-
Notifications
You must be signed in to change notification settings - Fork 0
/
architecture.py
601 lines (501 loc) · 27 KB
/
architecture.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
import torch
import torch.nn as nn
import torch.optim
import FrEIA.framework as Ff
import FrEIA.modules as Fm
from autoencoder import AutoEncoder
def get_model_by_params(params):
if params['architecture'].lower() == "glow":
model = baseline_glow(params['model_params'])
elif params['architecture'].lower() == "bw":
model = baseline_bw(params['model_params'])
elif params['architecture'].lower() == "color":
model = baseline_color(params['model_params'])
else:
raise(ValueError("Model architecture is not defined"))
if params["model_params"].get("pretrain_cond", False):
del model.cond_net
model.cond_net = AutoEncoder(params["model_params"]["encoder_sizes"], params["model_params"]["decoder_sizes"], bw = params['model_params'].get("bw")).encoder
return model
def random_orthog(n):
w = torch.randn(n, n)
w = w + w.T
w, S, V = torch.svd(w)
return w
def sub_conv(ch_hidden, kernel, num_hidden_layers=0, num_classes=0, batchnorm=True):
pad = kernel // 2
if batchnorm:
if num_classes < 2:
bn = lambda ch_out: nn.BatchNorm2d(ch_out)
else:
bn = lambda ch_out: ConditionalBatchNorm2d(ch_out, num_classes)
else:
bn = lambda ch_out: Identity()
return lambda ch_in, ch_out: nn.Sequential(
nn.Conv2d(ch_in, ch_hidden, kernel, padding=pad),
nn.LeakyReLU(),
*(nn.Conv2d(ch_hidden, ch_hidden, kernel, padding=pad),
nn.LeakyReLU()) * num_hidden_layers,
nn.Conv2d(ch_hidden, ch_out, kernel, padding=pad),
bn(ch_out)
)
def sub_fc(ch_hidden, num_hidden_layers=0, dropout=0.0, num_classes=0, batchnorm=True):
if batchnorm:
if num_classes < 2:
bn = lambda ch_out: nn.BatchNorm1d(ch_out)
else:
bn = lambda ch_out: ConditionalBatchNorm1d(ch_out, num_classes)
else:
bn = lambda ch_out: Identity()
return lambda ch_in, ch_out: nn.Sequential(
nn.Linear(ch_in, ch_hidden),
nn.LeakyReLU(),
*(nn.Linear(ch_hidden, ch_hidden),
nn.LeakyReLU()) * num_hidden_layers,
nn.Linear(ch_hidden, ch_out),
nn.Dropout(p=dropout),
bn(ch_out)
)
def baseline_color(m_params):
if not m_params['permute'] in ['soft', 'random', 'none', 'false', None, False]:
raise (RuntimeError("Erros in model params: No 'permute'' selected or not understood."))
if not m_params['act_norm'] in ['learnednorm', 'movingavg', 'none', 'false', None, False]:
raise (RuntimeError("Erros in model params: No 'act_norm' selected or not understood."))
cond = CondNet(m_params)
nodes = [Ff.InputNode(2, 64, 64)]
# outputs of the cond. net at different resolution levels
conditions = [Ff.ConditionNode(64, 64, 64),
Ff.ConditionNode(128, 32, 32),
Ff.ConditionNode(128, 16, 16),
Ff.ConditionNode(512)]
split_nodes = []
for k in range(m_params['blocks_per_group'][0]):
nodes.append(Ff.Node(nodes[-1], Fm.GLOWCouplingBlock,
{'subnet_constructor': sub_conv(64, m_params['kernel_size_per_group'][0],
m_params['hidden_layers_per_group'][0], batchnorm=m_params.get("bn", True)),
'clamp': m_params['clamping_per_group'][0]},
conditions=conditions[0],
name=F'block_{k}'))
if m_params['permute'] == 'random':
nodes.append(Ff.Node(nodes[-1], Fm.PermuteRandom, {'seed': k}))
elif m_params['permute'] == 'soft':
nodes.append(Ff.Node([nodes[-1].out0], Fm.conv_1x1, {'M':random_orthog(2)}))
if m_params['act_norm'] == 'learnednorm':
nodes.append(Ff.Node(nodes[-1], LearnedActNorm, {'M': torch.randn(1), "b": torch.randn(1)}))
elif m_params['act_norm'] == 'movingavg':
nodes.append(Ff.Node(nodes[-1], Fm.ActNorm, {}))
nodes.append(Ff.Node(nodes[-1], Fm.HaarDownsampling, {'rebalance': 0.5}))
for k in range(m_params['blocks_per_group'][1]):
nodes.append(Ff.Node(nodes[-1], Fm.GLOWCouplingBlock,
{
'subnet_constructor': sub_conv(128, m_params['kernel_size_per_group'][1],
m_params['hidden_layers_per_group'][1], batchnorm=m_params.get("bn", True)),
'clamp': m_params['clamping_per_group'][1],
},
conditions=conditions[1],
name=F'block_{k + 2}'))
if m_params['permute'] == 'random':
nodes.append(Ff.Node(nodes[-1], Fm.PermuteRandom, {'seed': k}))
elif m_params['permute'] == 'soft':
nodes.append(Ff.Node([nodes[-1].out0], Fm.conv_1x1, {'M':random_orthog(8)}))
if m_params['act_norm'] == 'learnednorm':
nodes.append(Ff.Node(nodes[-1], LearnedActNorm, {'M': torch.randn(1), "b": torch.randn(1)}))
elif m_params['act_norm'] == 'movingavg':
nodes.append(Ff.Node(nodes[-1], Fm.ActNorm, {}))
# split off 8/12 ch
if m_params.get("split", True):
nodes.append(Ff.Node(nodes[-1], Fm.Split1D,
{'split_size_or_sections': [6, 2], 'dim': 0}))
split_nodes.append(Ff.Node(nodes[-1].out1, Fm.Flatten, {}))
nodes.append(Ff.Node(nodes[-1], Fm.HaarDownsampling, {'rebalance': 0.5}))
for k in range(m_params['blocks_per_group'][2]):
nodes.append(Ff.Node(nodes[-1], Fm.GLOWCouplingBlock,
{
'subnet_constructor': sub_conv(256, m_params['kernel_size_per_group'][2][k],
m_params['hidden_layers_per_group'][2], batchnorm=m_params.get("bn", True)),
'clamp': m_params['clamping_per_group'][2],
},
conditions=conditions[2],
name=F'block_{k + 6}'))
if m_params['permute'] == 'random':
nodes.append(Ff.Node(nodes[-1], Fm.PermuteRandom, {'seed': k}))
elif m_params['permute'] == 'soft':
nodes.append(Ff.Node([nodes[-1].out0], Fm.conv_1x1, {'M':random_orthog(24 if m_params.get("split", True) else 32)}))
if m_params['act_norm'] == 'learnednorm':
nodes.append(Ff.Node(nodes[-1], LearnedActNorm , {'M': torch.randn(1), "b": torch.randn(1)}))
elif m_params['act_norm'] == 'movingavg':
nodes.append(Ff.Node(nodes[-1], Fm.ActNorm, {}))
# split off 8/16 ch
if m_params.get("split", True):
nodes.append(Ff.Node(nodes[-1], Fm.Split1D,
{'split_size_or_sections': [12, 12], 'dim': 0}))
split_nodes.append(Ff.Node(nodes[-1].out1, Fm.Flatten, {}))
nodes.append(Ff.Node(nodes[-1], Fm.Flatten, {}, name='flatten'))
# fully_connected part
for k in range(m_params['blocks_per_group'][3]):
nodes.append(Ff.Node(nodes[-1], Fm.GLOWCouplingBlock,
{
'clamp': m_params['clamping_per_group'][3],
'subnet_constructor': sub_fc(m_params['fc_size'], m_params['hidden_layers_per_group'][3], dropout=m_params['dropout_fc'], batchnorm=m_params.get("bn", True))
},
conditions=conditions[3],
name=F'block_{k + 10}'))
#nodes.append(Ff.Node(nodes[-1], Fm.PermuteRandom, {'seed': k}))
#nodes.append(Ff.Node(nodes[-1], LearnedActNorm , {'M': torch.randn(1), "b": torch.randn(1)}))
# concat everything
if m_params.get("split", True):
nodes.append(Ff.Node([s.out0 for s in split_nodes] + [nodes[-1].out0],
Fm.Concat1d, {'dim': 0}))
nodes.append(Ff.OutputNode(nodes[-1]))
inn = SketchINN(cond)
inn.build_inn(nodes, split_nodes, conditions)
return inn
def baseline_bw(m_params):
if not m_params['permute'] in ['soft', 'random', 'none', 'false', None, False]:
raise (RuntimeError("Erros in model params: No 'permute'' selected or not understood."))
if not m_params['act_norm'] in ['learnednorm', 'movingavg', 'none', 'false', None, False]:
raise (RuntimeError("Erros in model params: No 'act_norm' selected or not understood."))
cond = CondNet(m_params)
nodes = [Ff.InputNode(4, 32, 32)]
# outputs of the cond. net at different resolution levels
conditions = [Ff.ConditionNode(64, 32, 32),
Ff.ConditionNode(128, 16, 16),
Ff.ConditionNode(128, 8, 8),
Ff.ConditionNode(512)]
split_nodes = []
for k in range(m_params['blocks_per_group'][0]):
nodes.append(Ff.Node(nodes[-1], Fm.GLOWCouplingBlock,
{'subnet_constructor': sub_conv(64, m_params['kernel_size_per_group'][0],
m_params['hidden_layers_per_group'][0], batchnorm=m_params.get("bn", True)),
'clamp': m_params['clamping_per_group'][0]},
conditions=conditions[0],
name=F'block_{k}'))
if m_params['permute'] == 'random':
nodes.append(Ff.Node(nodes[-1], Fm.PermuteRandom, {'seed': k}))
elif m_params['permute'] == 'soft':
nodes.append(Ff.Node([nodes[-1].out0], Fm.conv_1x1, {'M':random_orthog(4)}))
if m_params['act_norm'] == 'learnednorm':
nodes.append(Ff.Node(nodes[-1], LearnedActNorm, {'M': torch.randn(1), "b": torch.randn(1)}))
elif m_params['act_norm'] == 'movingavg':
nodes.append(Ff.Node(nodes[-1], Fm.ActNorm, {}))
nodes.append(Ff.Node(nodes[-1], Fm.HaarDownsampling, {'rebalance': 0.5}))
for k in range(m_params['blocks_per_group'][1]):
nodes.append(Ff.Node(nodes[-1], Fm.GLOWCouplingBlock,
{
'subnet_constructor': sub_conv(128, m_params['kernel_size_per_group'][1],
m_params['hidden_layers_per_group'][1], batchnorm=m_params.get("bn", True)),
'clamp': m_params['clamping_per_group'][1],
},
conditions=conditions[1],
name=F'block_{k + 2}'))
if m_params['permute'] == 'random':
nodes.append(Ff.Node(nodes[-1], Fm.PermuteRandom, {'seed': k}))
elif m_params['permute'] == 'soft':
nodes.append(Ff.Node([nodes[-1].out0], Fm.conv_1x1, {'M':random_orthog(16)}))
if m_params['act_norm'] == 'learnednorm':
nodes.append(Ff.Node(nodes[-1], LearnedActNorm, {'M': torch.randn(1), "b": torch.randn(1)}))
elif m_params['act_norm'] == 'movingavg':
nodes.append(Ff.Node(nodes[-1], Fm.ActNorm, {}))
# split off 8/12 ch
if m_params.get("split", True):
nodes.append(Ff.Node(nodes[-1], Fm.Split1D,
{'split_size_or_sections': [6, 10], 'dim': 0}))
split_nodes.append(Ff.Node(nodes[-1].out1, Fm.Flatten, {}))
nodes.append(Ff.Node(nodes[-1], Fm.HaarDownsampling, {'rebalance': 0.5}))
for k in range(m_params['blocks_per_group'][2]):
nodes.append(Ff.Node(nodes[-1], Fm.GLOWCouplingBlock,
{
'subnet_constructor': sub_conv(256, m_params['kernel_size_per_group'][2][k],
m_params['hidden_layers_per_group'][2], batchnorm=m_params.get("bn", True)),
'clamp': m_params['clamping_per_group'][2],
},
conditions=conditions[2],
name=F'block_{k + 6}'))
if m_params['permute'] == 'random':
nodes.append(Ff.Node(nodes[-1], Fm.PermuteRandom, {'seed': k}))
elif m_params['permute'] == 'soft':
nodes.append(Ff.Node([nodes[-1].out0], Fm.conv_1x1, {'M':random_orthog(24 if m_params.get("split", True) else 64)}))
if m_params['act_norm'] == 'learnednorm':
nodes.append(Ff.Node(nodes[-1], LearnedActNorm , {'M': torch.randn(1), "b": torch.randn(1)}))
elif m_params['act_norm'] == 'movingavg':
nodes.append(Ff.Node(nodes[-1], Fm.ActNorm, {}))
# split off 8/16 ch
if m_params.get("split", True):
nodes.append(Ff.Node(nodes[-1], Fm.Split1D,
{'split_size_or_sections': [12, 12], 'dim': 0}))
split_nodes.append(Ff.Node(nodes[-1].out1, Fm.Flatten, {}))
nodes.append(Ff.Node(nodes[-1], Fm.Flatten, {}, name='flatten'))
# fully_connected part
for k in range(m_params['blocks_per_group'][3]):
nodes.append(Ff.Node(nodes[-1], Fm.GLOWCouplingBlock,
{
'clamp': m_params['clamping_per_group'][3],
'subnet_constructor': sub_fc(m_params['fc_size'], m_params['hidden_layers_per_group'][3], dropout=m_params['dropout_fc'], batchnorm=m_params.get("bn", True))
},
conditions=conditions[3],
name=F'block_{k + 10}'))
#nodes.append(Ff.Node(nodes[-1], Fm.PermuteRandom, {'seed': k}))
#nodes.append(Ff.Node(nodes[-1], LearnedActNorm , {'M': torch.randn(1), "b": torch.randn(1)}))
# concat everything
if m_params.get("split", True):
nodes.append(Ff.Node([s.out0 for s in split_nodes] + [nodes[-1].out0],
Fm.Concat1d, {'dim': 0}))
nodes.append(Ff.OutputNode(nodes[-1]))
inn = SketchINN(cond)
inn.build_inn(nodes, split_nodes, conditions)
return inn
def baseline_glow(m_params):
if not m_params['permute'] in ['soft', 'random', 'none', 'false', None, False]:
raise (RuntimeError("Erros in model params: No 'permute'' selected or not understood."))
if not m_params['act_norm'] in ['learnednorm', 'movingavg', 'none', 'false', None, False]:
raise (RuntimeError("Erros in model params: No 'act_norm' selected or not understood."))
cond = CondNet(m_params)
nodes = [Ff.InputNode(3, 64, 64)]
# outputs of the cond. net at different resolution levels
conditions = [Ff.ConditionNode(64, 64, 64),
Ff.ConditionNode(128, 32, 32),
Ff.ConditionNode(128, 16, 16),
Ff.ConditionNode(512)]
split_nodes = []
for k in range(m_params['blocks_per_group'][0]):
nodes.append(Ff.Node(nodes[-1], Fm.GLOWCouplingBlock,
{'subnet_constructor': sub_conv(64, m_params['kernel_size_per_group'][0],
m_params['hidden_layers_per_group'][0], batchnorm=m_params.get("bn", True)),
'clamp': m_params['clamping_per_group'][0]},
conditions=conditions[0],
name=F'block_{k}'))
if m_params['permute'] == 'random':
nodes.append(Ff.Node(nodes[-1], Fm.PermuteRandom, {'seed': k}))
elif m_params['permute'] == 'soft':
nodes.append(Ff.Node([nodes[-1].out0], Fm.conv_1x1, {'M':random_orthog(3)}))
if m_params['act_norm'] == 'learnednorm':
nodes.append(Ff.Node(nodes[-1], LearnedActNorm, {'M': torch.randn(1), "b": torch.randn(1)}))
elif m_params['act_norm'] == 'movingavg':
nodes.append(Ff.Node(nodes[-1], Fm.ActNorm, {}))
nodes.append(Ff.Node(nodes[-1], Fm.HaarDownsampling, {'rebalance': 0.5}))
for k in range(m_params['blocks_per_group'][1]):
nodes.append(Ff.Node(nodes[-1], Fm.GLOWCouplingBlock,
{
'subnet_constructor': sub_conv(128, m_params['kernel_size_per_group'][1],
m_params['hidden_layers_per_group'][1], batchnorm=m_params.get("bn", True)),
'clamp': m_params['clamping_per_group'][1],
},
conditions=conditions[1],
name=F'block_{k + 2}'))
if m_params['permute'] == 'random':
nodes.append(Ff.Node(nodes[-1], Fm.PermuteRandom, {'seed': k}))
elif m_params['permute'] == 'soft':
nodes.append(Ff.Node([nodes[-1].out0], Fm.conv_1x1, {'M':random_orthog(12)}))
if m_params['act_norm'] == 'learnednorm':
nodes.append(Ff.Node(nodes[-1], LearnedActNorm, {'M': torch.randn(1), "b": torch.randn(1)}))
elif m_params['act_norm'] == 'movingavg':
nodes.append(Ff.Node(nodes[-1], Fm.ActNorm, {}))
# split off 8/12 ch
if m_params.get("split", True):
nodes.append(Ff.Node(nodes[-1], Fm.Split1D,
{'split_size_or_sections': [4, 8], 'dim': 0}))
split_nodes.append(Ff.Node(nodes[-1].out1, Fm.Flatten, {}))
nodes.append(Ff.Node(nodes[-1], Fm.HaarDownsampling, {'rebalance': 0.5}))
for k in range(m_params['blocks_per_group'][2]):
nodes.append(Ff.Node(nodes[-1], Fm.GLOWCouplingBlock,
{
'subnet_constructor': sub_conv(256, m_params['kernel_size_per_group'][2][k],
m_params['hidden_layers_per_group'][2], batchnorm=m_params.get("bn", True)),
'clamp': m_params['clamping_per_group'][2],
},
conditions=conditions[2],
name=F'block_{k + 6}'))
if m_params['permute'] == 'random':
nodes.append(Ff.Node(nodes[-1], Fm.PermuteRandom, {'seed': k}))
elif m_params['permute'] == 'soft':
nodes.append(Ff.Node([nodes[-1].out0], Fm.conv_1x1, {'M':random_orthog(16 if m_params.get("split", True) else 48)}))
if m_params['act_norm'] == 'learnednorm':
nodes.append(Ff.Node(nodes[-1], LearnedActNorm , {'M': torch.randn(1), "b": torch.randn(1)}))
elif m_params['act_norm'] == 'movingavg':
nodes.append(Ff.Node(nodes[-1], Fm.ActNorm, {}))
# split off 8/16 ch
if m_params.get("split", True):
nodes.append(Ff.Node(nodes[-1], Fm.Split1D,
{'split_size_or_sections': [8, 8], 'dim': 0}))
split_nodes.append(Ff.Node(nodes[-1].out1, Fm.Flatten, {}))
nodes.append(Ff.Node(nodes[-1], Fm.Flatten, {}, name='flatten'))
# fully_connected part
for k in range(m_params['blocks_per_group'][3]):
nodes.append(Ff.Node(nodes[-1], Fm.GLOWCouplingBlock,
{
'clamp': m_params['clamping_per_group'][3],
'subnet_constructor': sub_fc(m_params['fc_size'], m_params['hidden_layers_per_group'][3], dropout=m_params['dropout_fc'], batchnorm=m_params.get("bn", True))
},
conditions=conditions[3],
name=F'block_{k + 10}'))
# concat everything
if m_params.get("split", True):
nodes.append(Ff.Node([s.out0 for s in split_nodes] + [nodes[-1].out0],
Fm.Concat1d, {'dim': 0}))
nodes.append(Ff.OutputNode(nodes[-1]))
inn = SketchINN(cond)
inn.build_inn(nodes, split_nodes, conditions)
return inn
class CondNet(nn.Module):
'''conditioning network'''
def __init__(self, params):
super().__init__()
bw_pool = Identity()
if params.get("bw"):
bw_pool = nn.AvgPool2d(2)
self.blocks = nn.ModuleList([nn.Sequential(nn.Conv2d(1, 64, 3, padding=1),
nn.LeakyReLU(),
nn.Conv2d(64, 64, 3, padding=1),
bw_pool),
nn.Sequential(nn.LeakyReLU(),
nn.Conv2d(64, 128, 3, padding=1),
nn.LeakyReLU(),
nn.Conv2d(128, 128, 3, padding=1, stride=2)),
nn.Sequential(nn.LeakyReLU(),
nn.Conv2d(128, 128, 3, padding=1, stride=2)),
nn.Sequential(nn.LeakyReLU(),
nn.AvgPool2d(4),
nn.BatchNorm2d(128),
Flatten(),
nn.Dropout(params.get("cond_dropout", 0.0)),
nn.Linear(512, 512))
]
)
def forward(self, c):
outputs = [c]
for m in self.blocks:
outputs.append(m(outputs[-1]))
return outputs[1:]
class SketchINN(nn.Module):
'''cINN, including the conditioning network'''
def __init__(self, cond):
super().__init__()
self.cond_net = cond
def build_inn(self, nodes, split_nodes, conditions):
self.model = Ff.ReversibleGraphNet(nodes + split_nodes + conditions, verbose=False)
self.trainable_parameters = [p for p in self.model.parameters() if p.requires_grad]
for p in self.trainable_parameters:
p.data = 0.02 * torch.randn_like(p)
def forward(self, x, c=None, rev=False, intermediate_outputs=False):
return self.model(x, c=self.cond_net(c), rev=rev, intermediate_outputs=intermediate_outputs)
def log_jacobian(self):
return self.model.log_jacobian(run_forward=False)
class Flatten(nn.Module):
def __init__(self, *args):
super().__init__()
def forward(self, x):
return x.view(x.shape[0], -1)
class SubnetConstructorFC(nn.Module):
"""This class constructs a subnet for the inner parts of the GLOWCouplingBlocks
as well as the condition preprocessor.
num_layers: number of layers to be constructed by this instance
size_in: input size of the subnet
size_out: output size of the subnet
internal_size: hidden size of the subnet. If None, set to 2*size
dropout: dropout chance of the subnet
"""
def __init__(self, num_layers, size_in, size_out, internal_size=None, dropout=0.0):
super().__init__()
if internal_size is None:
internal_size = size_out * 2
if num_layers < 1:
raise (ValueError("Subnet size has to be 1 or greater"))
self.layers = []
for n in range(num_layers):
input_dim, output_dim = internal_size, internal_size
if n == 0:
input_dim = size_in
if n == num_layers - 1:
output_dim = size_out
self.layers.append(nn.Linear(input_dim, output_dim))
if n < num_layers - 1:
self.layers.append(nn.Dropout(p=dropout))
self.layers.append(nn.ReLU())
self.layers = nn.ModuleList(self.layers)
def forward(self, x):
for l in self.layers:
x = l(x)
return x
class SubnetConstructorConv(nn.Module):
"""This class constructs a subnet for the inner parts of the GLOWCouplingBlocks
as well as the condition preprocessor.
num_layers: number of layers to be constructed by this instance
size_in: input size of the subnet
in_channels: number of channels accepted as network input
out_channels: number of channels to be put out
stride: convolution stride
conv_size: filter size
"""
def __init__(self, num_layers, size_in, in_channels, out_channels, stride=1, conv_size=3,
dropout=0.0):
super().__init__()
if num_layers < 1:
raise (ValueError("Subnet size has to be 1 or greater"))
self.layers = []
ch1 = in_channels
pad_size = self.get_pad_size(size_in, size_in, stride, conv_size)
for n in range(num_layers - 1):
self.layers.append(nn.Conv2d(ch1, out_channels, conv_size, padding=pad_size, stride=stride))
self.layers.append(nn.ReLU())
self.layers.append(nn.Dropout(dropout))
ch1 = out_channels
self.layers.append(nn.Conv2d(ch1, out_channels, conv_size, padding=pad_size, stride=stride))
self.layers = nn.ModuleList(self.layers)
def get_pad_size(self, size_out, size_in, stride, conv_size):
# size_out == size_in
# width == height
return int((size_in / (stride) + conv_size - size_in) // 2)
def forward(self, x):
for l in self.layers:
x = l(x)
return x
class LearnedActNorm(nn.Module):
'''Learned transformation according to y = Mx + b, with invertible
matrix M.'''
def __init__(self, dims_in, M, b):
super().__init__()
self.M = nn.Parameter(M, requires_grad=True)
self.b = nn.Parameter(b, requires_grad=True)
self.n_pixels = 1
self.activation = nn.Softplus(beta=0.5)
def forward(self, x, rev=False):
if not rev:
return [self.activation(self.M) * x[0] + self.b]
else:
return [(x[0] - self.b)/ self.activation(self.M) ]
def jacobian(self, x, rev=False):
return ((-1)**rev * self.n_pixels) * (torch.log(self.activation(self.M) + 1e-12).sum())
def output_dims(self, input_dims):
return input_dims
#taken from https://github.com/pytorch/pytorch/issues/8985#issuecomment-405080775
class ConditionalBatchNorm2d(nn.Module):
def __init__(self, num_features, num_classes):
super().__init__()
self.num_features = num_features
self.bn = nn.BatchNorm2d(num_features, affine=False)
self.embed = nn.Embedding(num_classes, num_features * 2)
self.embed.weight.data[:, :num_features].normal_(1, 0.02) # Initialise scale at N(1, 0.02)
self.embed.weight.data[:, num_features:].zero_() # Initialise bias at 0
def forward(self, x, y):
out = self.bn(x)
gamma, beta = self.embed(y).chunk(2, 1)
out = gamma.view(-1, self.num_features, 1, 1) * out + beta.view(-1, self.num_features, 1, 1)
return out
class ConditionalBatchNorm1d(nn.Module):
def __init__(self, num_features, num_classes):
super().__init__()
self.num_features = num_features
self.bn = nn.BatchNorm2d(num_features, affine=False)
self.embed = nn.Embedding(num_classes, num_features * 2)
self.embed.weight.data[:, :num_features].normal_(1, 0.02) # Initialise scale at N(1, 0.02)
self.embed.weight.data[:, num_features:].zero_() # Initialise bias at 0
def forward(self, x, y):
out = self.bn(x)
gamma, beta = self.embed(y).chunk(2, 1)
out = gamma.view(-1, self.num_features) * out + beta.view(-1, self.num_features)
return out
class Identity(nn.Module):
def __init__(self, *args, **kwargs):
super().__init__()
def forward(self, x):
return x