-
Notifications
You must be signed in to change notification settings - Fork 0
/
stylegan2.py
729 lines (629 loc) · 34.5 KB
/
stylegan2.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
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
# Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
#
# NVIDIA CORPORATION and its licensors retain all intellectual property
# and proprietary rights in and to this software, related documentation
# and any modifications thereto. Any use, reproduction, disclosure or
# distribution of this software and related documentation without an express
# license agreement from NVIDIA CORPORATION is strictly prohibited.
"""Network architectures from the paper
"Analyzing and Improving the Image Quality of StyleGAN".
Matches the original implementation of configs E-F by Karras et al. at
https://github.com/NVlabs/stylegan2/blob/master/training/networks_stylegan2.py"""
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch_utils import misc
from torch_utils.ops import conv2d_resample
from torch_utils.ops import upfirdn2d
from torch_utils.ops import bias_act
from torch_utils.ops import fma
latent_space = "w+"
#----------------------------------------------------------------------------
def normalize_2nd_moment(x, dim=1, eps=1e-8):
return x * (x.square().mean(dim=dim, keepdim=True) + eps).rsqrt()
#----------------------------------------------------------------------------
def modulated_conv2d(
x, # Input tensor of shape [batch_size, in_channels, in_height, in_width].
weight, # Weight tensor of shape [out_channels, in_channels, kernel_height, kernel_width].
styles, # Modulation coefficients of shape [batch_size, in_channels].
noise = None, # Optional noise tensor to add to the output activations.
up = 1, # Integer upsampling factor.
down = 1, # Integer downsampling factor.
padding = 0, # Padding with respect to the upsampled image.
resample_filter = None, # Low-pass filter to apply when resampling activations. Must be prepared beforehand by calling upfirdn2d.setup_filter().
demodulate = True, # Apply weight demodulation?
flip_weight = True, # False = convolution, True = correlation (matches torch.nn.functional.conv2d).
fused_modconv = True, # Perform modulation, convolution, and demodulation as a single fused operation?
):
batch_size = x.shape[0]
out_channels, in_channels, kh, kw = weight.shape
misc.assert_shape(weight, [out_channels, in_channels, kh, kw]) # [OIkk]
misc.assert_shape(x, [batch_size, in_channels, None, None]) # [NIHW]
misc.assert_shape(styles, [batch_size, in_channels]) # [NI]
# Pre-normalize inputs to avoid FP16 overflow.
if x.dtype == torch.float16 and demodulate:
weight = weight * (1 / np.sqrt(in_channels * kh * kw) / weight.norm(float('inf'), dim=[1,2,3], keepdim=True)) # max_Ikk
styles = styles / styles.norm(float('inf'), dim=1, keepdim=True) # max_I
# Calculate per-sample weights and demodulation coefficients.
w = None
dcoefs = None
if demodulate or fused_modconv:
w = weight.unsqueeze(0) # [NOIkk]
w = w * styles.reshape(batch_size, 1, -1, 1, 1) # [NOIkk]
if demodulate:
dcoefs = (w.square().sum(dim=[2,3,4]) + 1e-8).rsqrt() # [NO]
if demodulate and fused_modconv:
w = w * dcoefs.reshape(batch_size, -1, 1, 1, 1) # [NOIkk]
# Execute by scaling the activations before and after the convolution.
if not fused_modconv:
x = x * styles.to(x.dtype).reshape(batch_size, -1, 1, 1)
x = conv2d_resample.conv2d_resample(x=x, w=weight.to(x.dtype), f=resample_filter, up=up, down=down, padding=padding, flip_weight=flip_weight)
if demodulate and noise is not None:
x = fma.fma(x, dcoefs.to(x.dtype).reshape(batch_size, -1, 1, 1), noise.to(x.dtype))
elif demodulate:
x = x * dcoefs.to(x.dtype).reshape(batch_size, -1, 1, 1)
elif noise is not None:
x = x.add_(noise.to(x.dtype))
return x
# Execute as one fused op using grouped convolution.
with misc.suppress_tracer_warnings(): # this value will be treated as a constant
batch_size = int(batch_size)
misc.assert_shape(x, [batch_size, in_channels, None, None])
x = x.reshape(1, -1, *x.shape[2:])
w = w.reshape(-1, in_channels, kh, kw)
x = conv2d_resample.conv2d_resample(x=x, w=w.to(x.dtype), f=resample_filter, up=up, down=down, padding=padding, groups=batch_size, flip_weight=flip_weight)
x = x.reshape(batch_size, -1, *x.shape[2:])
if noise is not None:
x = x.add_(noise)
return x
#----------------------------------------------------------------------------
class FullyConnectedLayer(torch.nn.Module):
def __init__(self,
in_features, # Number of input features.
out_features, # Number of output features.
bias = True, # Apply additive bias before the activation function?
activation = 'linear', # Activation function: 'relu', 'lrelu', etc.
lr_multiplier = 1, # Learning rate multiplier.
bias_init = 0, # Initial value for the additive bias.
):
super().__init__()
self.in_features = in_features
self.out_features = out_features
self.activation = activation
self.weight = torch.nn.Parameter(torch.randn([out_features, in_features]) / lr_multiplier)
self.bias = torch.nn.Parameter(torch.full([out_features], np.float32(bias_init))) if bias else None
self.weight_gain = lr_multiplier / np.sqrt(in_features)
self.bias_gain = lr_multiplier
def forward(self, x):
w = self.weight.to(x.dtype) * self.weight_gain
b = self.bias
if b is not None:
b = b.to(x.dtype)
if self.bias_gain != 1:
b = b * self.bias_gain
if self.activation == 'linear' and b is not None:
x = torch.addmm(b.unsqueeze(0), x, w.t())
else:
x = x.matmul(w.t())
x = bias_act.bias_act(x, b, act=self.activation)
return x
def extra_repr(self):
return f'in_features={self.in_features:d}, out_features={self.out_features:d}, activation={self.activation:s}'
#----------------------------------------------------------------------------
class Conv2dLayer(torch.nn.Module):
def __init__(self,
in_channels, # Number of input channels.
out_channels, # Number of output channels.
kernel_size, # Width and height of the convolution kernel.
bias = True, # Apply additive bias before the activation function?
activation = 'linear', # Activation function: 'relu', 'lrelu', etc.
up = 1, # Integer upsampling factor.
down = 1, # Integer downsampling factor.
resample_filter = [1,3,3,1], # Low-pass filter to apply when resampling activations.
conv_clamp = None, # Clamp the output to +-X, None = disable clamping.
channels_last = False, # Expect the input to have memory_format=channels_last?
trainable = True, # Update the weights of this layer during training?
):
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.activation = activation
self.up = up
self.down = down
self.conv_clamp = conv_clamp
self.register_buffer('resample_filter', upfirdn2d.setup_filter(resample_filter))
self.padding = kernel_size // 2
self.weight_gain = 1 / np.sqrt(in_channels * (kernel_size ** 2))
self.act_gain = bias_act.activation_funcs[activation].def_gain
memory_format = torch.channels_last if channels_last else torch.contiguous_format
weight = torch.randn([out_channels, in_channels, kernel_size, kernel_size]).to(memory_format=memory_format)
bias = torch.zeros([out_channels]) if bias else None
if trainable:
self.weight = torch.nn.Parameter(weight)
self.bias = torch.nn.Parameter(bias) if bias is not None else None
else:
self.register_buffer('weight', weight)
if bias is not None:
self.register_buffer('bias', bias)
else:
self.bias = None
def forward(self, x, gain=1):
w = self.weight * self.weight_gain
b = self.bias.to(x.dtype) if self.bias is not None else None
flip_weight = (self.up == 1) # slightly faster
x = conv2d_resample.conv2d_resample(x=x, w=w.to(x.dtype), f=self.resample_filter, up=self.up, down=self.down, padding=self.padding, flip_weight=flip_weight)
act_gain = self.act_gain * gain
act_clamp = self.conv_clamp * gain if self.conv_clamp is not None else None
x = bias_act.bias_act(x, b, act=self.activation, gain=act_gain, clamp=act_clamp)
return x
def extra_repr(self):
return ' '.join([
f'in_channels={self.in_channels:d}, out_channels={self.out_channels:d}, activation={self.activation:s},',
f'up={self.up}, down={self.down}'])
#----------------------------------------------------------------------------
class MappingNetwork(torch.nn.Module):
def __init__(self,
z_dim, # Input latent (Z) dimensionality, 0 = no latent.
c_dim, # Conditioning label (C) dimensionality, 0 = no label.
w_dim, # Intermediate latent (W) dimensionality.
num_ws, # Number of intermediate latents to output, None = do not broadcast.
num_layers = 8, # Number of mapping layers.
embed_features = None, # Label embedding dimensionality, None = same as w_dim.
layer_features = None, # Number of intermediate features in the mapping layers, None = same as w_dim.
activation = 'lrelu', # Activation function: 'relu', 'lrelu', etc.
lr_multiplier = 0.01, # Learning rate multiplier for the mapping layers.
w_avg_beta = 0.998, # Decay for tracking the moving average of W during training, None = do not track.
):
super().__init__()
self.z_dim = z_dim
self.c_dim = c_dim
self.w_dim = w_dim
self.num_ws = num_ws
self.num_layers = num_layers
self.w_avg_beta = w_avg_beta
if embed_features is None:
embed_features = w_dim
if c_dim == 0:
embed_features = 0
if layer_features is None:
layer_features = w_dim
features_list = [z_dim + embed_features] + [layer_features] * (num_layers - 1) + [w_dim]
if c_dim > 0:
self.embed = FullyConnectedLayer(c_dim, embed_features)
for idx in range(num_layers):
in_features = features_list[idx]
out_features = features_list[idx + 1]
layer = FullyConnectedLayer(in_features, out_features, activation=activation, lr_multiplier=lr_multiplier)
setattr(self, f'fc{idx}', layer)
if num_ws is not None and w_avg_beta is not None:
self.register_buffer('w_avg', torch.zeros([w_dim]))
def forward(self, z, c, truncation_psi=1, truncation_cutoff=None, update_emas=False, repeat_w = False):
# Embed, normalize, and concat inputs.
x = None
with torch.autograd.profiler.record_function('input'):
if self.z_dim > 0:
misc.assert_shape(z, [None, self.z_dim])
x = normalize_2nd_moment(z.to(torch.float32))
if self.c_dim > 0:
misc.assert_shape(c, [None, self.c_dim])
y = normalize_2nd_moment(self.embed(c.to(torch.float32)))
x = torch.cat([x, y], dim=1) if x is not None else y
# Main layers.
for idx in range(self.num_layers):
layer = getattr(self, f'fc{idx}')
x = layer(x)
# Update moving average of W.
if update_emas and self.w_avg_beta is not None:
with torch.autograd.profiler.record_function('update_w_avg'):
self.w_avg.copy_(x.detach().mean(dim=0).lerp(self.w_avg, self.w_avg_beta))
# Broadcast.
#if self.num_ws is not None:
if repeat_w:
with torch.autograd.profiler.record_function('broadcast'):
x = x.unsqueeze(1).repeat([1, self.num_ws, 1])
# Apply truncation.
if truncation_psi != 1:
with torch.autograd.profiler.record_function('truncate'):
assert self.w_avg_beta is not None
if self.num_ws is None or truncation_cutoff is None:
x = self.w_avg.lerp(x, truncation_psi)
else:
x[:, :truncation_cutoff] = self.w_avg.lerp(x[:, :truncation_cutoff], truncation_psi)
return x
def extra_repr(self):
return f'z_dim={self.z_dim:d}, c_dim={self.c_dim:d}, w_dim={self.w_dim:d}, num_ws={self.num_ws:d}'
#----------------------------------------------------------------------------
class SynthesisLayer(torch.nn.Module):
def __init__(self,
in_channels, # Number of input channels.
out_channels, # Number of output channels.
w_dim, # Intermediate latent (W) dimensionality.
resolution, # Resolution of this layer.
kernel_size = 3, # Convolution kernel size.
up = 1, # Integer upsampling factor.
use_noise = True, # Enable noise input?
activation = 'lrelu', # Activation function: 'relu', 'lrelu', etc.
resample_filter = [1,3,3,1], # Low-pass filter to apply when resampling activations.
conv_clamp = None, # Clamp the output of convolution layers to +-X, None = disable clamping.
channels_last = False, # Use channels_last format for the weights?
):
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.w_dim = w_dim
self.resolution = resolution
self.up = up
self.use_noise = use_noise
self.activation = activation
self.conv_clamp = conv_clamp
self.register_buffer('resample_filter', upfirdn2d.setup_filter(resample_filter))
self.padding = kernel_size // 2
self.act_gain = bias_act.activation_funcs[activation].def_gain
self.affine = FullyConnectedLayer(w_dim, in_channels, bias_init=1)
memory_format = torch.channels_last if channels_last else torch.contiguous_format
self.weight = torch.nn.Parameter(torch.randn([out_channels, in_channels, kernel_size, kernel_size]).to(memory_format=memory_format))
if use_noise:
self.register_buffer('noise_const', torch.randn([resolution, resolution]))
self.noise_strength = torch.nn.Parameter(torch.zeros([]))
self.bias = torch.nn.Parameter(torch.zeros([out_channels]))
def forward(self, x, w, noise_mode='random',fused_modconv=True, gain=1, return_styles=False):
global latent_space
if latent_space == "w+":
styles = self.affine(w)
if return_styles:
return styles
else:
styles = w
assert noise_mode in ['random', 'const', 'none']
in_resolution = self.resolution // self.up
misc.assert_shape(x, [None, self.in_channels, in_resolution, in_resolution])
noise = None
if self.use_noise and noise_mode == 'random':
noise = torch.randn([x.shape[0], 1, self.resolution, self.resolution], device=x.device) * self.noise_strength
if self.use_noise and noise_mode == 'const':
noise = self.noise_const * self.noise_strength
flip_weight = (self.up == 1) # slightly faster
x = modulated_conv2d(x=x, weight=self.weight, styles=styles, noise=noise, up=self.up,
padding=self.padding, resample_filter=self.resample_filter, flip_weight=flip_weight, fused_modconv=fused_modconv)
act_gain = self.act_gain * gain
act_clamp = self.conv_clamp * gain if self.conv_clamp is not None else None
x = bias_act.bias_act(x, self.bias.to(x.dtype), act=self.activation, gain=act_gain, clamp=act_clamp)
return x
def extra_repr(self):
return ' '.join([
f'in_channels={self.in_channels:d}, out_channels={self.out_channels:d}, w_dim={self.w_dim:d},',
f'resolution={self.resolution:d}, up={self.up}, activation={self.activation:s}'])
#----------------------------------------------------------------------------
class ToRGBLayer(torch.nn.Module):
def __init__(self, in_channels, out_channels, w_dim, kernel_size=1, conv_clamp=None, channels_last=False):
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.w_dim = w_dim
self.conv_clamp = conv_clamp
self.affine = FullyConnectedLayer(w_dim, in_channels, bias_init=1)
memory_format = torch.channels_last if channels_last else torch.contiguous_format
self.weight = torch.nn.Parameter(torch.randn([out_channels, in_channels, kernel_size, kernel_size]).to(memory_format=memory_format))
self.bias = torch.nn.Parameter(torch.zeros([out_channels]))
self.weight_gain = 1 / np.sqrt(in_channels * (kernel_size ** 2))
def forward(self, x, w, fused_modconv=True, return_styles=False):
if latent_space == "w+":
styles = self.affine(w) * self.weight_gain
if return_styles:
return styles
else:
styles = w
x = modulated_conv2d(x=x, weight=self.weight, styles=styles, demodulate=False, fused_modconv=fused_modconv)
x = bias_act.bias_act(x, self.bias.to(x.dtype), clamp=self.conv_clamp)
return x
def extra_repr(self):
return f'in_channels={self.in_channels:d}, out_channels={self.out_channels:d}, w_dim={self.w_dim:d}'
#----------------------------------------------------------------------------
class ResLayers(nn.Module):
def __init__(self, in_channels, out_channels, stride=1):
super().__init__()
if (in_channels == out_channels) and stride==1:
self.shortcut_layer = nn.Identity()
else:
self.shortcut_layer = nn.Sequential(
nn.Conv2d(in_channels, out_channels, (1, 1), stride, bias=False))
self.res_layer = nn.Sequential(
nn.Conv2d(in_channels, out_channels, (3, 3), (1, 1), 1, bias=True), nn.LeakyReLU(0.2),
nn.Conv2d(out_channels, out_channels, (3, 3), stride, 1, bias=True) )
def forward(self, x):
shortcut = self.shortcut_layer(x)
res = self.res_layer(x)
return res + shortcut
class FeatureEdit(nn.Module):
def __init__(self, in_channels, out_channels):
super().__init__()
self.convs = nn.ModuleList()
iter_num = in_channels // out_channels
i_c = in_channels
for i in range(iter_num-1):
out_c = i_c - out_channels
self.convs.append( ResLayers(i_c,out_c,1) )
i_c = out_c
def forward(self, diff):
for block in self.convs:
diff = block(diff)
return diff
class FeatureAlignment(nn.Module):
def __init__(self, in_channels, out_channels):
super().__init__()
t_channel = 512
self.first_layer = nn.Conv2d(in_channels, t_channel, kernel_size=1, padding=0, bias=True)
self.conv1 = nn.Sequential(*[ResLayers(t_channel,t_channel,1)])
self.conv2 = nn.Sequential(*[ResLayers(t_channel,t_channel,2), ResLayers(t_channel,t_channel,1)])
self.conv3 = nn.Sequential(*[ResLayers(t_channel,t_channel,2), ResLayers(t_channel,t_channel,1)])
self.dconv1 = nn.Sequential(*[ResLayers(t_channel,t_channel,1), ResLayers(t_channel,t_channel,1)])
self.dconv2 = nn.Sequential(*[ResLayers(t_channel,t_channel,1), ResLayers(t_channel,t_channel,1)])
self.dconv3 = nn.Sequential(*[ResLayers(t_channel,t_channel,1), ResLayers(t_channel,t_channel,1)])
self.out_layer = nn.Conv2d(t_channel, out_channels, kernel_size=1, padding=0, bias=True)
def forward(self, encoder_feats, generator_feats):
x = torch.cat((encoder_feats,generator_feats), dim=1)
x = self.first_layer(x)
f1 = self.conv1(x)
f2 = self.conv2(f1)
f3 = self.conv3(f2)
shape = f3.shape[-1]
df1 = F.interpolate(f3, size=(shape*2,shape*2) , mode='bilinear', align_corners=True)
df2 = self.dconv1(df1 + f2)
df2 = F.interpolate(df2, size=(shape*4,shape*4) , mode='bilinear', align_corners=True)
df3 = self.dconv2(df2 + f1)
aligned_feats = self.out_layer(df3)
return aligned_feats
class FeatureExtraction(nn.Module):
def __init__(self,in_channels, out_channels ):
super().__init__()
t_channel = 512
self.first_layer = nn.Conv2d(in_channels, t_channel, kernel_size=1, padding=0, bias=True)
self.convs = nn.Sequential(*[ResLayers(t_channel,t_channel,1), ResLayers(t_channel,out_channels,1), ResLayers(out_channels,out_channels,1) ])
def forward(self, aligned_feats):
y = self.first_layer(aligned_feats)
y = self.convs(y)
return y
class GateNetwork(nn.Module):
def __init__(self, in_channels, out_channels):
super().__init__()
t_channel = 256
self.down1 = nn.Conv2d(in_channels, t_channel, kernel_size=3, padding=1, bias=True)
self.down2 = nn.Conv2d(in_channels, t_channel, kernel_size=3, padding=1, bias=True)
self.sigmoid = nn.Sigmoid()
self.convs = nn.Sequential(*[ResLayers(in_channels,in_channels,1), ResLayers(in_channels,out_channels,1), ResLayers(out_channels,out_channels,1) ])
self.convs2 = nn.Sequential(*[ResLayers(in_channels,in_channels,1), ResLayers(in_channels,out_channels,1), ResLayers(out_channels,1,1) ])
def forward(self, generator_feats, y):
generator_feats = self.down1(generator_feats)
y = self.down2(y)
x = torch.cat((generator_feats, y), dim=1)
deltaF = self.convs(x)
gate = self.convs2(x)
gate = self.sigmoid(gate)
return deltaF, gate
g_e_concat_shape={64: 640, 32:768}
e_shape = {64: 128, 32:256}
class SynthesisBlock(torch.nn.Module):
def __init__(self,
in_channels, # Number of input channels, 0 = first block.
out_channels, # Number of output channels.
w_dim, # Intermediate latent (W) dimensionality.
resolution, # Resolution of this block.
img_channels, # Number of output color channels.
is_last, # Is this the last block?
architecture = 'skip', # Architecture: 'orig', 'skip', 'resnet'.
resample_filter = [1,3,3,1], # Low-pass filter to apply when resampling activations.
conv_clamp = 256, # Clamp the output of convolution layers to +-X, None = disable clamping.
use_fp16 = False, # Use FP16 for this block?
fp16_channels_last = False, # Use channels-last memory format with FP16?
fused_modconv_default = True, # Default value of fused_modconv. 'inference_only' = True for inference, False for training.
embed_res = 64, # Which resolution we embed the images
**layer_kwargs, # Arguments for SynthesisLayer.
):
assert architecture in ['orig', 'skip', 'resnet']
super().__init__()
self.in_channels = in_channels
self.w_dim = w_dim
self.resolution = resolution
self.img_channels = img_channels
self.is_last = is_last
self.architecture = architecture
self.use_fp16 = use_fp16
self.channels_last = (use_fp16 and fp16_channels_last)
self.fused_modconv_default = fused_modconv_default
self.register_buffer('resample_filter', upfirdn2d.setup_filter(resample_filter))
self.num_conv = 0
self.num_torgb = 0
if in_channels == 0:
self.const = torch.nn.Parameter(torch.randn([out_channels, resolution, resolution]))
if in_channels != 0:
self.conv0 = SynthesisLayer(in_channels, out_channels, w_dim=w_dim, resolution=resolution, up=2,
resample_filter=resample_filter, conv_clamp=conv_clamp, channels_last=self.channels_last, **layer_kwargs)
self.num_conv += 1
self.conv1 = SynthesisLayer(out_channels, out_channels, w_dim=w_dim, resolution=resolution,
conv_clamp=conv_clamp, channels_last=self.channels_last, **layer_kwargs)
self.num_conv += 1
if is_last or architecture == 'skip':
self.torgb = ToRGBLayer(out_channels, img_channels, w_dim=w_dim,
conv_clamp=conv_clamp, channels_last=self.channels_last)
self.num_torgb += 1
if in_channels != 0 and architecture == 'resnet':
self.skip = Conv2dLayer(in_channels, out_channels, kernel_size=1, bias=False, up=2,
resample_filter=resample_filter, channels_last=self.channels_last)
if resolution == embed_res:
in_c = g_e_concat_shape.get(embed_res)
# E1
self.modify_feature_alignment = FeatureAlignment(in_channels=in_c, out_channels=512)
self.modify_feature_extraction = FeatureExtraction(in_channels=512, out_channels=512)
# E2
self.modify_feature_gates = GateNetwork(in_channels=512, out_channels=512)
self.embed_res = embed_res
def forward(self, x, img, ws, highres_outs=None, return_f = False, return_styles=False,
force_fp32=False, fused_modconv=None, update_emas=False, **layer_kwargs):
_ = update_emas # unused
global latent_space
if latent_space == "w+":
misc.assert_shape(ws, [None, self.num_conv + self.num_torgb, self.w_dim])
w_iter = iter(ws.unbind(dim=1))
if ws.device.type != 'cuda':
force_fp32 = True
else:
w_iter = iter(ws)
dtype = torch.float16 if self.use_fp16 and not force_fp32 else torch.float32
memory_format = torch.channels_last if self.channels_last and not force_fp32 else torch.contiguous_format
if fused_modconv is None:
fused_modconv = self.fused_modconv_default
if fused_modconv == 'inference_only':
fused_modconv = (not self.training)
# Input.
if self.in_channels == 0:
batch_size = ws.shape[0] if latent_space == 'w+' else ws[0].shape[0]
x = self.const.to(dtype=dtype, memory_format=memory_format)
x = x.unsqueeze(0).repeat([batch_size, 1, 1, 1])
else:
if not return_styles:
misc.assert_shape(x, [None, self.in_channels, self.resolution // 2, self.resolution // 2])
x = x.to(dtype=dtype, memory_format=memory_format)
gouts = {}
style_outs = []
# Main layers.
if self.in_channels == 0:
x = self.conv1(x, next(w_iter), fused_modconv=fused_modconv, return_styles=return_styles, **layer_kwargs)
if return_styles: style_outs.append(x)
elif self.architecture == 'resnet':
y = self.skip(x, gain=np.sqrt(0.5))
x = self.conv0(x, next(w_iter), fused_modconv=fused_modconv, **layer_kwargs)
x = self.conv1(x, next(w_iter), fused_modconv=fused_modconv, gain=np.sqrt(0.5), **layer_kwargs)
x = y.add_(x)
else:
x = self.conv0(x, next(w_iter), fused_modconv=fused_modconv, return_styles=return_styles, **layer_kwargs)
if return_styles: style_outs.append(x)
if x.shape[-1] == self.embed_res and return_f:
return x, None
#HighResFeat Generator Modification
if x.shape[-1] == self.embed_res and highres_outs is not None:
# Concat Skip Connections with Generator F features. (E1)
aligned_feats = self.modify_feature_alignment(highres_outs[f'{self.embed_res}x{self.embed_res}'], highres_outs['F_feats'])
aligned_feats = self.modify_feature_extraction(aligned_feats)
# Adopt to random edits (E2)
deltaF, gate = self.modify_feature_gates(x, aligned_feats)
x = (x * (1-gate) ) + ( (x + deltaF) * gate )
x = self.conv1(x, next(w_iter), fused_modconv=fused_modconv, return_styles=return_styles, **layer_kwargs)
if return_styles: style_outs.append(x)
# ToRGB.
if img is not None:
misc.assert_shape(img, [None, self.img_channels, self.resolution // 2, self.resolution // 2])
img = upfirdn2d.upsample2d(img, self.resample_filter)
if self.is_last or self.architecture == 'skip':
y = self.torgb(x, next(w_iter), fused_modconv=fused_modconv, return_styles=return_styles)
if return_styles:
style_outs.append(y)
return style_outs, None
y = y.to(dtype=torch.float32, memory_format=torch.contiguous_format)
img = img.add_(y) if img is not None else y
assert x.dtype == dtype
assert img is None or img.dtype == torch.float32
return x, img
def extra_repr(self):
return f'resolution={self.resolution:d}, architecture={self.architecture:s}'
#----------------------------------------------------------------------------
class SynthesisNetwork(torch.nn.Module):
def __init__(self,
w_dim, # Intermediate latent (W) dimensionality.
img_resolution, # Output image resolution.
img_channels, # Number of color channels.
channel_base = 32768, # Overall multiplier for the number of channels.
channel_max = 512, # Maximum number of channels in any layer.
num_fp16_res = 4, # Use FP16 for the N highest resolutions.
**block_kwargs, # Arguments for SynthesisBlock.
):
assert img_resolution >= 4 and img_resolution & (img_resolution - 1) == 0
super().__init__()
self.w_dim = w_dim
self.img_resolution = img_resolution
self.img_resolution_log2 = int(np.log2(img_resolution))
self.img_channels = img_channels
self.num_fp16_res = num_fp16_res
self.block_resolutions = [2 ** i for i in range(2, self.img_resolution_log2 + 1)]
channels_dict = {res: min(channel_base // res, channel_max) for res in self.block_resolutions}
fp16_resolution = max(2 ** (self.img_resolution_log2 + 1 - num_fp16_res), 8)
self.num_ws = 0
for res in self.block_resolutions:
in_channels = channels_dict[res // 2] if res > 4 else 0
out_channels = channels_dict[res]
use_fp16 = (res >= fp16_resolution)
is_last = (res == self.img_resolution)
block = SynthesisBlock(in_channels, out_channels, w_dim=w_dim, resolution=res,
img_channels=img_channels, is_last=is_last, use_fp16=use_fp16, **block_kwargs)
self.num_ws += block.num_conv
if is_last:
self.num_ws += block.num_torgb
setattr(self, f'b{res}', block)
def forward(self, ws, highres_outs=None, return_f = False, return_styles=False, **block_kwargs):
block_ws = []
global latent_space
with torch.autograd.profiler.record_function('split_ws'):
if latent_space == 'w+':
misc.assert_shape(ws, [None, self.num_ws, self.w_dim])
ws = ws.to(torch.float32)
w_idx = 0
s_idx = 0
for res in self.block_resolutions:
block = getattr(self, f'b{res}')
if latent_space == 'w+':
block_ws.append(ws.narrow(1, w_idx, block.num_conv + block.num_torgb))
else:
block_ws.append(ws[s_idx: s_idx + block.num_conv + block.num_torgb])
w_idx += block.num_conv
s_idx += block.num_conv + block.num_torgb
x = img = None
styles = []
conv_idx = 0
for res, cur_ws in zip(self.block_resolutions, block_ws):
block = getattr(self, f'b{res}')
x, img = block(x, img, cur_ws, highres_outs, return_f, return_styles, **block_kwargs)
if return_f and img is None:
return x
if return_styles:
styles.extend(x)
conv_idx += block.num_conv
if return_styles: return styles
return img
def extra_repr(self):
return ' '.join([
f'w_dim={self.w_dim:d}, num_ws={self.num_ws:d},',
f'img_resolution={self.img_resolution:d}, img_channels={self.img_channels:d},',
f'num_fp16_res={self.num_fp16_res:d}'])
#----------------------------------------------------------------------------
class Generator(torch.nn.Module):
def __init__(self,
z_dim, # Input latent (Z) dimensionality.
c_dim, # Conditioning label (C) dimensionality.
w_dim, # Intermediate latent (W) dimensionality.
resolution, # Output resolution.
img_channels, # Number of output color channels.
mapping_kwargs = {}, # Arguments for MappingNetwork.
**synthesis_kwargs, # Arguments for SynthesisNetwork.
):
super().__init__()
self.z_dim = z_dim
self.c_dim = c_dim
self.w_dim = w_dim
self.resolution = resolution
self.img_channels = img_channels
self.synthesis = SynthesisNetwork(w_dim=w_dim, img_resolution=resolution, img_channels=img_channels, **synthesis_kwargs)
self.num_ws = self.synthesis.num_ws
self.mapping = MappingNetwork(z_dim=z_dim, c_dim=c_dim, w_dim=w_dim, num_ws=self.num_ws, **mapping_kwargs)
def forward(self, lat, c, truncation_psi=1, truncation_cutoff=None, update_emas=False, mode='synthesis', return_f = False,
return_styles=False, **synthesis_kwargs):
global latent_space
if type(lat) is list:
latent_space = 's'
else:
latent_space = 'w+'
if mode == 'mapping':
ws = self.mapping(lat, c, truncation_psi=truncation_psi, truncation_cutoff=truncation_cutoff, update_emas=update_emas)
return ws
if mode == 'synthesis':
img = self.synthesis(lat, highres_outs = c, return_f=return_f, update_emas=False, return_styles=return_styles, **synthesis_kwargs)
return img