-
Notifications
You must be signed in to change notification settings - Fork 1
/
pseudo_labeler.py
338 lines (258 loc) · 13.1 KB
/
pseudo_labeler.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
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
import math
import torch
import torch.nn.functional as F
import tqdm
from scipy.optimize import linear_sum_assignment
from model.utils import tanh_clip
from utils import to_one_hot
class BasePseudoLabeler(object):
def __init__(self, num_classes):
self.num_classes = num_classes
def pseudo_label_tgt(self, src_test_collection, tgt_test_collection):
raise NotImplementedError
class KMeansPseudoLabeler(BasePseudoLabeler):
def __init__(self, num_classes, batch_size=4096, eps=0.0005):
super().__init__(num_classes)
self.batch_size = batch_size
self.eps = eps
self.init_centers = None
self.centers = None
self.stop = False
@staticmethod
def get_dist(point_a, point_b, cross=False):
point_a = F.normalize(point_a, dim=1)
point_b = F.normalize(point_b, dim=1)
if not cross:
return 0.5 * (torch.tensor(1.0).cuda() - torch.sum(point_a * point_b, dim=1))
else:
assert (point_a.size(1) == point_b.size(1))
return 0.5 * (torch.tensor(1.0).cuda() - torch.mm(point_a, point_b.transpose(0, 1)))
def get_src_centers(self, src_features, src_true_labels):
centers = 0
refs = torch.LongTensor(range(self.num_classes)).unsqueeze(1).cuda()
num_batches = src_features.size(0) // self.batch_size + 1
src_index = 0
for i in range(num_batches):
cur_len = min(self.batch_size, src_features.size(0) - src_index)
cur_features = src_features.narrow(0, src_index, cur_len)
cur_true_labels = src_true_labels.narrow(0, src_index, cur_len)
cur_true_labels = cur_true_labels.unsqueeze(0).expand(self.num_classes, -1)
mask = (cur_true_labels == refs).unsqueeze(2).float()
cur_features = cur_features.unsqueeze(0)
centers += torch.sum(cur_features * mask, dim=1)
src_index += cur_len
return centers
def clustering_stop(self, centers):
if centers is None:
self.stop = False
else:
dist = self.get_dist(centers, self.centers)
dist = torch.mean(dist, dim=0)
print('dist %.4f' % dist.item())
self.stop = dist.item() < self.eps
def assign_labels(self, feats):
dists = self.get_dist(feats, self.centers, cross=True)
labels = torch.min(dists, dim=1)[1]
return dists, labels
def align_centers(self):
cost = self.get_dist(self.centers, self.init_centers, cross=True)
cost = cost.cpu().numpy()
_, col_ind = linear_sum_assignment(cost)
return col_ind
def pseudo_label_tgt(self, src_test_collection, tgt_test_collection):
"""
pseudo label target samples.
Args:
src_test_collection['features']: (n_src, n_rkhs)
tgt_test_collection['features']: (n_tgt, n_rkhs)
Returns:
tgt_pseudo_label : (n_tgt, num_classes) contains pseudo label of entire target samples
"""
src_features = src_test_collection['features']
tgt_features = tgt_test_collection['features']
src_true_labels = src_test_collection['true_labels']
assert src_features.size(1) == tgt_features.size(1)
src_centers = self.get_src_centers(src_features, src_true_labels)
self.init_centers = src_centers
self.centers = src_centers
centers = None
self.stop = False
refs = torch.LongTensor(range(self.num_classes)).unsqueeze(1).cuda()
num_samples = tgt_features.size(0)
num_split = math.ceil(1.0 * num_samples / self.batch_size)
while True:
self.clustering_stop(centers)
if centers is not None:
self.centers = centers
if self.stop: break
centers = 0
count = 0
start = 0
for _ in range(num_split):
cur_len = min(self.batch_size, num_samples - start)
cur_feature = tgt_features.narrow(0, start, cur_len)
dist2center, labels = self.assign_labels(cur_feature)
labels_one_hot = to_one_hot(labels, self.num_classes)
count += torch.sum(labels_one_hot, dim=0)
labels = labels.unsqueeze(0)
mask = (labels == refs).unsqueeze(2).float()
reshaped_feature = cur_feature.unsqueeze(0)
# update centers
centers += torch.sum(reshaped_feature * mask, dim=1)
start += cur_len
mask = (count.unsqueeze(1) > 0).float()
centers = mask * centers + (1 - mask) * self.init_centers
dist2center = []
start = 0
for N in range(num_split):
cur_len = min(self.batch_size, num_samples - start)
cur_feature = tgt_features.narrow(0, start, cur_len)
cur_dist2center, _ = self.assign_labels(cur_feature)
dist2center += [cur_dist2center]
start += cur_len
tgt_dist2center = torch.cat(dist2center, dim=0)
cluster2label = self.align_centers()
# reorder the centers
self.centers = self.centers[cluster2label, :]
# re-label the data according to the index
num_samples = len(tgt_features)
for k in range(num_samples):
tgt_dist2center[k] = tgt_dist2center[k][cluster2label]
return torch.tensor(1.0).cuda() - tgt_dist2center #torch.softmax(torch.tensor(1.0).cuda() - tgt_dist2center, dim=1)
class ClassifierPseudoLabeler(BasePseudoLabeler):
def __init__(self, num_classes):
super().__init__(num_classes)
def pseudo_label_tgt(self, src_test_collection, tgt_test_collection):
del src_test_collection
tgt_logits = tgt_test_collection['logits']
tgt_pseudo_labels = F.softmax(tgt_logits, dim=1)
# tgt_pseudo_confidences = torch.max(tgt_pseudo_labels, dim=1)[0]
# return tgt_pseudo_labels, tgt_pseudo_confidences
return tgt_pseudo_labels
class InfoPseudoLabeler(BasePseudoLabeler):
def __init__(self, num_classes, batch_size=4096, tclip=20., normalize=True):
super().__init__(num_classes)
self.batch_size = batch_size
self.tclip = tclip
self.normalize = normalize
def pseudo_label_tgt(self, src_test_collection, tgt_test_collection):
"""
pseudo label target samples.
Args:
src_test_collection: (n_src, n_rkhs)
tgt_test_collection: (n_tgt, n_rkhs)
Returns:
tgt_pseudo_label : (n_tgt, num_classes) contains pseudo label of entire target samples
"""
src_features = src_test_collection['features']
tgt_features = tgt_test_collection['features']
src_true_labels = src_test_collection['true_labels']
n_src = src_features.size(0)
n_tgt = tgt_features.size(0)
assert src_features.size(1) == tgt_features.size(1)
n_rkhs = src_features.size(1)
num_batches = (n_src + n_tgt) // self.batch_size + 1
src_batch_sizes = [n_src // num_batches + 1] * (n_src % num_batches) \
+ [n_src // num_batches] * (num_batches - n_src % num_batches)
tgt_batch_sizes = [n_tgt // num_batches + 1] * (n_tgt % num_batches) \
+ [n_tgt // num_batches] * (num_batches - n_tgt % num_batches)
src_perm_indices = torch.randperm(n_src).cuda()
tgt_perm_indices = torch.randperm(n_tgt).cuda()
tgt_pseudo_labels = torch.zeros(n_tgt, self.num_classes).cuda()
src_index = 0
tgt_index = 0
for i in tqdm.tqdm(range(num_batches), desc='Target pseudo labeling', leave=False, ascii=True):
src_batch_indices = src_perm_indices[src_index:src_index + src_batch_sizes[i]]
tgt_batch_indices = tgt_perm_indices[tgt_index:tgt_index + tgt_batch_sizes[i]]
src_index += src_batch_sizes[i]
tgt_index += tgt_batch_sizes[i]
src_batch_features = src_features[src_batch_indices]
tgt_batch_features = tgt_features[tgt_batch_indices]
src_batch_true_labels = src_true_labels[src_batch_indices]
if self.normalize:
src_batch_features = F.normalize(src_batch_features, dim=1)
tgt_batch_features = F.normalize(tgt_batch_features, dim=1)
raw_scores = torch.mm(tgt_batch_features, src_batch_features.transpose(0, 1)).float()
if self.normalize:
raw_scores *= self.tclip
else:
raw_scores = raw_scores / n_rkhs ** 0.5
raw_scores = tanh_clip(raw_scores, clip_val=self.tclip)
prop_scores = F.softmax(raw_scores, dim=1)
src_one_hot_label = to_one_hot(src_batch_true_labels, self.num_classes)
tgt_batch_pseudo_label = torch.mm(prop_scores, src_one_hot_label)
# balance class
src_class_count = torch.sum(src_one_hot_label, dim=0)
src_class_count = torch.max(src_class_count, torch.ones_like(src_class_count).cuda())
tgt_batch_pseudo_label /= src_class_count
tgt_batch_pseudo_label = F.normalize(tgt_batch_pseudo_label, p=1, dim=1)
tgt_pseudo_labels[tgt_batch_indices] = tgt_batch_pseudo_label
tgt_pseudo_confidences = torch.max(tgt_pseudo_labels, dim=1)[0]
return tgt_pseudo_labels, tgt_pseudo_confidences
class PropagatePseudoLabeler(BasePseudoLabeler):
def __init__(self, num_classes, batch_size=4096, tclip=20., normalize=False):
super().__init__(num_classes)
self.batch_size = batch_size
self.tclip = tclip
self.normalize = normalize
def pseudo_label_tgt(self, src_test_collection, tgt_test_collection):
"""
pseudo label target samples.
Args:
src_test_collection: (n_src, n_rkhs)
tgt_test_collection: (n_tgt, n_rkhs)
Returns:
tgt_pseudo_label : (n_tgt, num_classes) contains pseudo label of entire target samples
"""
src_features = src_test_collection['features']
tgt_features = tgt_test_collection['features']
src_true_labels = src_test_collection['true_labels']
n_src = src_features.size(0)
n_tgt = tgt_features.size(0)
assert src_features.size(1) == tgt_features.size(1)
n_rkhs = src_features.size(1)
num_batches = (n_src + n_tgt) // self.batch_size + 1
src_batch_sizes = [n_src // num_batches + 1] * (n_src % num_batches) \
+ [n_src // num_batches] * (num_batches - n_src % num_batches)
tgt_batch_sizes = [n_tgt // num_batches + 1] * (n_tgt % num_batches) \
+ [n_tgt // num_batches] * (num_batches - n_tgt % num_batches)
src_perm_indices = torch.randperm(n_src).cuda()
tgt_perm_indices = torch.randperm(n_tgt).cuda()
tgt_pseudo_labels = torch.zeros(n_tgt, self.num_classes).cuda()
src_index = 0
tgt_index = 0
for i in tqdm.tqdm(range(num_batches), desc='Target pseudo labeling', leave=False, ascii=True):
src_batch_indices = src_perm_indices[src_index:src_index + src_batch_sizes[i]]
tgt_batch_indices = tgt_perm_indices[tgt_index:tgt_index + tgt_batch_sizes[i]]
src_index += src_batch_sizes[i]
tgt_index += tgt_batch_sizes[i]
src_batch_features = src_features[src_batch_indices]
tgt_batch_features = tgt_features[tgt_batch_indices]
src_batch_true_labels = src_true_labels[src_batch_indices]
n_src_batch = src_batch_features.size(0)
n_tgt_batch = tgt_batch_features.size(0)
batch_features = torch.cat([src_batch_features, tgt_batch_features], dim=0)
if self.normalize:
batch_features = F.normalize(batch_features, dim=1)
raw_scores = torch.mm(batch_features, batch_features.transpose(0, 1)).float()
if self.normalize:
raw_scores *= self.tclip
else:
raw_scores = raw_scores / n_rkhs ** 0.5
raw_scores = tanh_clip(raw_scores, clip_val=self.tclip)
prop_scores = F.softmax(raw_scores, dim=1)
prop_ts = prop_scores[n_src_batch:, :n_src_batch]
prop_tt = prop_scores[n_src_batch:, n_src_batch:]
src_one_hot_label = to_one_hot(src_batch_true_labels, self.num_classes)
# initialize tgt pseudo label
# tgt_batch_pseudo_label = torch.ones(n_tgt_batch, self.num_classes).cuda() / self.num_classes
# label propagation
# for _ in range(self.lp_iterations):
# tgt_batch_pseudo_label = torch.mm(prop_ts, src_one_hot_label) \
# + torch.mm(prop_tt, tgt_batch_pseudo_label)
tgt_batch_pseudo_label = torch.mm(
torch.inverse(torch.eye(n_tgt_batch).cuda() - prop_tt),
torch.mm(prop_ts, src_one_hot_label))
tgt_pseudo_labels[tgt_batch_indices] = tgt_batch_pseudo_label
tgt_pseudo_confidences = torch.max(tgt_pseudo_labels, dim=1)[0]
return tgt_pseudo_labels, tgt_pseudo_confidences