidates = []
#set of predicted bins
tmp = set()
for j in range(num_probes):
candidates.append(cnt[probes[i, j]])
tmp.add(int(probes[i, j]))
overall_candidates = sum(candidates)
score = 0
for j in range(k):
if int(codes[answers[i, j]]) in tmp:
score += 1
total_score += score
total_candidates += overall_candidates
print(num_probes, float(total_score) / float(k * num_queries), float(total_candidates) / float(num_queries))
'''
Input:
-dataset
Returns:
-centers. MUST ensure num_centers < len(dataset)
-codes.
'''
def build_kmeans(dataset, num_centers):
num_points = dataset.size()[0]
if num_centers > num_points:
print('WARNING: num_centers > num_points! Setting num_centers = num_points')
num_centers = num_points
dimension = dataset.size()[1]
centers = torch.zeros(num_centers, dimension, dtype=torch.float).to(device)
used = torch.zeros(num_points, dtype=torch.long)
for i in range(num_centers):
while True:
cur_id = random.randint(0, num_points - 1)
if used[cur_id] > 0:
continue
used[cur_id] = 1
centers[i] = dataset[cur_id]
break
centers = torch.transpose(centers, 0, 1)
new_centers = torch.zeros(num_centers, dimension, dtype=torch.float).to(device)
cnt = torch.zeros(num_centers, dtype=torch.float).to(device)
all_ones = torch.ones(chunk_size, dtype=torch.float).to(device)
if num_points % chunk_size != 0:
all_ones_last = torch.ones(num_points % chunk_size, dtype=torch.float).to(device)
all_ones_cnt = torch.ones(num_centers, dtype=torch.float).to(device)
codes = torch.zeros(num_points, dtype=torch.long).to(device)
for it in range(num_iterations):
centers_norms = torch.sum(centers ** 2, dim=0).view(1, -1)
new_centers.fill_(0.0)
cnt.fill_(0.0)
for i in range(0, num_points, chunk_size):
begin = i
end = min(i + chunk_size, num_points)
dataset_piece = dataset[begin:end, :]
dataset_norms = torch.sum(dataset_piece ** 2, dim=1).view(-1, 1)
distances = torch.mm(dataset_piece, centers)
distances *= -2.0
distances += dataset_norms
distances += centers_norms
_, min_ind = torch.min(distances, dim=1)
codes[begin:end] = min_ind
new_centers.scatter_add_(0, min_ind.view(-1, 1).expand(-1, dimension), dataset_piece)
if end - begin == chunk_size:
cnt.scatter_add_(0, min_ind, all_ones)
else:
cnt.scatter_add_(0, min_ind, all_ones_last)
if it + 1 == num_iterations:
break
cnt = torch.where(cnt > 1e-3, cnt, all_ones_cnt)
new_centers /= cnt.view(-1, 1)
centers = torch.transpose(new_centers, 0, 1).clone()
#eval_kmeans(queries, centers, codes)
return centers, codes
if __name__ == '__main__':
dataset_numpy = np.load('dataset.npy')
queries_numpy = np.load('queries.npy')
answers_numpy = np.load('answers.npy')
dataset = torch.from_numpy(dataset_numpy).to(device)
queries = torch.from_numpy(queries_numpy).to(device)
answers = torch.from_numpy(answers_numpy)