-
Notifications
You must be signed in to change notification settings - Fork 14
/
loss_functions.py
449 lines (371 loc) · 15.5 KB
/
loss_functions.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
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
import math
from collections import OrderedDict
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
'''
AdaCos and Ad margin loss taken from https://github.com/4uiiurz1/pytorch-adacos
'''
class DropClassBase(nn.Module):
def __init__(self, num_classes):
'''
DropClass class which other classifier heads should inherit from
This is to package the useful wrapper scripts for which classes to include/ignore
The class has two main modes, called via .drop() and .nodrop(), which sets which method will be
called by .forward()
forward_drop defines the ordinary behaviour
forward_nodrop defines the behaviour in which only the remaining class columns are used
'''
super(DropClassBase, self).__init__()
self.n_classes = num_classes
self.dropmode = False # Default is the normal behaviour
self.set_ignored_classes([])
self.combined_class_label = None
def forward(self, input, label=None):
'''
input: (batch_size, num_features): FloatTensor
label (optional): (batch_size): LongTensor
'''
if self.dropmode:
if label is not None:
assert (torch.max(label) < len(self.rem_classes)), 'Contains label out of range of allowed classes: Have they been converted?'
return self.forward_drop(input, label=label)
else:
return self.forward_nodrop(input, label=label)
def drop(self):
self.dropmode = True
def nodrop(self):
self.dropmode = False
def forward_drop(self, input, label=None):
raise NotImplementedError
def forward_nodrop(self, input, label=None):
raise NotImplementedError
def set_ignored_classes(self, ignored:list):
if len(ignored) != 0:
assert min(ignored) >= 0
assert max(ignored) < self.n_classes
self.ignored = sorted(list(set(ignored)))
self.rem_classes = sorted(set(np.arange(self.n_classes)) - set(ignored))
self.ldict = OrderedDict({k:v for v, k in enumerate(self.rem_classes)}) #mapping of original label to new index
self.idict = OrderedDict({k:v for k, v in enumerate(self.rem_classes)}) #mapping of remaining indexes to original label
def set_remaining_classes(self, remaining:list):
assert min(remaining) >= 0
assert max(remaining) < self.n_classes
self.rem_classes = sorted(set(remaining))
self.ignored = sorted(set(np.arange(self.n_classes)) - set(remaining))
self.ldict = OrderedDict({k:v for v, k in enumerate(self.rem_classes)}) #mapping of original label to new index
self.idict = OrderedDict({k:v for k, v in enumerate(self.rem_classes)}) #mapping of remaining indexes to original label
def get_mini_labels(self, label:list):
# convert list of labels into new indexes for ignored classes
mini_labels = torch.LongTensor(list(map(lambda x: self.ldict[x], label)))
return mini_labels
def get_orig_labels(self, label:list):
# convert list of mini_labels into original class labels
# assert not self.combined_class_label, 'Combined classes means original labels not recoverable'
orig_labels = list(map(lambda x: self.idict[x], label))
return orig_labels
def set_remaining_classes_comb(self, remaining:list):
# remaining must not include the combined class
assert self.combined_class_label is not None, 'combined_class_label has not been set'
assert min(remaining) >= 0
assert max(remaining) < self.n_classes
remaining.append(self.combined_class_label)
self.rem_classes = sorted(set(remaining))
self.ignored = sorted(set(np.arange(self.n_classes)) - set(remaining)) # not really ignored, just combined
self.ldict = OrderedDict({k:v for v, k in enumerate(self.rem_classes)})
for k in self.ignored:
self.ldict[k] = self.combined_class_label # set all ignored classes to the combined class label
self.idict = OrderedDict({k:v for k, v in enumerate(self.rem_classes)}) # not the original mapping for comb classes
class DropAffine(DropClassBase):
def __init__(self, num_features, num_classes):
super(DropAffine, self).__init__(num_classes)
self.fc = nn.Linear(num_features, num_classes)
self.reset_parameters()
def reset_parameters(self):
self.fc.reset_parameters()
def forward_nodrop(self, input, label=None):
W = self.fc.weight
b = self.fc.bias
logits = F.linear(input, W, b)
return logits
def forward_drop(self, input, label=None):
W = self.fc.weight[self.rem_classes]
b = self.fc.bias[self.rem_classes]
logits = F.linear(input, W, b)
return logits
class L2SoftMax(DropClassBase):
def __init__(self, num_features, num_classes):
super(L2SoftMax, self).__init__(num_classes)
self.W = nn.Parameter(torch.FloatTensor(num_classes, num_features))
self.reset_parameters()
def reset_parameters(self):
nn.init.xavier_uniform_(self.W)
def forward_nodrop(self, input, label=None):
x = F.normalize(input)
W = F.normalize(self.W)
logits = F.linear(x, W)
return logits
def forward_drop(self, input, label=None):
x = F.normalize(input)
W = F.normalize(self.W[self.rem_classes])
logits = F.linear(x, W)
return logits
class SoftMax(DropClassBase):
def __init__(self, num_features, num_classes):
super(SoftMax, self).__init__(num_classes)
self.W = nn.Parameter(torch.FloatTensor(num_classes, num_features))
self.reset_parameters()
def reset_parameters(self):
nn.init.xavier_uniform_(self.W)
def forward_nodrop(self, input, label=None):
x = input
W = self.W
logits = F.linear(x, W)
return logits
def forward_drop(self, input, label=None):
x = input
W = self.W[self.rem_classes]
logits = F.linear(x, W)
return logits
class XVecHead(DropClassBase):
def __init__(self, num_features, num_classes, hidden_features=None):
super(XVecHead, self).__init__(num_classes)
hidden_features = num_features if not hidden_features else hidden_features
self.fc_hidden = nn.Linear(num_features, hidden_features)
self.nl = nn.LeakyReLU()
self.bn = nn.BatchNorm1d(hidden_features)
self.fc = nn.Linear(hidden_features, num_classes)
self.reset_parameters()
def reset_parameters(self):
self.fc.reset_parameters()
def forward_nodrop(self, input, label=None):
input = self.fc_hidden(input)
input = self.nl(input)
input = self.bn(input)
W = self.fc.weight
b = self.fc.bias
logits = F.linear(input, W, b)
return logits
def forward_drop(self, input, label=None):
input = self.fc_hidden(input)
input = self.nl(input)
input = self.bn(input)
W = self.fc.weight[self.rem_classes]
b = self.fc.bias[self.rem_classes]
logits = F.linear(input, W, b)
return logits
class AMSMLoss(DropClassBase):
def __init__(self, num_features, num_classes, s=30.0, m=0.4):
super(AMSMLoss, self).__init__(num_classes)
self.num_features = num_features
self.n_classes = num_classes
self.s = s
self.m = m
self.W = nn.Parameter(torch.FloatTensor(num_classes, num_features))
self.reset_parameters()
def reset_parameters(self):
nn.init.xavier_uniform_(self.W)
def forward_nodrop(self, input, label=None):
# normalize features
x = F.normalize(input)
# normalize weights
W = F.normalize(self.W)
# dot product
logits = F.linear(x, W)
if label is None:
return logits
# add margin
target_logits = logits - self.m
one_hot = torch.zeros_like(logits)
one_hot.scatter_(1, label.view(-1, 1).long(), 1)
output = logits * (1 - one_hot) + target_logits * one_hot
# feature re-scale
output *= self.s
return output
def forward_drop(self, input, label=None):
# normalize features
x = F.normalize(input)
# normalize weights
W = F.normalize(self.W[self.rem_classes])
# dot product
logits = F.linear(x, W)
if label is None:
return logits
# add margin
target_logits = logits - self.m
one_hot = torch.zeros_like(logits)
one_hot.scatter_(1, label.view(-1, 1).long(), 1)
output = logits * (1 - one_hot) + target_logits * one_hot
# feature re-scale
output *= self.s
return output
class SphereFace(DropClassBase):
def __init__(self, num_features, num_classes, s=30.0, m=1.35):
super(SphereFace, self).__init__(num_classes)
self.num_features = num_features
self.n_classes = num_classes
self.s = s
self.m = m
self.W = nn.Parameter(torch.FloatTensor(num_classes, num_features))
self.reset_parameters()
def reset_parameters(self):
nn.init.xavier_uniform_(self.W)
def forward_nodrop(self, input, label=None):
# normalize features
x = F.normalize(input)
# normalize weights
W = F.normalize(self.W)
# dot product
logits = F.linear(x, W)
if label is None:
return logits
# add margin
theta = torch.acos(torch.clamp(logits, -1.0 + 1e-7, 1.0 - 1e-7))
target_logits = torch.cos(self.m * theta)
one_hot = torch.zeros_like(logits)
one_hot.scatter_(1, label.view(-1, 1).long(), 1)
output = logits * (1 - one_hot) + target_logits * one_hot
# feature re-scale
output *= self.s
return output
def forward_drop(self, input, label=None):
# normalize features
x = F.normalize(input)
# normalize weights
W = F.normalize(self.W[self.rem_classes])
# dot product
logits = F.linear(x, W)
if label is None:
return logits
# add margin
theta = torch.acos(torch.clamp(logits, -1.0 + 1e-7, 1.0 - 1e-7))
target_logits = torch.cos(self.m * theta)
one_hot = torch.zeros_like(logits)
one_hot.scatter_(1, label.view(-1, 1).long(), 1)
output = logits * (1 - one_hot) + target_logits * one_hot
# feature re-scale
output *= self.s
return output
class ArcFace(DropClassBase):
def __init__(self, num_features, num_classes, s=30.0, m=0.50):
super(ArcFace, self).__init__(num_classes)
self.num_features = num_features
self.n_classes = num_classes
self.s = s
self.m = m
self.W = nn.Parameter(torch.FloatTensor(num_classes, num_features))
self.reset_parameters()
def reset_parameters(self):
nn.init.xavier_uniform_(self.W)
def forward_nodrop(self, input, label=None):
# normalize features
x = F.normalize(input)
# normalize weights
W = F.normalize(self.W)
# dot product
logits = F.linear(x, W)
if label is None:
return logits
# add margin
theta = torch.acos(torch.clamp(logits, -1.0 + 1e-7, 1.0 - 1e-7))
target_logits = torch.cos(theta + self.m)
one_hot = torch.zeros_like(logits)
one_hot.scatter_(1, label.view(-1, 1).long(), 1)
output = logits * (1 - one_hot) + target_logits * one_hot
# feature re-scale
output *= self.s
return output
def forward_drop(self, input, label=None):
# normalize features
x = F.normalize(input)
# normalize weights
W = F.normalize(self.W[self.rem_classes])
# dot product
logits = F.linear(x, W)
if label is None:
return logits
# add margin
theta = torch.acos(torch.clamp(logits, -1.0 + 1e-7, 1.0 - 1e-7))
target_logits = torch.cos(theta + self.m)
one_hot = torch.zeros_like(logits)
one_hot.scatter_(1, label.view(-1, 1).long(), 1)
output = logits * (1 - one_hot) + target_logits * one_hot
# feature re-scale
output *= self.s
return output
class AdaCos(DropClassBase):
def __init__(self, num_features, num_classes, m=0.50):
super(AdaCos, self).__init__(num_classes)
self.num_features = num_features
self.n_classes = num_classes
self.s = math.sqrt(2) * math.log(num_classes - 1)
self.m = m
self.W = nn.Parameter(torch.FloatTensor(num_classes, num_features))
self.reset_parameters()
def reset_parameters(self):
nn.init.xavier_uniform_(self.W)
def forward_nodrop(self, input, label=None):
# normalize features
x = F.normalize(input)
# normalize weights
W = F.normalize(self.W)
# dot product
logits = F.linear(x, W)
if label is None:
return logits
# feature re-scale
theta = torch.acos(torch.clamp(logits, -1.0 + 1e-7, 1.0 - 1e-7))
one_hot = torch.zeros_like(logits)
one_hot.scatter_(1, label.view(-1, 1).long(), 1)
with torch.no_grad():
B_avg = torch.where(one_hot < 1, torch.exp(self.s * logits), torch.zeros_like(logits))
B_avg = torch.sum(B_avg) / input.size(0)
theta_med = torch.median(theta[one_hot == 1])
self.s = torch.log(B_avg) / torch.cos(torch.min(math.pi/4 * torch.ones_like(theta_med), theta_med))
output = self.s * logits
return output
def forward_drop(self, input, label=None):
# normalize features
x = F.normalize(input)
# normalize weights
W = F.normalize(self.W[self.rem_classes])
# dot product
logits = F.linear(x, W)
if label is None:
return logits
# feature re-scale
theta = torch.acos(torch.clamp(logits, -1.0 + 1e-7, 1.0 - 1e-7))
one_hot = torch.zeros_like(logits)
one_hot.scatter_(1, label.view(-1, 1).long(), 1)
with torch.no_grad():
B_avg = torch.where(one_hot < 1, torch.exp(self.s * logits), torch.zeros_like(logits))
B_avg = torch.sum(B_avg) / input.size(0)
theta_med = torch.median(theta[one_hot == 1])
self.s = torch.log(B_avg) / torch.cos(torch.min(math.pi/4 * torch.ones_like(theta_med), theta_med))
output = self.s * logits
return output
class DisturbLabelLoss(nn.Module):
def __init__(self, device, disturb_prob=0.1):
super(DisturbLabelLoss, self).__init__()
self.disturb_prob = disturb_prob
self.ce = nn.CrossEntropyLoss()
self.device = device
def forward(self, pred, target):
with torch.no_grad():
disturb_indexes = torch.rand(len(pred)) < self.disturb_prob
target[disturb_indexes] = torch.randint(pred.shape[-1], (int(disturb_indexes.sum()),)).to(self.device)
return self.ce(pred, target)
class LabelSmoothingLoss(nn.Module):
def __init__(self, smoothing=0.1, dim=-1):
super(LabelSmoothingLoss, self).__init__()
self.confidence = 1.0 - smoothing
self.smoothing = smoothing
self.dim = dim
def forward(self, pred, target):
pred = pred.log_softmax(dim=self.dim)
with torch.no_grad():
true_dist = torch.zeros_like(pred)
true_dist.fill_(self.smoothing / (pred.shape[-1] - 1))
true_dist.scatter_(1, target.data.unsqueeze(1), self.confidence)
return torch.mean(torch.sum(-true_dist * pred, dim=self.dim))