-
Notifications
You must be signed in to change notification settings - Fork 3
/
modules.py
654 lines (470 loc) · 21.2 KB
/
modules.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
import torch
from torch import nn
from torchmeta.modules import (MetaModule, MetaSequential)
from torchmeta.modules.utils import get_subdict
import numpy as np
from collections import OrderedDict
import math
import torch.nn.functional as F
class BatchLinear(nn.Linear, MetaModule):
'''A linear meta-layer that can deal with batched weight matrices and biases, as for instance output by a
hypernetwork.'''
__doc__ = nn.Linear.__doc__
def forward(self, input, params=None):
if params is None:
params = OrderedDict(self.named_parameters())
bias = params.get('bias', None)
weight = params['weight']
output = input.matmul(weight.permute(*[i for i in range(len(weight.shape) - 2)], -1, -2))
output += bias.unsqueeze(-2)
return output
class Sine(nn.Module):
def __init(self):
super().__init__()
def forward(self, input):
return torch.sin(30 * input)
class FCBlock(MetaModule):
'''A fully connected neural network that also allows swapping out the weights when used with a hypernetwork.
Can be used just as a normal neural network though, as well.
'''
def __init__(self, in_features, out_features, num_hidden_layers, hidden_features,
outermost_linear=False, nonlinearity='relu', weight_init=None):
super().__init__()
self.first_layer_init = None
nls_and_inits = {'sine':(Sine(), sine_init, first_layer_sine_init),
'relu':(nn.ReLU(inplace=True), init_weights_normal, None),
'sigmoid':(nn.Sigmoid(), init_weights_xavier, None),
'tanh':(nn.Tanh(), init_weights_xavier, None),
'selu':(nn.SELU(inplace=True), init_weights_selu, None),
'softplus':(nn.Softplus(), init_weights_normal, None),
'elu':(nn.ELU(inplace=True), init_weights_elu, None)}
nl, nl_weight_init, first_layer_init = nls_and_inits[nonlinearity]
if weight_init is not None:
self.weight_init = weight_init
else:
self.weight_init = nl_weight_init
self.net = []
self.net.append(MetaSequential(
BatchLinear(in_features, hidden_features), nl
))
for i in range(num_hidden_layers):
self.net.append(MetaSequential(
BatchLinear(hidden_features, hidden_features), nl
))
if outermost_linear:
self.net.append(MetaSequential(BatchLinear(hidden_features, out_features)))
else:
self.net.append(MetaSequential(
BatchLinear(hidden_features, out_features), nl
))
self.net = MetaSequential(*self.net)
if self.weight_init is not None:
self.net.apply(self.weight_init)
if first_layer_init is not None:
self.net[0].apply(first_layer_init)
def forward(self, coords, params=None, **kwargs):
if params is None:
params = OrderedDict(self.named_parameters())
output = self.net(coords, params=get_subdict(params, 'net'))
return output
def forward_with_activations(self, coords, params=None, retain_grad=False):
'''Returns not only model output, but also intermediate activations.'''
if params is None:
params = OrderedDict(self.named_parameters())
activations = OrderedDict()
x = coords.clone().detach().requires_grad_(True)
activations['input'] = x
for i, layer in enumerate(self.net):
subdict = get_subdict(params, 'net.%d' % i)
for j, sublayer in enumerate(layer):
if isinstance(sublayer, BatchLinear):
x = sublayer(x, params=get_subdict(subdict, '%d' % j))
else:
x = sublayer(x)
if retain_grad:
x.retain_grad()
activations['_'.join((str(sublayer.__class__), "%d" % i))] = x
return activations
class SingleBVPNet(MetaModule):
'''A canonical representation network for a BVP.'''
def __init__(self, out_features=1, type='sine', in_features=2,
mode='mlp', hidden_features=256, num_hidden_layers=3, **kwargs):
super().__init__()
self.mode = mode
if self.mode == 'rbf':
self.rbf_layer = RBFLayer(in_features=in_features, out_features=kwargs.get('rbf_centers', 1024))
in_features = kwargs.get('rbf_centers', 1024)
elif self.mode == 'nerf':
self.positional_encoding = PosEncodingNeRF(in_features=in_features,
sidelength=kwargs.get('sidelength', None),
fn_samples=kwargs.get('fn_samples', None),
use_nyquist=kwargs.get('use_nyquist', True))
in_features = self.positional_encoding.out_dim
self.image_downsampling = ImageDownsampling(sidelength=kwargs.get('sidelength', None),
downsample=kwargs.get('downsample', False))
self.net = FCBlock(in_features=in_features, out_features=out_features, num_hidden_layers=num_hidden_layers,
hidden_features=hidden_features, outermost_linear=True, nonlinearity=type)
print(self)
def forward(self, model_input, params=None):
if params is None:
params = OrderedDict(self.named_parameters())
coords_org = model_input['coords'].clone().detach().requires_grad_(True)
coords = coords_org
if self.image_downsampling.downsample:
coords = self.image_downsampling(coords)
if self.mode == 'rbf':
coords = self.rbf_layer(coords)
elif self.mode == 'nerf':
coords = self.positional_encoding(coords)
output = self.net(coords, get_subdict(params, 'net'))
return {'model_in': coords_org, 'model_out': output}
def forward_with_activations(self, model_input):
'''Returns not only model output, but also intermediate activations.'''
coords = model_input['coords'].clone().detach().requires_grad_(True)
activations = self.net.forward_with_activations(coords)
return {'model_in': coords, 'model_out': activations.popitem(), 'activations': activations}
class PINNet(nn.Module):
def __init__(self, out_features=1, type='tanh', in_features=2, mode='mlp'):
super().__init__()
self.mode = mode
self.net = FCBlock(in_features=in_features, out_features=out_features, num_hidden_layers=8,
hidden_features=20, outermost_linear=True, nonlinearity=type,
weight_init=init_weights_trunc_normal)
print(self)
def forward(self, model_input):
coords = model_input['coords'].clone().detach().requires_grad_(True)
output = self.net(coords)
return {'model_in': coords, 'model_out': output}
class ImageDownsampling(nn.Module):
def __init__(self, sidelength, downsample=False):
super().__init__()
if isinstance(sidelength, int):
self.sidelength = (sidelength, sidelength)
else:
self.sidelength = sidelength
if self.sidelength is not None:
self.sidelength = torch.Tensor(self.sidelength).cuda().float()
else:
assert downsample is False
self.downsample = downsample
def forward(self, coords):
if self.downsample:
return coords + self.forward_bilinear(coords)
else:
return coords
def forward_box(self, coords):
return 2 * (torch.rand_like(coords) - 0.5) / self.sidelength
def forward_bilinear(self, coords):
Y = torch.sqrt(torch.rand_like(coords)) - 1
Z = 1 - torch.sqrt(torch.rand_like(coords))
b = torch.rand_like(coords) < 0.5
Q = (b * Y + ~b * Z) / self.sidelength
return Q
class PosEncodingNeRF(nn.Module):
def __init__(self, in_features, sidelength=None, fn_samples=None, use_nyquist=True):
super().__init__()
self.in_features = in_features
if self.in_features == 3:
self.num_frequencies = 10
elif self.in_features == 2:
assert sidelength is not None
if isinstance(sidelength, int):
sidelength = (sidelength, sidelength)
self.num_frequencies = 4
if use_nyquist:
self.num_frequencies = self.get_num_frequencies_nyquist(min(sidelength[0], sidelength[1]))
elif self.in_features == 1:
assert fn_samples is not None
self.num_frequencies = 4
if use_nyquist:
self.num_frequencies = self.get_num_frequencies_nyquist(fn_samples)
self.out_dim = in_features + 2 * in_features * self.num_frequencies
def get_num_frequencies_nyquist(self, samples):
nyquist_rate = 1 / (2 * (2 * 1 / samples))
return int(math.floor(math.log(nyquist_rate, 2)))
def forward(self, coords):
coords = coords.view(coords.shape[0], -1, self.in_features)
coords_pos_enc = coords
for i in range(self.num_frequencies):
for j in range(self.in_features):
c = coords[..., j]
sin = torch.unsqueeze(torch.sin((2 ** i) * np.pi * c), -1)
cos = torch.unsqueeze(torch.cos((2 ** i) * np.pi * c), -1)
coords_pos_enc = torch.cat((coords_pos_enc, sin, cos), axis=-1)
return coords_pos_enc.reshape(coords.shape[0], -1, self.out_dim)
class RBFLayer(nn.Module):
def __init__(self, in_features, out_features):
super().__init__()
self.in_features = in_features
self.out_features = out_features
self.centres = nn.Parameter(torch.Tensor(out_features, in_features))
self.sigmas = nn.Parameter(torch.Tensor(out_features))
self.reset_parameters()
self.freq = nn.Parameter(np.pi * torch.ones((1, self.out_features)))
def reset_parameters(self):
nn.init.uniform_(self.centres, -1, 1)
nn.init.constant_(self.sigmas, 10)
def forward(self, input):
input = input[0, ...]
size = (input.size(0), self.out_features, self.in_features)
x = input.unsqueeze(1).expand(size)
c = self.centres.unsqueeze(0).expand(size)
distances = (x - c).pow(2).sum(-1) * self.sigmas.unsqueeze(0)
return self.gaussian(distances).unsqueeze(0)
def gaussian(self, alpha):
phi = torch.exp(-1 * alpha.pow(2))
return phi
class SetEncoder(nn.Module):
def __init__(self, in_features, out_features,
num_hidden_layers, hidden_features, nonlinearity='relu'):
super().__init__()
assert nonlinearity in ['relu', 'sine'], 'Unknown nonlinearity type'
if nonlinearity == 'relu':
nl = nn.ReLU(inplace=True)
weight_init = init_weights_normal
elif nonlinearity == 'sine':
nl = Sine()
weight_init = sine_init
self.net = [nn.Linear(in_features, hidden_features), nl]
self.net.extend([nn.Sequential(nn.Linear(hidden_features, hidden_features), nl)
for _ in range(num_hidden_layers)])
self.net.extend([nn.Linear(hidden_features, out_features), nl])
self.net = nn.Sequential(*self.net)
self.net.apply(weight_init)
def forward(self, context_x, context_y, ctxt_mask=None, **kwargs):
input = torch.cat((context_x, context_y), dim=-1)
embeddings = self.net(input)
if ctxt_mask is not None:
embeddings = embeddings * ctxt_mask
embedding = embeddings.mean(dim=-2) * (embeddings.shape[-2] / torch.sum(ctxt_mask, dim=-2))
return embedding
return embeddings.mean(dim=-2)
class ConvImgEncoder(nn.Module):
def __init__(self, channel, image_resolution):
super().__init__()
self.conv_theta = nn.Conv2d(channel, 128, 3, 1, 1)
self.relu = nn.ReLU(inplace=True)
self.cnn = nn.Sequential(
nn.Conv2d(128, 256, 3, 1, 1),
nn.ReLU(),
Conv2dResBlock(256, 256),
Conv2dResBlock(256, 256),
Conv2dResBlock(256, 256),
Conv2dResBlock(256, 256),
nn.Conv2d(256, 256, 1, 1, 0)
)
self.relu_2 = nn.ReLU(inplace=True)
self.fc = nn.Linear(1024, 1)
self.image_resolution = image_resolution
def forward(self, I):
o = self.relu(self.conv_theta(I))
o = self.cnn(o)
o = self.fc(self.relu_2(o).view(o.shape[0], 256, -1)).squeeze(-1)
return o
class PartialConvImgEncoder(nn.Module):
'''Adapted from https://github.com/NVIDIA/partialconv/blob/master/models/partialconv2d.py
'''
def __init__(self, channel, image_resolution):
super().__init__()
self.conv1 = PartialConv2d(channel, 256, kernel_size=3, stride=1, padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(256)
self.relu = nn.ReLU(inplace=True)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
self.layer1 = BasicBlock(256, 256)
self.layer2 = BasicBlock(256, 256)
self.layer3 = BasicBlock(256, 256)
self.layer4 = BasicBlock(256, 256)
self.image_resolution = image_resolution
self.channel = channel
self.relu_2 = nn.ReLU(inplace=True)
self.fc = nn.Linear(1024, 1)
for m in self.modules():
if isinstance(m, PartialConv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
elif isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
def forward(self, I):
M_c = I.clone().detach()
M_c = M_c > 0.
M_c = M_c[:,0,...]
M_c = M_c.unsqueeze(1)
M_c = M_c.float()
x = self.conv1(I, M_c)
x = self.bn1(x)
x = self.relu(x)
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.layer4(x)
o = self.fc(x.view(x.shape[0], 256, -1)).squeeze(-1)
return o
class Conv2dResBlock(nn.Module):
def __init__(self, in_channel, out_channel=128):
super().__init__()
self.convs = nn.Sequential(
nn.Conv2d(in_channel, out_channel, 5, 1, 2),
nn.ReLU(),
nn.Conv2d(out_channel, out_channel, 5, 1, 2),
nn.ReLU()
)
self.final_relu = nn.ReLU()
def forward(self, x):
shortcut = x
output = self.convs(x)
output = self.final_relu(output + shortcut)
return output
def channel_last(x):
return x.transpose(1, 2).transpose(2, 3)
class PartialConv2d(nn.Conv2d):
def __init__(self, *args, **kwargs):
if 'multi_channel' in kwargs:
self.multi_channel = kwargs['multi_channel']
kwargs.pop('multi_channel')
else:
self.multi_channel = False
if 'return_mask' in kwargs:
self.return_mask = kwargs['return_mask']
kwargs.pop('return_mask')
else:
self.return_mask = False
super(PartialConv2d, self).__init__(*args, **kwargs)
if self.multi_channel:
self.weight_maskUpdater = torch.ones(self.out_channels, self.in_channels, self.kernel_size[0], self.kernel_size[1])
else:
self.weight_maskUpdater = torch.ones(1, 1, self.kernel_size[0], self.kernel_size[1])
self.slide_winsize = self.weight_maskUpdater.shape[1] * self.weight_maskUpdater.shape[2] * self.weight_maskUpdater.shape[3]
self.last_size = (None, None, None, None)
self.update_mask = None
self.mask_ratio = None
def forward(self, input, mask_in=None):
assert len(input.shape) == 4
if mask_in is not None or self.last_size != tuple(input.shape):
self.last_size = tuple(input.shape)
with torch.no_grad():
if self.weight_maskUpdater.type() != input.type():
self.weight_maskUpdater = self.weight_maskUpdater.to(input)
if mask_in is None:
if self.multi_channel:
mask = torch.ones(input.data.shape[0], input.data.shape[1], input.data.shape[2], input.data.shape[3]).to(input)
else:
mask = torch.ones(1, 1, input.data.shape[2], input.data.shape[3]).to(input)
else:
mask = mask_in
self.update_mask = F.conv2d(mask, self.weight_maskUpdater, bias=None, stride=self.stride, padding=self.padding, dilation=self.dilation, groups=1)
self.mask_ratio = self.slide_winsize / (self.update_mask + 1e-8)
self.update_mask = torch.clamp(self.update_mask, 0, 1)
self.mask_ratio = torch.mul(self.mask_ratio, self.update_mask)
raw_out = super(PartialConv2d, self).forward(torch.mul(input, mask) if mask_in is not None else input)
if self.bias is not None:
bias_view = self.bias.view(1, self.out_channels, 1, 1)
output = torch.mul(raw_out - bias_view, self.mask_ratio) + bias_view
output = torch.mul(output, self.update_mask)
else:
output = torch.mul(raw_out, self.mask_ratio)
if self.return_mask:
return output, self.update_mask
else:
return output
def conv3x3(in_planes, out_planes, stride=1):
return PartialConv2d(in_planes, out_planes, kernel_size=3, stride=stride,
padding=1, bias=False)
class BasicBlock(nn.Module):
expansion = 1
def __init__(self, inplanes, planes, stride=1, downsample=None):
super(BasicBlock, self).__init__()
self.conv1 = conv3x3(inplanes, planes, stride)
self.bn1 = nn.BatchNorm2d(planes)
self.relu = nn.ReLU(inplace=True)
self.conv2 = conv3x3(planes, planes)
self.bn2 = nn.BatchNorm2d(planes)
self.downsample = downsample
self.stride = stride
def forward(self, x):
residual = x
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
out = self.conv2(out)
out = self.bn2(out)
if self.downsample is not None:
residual = self.downsample(x)
out += residual
out = self.relu(out)
return out
def _no_grad_trunc_normal_(tensor, mean, std, a, b):
def norm_cdf(x):
return (1. + math.erf(x / math.sqrt(2.))) / 2.
with torch.no_grad():
l = norm_cdf((a - mean) / std)
u = norm_cdf((b - mean) / std)
tensor.uniform_(2 * l - 1, 2 * u - 1)
tensor.erfinv_()
tensor.mul_(std * math.sqrt(2.))
tensor.add_(mean)
tensor.clamp_(min=a, max=b)
return tensor
def init_weights_trunc_normal(m):
if type(m) == BatchLinear or type(m) == nn.Linear:
if hasattr(m, 'weight'):
fan_in = m.weight.size(1)
fan_out = m.weight.size(0)
std = math.sqrt(2.0 / float(fan_in + fan_out))
mean = 0.
_no_grad_trunc_normal_(m.weight, mean, std, -2 * std, 2 * std)
def init_weights_normal(m):
if type(m) == BatchLinear or type(m) == nn.Linear:
if hasattr(m, 'weight'):
nn.init.kaiming_normal_(m.weight, a=0.0, nonlinearity='relu', mode='fan_in')
def init_weights_selu(m):
if type(m) == BatchLinear or type(m) == nn.Linear:
if hasattr(m, 'weight'):
num_input = m.weight.size(-1)
nn.init.normal_(m.weight, std=1 / math.sqrt(num_input))
def init_weights_elu(m):
if type(m) == BatchLinear or type(m) == nn.Linear:
if hasattr(m, 'weight'):
num_input = m.weight.size(-1)
nn.init.normal_(m.weight, std=math.sqrt(1.5505188080679277) / math.sqrt(num_input))
def init_weights_xavier(m):
if type(m) == BatchLinear or type(m) == nn.Linear:
if hasattr(m, 'weight'):
nn.init.xavier_normal_(m.weight)
def sine_init(m):
with torch.no_grad():
if hasattr(m, 'weight'):
num_input = m.weight.size(-1)
m.weight.uniform_(-np.sqrt(6 / num_input) / 30, np.sqrt(6 / num_input) / 30)
def first_layer_sine_init(m):
with torch.no_grad():
if hasattr(m, 'weight'):
num_input = m.weight.size(-1)
m.weight.uniform_(-1 / num_input, 1 / num_input)
def compl_conj(x):
y = x.clone()
y[..., 1::2] = -1 * y[..., 1::2]
return y
def compl_div(x, y):
''' x / y '''
a = x[..., ::2]
b = x[..., 1::2]
c = y[..., ::2]
d = y[..., 1::2]
outr = (a * c + b * d) / (c ** 2 + d ** 2)
outi = (b * c - a * d) / (c ** 2 + d ** 2)
out = torch.zeros_like(x)
out[..., ::2] = outr
out[..., 1::2] = outi
return out
def compl_mul(x, y):
''' x * y '''
a = x[..., ::2]
b = x[..., 1::2]
c = y[..., ::2]
d = y[..., 1::2]
outr = a * c - b * d
outi = (a + b) * (c + d) - a * c - b * d
out = torch.zeros_like(x)
out[..., ::2] = outr
out[..., 1::2] = outi
return out