-
Notifications
You must be signed in to change notification settings - Fork 0
/
models.py
695 lines (598 loc) · 27.8 KB
/
models.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
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
import torch
from torch import nn
from src.choquet_utils import (
generate_interpolates, get_model_grad_wrt_gen_params, get_model_grad_wrt_interpolates,
count_nonzero_weights, get_weight_norms, project_weights_to_positive, test_model_convexity,
)
class MaxOut2D(nn.Module):
"""
Pytorch implementation of MaxOut on channels for an input that is C x H x W.
Reshape input from N x C x H x W --> N x H*W x C --> perform MaxPool1D on dim 2, i.e. channels --> reshape back to
N x C//maxout_kernel x H x W.
"""
def __init__(self, max_out):
super(MaxOut2D, self).__init__()
self.max_out = max_out
self.max_pool = nn.MaxPool1d(max_out)
def forward(self, x):
batch_size = x.shape[0]
channels = x.shape[1]
height = x.shape[2]
width = x.shape[3]
# Reshape input from N x C x H x W --> N x H*W x C
x_reshape = torch.permute(x, (0, 2, 3, 1)).view(batch_size, height * width, channels)
# Pool along channel dims
x_pooled = self.max_pool(x_reshape)
# Reshape back to N x C//maxout_kernel x H x W.
return torch.permute(x_pooled, (0, 2, 1)).view(batch_size, channels // self.max_out, height, width).contiguous()
class DistributionGenerator(nn.Module):
def __init__(self, dim=32, dimh=64, num_layers=2, activation='relu', max_out=4, dropout=False):
super(DistributionGenerator, self).__init__()
non_linearity = nn.MaxPool1d(kernel_size=max_out) if activation == 'max_out' else nn.ReLU()
dimh_adjusted = dimh // max_out if activation == 'max_out' else dimh
if dropout:
input_to_hidden = nn.Sequential(
nn.Linear(dim, dimh),
non_linearity,
nn.Dropout()
)
layers = []
for _ in range(num_layers - 1):
layers += [nn.Sequential(nn.Linear(dimh_adjusted, dimh), non_linearity, nn.Dropout())]
main = nn.Sequential(*layers)
else:
input_to_hidden = nn.Sequential(
nn.Linear(dim, dimh),
non_linearity,
)
layers = []
for _ in range(num_layers - 1):
layers += [nn.Sequential(nn.Linear(dimh_adjusted, dimh), non_linearity)]
main = nn.Sequential(*layers)
self.residual = nn.Linear(dim, dimh_adjusted)
self.input_to_hidden = input_to_hidden
self.main = main
self.main_output = nn.Linear(dimh_adjusted, 2)
def forward(self, x):
output = self.input_to_hidden(x)
for i, layer in enumerate(self.main):
output = layer(output) + self.residual(x)
output = self.main_output(output)
return output
@staticmethod
def get_model_args_as_dict(args):
return {
"dim": args.z_dim,
"dimh": args.g_hidden_dim,
"num_layers": args.g_n_layers,
'activation': args.activation,
'max_out': args.max_out,
'dropout': args.dropout,
}
class DistributionDiscriminator(nn.Module):
def __init__(self, dim=2, dimh=64, num_layers=2, activation='relu', max_out=4, dropout=False):
super(DistributionDiscriminator, self).__init__()
non_linearity = nn.MaxPool1d(kernel_size=max_out) if activation == 'max_out' else nn.ReLU()
dimh_adjusted = dimh // max_out if activation == 'max_out' else dimh
if dropout:
input_to_hidden = nn.Sequential(
nn.Linear(dim, dimh),
non_linearity,
nn.Dropout()
)
layers = []
for _ in range(num_layers - 1):
layers += [nn.Sequential(nn.Linear(dimh_adjusted, dimh), non_linearity, nn.Dropout())]
main = nn.Sequential(*layers)
else:
input_to_hidden = nn.Sequential(
nn.Linear(dim, dimh),
non_linearity,
)
layers = []
for _ in range(num_layers - 1):
layers += [nn.Sequential(nn.Linear(dimh_adjusted, dimh), non_linearity)]
main = nn.Sequential(*layers)
self.residual = nn.Linear(dim, dimh_adjusted)
self.input_to_hidden = input_to_hidden
self.main = main
self.main_output = nn.Linear(dimh_adjusted, 1)
def forward(self, x):
output = self.input_to_hidden(x)
for layer in self.main:
output = layer(output) + self.residual(x)
output = self.main_output(output)
return output
@staticmethod
def get_model_args_as_dict(args):
return {
'dim': 2,
'dimh': args.d_hidden_dim,
'num_layers': args.d_n_layers,
'activation': args.activation,
'max_out': args.max_out,
'dropout': args.dropout,
}
class MnistGenerator(nn.Module):
"""
This class is largely based on the generator from:
https://github.com/caogang/wgan-gp/blob/master/gan_mnist.py
"""
def __init__(self, dim=32, dimh=64, output_dim=(1, 28, 28), activation='relu', max_out=0, dropout=False):
super(MnistGenerator, self).__init__()
self.dimh_adjusted = dimh if activation == 'relu' else dimh // max_out
self.output_dim = (-1, *output_dim)
if dropout:
preprocess = nn.Sequential(
nn.Linear(dim, 4 * 4 * 4 * dimh),
nn.ReLU(True) if activation == 'relu' else nn.MaxPool1d(kernel_size=max_out),
nn.Dropout()
)
block1 = nn.Sequential(
nn.ConvTranspose2d(4 * self.dimh_adjusted, 2 * dimh, (5, 5)),
nn.ReLU(True) if activation == 'relu' else MaxOut2D(max_out=max_out),
nn.Dropout2d()
)
block2 = nn.Sequential(
nn.ConvTranspose2d(2 * self.dimh_adjusted, dimh, (5, 5)),
nn.ReLU(True) if activation == 'relu' else MaxOut2D(max_out=max_out),
nn.Dropout2d()
)
else:
preprocess = nn.Sequential(
nn.Linear(dim, 4 * 4 * 4 * dimh),
nn.ReLU(True) if activation == 'relu' else MaxOut2D(max_out=max_out),
)
block1 = nn.Sequential(
nn.ConvTranspose2d(4 * self.dimh_adjusted, 2 * dimh, (5, 5)),
nn.ReLU(True) if activation == 'relu' else MaxOut2D(max_out=max_out),
)
block2 = nn.Sequential(
nn.ConvTranspose2d(2 * self.dimh_adjusted, dimh, (5, 5)),
nn.ReLU(True) if activation == 'relu' else MaxOut2D(max_out=max_out),
)
deconv_out = nn.ConvTranspose2d(self.dimh_adjusted, 1, (8, 8), stride=(2, 2))
self.block1 = block1
self.block2 = block2
self.deconv_out = deconv_out
self.preprocess = preprocess
self.sigmoid = nn.Sigmoid()
def forward(self, x):
output = self.preprocess(x)
output = output.view(-1, 4 * self.dimh_adjusted, 4, 4)
output = self.block1(output)
output = output[:, :, :7, :7]
output = self.block2(output)
output = self.deconv_out(output)
output = self.sigmoid(output)
return output.view(self.output_dim)
@staticmethod
def get_model_args_as_dict(args):
return {
'dim': args.z_dim,
'dimh': args.g_hidden_dim,
'output_dim': (1, 28, 28),
'activation': args.activation,
'max_out': args.max_out,
'dropout': args.dropout,
}
class MnistDiscriminator(nn.Module):
"""
This class is largely based on the generator from:
https://github.com/caogang/wgan-gp/blob/master/gan_mnist.py
"""
def __init__(self, dim=1, dimh=64, activation='relu', max_out=0, dropout=False):
super(MnistDiscriminator, self).__init__()
self.dimh_adjusted = dimh if activation == 'relu' else dimh // max_out
if dropout:
input_to_hidden = nn.Sequential(
nn.Conv2d(dim, dimh, (5, 5), stride=(2, 2), padding=2),
nn.ReLU(True) if activation == 'relu' else MaxOut2D(max_out=max_out),
nn.Dropout2d(),
)
main = nn.Sequential(
nn.Conv2d(self.dimh_adjusted, 2*dimh, (5, 5), stride=(2, 2), padding=2),
nn.ReLU(True) if activation == 'relu' else MaxOut2D(max_out=max_out),
nn.Dropout2d(),
nn.Conv2d(2*self.dimh_adjusted, 4 * dimh, (5, 5), stride=(2, 2), padding=2),
nn.ReLU(True) if activation == 'relu' else MaxOut2D(max_out=max_out),
nn.Dropout2d(),
)
else:
input_to_hidden = nn.Sequential(
nn.Conv2d(dim, dimh, (5, 5), stride=(2, 2), padding=2),
nn.ReLU(True) if activation == 'relu' else MaxOut2D(max_out=max_out),
)
main = nn.Sequential(
nn.Conv2d(dimh, 2*self.dimh_adjusted, (5, 5), stride=(2, 2), padding=2),
nn.ReLU(True) if activation == 'relu' else MaxOut2D(max_out=max_out),
nn.Conv2d(2*self.dimh_adjusted, 4 * dimh, (5, 5), stride=(2, 2), padding=2),
nn.ReLU(True) if activation == 'relu' else MaxOut2D(max_out=max_out),
)
self.input_to_hidden = input_to_hidden
self.main = main
self.main_output = nn.Linear(64 * self.dimh_adjusted, 1)
def forward(self, x):
out = self.input_to_hidden(x)
out = self.main(out)
out = out.view(out.shape[0], -1)
out = self.main_output(out)
return out
@staticmethod
def get_model_args_as_dict(args):
return {
'dim': 1,
'dimh': args.d_hidden_dim,
'activation': args.activation,
'max_out': args.max_out,
'dropout': args.dropout,
}
class UpsampleConv(nn.Module):
"""
Code from:
https://github.com/ozanciga/gans-with-pytorch/blob/2071efd166935f0b4fb321227e94aa2ad1cfa273/wgan-gp/models.py#L29
"""
def __init__(self, n_input, n_output, k_size):
super(UpsampleConv, self).__init__()
self.model = nn.Sequential(
nn.PixelShuffle(2),
nn.Conv2d(n_input, n_output, k_size, stride=(1, 1), padding=(k_size - 1) // 2, bias=True)
)
def forward(self, x):
x = x.repeat((1, 4, 1, 1)) # Weird concat of WGAN-GPs upsampling process.
out = self.model(x)
return out
class ResidualBlock(nn.Module):
"""
Code from:
https://github.com/ozanciga/gans-with-pytorch/blob/2071efd166935f0b4fb321227e94aa2ad1cfa273/wgan-gp/models.py#L29
"""
def __init__(self, n_input, n_output, k_size, resample='up', bn=True, spatial_dim=None):
super(ResidualBlock, self).__init__()
self.resample = resample
if resample == 'up':
self.conv1 = UpsampleConv(n_input, n_output, k_size)
self.conv2 = nn.Conv2d(n_output, n_output, k_size, padding=(k_size - 1) // 2)
self.conv_shortcut = UpsampleConv(n_input, n_output, k_size)
self.out_dim = n_output
else:
self.conv1 = nn.Conv2d(n_input, n_input, k_size, padding=(k_size - 1) // 2)
self.conv2 = nn.Conv2d(n_input, n_input, k_size, padding=(k_size - 1) // 2)
self.conv_shortcut = None # Identity
self.out_dim = n_input
self.ln_dims = [n_input, spatial_dim, spatial_dim]
self.model = nn.Sequential(
nn.BatchNorm2d(n_input) if bn else nn.LayerNorm(self.ln_dims),
nn.ReLU(inplace=True),
self.conv1,
nn.BatchNorm2d(self.out_dim) if bn else nn.LayerNorm(self.ln_dims),
nn.ReLU(inplace=True),
self.conv2,
)
def forward(self, x):
if self.conv_shortcut is None:
return x + self.model(x)
else:
return self.conv_shortcut(x) + self.model(x)
class MaxOutResidualBlock(nn.Module):
def __init__(self, n_input, k_size, max_out, upsample_dim=None, dropout=False):
super(MaxOutResidualBlock, self).__init__()
if upsample_dim:
self.conv1 = UpsampleConv(n_input // max_out, upsample_dim, k_size)
self.conv2 = nn.Conv2d(upsample_dim // max_out, upsample_dim, k_size, padding=(k_size - 1) // 2)
self.conv_shortcut = UpsampleConv(n_input, upsample_dim, k_size)
self.out_dim = upsample_dim
else:
self.conv1 = nn.Conv2d(n_input // max_out, n_input, k_size, padding=(k_size - 1) // 2)
self.conv2 = nn.Conv2d(n_input // max_out, n_input, k_size, padding=(k_size - 1) // 2)
self.conv_shortcut = None
if dropout:
self.model = nn.Sequential(
MaxOut2D(max_out=max_out),
nn.Dropout2d(),
self.conv1,
MaxOut2D(max_out=max_out),
nn.Dropout2d(),
self.conv2,
)
else:
self.model = nn.Sequential(
MaxOut2D(max_out=max_out),
self.conv1,
MaxOut2D(max_out=max_out),
self.conv2,
)
def forward(self, x):
if self.conv_shortcut is None:
return x + self.model(x)
else:
return self.conv_shortcut(x) + self.model(x)
def register_convex_modules(self):
return self.conv1, self.conv2
class Cifar10ResidualGenerator(nn.Module):
"""
Code from:
https://github.com/ozanciga/gans-with-pytorch/blob/2071efd166935f0b4fb321227e94aa2ad1cfa273/wgan-gp/models.py#L29
"""
def __init__(self, dim, dimh, activation='max_out', max_out=4, dropout=False):
super(Cifar10ResidualGenerator, self).__init__()
self.model = nn.Sequential(
nn.ConvTranspose2d(dim, dimh, (4, 4), (1, 1), (0, 0)),
ResidualBlock(dimh, dimh, 3, resample='up') if activation == 'relu'
else MaxOutResidualBlock(n_input=dimh, k_size=3, max_out=max_out, upsample_dim=dimh, dropout=dropout),
ResidualBlock(dimh, dimh, 3, resample='up') if activation == 'relu'
else MaxOutResidualBlock(n_input=dimh, k_size=3, max_out=max_out, upsample_dim=dimh, dropout=dropout),
ResidualBlock(dimh, dimh, 3, resample='up') if activation == 'relu'
else MaxOutResidualBlock(n_input=dimh, k_size=3, max_out=max_out, upsample_dim=dimh, dropout=dropout),
nn.BatchNorm2d(dimh),
nn.ReLU(inplace=True) if activation == 'relu' else MaxOut2D(max_out=max_out),
nn.Dropout(),
nn.Conv2d(dimh // (1 if activation == 'relu' else max_out), 3, (3, 3), padding=(3 - 1) // 2), # 3 x 32 x 32
nn.Tanh()
)
def forward(self, z):
img = self.model(z.unsqueeze(2).unsqueeze(3))
return img
@staticmethod
def get_model_args_as_dict(args):
return {
'dim': args.z_dim,
'dimh': args.g_hidden_dim,
'activation': args.activation,
'max_out': args.max_out,
'dropout': args.dropout,
}
class Cifar10Discriminator(nn.Module):
"""
This class is largely based on the generator from:
https://github.com/caogang/wgan-gp/blob/master/gan_cifar10.py
"""
def __init__(self, dim=3, dimh=64, activation='relu',
max_out=4, dropout=False):
super(Cifar10Discriminator, self).__init__()
non_linearity = MaxOut2D(max_out=max_out) if activation == 'max_out' else nn.ReLU()
dimh_adjusted = dimh // max_out if activation == 'max_out' else dimh
if dropout:
input_to_hidden = nn.Sequential(
nn.Conv2d(dim, dimh, (3, 3), (2, 2), padding=1), # output: [bsz, dimh, 16, 16]
non_linearity,
nn.Dropout2d()
)
main = nn.Sequential(
nn.Conv2d(dimh_adjusted, 2 * dimh, (3, 3), (2, 2), padding=1), # output: [bsz, 2*dimh, 8, 8]
non_linearity,
nn.Dropout2d(),
nn.Conv2d(2 * dimh_adjusted, 4 * dimh, (3, 3), (2, 2), padding=1), # output: [bsz, 4*dimh, 4, 4]
non_linearity,
nn.Dropout2d()
)
else:
input_to_hidden = nn.Sequential(
nn.Conv2d(dim, dimh, (3, 3), (2, 2), padding=1), # output: [bsz, dimh, 16, 16]
non_linearity
)
main = nn.Sequential(
nn.Conv2d(dimh_adjusted, 2 * dimh, (3, 3), (2, 2), padding=1), # output: [bsz, 2*dimh, 8, 8]
non_linearity,
nn.Conv2d(2 * dimh_adjusted, 4 * dimh, (3, 3), (2, 2), padding=1), # output: [bsz, 4*dimh, 4, 4]
non_linearity
)
self.input_to_hidden = input_to_hidden
self.main = main
self.main_output = nn.Linear(4 * 4 * 4 * dimh_adjusted, 1)
def forward(self, x):
output = self.input_to_hidden(x)
output = self.main(output) # output: [bsz, 64*dimh]
output = self.main_output(output.view(output.shape[0], -1))
return output
@staticmethod
def get_model_args_as_dict(args):
return {
'dim': 3,
'dimh': args.d_hidden_dim,
'activation': args.activation,
'max_out': args.max_out,
'dropout': args.dropout,
}
class CTDiscrepancy(nn.Module):
"""
D_CT = inf_{u in cvx} int u d(mu_{+} - mu_{-}) + 0.5*int ||x||^2 d(mu_{-} - mu_{+})
where mu_{+} is distribution on generated data and mu_{-} is distribution on real data.
When training the discriminator (u), the norm term should not carry torch gradient.
"""
def __init__(self, critic, name='ct_disc'):
super(CTDiscrepancy, self).__init__()
self.critic = critic
self.positive_weight_module_names = ['main', 'main_output']
self.name = name
def forward(self, x):
return self.critic(x)
def objective(self, gen_data, real_data):
"""
Choquet objective
:param gen_data: Generated data
:param real_data: Ground truth / Real data
:return: dict with objective values. key `objective` contains term for .backward()
"""
u_integral = self.critic(gen_data).mean() - self.critic(real_data).mean()
objective = u_integral
if self.critic.training:
with torch.no_grad():
data_norm = 0.5 * (torch.linalg.vector_norm(real_data, ord=2, dim=1) ** 2).mean()
data_norm -= 0.5 * (torch.linalg.vector_norm(gen_data, ord=2, dim=1) ** 2).mean()
else:
data_norm = 0.5 * (torch.linalg.vector_norm(real_data, ord=2, dim=1) ** 2).mean()
data_norm -= 0.5 * (torch.linalg.vector_norm(gen_data, ord=2, dim=1) ** 2).mean()
objective += data_norm
return {'objective': objective, 'u_integral': u_integral, 'data_norm': data_norm}
def grad_reg_wrt_gen_params(self, z, generator, grad_reg_lambda):
grad_norm = get_model_grad_wrt_gen_params(self.critic, z, generator)
return grad_norm * grad_reg_lambda
def grad_reg_wrt_interpolates(self, real_data, fake_data, grad_reg_lambda):
grad_norm = get_model_grad_wrt_interpolates(self.critic, real_data, fake_data)
return grad_norm * grad_reg_lambda
def get_non_pos_params(self):
params = {}
for mod in self.non_positive_weight_module_names:
params[f'critic_{mod}'] = dict(self.critic.named_modules())[mod]
return nn.ModuleDict(params).parameters()
def get_pos_params(self):
pos_params = {}
for pos_mod in self.positive_weight_module_names:
pos_params[f'critic_{pos_mod}'] = dict(self.critic.named_modules())[pos_mod]
return nn.ModuleDict(pos_params).parameters()
def project_critic_weights_to_positive(self):
project_weights_to_positive(self.critic, self.positive_weight_module_names)
def log_critic_weight_norms(self, log):
weight_norms = get_weight_norms(self.critic)
for k, v in weight_norms.items():
if v:
log(f'weights/{self.name}_{k}', v)
def log_critic_convexity(self, batch, log):
convex_percentage = test_model_convexity(self.critic, batch)
log(f'convexity/{self.name}_u', convex_percentage)
return {f'{self.name}_u': convex_percentage}
def log_critic_nonzero_weights(self, log):
nonzero_percentage = count_nonzero_weights(self.critic, self.positive_weight_module_names)
log(f'nonzero/{self.name}_u', nonzero_percentage)
return {f'{self.name}_u': nonzero_percentage}
class VariationalDominanceCriterion(CTDiscrepancy):
"""
VDC = inf_{u in cvx} int u d(mu_{+} - mu_{-})
where mu_{+} is distribution on generated data and mu_{-} is distribution on real data.
Same objective as d_CT, but without data norm term.
"""
def __init__(self, critic, name='vdc'):
super().__init__(critic, name)
def objective(self, dominating_data, dominated_data):
"""
Choquet objective
:param dominating_data: Data from dominating distribution
:param dominated_data: Data from distribution that is to be dominated
:return: E[u(dominating)] - E[u(dominated)]
"""
critic_on_dominating = self.critic(dominating_data).mean()
critic_on_dominated = self.critic(dominated_data).mean()
u_integral = critic_on_dominating - critic_on_dominated
return u_integral
def grad_reg_wrt_interpolates(self, data0, data1, grad_reg_lambda):
grad_norm = get_model_grad_wrt_interpolates(self.critic, data0, data1)
return grad_norm * grad_reg_lambda
def reg_u_squared(self, data0, data1, reg_lambda):
critic_squared_on_data0 = torch.square(self.critic(data0)).mean()
critic_squared_on_data1 = torch.square(self.critic(data1)).mean()
return reg_lambda*(critic_squared_on_data0 + critic_squared_on_data1)
class CTDistance(nn.Module):
"""
d_CT = inf_{u_0 in cvx} int u_0 d(mu_{+} - mu_{-}) + inf_{u_1 in cvx} int u_1 d(mu_{-} - mu_{+})
where mu_{+} is distribution on generated data and mu_{-} is distribution on real data.
We experiment with different ways of combining the integral terms above:
- `sum`
- `min`
"""
def __init__(self, critics, how_to_combine_integral_terms='sum', split_regularization=False, name='ct_dist'):
super(CTDistance, self).__init__()
self.critic_0 = critics[0]
self.critic_1 = critics[1]
self.how_to_combine_integral_terms = how_to_combine_integral_terms
self.split_regularization = split_regularization
self.non_positive_weight_module_names = ['input_to_hidden']
self.positive_weight_module_names = ['main', 'main_output']
self.name = name
def forward(self, x):
return self.critic_0(x), self.critic_1(x)
def objective(self, gen_data, real_data):
"""
Choquet objective
:param gen_data: Generated data
:param real_data: Ground truth / Real data
:return: dict with objective values. Each critic receives its own objective value (i.e., u{i}_integral) for
.backward() calls
"""
critics_on_gen_data = self(gen_data)
critics_on_real_data = self(real_data)
u0_integral = critics_on_gen_data[0].mean() - critics_on_real_data[0].mean()
u1_integral = critics_on_real_data[1].mean() - critics_on_gen_data[1].mean()
if self.how_to_combine_integral_terms == 'sum':
objective = u0_integral + u1_integral
elif self.how_to_combine_integral_terms == 'min':
if u0_integral < u1_integral:
objective = u0_integral
u1_integral = u1_integral.detach()
else:
objective = u1_integral
u0_integral = u0_integral.detach()
else:
raise NotImplementedError
return {'objective': objective, 'u0_integral': u0_integral, 'u1_integral': u1_integral}
def grad_reg_wrt_gen_params(self, z, generator, grad_reg_lambda):
gen_z = generator(z)
crit_out = self(gen_z)
gr_norm_sq = None
gen_params_with_grad = [g for g in generator.parameters() if g.requires_grad]
if self.split_regularization:
gr_norm_sq = get_model_grad_wrt_gen_params(self.critic_0, z, generator)
gr_norm_sq += get_model_grad_wrt_gen_params(self.critic_1, z, generator)
else:
grads = torch.autograd.grad(crit_out[0].mean() - crit_out[1].mean(), gen_params_with_grad,
create_graph=True, retain_graph=True)[0]
for gr in grads:
if gr_norm_sq is None:
gr_norm_sq = (gr ** 2).sum()
else:
gr_norm_sq += (gr ** 2).sum()
return grad_reg_lambda * gr_norm_sq
def grad_reg_wrt_interpolates(self, real_data, fake_data, grad_reg_lambda):
interpolates = generate_interpolates(real_data, fake_data)
if self.split_regularization:
# critic_0
grad_penalty = get_model_grad_wrt_interpolates(self.critic_0,
real_data=None, fake_data=None, interpolates=interpolates)
# critic_1
grad_penalty += get_model_grad_wrt_interpolates(self.critic_1,
real_data=None, fake_data=None, interpolates=interpolates)
else:
interpolates = interpolates.detach().clone().requires_grad_(True)
crit_interpolates_diff = self.critic_0(interpolates) - self.critic_1(interpolates)
gradients = torch.autograd.grad(outputs=crit_interpolates_diff, inputs=interpolates,
grad_outputs=torch.ones(crit_interpolates_diff.size()).type_as(
crit_interpolates_diff),
create_graph=True, retain_graph=True, only_inputs=True)[0]
gradients = gradients.view(gradients.size(0), -1)
grad_penalty = ((gradients.norm(2, dim=1) - 1) ** 2).mean()
return grad_penalty * grad_reg_lambda
def get_non_pos_params(self):
params = {}
for i, crit in enumerate([self.critic_0, self.critic_1]):
for mod in self.non_positive_weight_module_names:
params[f'critic{i}_{mod}'] = dict(crit.named_modules())[mod]
return nn.ModuleDict(params).parameters()
def get_pos_params(self):
pos_params = {}
for i, crit in enumerate([self.critic_0, self.critic_1]):
for pos_mod in self.positive_weight_module_names:
pos_params[f'critic{i}_{pos_mod}'] = dict(crit.named_modules())[pos_mod]
return nn.ModuleDict(pos_params).parameters()
def project_critic_weights_to_positive(self):
for crit in [self.critic_0, self.critic_1]:
project_weights_to_positive(crit, self.positive_weight_module_names)
def log_critic_convexity(self, batch, log):
convex_percentages = {}
for i, crit in enumerate([self.critic_0, self.critic_1]):
convex_percentage = test_model_convexity(crit, batch)
convex_percentages[f'{self.name}_u_{i}'] = convex_percentage
log(f'convexity/{self.name}_u_{i}', convex_percentage)
return convex_percentages
def log_critic_weight_norms(self, log):
for i, crit in enumerate([self.critic_0, self.critic_1]):
weight_norms = get_weight_norms(crit)
for k, v in weight_norms.items():
if v:
log(f'weights/{self.name}_u_{i}/{k}', v)
def log_critic_nonzero_weights(self, log):
nonzero_percentages = {}
for i, crit in enumerate([self.critic_0, self.critic_1]):
nonzero_percentage = count_nonzero_weights(crit, self.positive_weight_module_names)
nonzero_percentages[f'{self.name}_u_{i}'] = nonzero_percentage
log(f'nonzero/{self.name}_u_{i}', nonzero_percentage)
return nonzero_percentages