Spaces:
Sleeping
Sleeping
Update videoretalking/models/base_blocks.py
Browse files- videoretalking/models/base_blocks.py +553 -553
videoretalking/models/base_blocks.py
CHANGED
@@ -1,554 +1,554 @@
|
|
1 |
-
import math
|
2 |
-
import torch
|
3 |
-
import torch.nn as nn
|
4 |
-
import torch.nn.functional as F
|
5 |
-
from torch.nn.modules.batchnorm import BatchNorm2d
|
6 |
-
from torch.nn.utils.spectral_norm import spectral_norm as SpectralNorm
|
7 |
-
|
8 |
-
from models.ffc import FFC
|
9 |
-
from basicsr.archs.arch_util import default_init_weights
|
10 |
-
|
11 |
-
|
12 |
-
class Conv2d(nn.Module):
|
13 |
-
def __init__(self, cin, cout, kernel_size, stride, padding, residual=False, *args, **kwargs):
|
14 |
-
super().__init__(*args, **kwargs)
|
15 |
-
self.conv_block = nn.Sequential(
|
16 |
-
nn.Conv2d(cin, cout, kernel_size, stride, padding),
|
17 |
-
nn.BatchNorm2d(cout)
|
18 |
-
)
|
19 |
-
self.act = nn.ReLU()
|
20 |
-
self.residual = residual
|
21 |
-
|
22 |
-
def forward(self, x):
|
23 |
-
out = self.conv_block(x)
|
24 |
-
if self.residual:
|
25 |
-
out += x
|
26 |
-
return self.act(out)
|
27 |
-
|
28 |
-
|
29 |
-
class ResBlock(nn.Module):
|
30 |
-
def __init__(self, in_channels, out_channels, mode='down'):
|
31 |
-
super(ResBlock, self).__init__()
|
32 |
-
self.conv1 = nn.Conv2d(in_channels, in_channels, 3, 1, 1)
|
33 |
-
self.conv2 = nn.Conv2d(in_channels, out_channels, 3, 1, 1)
|
34 |
-
self.skip = nn.Conv2d(in_channels, out_channels, 1, bias=False)
|
35 |
-
if mode == 'down':
|
36 |
-
self.scale_factor = 0.5
|
37 |
-
elif mode == 'up':
|
38 |
-
self.scale_factor = 2
|
39 |
-
|
40 |
-
def forward(self, x):
|
41 |
-
out = F.leaky_relu_(self.conv1(x), negative_slope=0.2)
|
42 |
-
# upsample/downsample
|
43 |
-
out = F.interpolate(out, scale_factor=self.scale_factor, mode='bilinear', align_corners=False)
|
44 |
-
out = F.leaky_relu_(self.conv2(out), negative_slope=0.2)
|
45 |
-
# skip
|
46 |
-
x = F.interpolate(x, scale_factor=self.scale_factor, mode='bilinear', align_corners=False)
|
47 |
-
skip = self.skip(x)
|
48 |
-
out = out + skip
|
49 |
-
return out
|
50 |
-
|
51 |
-
|
52 |
-
class LayerNorm2d(nn.Module):
|
53 |
-
def __init__(self, n_out, affine=True):
|
54 |
-
super(LayerNorm2d, self).__init__()
|
55 |
-
self.n_out = n_out
|
56 |
-
self.affine = affine
|
57 |
-
|
58 |
-
if self.affine:
|
59 |
-
self.weight = nn.Parameter(torch.ones(n_out, 1, 1))
|
60 |
-
self.bias = nn.Parameter(torch.zeros(n_out, 1, 1))
|
61 |
-
|
62 |
-
def forward(self, x):
|
63 |
-
normalized_shape = x.size()[1:]
|
64 |
-
if self.affine:
|
65 |
-
return F.layer_norm(x, normalized_shape, \
|
66 |
-
self.weight.expand(normalized_shape),
|
67 |
-
self.bias.expand(normalized_shape))
|
68 |
-
else:
|
69 |
-
return F.layer_norm(x, normalized_shape)
|
70 |
-
|
71 |
-
|
72 |
-
def spectral_norm(module, use_spect=True):
|
73 |
-
if use_spect:
|
74 |
-
return SpectralNorm(module)
|
75 |
-
else:
|
76 |
-
return module
|
77 |
-
|
78 |
-
|
79 |
-
class FirstBlock2d(nn.Module):
|
80 |
-
def __init__(self, input_nc, output_nc, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
|
81 |
-
super(FirstBlock2d, self).__init__()
|
82 |
-
kwargs = {'kernel_size': 7, 'stride': 1, 'padding': 3}
|
83 |
-
conv = spectral_norm(nn.Conv2d(input_nc, output_nc, **kwargs), use_spect)
|
84 |
-
|
85 |
-
if type(norm_layer) == type(None):
|
86 |
-
self.model = nn.Sequential(conv, nonlinearity)
|
87 |
-
else:
|
88 |
-
self.model = nn.Sequential(conv, norm_layer(output_nc), nonlinearity)
|
89 |
-
|
90 |
-
def forward(self, x):
|
91 |
-
out = self.model(x)
|
92 |
-
return out
|
93 |
-
|
94 |
-
|
95 |
-
class DownBlock2d(nn.Module):
|
96 |
-
def __init__(self, input_nc, output_nc, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
|
97 |
-
super(DownBlock2d, self).__init__()
|
98 |
-
kwargs = {'kernel_size': 3, 'stride': 1, 'padding': 1}
|
99 |
-
conv = spectral_norm(nn.Conv2d(input_nc, output_nc, **kwargs), use_spect)
|
100 |
-
pool = nn.AvgPool2d(kernel_size=(2, 2))
|
101 |
-
|
102 |
-
if type(norm_layer) == type(None):
|
103 |
-
self.model = nn.Sequential(conv, nonlinearity, pool)
|
104 |
-
else:
|
105 |
-
self.model = nn.Sequential(conv, norm_layer(output_nc), nonlinearity, pool)
|
106 |
-
|
107 |
-
def forward(self, x):
|
108 |
-
out = self.model(x)
|
109 |
-
return out
|
110 |
-
|
111 |
-
|
112 |
-
class UpBlock2d(nn.Module):
|
113 |
-
def __init__(self, input_nc, output_nc, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
|
114 |
-
super(UpBlock2d, self).__init__()
|
115 |
-
kwargs = {'kernel_size': 3, 'stride': 1, 'padding': 1}
|
116 |
-
conv = spectral_norm(nn.Conv2d(input_nc, output_nc, **kwargs), use_spect)
|
117 |
-
if type(norm_layer) == type(None):
|
118 |
-
self.model = nn.Sequential(conv, nonlinearity)
|
119 |
-
else:
|
120 |
-
self.model = nn.Sequential(conv, norm_layer(output_nc), nonlinearity)
|
121 |
-
|
122 |
-
def forward(self, x):
|
123 |
-
out = self.model(F.interpolate(x, scale_factor=2))
|
124 |
-
return out
|
125 |
-
|
126 |
-
|
127 |
-
class ADAIN(nn.Module):
|
128 |
-
def __init__(self, norm_nc, feature_nc):
|
129 |
-
super().__init__()
|
130 |
-
|
131 |
-
self.param_free_norm = nn.InstanceNorm2d(norm_nc, affine=False)
|
132 |
-
|
133 |
-
nhidden = 128
|
134 |
-
use_bias=True
|
135 |
-
|
136 |
-
self.mlp_shared = nn.Sequential(
|
137 |
-
nn.Linear(feature_nc, nhidden, bias=use_bias),
|
138 |
-
nn.ReLU()
|
139 |
-
)
|
140 |
-
self.mlp_gamma = nn.Linear(nhidden, norm_nc, bias=use_bias)
|
141 |
-
self.mlp_beta = nn.Linear(nhidden, norm_nc, bias=use_bias)
|
142 |
-
|
143 |
-
def forward(self, x, feature):
|
144 |
-
|
145 |
-
# Part 1. generate parameter-free normalized activations
|
146 |
-
normalized = self.param_free_norm(x)
|
147 |
-
# Part 2. produce scaling and bias conditioned on feature
|
148 |
-
feature = feature.view(feature.size(0), -1)
|
149 |
-
actv = self.mlp_shared(feature)
|
150 |
-
gamma = self.mlp_gamma(actv)
|
151 |
-
beta = self.mlp_beta(actv)
|
152 |
-
|
153 |
-
# apply scale and bias
|
154 |
-
gamma = gamma.view(*gamma.size()[:2], 1,1)
|
155 |
-
beta = beta.view(*beta.size()[:2], 1,1)
|
156 |
-
out = normalized * (1 + gamma) + beta
|
157 |
-
return out
|
158 |
-
|
159 |
-
|
160 |
-
class FineADAINResBlock2d(nn.Module):
|
161 |
-
"""
|
162 |
-
Define an Residual block for different types
|
163 |
-
"""
|
164 |
-
def __init__(self, input_nc, feature_nc, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
|
165 |
-
super(FineADAINResBlock2d, self).__init__()
|
166 |
-
kwargs = {'kernel_size': 3, 'stride': 1, 'padding': 1}
|
167 |
-
self.conv1 = spectral_norm(nn.Conv2d(input_nc, input_nc, **kwargs), use_spect)
|
168 |
-
self.conv2 = spectral_norm(nn.Conv2d(input_nc, input_nc, **kwargs), use_spect)
|
169 |
-
self.norm1 = ADAIN(input_nc, feature_nc)
|
170 |
-
self.norm2 = ADAIN(input_nc, feature_nc)
|
171 |
-
self.actvn = nonlinearity
|
172 |
-
|
173 |
-
def forward(self, x, z):
|
174 |
-
dx = self.actvn(self.norm1(self.conv1(x), z))
|
175 |
-
dx = self.norm2(self.conv2(x), z)
|
176 |
-
out = dx + x
|
177 |
-
return out
|
178 |
-
|
179 |
-
|
180 |
-
class FineADAINResBlocks(nn.Module):
|
181 |
-
def __init__(self, num_block, input_nc, feature_nc, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
|
182 |
-
super(FineADAINResBlocks, self).__init__()
|
183 |
-
self.num_block = num_block
|
184 |
-
for i in range(num_block):
|
185 |
-
model = FineADAINResBlock2d(input_nc, feature_nc, norm_layer, nonlinearity, use_spect)
|
186 |
-
setattr(self, 'res'+str(i), model)
|
187 |
-
|
188 |
-
def forward(self, x, z):
|
189 |
-
for i in range(self.num_block):
|
190 |
-
model = getattr(self, 'res'+str(i))
|
191 |
-
x = model(x, z)
|
192 |
-
return x
|
193 |
-
|
194 |
-
|
195 |
-
class ADAINEncoderBlock(nn.Module):
|
196 |
-
def __init__(self, input_nc, output_nc, feature_nc, nonlinearity=nn.LeakyReLU(), use_spect=False):
|
197 |
-
super(ADAINEncoderBlock, self).__init__()
|
198 |
-
kwargs_down = {'kernel_size': 4, 'stride': 2, 'padding': 1}
|
199 |
-
kwargs_fine = {'kernel_size': 3, 'stride': 1, 'padding': 1}
|
200 |
-
|
201 |
-
self.conv_0 = spectral_norm(nn.Conv2d(input_nc, output_nc, **kwargs_down), use_spect)
|
202 |
-
self.conv_1 = spectral_norm(nn.Conv2d(output_nc, output_nc, **kwargs_fine), use_spect)
|
203 |
-
|
204 |
-
|
205 |
-
self.norm_0 = ADAIN(input_nc, feature_nc)
|
206 |
-
self.norm_1 = ADAIN(output_nc, feature_nc)
|
207 |
-
self.actvn = nonlinearity
|
208 |
-
|
209 |
-
def forward(self, x, z):
|
210 |
-
x = self.conv_0(self.actvn(self.norm_0(x, z)))
|
211 |
-
x = self.conv_1(self.actvn(self.norm_1(x, z)))
|
212 |
-
return x
|
213 |
-
|
214 |
-
|
215 |
-
class ADAINDecoderBlock(nn.Module):
|
216 |
-
def __init__(self, input_nc, output_nc, hidden_nc, feature_nc, use_transpose=True, nonlinearity=nn.LeakyReLU(), use_spect=False):
|
217 |
-
super(ADAINDecoderBlock, self).__init__()
|
218 |
-
# Attributes
|
219 |
-
self.actvn = nonlinearity
|
220 |
-
hidden_nc = min(input_nc, output_nc) if hidden_nc is None else hidden_nc
|
221 |
-
|
222 |
-
kwargs_fine = {'kernel_size':3, 'stride':1, 'padding':1}
|
223 |
-
if use_transpose:
|
224 |
-
kwargs_up = {'kernel_size':3, 'stride':2, 'padding':1, 'output_padding':1}
|
225 |
-
else:
|
226 |
-
kwargs_up = {'kernel_size':3, 'stride':1, 'padding':1}
|
227 |
-
|
228 |
-
# create conv layers
|
229 |
-
self.conv_0 = spectral_norm(nn.Conv2d(input_nc, hidden_nc, **kwargs_fine), use_spect)
|
230 |
-
if use_transpose:
|
231 |
-
self.conv_1 = spectral_norm(nn.ConvTranspose2d(hidden_nc, output_nc, **kwargs_up), use_spect)
|
232 |
-
self.conv_s = spectral_norm(nn.ConvTranspose2d(input_nc, output_nc, **kwargs_up), use_spect)
|
233 |
-
else:
|
234 |
-
self.conv_1 = nn.Sequential(spectral_norm(nn.Conv2d(hidden_nc, output_nc, **kwargs_up), use_spect),
|
235 |
-
nn.Upsample(scale_factor=2))
|
236 |
-
self.conv_s = nn.Sequential(spectral_norm(nn.Conv2d(input_nc, output_nc, **kwargs_up), use_spect),
|
237 |
-
nn.Upsample(scale_factor=2))
|
238 |
-
# define normalization layers
|
239 |
-
self.norm_0 = ADAIN(input_nc, feature_nc)
|
240 |
-
self.norm_1 = ADAIN(hidden_nc, feature_nc)
|
241 |
-
self.norm_s = ADAIN(input_nc, feature_nc)
|
242 |
-
|
243 |
-
def forward(self, x, z):
|
244 |
-
x_s = self.shortcut(x, z)
|
245 |
-
dx = self.conv_0(self.actvn(self.norm_0(x, z)))
|
246 |
-
dx = self.conv_1(self.actvn(self.norm_1(dx, z)))
|
247 |
-
out = x_s + dx
|
248 |
-
return out
|
249 |
-
|
250 |
-
def shortcut(self, x, z):
|
251 |
-
x_s = self.conv_s(self.actvn(self.norm_s(x, z)))
|
252 |
-
return x_s
|
253 |
-
|
254 |
-
|
255 |
-
class FineEncoder(nn.Module):
|
256 |
-
"""docstring for Encoder"""
|
257 |
-
def __init__(self, image_nc, ngf, img_f, layers, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
|
258 |
-
super(FineEncoder, self).__init__()
|
259 |
-
self.layers = layers
|
260 |
-
self.first = FirstBlock2d(image_nc, ngf, norm_layer, nonlinearity, use_spect)
|
261 |
-
for i in range(layers):
|
262 |
-
in_channels = min(ngf*(2**i), img_f)
|
263 |
-
out_channels = min(ngf*(2**(i+1)), img_f)
|
264 |
-
model = DownBlock2d(in_channels, out_channels, norm_layer, nonlinearity, use_spect)
|
265 |
-
setattr(self, 'down' + str(i), model)
|
266 |
-
self.output_nc = out_channels
|
267 |
-
|
268 |
-
def forward(self, x):
|
269 |
-
x = self.first(x)
|
270 |
-
out=[x]
|
271 |
-
for i in range(self.layers):
|
272 |
-
model = getattr(self, 'down'+str(i))
|
273 |
-
x = model(x)
|
274 |
-
out.append(x)
|
275 |
-
return out
|
276 |
-
|
277 |
-
|
278 |
-
class FineDecoder(nn.Module):
|
279 |
-
"""docstring for FineDecoder"""
|
280 |
-
def __init__(self, image_nc, feature_nc, ngf, img_f, layers, num_block, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
|
281 |
-
super(FineDecoder, self).__init__()
|
282 |
-
self.layers = layers
|
283 |
-
for i in range(layers)[::-1]:
|
284 |
-
in_channels = min(ngf*(2**(i+1)), img_f)
|
285 |
-
out_channels = min(ngf*(2**i), img_f)
|
286 |
-
up = UpBlock2d(in_channels, out_channels, norm_layer, nonlinearity, use_spect)
|
287 |
-
res = FineADAINResBlocks(num_block, in_channels, feature_nc, norm_layer, nonlinearity, use_spect)
|
288 |
-
jump = Jump(out_channels, norm_layer, nonlinearity, use_spect)
|
289 |
-
setattr(self, 'up' + str(i), up)
|
290 |
-
setattr(self, 'res' + str(i), res)
|
291 |
-
setattr(self, 'jump' + str(i), jump)
|
292 |
-
self.final = FinalBlock2d(out_channels, image_nc, use_spect, 'tanh')
|
293 |
-
self.output_nc = out_channels
|
294 |
-
|
295 |
-
def forward(self, x, z):
|
296 |
-
out = x.pop()
|
297 |
-
for i in range(self.layers)[::-1]:
|
298 |
-
res_model = getattr(self, 'res' + str(i))
|
299 |
-
up_model = getattr(self, 'up' + str(i))
|
300 |
-
jump_model = getattr(self, 'jump' + str(i))
|
301 |
-
out = res_model(out, z)
|
302 |
-
out = up_model(out)
|
303 |
-
out = jump_model(x.pop()) + out
|
304 |
-
out_image = self.final(out)
|
305 |
-
return out_image
|
306 |
-
|
307 |
-
|
308 |
-
class ADAINEncoder(nn.Module):
|
309 |
-
def __init__(self, image_nc, pose_nc, ngf, img_f, layers, nonlinearity=nn.LeakyReLU(), use_spect=False):
|
310 |
-
super(ADAINEncoder, self).__init__()
|
311 |
-
self.layers = layers
|
312 |
-
self.input_layer = nn.Conv2d(image_nc, ngf, kernel_size=7, stride=1, padding=3)
|
313 |
-
for i in range(layers):
|
314 |
-
in_channels = min(ngf * (2**i), img_f)
|
315 |
-
out_channels = min(ngf *(2**(i+1)), img_f)
|
316 |
-
model = ADAINEncoderBlock(in_channels, out_channels, pose_nc, nonlinearity, use_spect)
|
317 |
-
setattr(self, 'encoder' + str(i), model)
|
318 |
-
self.output_nc = out_channels
|
319 |
-
|
320 |
-
def forward(self, x, z):
|
321 |
-
out = self.input_layer(x)
|
322 |
-
out_list = [out]
|
323 |
-
for i in range(self.layers):
|
324 |
-
model = getattr(self, 'encoder' + str(i))
|
325 |
-
out = model(out, z)
|
326 |
-
out_list.append(out)
|
327 |
-
return out_list
|
328 |
-
|
329 |
-
|
330 |
-
class ADAINDecoder(nn.Module):
|
331 |
-
"""docstring for ADAINDecoder"""
|
332 |
-
def __init__(self, pose_nc, ngf, img_f, encoder_layers, decoder_layers, skip_connect=True,
|
333 |
-
nonlinearity=nn.LeakyReLU(), use_spect=False):
|
334 |
-
|
335 |
-
super(ADAINDecoder, self).__init__()
|
336 |
-
self.encoder_layers = encoder_layers
|
337 |
-
self.decoder_layers = decoder_layers
|
338 |
-
self.skip_connect = skip_connect
|
339 |
-
use_transpose = True
|
340 |
-
for i in range(encoder_layers-decoder_layers, encoder_layers)[::-1]:
|
341 |
-
in_channels = min(ngf * (2**(i+1)), img_f)
|
342 |
-
in_channels = in_channels*2 if i != (encoder_layers-1) and self.skip_connect else in_channels
|
343 |
-
out_channels = min(ngf * (2**i), img_f)
|
344 |
-
model = ADAINDecoderBlock(in_channels, out_channels, out_channels, pose_nc, use_transpose, nonlinearity, use_spect)
|
345 |
-
setattr(self, 'decoder' + str(i), model)
|
346 |
-
self.output_nc = out_channels*2 if self.skip_connect else out_channels
|
347 |
-
|
348 |
-
def forward(self, x, z):
|
349 |
-
out = x.pop() if self.skip_connect else x
|
350 |
-
for i in range(self.encoder_layers-self.decoder_layers, self.encoder_layers)[::-1]:
|
351 |
-
model = getattr(self, 'decoder' + str(i))
|
352 |
-
out = model(out, z)
|
353 |
-
out = torch.cat([out, x.pop()], 1) if self.skip_connect else out
|
354 |
-
return out
|
355 |
-
|
356 |
-
|
357 |
-
class ADAINHourglass(nn.Module):
|
358 |
-
def __init__(self, image_nc, pose_nc, ngf, img_f, encoder_layers, decoder_layers, nonlinearity, use_spect):
|
359 |
-
super(ADAINHourglass, self).__init__()
|
360 |
-
self.encoder = ADAINEncoder(image_nc, pose_nc, ngf, img_f, encoder_layers, nonlinearity, use_spect)
|
361 |
-
self.decoder = ADAINDecoder(pose_nc, ngf, img_f, encoder_layers, decoder_layers, True, nonlinearity, use_spect)
|
362 |
-
self.output_nc = self.decoder.output_nc
|
363 |
-
|
364 |
-
def forward(self, x, z):
|
365 |
-
return self.decoder(self.encoder(x, z), z)
|
366 |
-
|
367 |
-
|
368 |
-
class FineADAINLama(nn.Module):
|
369 |
-
def __init__(self, input_nc, feature_nc, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
|
370 |
-
super(FineADAINLama, self).__init__()
|
371 |
-
kwargs = {'kernel_size': 3, 'stride': 1, 'padding': 1}
|
372 |
-
self.actvn = nonlinearity
|
373 |
-
ratio_gin = 0.75
|
374 |
-
ratio_gout = 0.75
|
375 |
-
self.ffc = FFC(input_nc, input_nc, 3,
|
376 |
-
ratio_gin, ratio_gout, 1, 1, 1,
|
377 |
-
1, False, False, padding_type='reflect')
|
378 |
-
global_channels = int(input_nc * ratio_gout)
|
379 |
-
self.bn_l = ADAIN(input_nc - global_channels, feature_nc)
|
380 |
-
self.bn_g = ADAIN(global_channels, feature_nc)
|
381 |
-
|
382 |
-
def forward(self, x, z):
|
383 |
-
x_l, x_g = self.ffc(x)
|
384 |
-
x_l = self.actvn(self.bn_l(x_l,z))
|
385 |
-
x_g = self.actvn(self.bn_g(x_g,z))
|
386 |
-
return x_l, x_g
|
387 |
-
|
388 |
-
|
389 |
-
class FFCResnetBlock(nn.Module):
|
390 |
-
def __init__(self, dim, feature_dim, padding_type='reflect', norm_layer=BatchNorm2d, activation_layer=nn.ReLU, dilation=1,
|
391 |
-
spatial_transform_kwargs=None, inline=False, **conv_kwargs):
|
392 |
-
super().__init__()
|
393 |
-
self.conv1 = FineADAINLama(dim, feature_dim, **conv_kwargs)
|
394 |
-
self.conv2 = FineADAINLama(dim, feature_dim, **conv_kwargs)
|
395 |
-
self.inline = True
|
396 |
-
|
397 |
-
def forward(self, x, z):
|
398 |
-
if self.inline:
|
399 |
-
x_l, x_g = x[:, :-self.conv1.ffc.global_in_num], x[:, -self.conv1.ffc.global_in_num:]
|
400 |
-
else:
|
401 |
-
x_l, x_g = x if type(x) is tuple else (x, 0)
|
402 |
-
|
403 |
-
id_l, id_g = x_l, x_g
|
404 |
-
x_l, x_g = self.conv1((x_l, x_g), z)
|
405 |
-
x_l, x_g = self.conv2((x_l, x_g), z)
|
406 |
-
|
407 |
-
x_l, x_g = id_l + x_l, id_g + x_g
|
408 |
-
out = x_l, x_g
|
409 |
-
if self.inline:
|
410 |
-
out = torch.cat(out, dim=1)
|
411 |
-
return out
|
412 |
-
|
413 |
-
|
414 |
-
class FFCADAINResBlocks(nn.Module):
|
415 |
-
def __init__(self, num_block, input_nc, feature_nc, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
|
416 |
-
super(FFCADAINResBlocks, self).__init__()
|
417 |
-
self.num_block = num_block
|
418 |
-
for i in range(num_block):
|
419 |
-
model = FFCResnetBlock(input_nc, feature_nc, norm_layer, nonlinearity, use_spect)
|
420 |
-
setattr(self, 'res'+str(i), model)
|
421 |
-
|
422 |
-
def forward(self, x, z):
|
423 |
-
for i in range(self.num_block):
|
424 |
-
model = getattr(self, 'res'+str(i))
|
425 |
-
x = model(x, z)
|
426 |
-
return x
|
427 |
-
|
428 |
-
|
429 |
-
class Jump(nn.Module):
|
430 |
-
def __init__(self, input_nc, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
|
431 |
-
super(Jump, self).__init__()
|
432 |
-
kwargs = {'kernel_size': 3, 'stride': 1, 'padding': 1}
|
433 |
-
conv = spectral_norm(nn.Conv2d(input_nc, input_nc, **kwargs), use_spect)
|
434 |
-
if type(norm_layer) == type(None):
|
435 |
-
self.model = nn.Sequential(conv, nonlinearity)
|
436 |
-
else:
|
437 |
-
self.model = nn.Sequential(conv, norm_layer(input_nc), nonlinearity)
|
438 |
-
|
439 |
-
def forward(self, x):
|
440 |
-
out = self.model(x)
|
441 |
-
return out
|
442 |
-
|
443 |
-
|
444 |
-
class FinalBlock2d(nn.Module):
|
445 |
-
def __init__(self, input_nc, output_nc, use_spect=False, tanh_or_sigmoid='tanh'):
|
446 |
-
super(FinalBlock2d, self).__init__()
|
447 |
-
kwargs = {'kernel_size': 7, 'stride': 1, 'padding':3}
|
448 |
-
conv = spectral_norm(nn.Conv2d(input_nc, output_nc, **kwargs), use_spect)
|
449 |
-
if tanh_or_sigmoid == 'sigmoid':
|
450 |
-
out_nonlinearity = nn.Sigmoid()
|
451 |
-
else:
|
452 |
-
out_nonlinearity = nn.Tanh()
|
453 |
-
self.model = nn.Sequential(conv, out_nonlinearity)
|
454 |
-
|
455 |
-
def forward(self, x):
|
456 |
-
out = self.model(x)
|
457 |
-
return out
|
458 |
-
|
459 |
-
|
460 |
-
class ModulatedConv2d(nn.Module):
|
461 |
-
def __init__(self,
|
462 |
-
in_channels,
|
463 |
-
out_channels,
|
464 |
-
kernel_size,
|
465 |
-
num_style_feat,
|
466 |
-
demodulate=True,
|
467 |
-
sample_mode=None,
|
468 |
-
eps=1e-8):
|
469 |
-
super(ModulatedConv2d, self).__init__()
|
470 |
-
self.in_channels = in_channels
|
471 |
-
self.out_channels = out_channels
|
472 |
-
self.kernel_size = kernel_size
|
473 |
-
self.demodulate = demodulate
|
474 |
-
self.sample_mode = sample_mode
|
475 |
-
self.eps = eps
|
476 |
-
|
477 |
-
# modulation inside each modulated conv
|
478 |
-
self.modulation = nn.Linear(num_style_feat, in_channels, bias=True)
|
479 |
-
# initialization
|
480 |
-
default_init_weights(self.modulation, scale=1, bias_fill=1, a=0, mode='fan_in', nonlinearity='linear')
|
481 |
-
|
482 |
-
self.weight = nn.Parameter(
|
483 |
-
torch.randn(1, out_channels, in_channels, kernel_size, kernel_size) /
|
484 |
-
math.sqrt(in_channels * kernel_size**2))
|
485 |
-
self.padding = kernel_size // 2
|
486 |
-
|
487 |
-
def forward(self, x, style):
|
488 |
-
b, c, h, w = x.shape
|
489 |
-
style = self.modulation(style).view(b, 1, c, 1, 1)
|
490 |
-
weight = self.weight * style
|
491 |
-
|
492 |
-
if self.demodulate:
|
493 |
-
demod = torch.rsqrt(weight.pow(2).sum([2, 3, 4]) + self.eps)
|
494 |
-
weight = weight * demod.view(b, self.out_channels, 1, 1, 1)
|
495 |
-
|
496 |
-
weight = weight.view(b * self.out_channels, c, self.kernel_size, self.kernel_size)
|
497 |
-
|
498 |
-
# upsample or downsample if necessary
|
499 |
-
if self.sample_mode == 'upsample':
|
500 |
-
x = F.interpolate(x, scale_factor=2, mode='bilinear', align_corners=False)
|
501 |
-
elif self.sample_mode == 'downsample':
|
502 |
-
x = F.interpolate(x, scale_factor=0.5, mode='bilinear', align_corners=False)
|
503 |
-
|
504 |
-
b, c, h, w = x.shape
|
505 |
-
x = x.view(1, b * c, h, w)
|
506 |
-
out = F.conv2d(x, weight, padding=self.padding, groups=b)
|
507 |
-
out = out.view(b, self.out_channels, *out.shape[2:4])
|
508 |
-
return out
|
509 |
-
|
510 |
-
def __repr__(self):
|
511 |
-
return (f'{self.__class__.__name__}(in_channels={self.in_channels}, out_channels={self.out_channels}, '
|
512 |
-
f'kernel_size={self.kernel_size}, demodulate={self.demodulate}, sample_mode={self.sample_mode})')
|
513 |
-
|
514 |
-
|
515 |
-
class StyleConv(nn.Module):
|
516 |
-
def __init__(self, in_channels, out_channels, kernel_size, num_style_feat, demodulate=True, sample_mode=None):
|
517 |
-
super(StyleConv, self).__init__()
|
518 |
-
self.modulated_conv = ModulatedConv2d(
|
519 |
-
in_channels, out_channels, kernel_size, num_style_feat, demodulate=demodulate, sample_mode=sample_mode)
|
520 |
-
self.weight = nn.Parameter(torch.zeros(1)) # for noise injection
|
521 |
-
self.bias = nn.Parameter(torch.zeros(1, out_channels, 1, 1))
|
522 |
-
self.activate = nn.LeakyReLU(negative_slope=0.2, inplace=True)
|
523 |
-
|
524 |
-
def forward(self, x, style, noise=None):
|
525 |
-
# modulate
|
526 |
-
out = self.modulated_conv(x, style) * 2**0.5 # for conversion
|
527 |
-
# noise injection
|
528 |
-
if noise is None:
|
529 |
-
b, _, h, w = out.shape
|
530 |
-
noise = out.new_empty(b, 1, h, w).normal_()
|
531 |
-
out = out + self.weight * noise
|
532 |
-
# add bias
|
533 |
-
out = out + self.bias
|
534 |
-
# activation
|
535 |
-
out = self.activate(out)
|
536 |
-
return out
|
537 |
-
|
538 |
-
|
539 |
-
class ToRGB(nn.Module):
|
540 |
-
def __init__(self, in_channels, num_style_feat, upsample=True):
|
541 |
-
super(ToRGB, self).__init__()
|
542 |
-
self.upsample = upsample
|
543 |
-
self.modulated_conv = ModulatedConv2d(
|
544 |
-
in_channels, 3, kernel_size=1, num_style_feat=num_style_feat, demodulate=False, sample_mode=None)
|
545 |
-
self.bias = nn.Parameter(torch.zeros(1, 3, 1, 1))
|
546 |
-
|
547 |
-
def forward(self, x, style, skip=None):
|
548 |
-
out = self.modulated_conv(x, style)
|
549 |
-
out = out + self.bias
|
550 |
-
if skip is not None:
|
551 |
-
if self.upsample:
|
552 |
-
skip = F.interpolate(skip, scale_factor=2, mode='bilinear', align_corners=False)
|
553 |
-
out = out + skip
|
554 |
return out
|
|
|
1 |
+
import math
|
2 |
+
import torch
|
3 |
+
import torch.nn as nn
|
4 |
+
import torch.nn.functional as F
|
5 |
+
from torch.nn.modules.batchnorm import BatchNorm2d
|
6 |
+
from torch.nn.utils.spectral_norm import spectral_norm as SpectralNorm
|
7 |
+
|
8 |
+
from videoretalking.models.ffc import FFC
|
9 |
+
from basicsr.archs.arch_util import default_init_weights
|
10 |
+
|
11 |
+
|
12 |
+
class Conv2d(nn.Module):
|
13 |
+
def __init__(self, cin, cout, kernel_size, stride, padding, residual=False, *args, **kwargs):
|
14 |
+
super().__init__(*args, **kwargs)
|
15 |
+
self.conv_block = nn.Sequential(
|
16 |
+
nn.Conv2d(cin, cout, kernel_size, stride, padding),
|
17 |
+
nn.BatchNorm2d(cout)
|
18 |
+
)
|
19 |
+
self.act = nn.ReLU()
|
20 |
+
self.residual = residual
|
21 |
+
|
22 |
+
def forward(self, x):
|
23 |
+
out = self.conv_block(x)
|
24 |
+
if self.residual:
|
25 |
+
out += x
|
26 |
+
return self.act(out)
|
27 |
+
|
28 |
+
|
29 |
+
class ResBlock(nn.Module):
|
30 |
+
def __init__(self, in_channels, out_channels, mode='down'):
|
31 |
+
super(ResBlock, self).__init__()
|
32 |
+
self.conv1 = nn.Conv2d(in_channels, in_channels, 3, 1, 1)
|
33 |
+
self.conv2 = nn.Conv2d(in_channels, out_channels, 3, 1, 1)
|
34 |
+
self.skip = nn.Conv2d(in_channels, out_channels, 1, bias=False)
|
35 |
+
if mode == 'down':
|
36 |
+
self.scale_factor = 0.5
|
37 |
+
elif mode == 'up':
|
38 |
+
self.scale_factor = 2
|
39 |
+
|
40 |
+
def forward(self, x):
|
41 |
+
out = F.leaky_relu_(self.conv1(x), negative_slope=0.2)
|
42 |
+
# upsample/downsample
|
43 |
+
out = F.interpolate(out, scale_factor=self.scale_factor, mode='bilinear', align_corners=False)
|
44 |
+
out = F.leaky_relu_(self.conv2(out), negative_slope=0.2)
|
45 |
+
# skip
|
46 |
+
x = F.interpolate(x, scale_factor=self.scale_factor, mode='bilinear', align_corners=False)
|
47 |
+
skip = self.skip(x)
|
48 |
+
out = out + skip
|
49 |
+
return out
|
50 |
+
|
51 |
+
|
52 |
+
class LayerNorm2d(nn.Module):
|
53 |
+
def __init__(self, n_out, affine=True):
|
54 |
+
super(LayerNorm2d, self).__init__()
|
55 |
+
self.n_out = n_out
|
56 |
+
self.affine = affine
|
57 |
+
|
58 |
+
if self.affine:
|
59 |
+
self.weight = nn.Parameter(torch.ones(n_out, 1, 1))
|
60 |
+
self.bias = nn.Parameter(torch.zeros(n_out, 1, 1))
|
61 |
+
|
62 |
+
def forward(self, x):
|
63 |
+
normalized_shape = x.size()[1:]
|
64 |
+
if self.affine:
|
65 |
+
return F.layer_norm(x, normalized_shape, \
|
66 |
+
self.weight.expand(normalized_shape),
|
67 |
+
self.bias.expand(normalized_shape))
|
68 |
+
else:
|
69 |
+
return F.layer_norm(x, normalized_shape)
|
70 |
+
|
71 |
+
|
72 |
+
def spectral_norm(module, use_spect=True):
|
73 |
+
if use_spect:
|
74 |
+
return SpectralNorm(module)
|
75 |
+
else:
|
76 |
+
return module
|
77 |
+
|
78 |
+
|
79 |
+
class FirstBlock2d(nn.Module):
|
80 |
+
def __init__(self, input_nc, output_nc, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
|
81 |
+
super(FirstBlock2d, self).__init__()
|
82 |
+
kwargs = {'kernel_size': 7, 'stride': 1, 'padding': 3}
|
83 |
+
conv = spectral_norm(nn.Conv2d(input_nc, output_nc, **kwargs), use_spect)
|
84 |
+
|
85 |
+
if type(norm_layer) == type(None):
|
86 |
+
self.model = nn.Sequential(conv, nonlinearity)
|
87 |
+
else:
|
88 |
+
self.model = nn.Sequential(conv, norm_layer(output_nc), nonlinearity)
|
89 |
+
|
90 |
+
def forward(self, x):
|
91 |
+
out = self.model(x)
|
92 |
+
return out
|
93 |
+
|
94 |
+
|
95 |
+
class DownBlock2d(nn.Module):
|
96 |
+
def __init__(self, input_nc, output_nc, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
|
97 |
+
super(DownBlock2d, self).__init__()
|
98 |
+
kwargs = {'kernel_size': 3, 'stride': 1, 'padding': 1}
|
99 |
+
conv = spectral_norm(nn.Conv2d(input_nc, output_nc, **kwargs), use_spect)
|
100 |
+
pool = nn.AvgPool2d(kernel_size=(2, 2))
|
101 |
+
|
102 |
+
if type(norm_layer) == type(None):
|
103 |
+
self.model = nn.Sequential(conv, nonlinearity, pool)
|
104 |
+
else:
|
105 |
+
self.model = nn.Sequential(conv, norm_layer(output_nc), nonlinearity, pool)
|
106 |
+
|
107 |
+
def forward(self, x):
|
108 |
+
out = self.model(x)
|
109 |
+
return out
|
110 |
+
|
111 |
+
|
112 |
+
class UpBlock2d(nn.Module):
|
113 |
+
def __init__(self, input_nc, output_nc, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
|
114 |
+
super(UpBlock2d, self).__init__()
|
115 |
+
kwargs = {'kernel_size': 3, 'stride': 1, 'padding': 1}
|
116 |
+
conv = spectral_norm(nn.Conv2d(input_nc, output_nc, **kwargs), use_spect)
|
117 |
+
if type(norm_layer) == type(None):
|
118 |
+
self.model = nn.Sequential(conv, nonlinearity)
|
119 |
+
else:
|
120 |
+
self.model = nn.Sequential(conv, norm_layer(output_nc), nonlinearity)
|
121 |
+
|
122 |
+
def forward(self, x):
|
123 |
+
out = self.model(F.interpolate(x, scale_factor=2))
|
124 |
+
return out
|
125 |
+
|
126 |
+
|
127 |
+
class ADAIN(nn.Module):
|
128 |
+
def __init__(self, norm_nc, feature_nc):
|
129 |
+
super().__init__()
|
130 |
+
|
131 |
+
self.param_free_norm = nn.InstanceNorm2d(norm_nc, affine=False)
|
132 |
+
|
133 |
+
nhidden = 128
|
134 |
+
use_bias=True
|
135 |
+
|
136 |
+
self.mlp_shared = nn.Sequential(
|
137 |
+
nn.Linear(feature_nc, nhidden, bias=use_bias),
|
138 |
+
nn.ReLU()
|
139 |
+
)
|
140 |
+
self.mlp_gamma = nn.Linear(nhidden, norm_nc, bias=use_bias)
|
141 |
+
self.mlp_beta = nn.Linear(nhidden, norm_nc, bias=use_bias)
|
142 |
+
|
143 |
+
def forward(self, x, feature):
|
144 |
+
|
145 |
+
# Part 1. generate parameter-free normalized activations
|
146 |
+
normalized = self.param_free_norm(x)
|
147 |
+
# Part 2. produce scaling and bias conditioned on feature
|
148 |
+
feature = feature.view(feature.size(0), -1)
|
149 |
+
actv = self.mlp_shared(feature)
|
150 |
+
gamma = self.mlp_gamma(actv)
|
151 |
+
beta = self.mlp_beta(actv)
|
152 |
+
|
153 |
+
# apply scale and bias
|
154 |
+
gamma = gamma.view(*gamma.size()[:2], 1,1)
|
155 |
+
beta = beta.view(*beta.size()[:2], 1,1)
|
156 |
+
out = normalized * (1 + gamma) + beta
|
157 |
+
return out
|
158 |
+
|
159 |
+
|
160 |
+
class FineADAINResBlock2d(nn.Module):
|
161 |
+
"""
|
162 |
+
Define an Residual block for different types
|
163 |
+
"""
|
164 |
+
def __init__(self, input_nc, feature_nc, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
|
165 |
+
super(FineADAINResBlock2d, self).__init__()
|
166 |
+
kwargs = {'kernel_size': 3, 'stride': 1, 'padding': 1}
|
167 |
+
self.conv1 = spectral_norm(nn.Conv2d(input_nc, input_nc, **kwargs), use_spect)
|
168 |
+
self.conv2 = spectral_norm(nn.Conv2d(input_nc, input_nc, **kwargs), use_spect)
|
169 |
+
self.norm1 = ADAIN(input_nc, feature_nc)
|
170 |
+
self.norm2 = ADAIN(input_nc, feature_nc)
|
171 |
+
self.actvn = nonlinearity
|
172 |
+
|
173 |
+
def forward(self, x, z):
|
174 |
+
dx = self.actvn(self.norm1(self.conv1(x), z))
|
175 |
+
dx = self.norm2(self.conv2(x), z)
|
176 |
+
out = dx + x
|
177 |
+
return out
|
178 |
+
|
179 |
+
|
180 |
+
class FineADAINResBlocks(nn.Module):
|
181 |
+
def __init__(self, num_block, input_nc, feature_nc, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
|
182 |
+
super(FineADAINResBlocks, self).__init__()
|
183 |
+
self.num_block = num_block
|
184 |
+
for i in range(num_block):
|
185 |
+
model = FineADAINResBlock2d(input_nc, feature_nc, norm_layer, nonlinearity, use_spect)
|
186 |
+
setattr(self, 'res'+str(i), model)
|
187 |
+
|
188 |
+
def forward(self, x, z):
|
189 |
+
for i in range(self.num_block):
|
190 |
+
model = getattr(self, 'res'+str(i))
|
191 |
+
x = model(x, z)
|
192 |
+
return x
|
193 |
+
|
194 |
+
|
195 |
+
class ADAINEncoderBlock(nn.Module):
|
196 |
+
def __init__(self, input_nc, output_nc, feature_nc, nonlinearity=nn.LeakyReLU(), use_spect=False):
|
197 |
+
super(ADAINEncoderBlock, self).__init__()
|
198 |
+
kwargs_down = {'kernel_size': 4, 'stride': 2, 'padding': 1}
|
199 |
+
kwargs_fine = {'kernel_size': 3, 'stride': 1, 'padding': 1}
|
200 |
+
|
201 |
+
self.conv_0 = spectral_norm(nn.Conv2d(input_nc, output_nc, **kwargs_down), use_spect)
|
202 |
+
self.conv_1 = spectral_norm(nn.Conv2d(output_nc, output_nc, **kwargs_fine), use_spect)
|
203 |
+
|
204 |
+
|
205 |
+
self.norm_0 = ADAIN(input_nc, feature_nc)
|
206 |
+
self.norm_1 = ADAIN(output_nc, feature_nc)
|
207 |
+
self.actvn = nonlinearity
|
208 |
+
|
209 |
+
def forward(self, x, z):
|
210 |
+
x = self.conv_0(self.actvn(self.norm_0(x, z)))
|
211 |
+
x = self.conv_1(self.actvn(self.norm_1(x, z)))
|
212 |
+
return x
|
213 |
+
|
214 |
+
|
215 |
+
class ADAINDecoderBlock(nn.Module):
|
216 |
+
def __init__(self, input_nc, output_nc, hidden_nc, feature_nc, use_transpose=True, nonlinearity=nn.LeakyReLU(), use_spect=False):
|
217 |
+
super(ADAINDecoderBlock, self).__init__()
|
218 |
+
# Attributes
|
219 |
+
self.actvn = nonlinearity
|
220 |
+
hidden_nc = min(input_nc, output_nc) if hidden_nc is None else hidden_nc
|
221 |
+
|
222 |
+
kwargs_fine = {'kernel_size':3, 'stride':1, 'padding':1}
|
223 |
+
if use_transpose:
|
224 |
+
kwargs_up = {'kernel_size':3, 'stride':2, 'padding':1, 'output_padding':1}
|
225 |
+
else:
|
226 |
+
kwargs_up = {'kernel_size':3, 'stride':1, 'padding':1}
|
227 |
+
|
228 |
+
# create conv layers
|
229 |
+
self.conv_0 = spectral_norm(nn.Conv2d(input_nc, hidden_nc, **kwargs_fine), use_spect)
|
230 |
+
if use_transpose:
|
231 |
+
self.conv_1 = spectral_norm(nn.ConvTranspose2d(hidden_nc, output_nc, **kwargs_up), use_spect)
|
232 |
+
self.conv_s = spectral_norm(nn.ConvTranspose2d(input_nc, output_nc, **kwargs_up), use_spect)
|
233 |
+
else:
|
234 |
+
self.conv_1 = nn.Sequential(spectral_norm(nn.Conv2d(hidden_nc, output_nc, **kwargs_up), use_spect),
|
235 |
+
nn.Upsample(scale_factor=2))
|
236 |
+
self.conv_s = nn.Sequential(spectral_norm(nn.Conv2d(input_nc, output_nc, **kwargs_up), use_spect),
|
237 |
+
nn.Upsample(scale_factor=2))
|
238 |
+
# define normalization layers
|
239 |
+
self.norm_0 = ADAIN(input_nc, feature_nc)
|
240 |
+
self.norm_1 = ADAIN(hidden_nc, feature_nc)
|
241 |
+
self.norm_s = ADAIN(input_nc, feature_nc)
|
242 |
+
|
243 |
+
def forward(self, x, z):
|
244 |
+
x_s = self.shortcut(x, z)
|
245 |
+
dx = self.conv_0(self.actvn(self.norm_0(x, z)))
|
246 |
+
dx = self.conv_1(self.actvn(self.norm_1(dx, z)))
|
247 |
+
out = x_s + dx
|
248 |
+
return out
|
249 |
+
|
250 |
+
def shortcut(self, x, z):
|
251 |
+
x_s = self.conv_s(self.actvn(self.norm_s(x, z)))
|
252 |
+
return x_s
|
253 |
+
|
254 |
+
|
255 |
+
class FineEncoder(nn.Module):
|
256 |
+
"""docstring for Encoder"""
|
257 |
+
def __init__(self, image_nc, ngf, img_f, layers, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
|
258 |
+
super(FineEncoder, self).__init__()
|
259 |
+
self.layers = layers
|
260 |
+
self.first = FirstBlock2d(image_nc, ngf, norm_layer, nonlinearity, use_spect)
|
261 |
+
for i in range(layers):
|
262 |
+
in_channels = min(ngf*(2**i), img_f)
|
263 |
+
out_channels = min(ngf*(2**(i+1)), img_f)
|
264 |
+
model = DownBlock2d(in_channels, out_channels, norm_layer, nonlinearity, use_spect)
|
265 |
+
setattr(self, 'down' + str(i), model)
|
266 |
+
self.output_nc = out_channels
|
267 |
+
|
268 |
+
def forward(self, x):
|
269 |
+
x = self.first(x)
|
270 |
+
out=[x]
|
271 |
+
for i in range(self.layers):
|
272 |
+
model = getattr(self, 'down'+str(i))
|
273 |
+
x = model(x)
|
274 |
+
out.append(x)
|
275 |
+
return out
|
276 |
+
|
277 |
+
|
278 |
+
class FineDecoder(nn.Module):
|
279 |
+
"""docstring for FineDecoder"""
|
280 |
+
def __init__(self, image_nc, feature_nc, ngf, img_f, layers, num_block, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
|
281 |
+
super(FineDecoder, self).__init__()
|
282 |
+
self.layers = layers
|
283 |
+
for i in range(layers)[::-1]:
|
284 |
+
in_channels = min(ngf*(2**(i+1)), img_f)
|
285 |
+
out_channels = min(ngf*(2**i), img_f)
|
286 |
+
up = UpBlock2d(in_channels, out_channels, norm_layer, nonlinearity, use_spect)
|
287 |
+
res = FineADAINResBlocks(num_block, in_channels, feature_nc, norm_layer, nonlinearity, use_spect)
|
288 |
+
jump = Jump(out_channels, norm_layer, nonlinearity, use_spect)
|
289 |
+
setattr(self, 'up' + str(i), up)
|
290 |
+
setattr(self, 'res' + str(i), res)
|
291 |
+
setattr(self, 'jump' + str(i), jump)
|
292 |
+
self.final = FinalBlock2d(out_channels, image_nc, use_spect, 'tanh')
|
293 |
+
self.output_nc = out_channels
|
294 |
+
|
295 |
+
def forward(self, x, z):
|
296 |
+
out = x.pop()
|
297 |
+
for i in range(self.layers)[::-1]:
|
298 |
+
res_model = getattr(self, 'res' + str(i))
|
299 |
+
up_model = getattr(self, 'up' + str(i))
|
300 |
+
jump_model = getattr(self, 'jump' + str(i))
|
301 |
+
out = res_model(out, z)
|
302 |
+
out = up_model(out)
|
303 |
+
out = jump_model(x.pop()) + out
|
304 |
+
out_image = self.final(out)
|
305 |
+
return out_image
|
306 |
+
|
307 |
+
|
308 |
+
class ADAINEncoder(nn.Module):
|
309 |
+
def __init__(self, image_nc, pose_nc, ngf, img_f, layers, nonlinearity=nn.LeakyReLU(), use_spect=False):
|
310 |
+
super(ADAINEncoder, self).__init__()
|
311 |
+
self.layers = layers
|
312 |
+
self.input_layer = nn.Conv2d(image_nc, ngf, kernel_size=7, stride=1, padding=3)
|
313 |
+
for i in range(layers):
|
314 |
+
in_channels = min(ngf * (2**i), img_f)
|
315 |
+
out_channels = min(ngf *(2**(i+1)), img_f)
|
316 |
+
model = ADAINEncoderBlock(in_channels, out_channels, pose_nc, nonlinearity, use_spect)
|
317 |
+
setattr(self, 'encoder' + str(i), model)
|
318 |
+
self.output_nc = out_channels
|
319 |
+
|
320 |
+
def forward(self, x, z):
|
321 |
+
out = self.input_layer(x)
|
322 |
+
out_list = [out]
|
323 |
+
for i in range(self.layers):
|
324 |
+
model = getattr(self, 'encoder' + str(i))
|
325 |
+
out = model(out, z)
|
326 |
+
out_list.append(out)
|
327 |
+
return out_list
|
328 |
+
|
329 |
+
|
330 |
+
class ADAINDecoder(nn.Module):
|
331 |
+
"""docstring for ADAINDecoder"""
|
332 |
+
def __init__(self, pose_nc, ngf, img_f, encoder_layers, decoder_layers, skip_connect=True,
|
333 |
+
nonlinearity=nn.LeakyReLU(), use_spect=False):
|
334 |
+
|
335 |
+
super(ADAINDecoder, self).__init__()
|
336 |
+
self.encoder_layers = encoder_layers
|
337 |
+
self.decoder_layers = decoder_layers
|
338 |
+
self.skip_connect = skip_connect
|
339 |
+
use_transpose = True
|
340 |
+
for i in range(encoder_layers-decoder_layers, encoder_layers)[::-1]:
|
341 |
+
in_channels = min(ngf * (2**(i+1)), img_f)
|
342 |
+
in_channels = in_channels*2 if i != (encoder_layers-1) and self.skip_connect else in_channels
|
343 |
+
out_channels = min(ngf * (2**i), img_f)
|
344 |
+
model = ADAINDecoderBlock(in_channels, out_channels, out_channels, pose_nc, use_transpose, nonlinearity, use_spect)
|
345 |
+
setattr(self, 'decoder' + str(i), model)
|
346 |
+
self.output_nc = out_channels*2 if self.skip_connect else out_channels
|
347 |
+
|
348 |
+
def forward(self, x, z):
|
349 |
+
out = x.pop() if self.skip_connect else x
|
350 |
+
for i in range(self.encoder_layers-self.decoder_layers, self.encoder_layers)[::-1]:
|
351 |
+
model = getattr(self, 'decoder' + str(i))
|
352 |
+
out = model(out, z)
|
353 |
+
out = torch.cat([out, x.pop()], 1) if self.skip_connect else out
|
354 |
+
return out
|
355 |
+
|
356 |
+
|
357 |
+
class ADAINHourglass(nn.Module):
|
358 |
+
def __init__(self, image_nc, pose_nc, ngf, img_f, encoder_layers, decoder_layers, nonlinearity, use_spect):
|
359 |
+
super(ADAINHourglass, self).__init__()
|
360 |
+
self.encoder = ADAINEncoder(image_nc, pose_nc, ngf, img_f, encoder_layers, nonlinearity, use_spect)
|
361 |
+
self.decoder = ADAINDecoder(pose_nc, ngf, img_f, encoder_layers, decoder_layers, True, nonlinearity, use_spect)
|
362 |
+
self.output_nc = self.decoder.output_nc
|
363 |
+
|
364 |
+
def forward(self, x, z):
|
365 |
+
return self.decoder(self.encoder(x, z), z)
|
366 |
+
|
367 |
+
|
368 |
+
class FineADAINLama(nn.Module):
|
369 |
+
def __init__(self, input_nc, feature_nc, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
|
370 |
+
super(FineADAINLama, self).__init__()
|
371 |
+
kwargs = {'kernel_size': 3, 'stride': 1, 'padding': 1}
|
372 |
+
self.actvn = nonlinearity
|
373 |
+
ratio_gin = 0.75
|
374 |
+
ratio_gout = 0.75
|
375 |
+
self.ffc = FFC(input_nc, input_nc, 3,
|
376 |
+
ratio_gin, ratio_gout, 1, 1, 1,
|
377 |
+
1, False, False, padding_type='reflect')
|
378 |
+
global_channels = int(input_nc * ratio_gout)
|
379 |
+
self.bn_l = ADAIN(input_nc - global_channels, feature_nc)
|
380 |
+
self.bn_g = ADAIN(global_channels, feature_nc)
|
381 |
+
|
382 |
+
def forward(self, x, z):
|
383 |
+
x_l, x_g = self.ffc(x)
|
384 |
+
x_l = self.actvn(self.bn_l(x_l,z))
|
385 |
+
x_g = self.actvn(self.bn_g(x_g,z))
|
386 |
+
return x_l, x_g
|
387 |
+
|
388 |
+
|
389 |
+
class FFCResnetBlock(nn.Module):
|
390 |
+
def __init__(self, dim, feature_dim, padding_type='reflect', norm_layer=BatchNorm2d, activation_layer=nn.ReLU, dilation=1,
|
391 |
+
spatial_transform_kwargs=None, inline=False, **conv_kwargs):
|
392 |
+
super().__init__()
|
393 |
+
self.conv1 = FineADAINLama(dim, feature_dim, **conv_kwargs)
|
394 |
+
self.conv2 = FineADAINLama(dim, feature_dim, **conv_kwargs)
|
395 |
+
self.inline = True
|
396 |
+
|
397 |
+
def forward(self, x, z):
|
398 |
+
if self.inline:
|
399 |
+
x_l, x_g = x[:, :-self.conv1.ffc.global_in_num], x[:, -self.conv1.ffc.global_in_num:]
|
400 |
+
else:
|
401 |
+
x_l, x_g = x if type(x) is tuple else (x, 0)
|
402 |
+
|
403 |
+
id_l, id_g = x_l, x_g
|
404 |
+
x_l, x_g = self.conv1((x_l, x_g), z)
|
405 |
+
x_l, x_g = self.conv2((x_l, x_g), z)
|
406 |
+
|
407 |
+
x_l, x_g = id_l + x_l, id_g + x_g
|
408 |
+
out = x_l, x_g
|
409 |
+
if self.inline:
|
410 |
+
out = torch.cat(out, dim=1)
|
411 |
+
return out
|
412 |
+
|
413 |
+
|
414 |
+
class FFCADAINResBlocks(nn.Module):
|
415 |
+
def __init__(self, num_block, input_nc, feature_nc, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
|
416 |
+
super(FFCADAINResBlocks, self).__init__()
|
417 |
+
self.num_block = num_block
|
418 |
+
for i in range(num_block):
|
419 |
+
model = FFCResnetBlock(input_nc, feature_nc, norm_layer, nonlinearity, use_spect)
|
420 |
+
setattr(self, 'res'+str(i), model)
|
421 |
+
|
422 |
+
def forward(self, x, z):
|
423 |
+
for i in range(self.num_block):
|
424 |
+
model = getattr(self, 'res'+str(i))
|
425 |
+
x = model(x, z)
|
426 |
+
return x
|
427 |
+
|
428 |
+
|
429 |
+
class Jump(nn.Module):
|
430 |
+
def __init__(self, input_nc, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
|
431 |
+
super(Jump, self).__init__()
|
432 |
+
kwargs = {'kernel_size': 3, 'stride': 1, 'padding': 1}
|
433 |
+
conv = spectral_norm(nn.Conv2d(input_nc, input_nc, **kwargs), use_spect)
|
434 |
+
if type(norm_layer) == type(None):
|
435 |
+
self.model = nn.Sequential(conv, nonlinearity)
|
436 |
+
else:
|
437 |
+
self.model = nn.Sequential(conv, norm_layer(input_nc), nonlinearity)
|
438 |
+
|
439 |
+
def forward(self, x):
|
440 |
+
out = self.model(x)
|
441 |
+
return out
|
442 |
+
|
443 |
+
|
444 |
+
class FinalBlock2d(nn.Module):
|
445 |
+
def __init__(self, input_nc, output_nc, use_spect=False, tanh_or_sigmoid='tanh'):
|
446 |
+
super(FinalBlock2d, self).__init__()
|
447 |
+
kwargs = {'kernel_size': 7, 'stride': 1, 'padding':3}
|
448 |
+
conv = spectral_norm(nn.Conv2d(input_nc, output_nc, **kwargs), use_spect)
|
449 |
+
if tanh_or_sigmoid == 'sigmoid':
|
450 |
+
out_nonlinearity = nn.Sigmoid()
|
451 |
+
else:
|
452 |
+
out_nonlinearity = nn.Tanh()
|
453 |
+
self.model = nn.Sequential(conv, out_nonlinearity)
|
454 |
+
|
455 |
+
def forward(self, x):
|
456 |
+
out = self.model(x)
|
457 |
+
return out
|
458 |
+
|
459 |
+
|
460 |
+
class ModulatedConv2d(nn.Module):
|
461 |
+
def __init__(self,
|
462 |
+
in_channels,
|
463 |
+
out_channels,
|
464 |
+
kernel_size,
|
465 |
+
num_style_feat,
|
466 |
+
demodulate=True,
|
467 |
+
sample_mode=None,
|
468 |
+
eps=1e-8):
|
469 |
+
super(ModulatedConv2d, self).__init__()
|
470 |
+
self.in_channels = in_channels
|
471 |
+
self.out_channels = out_channels
|
472 |
+
self.kernel_size = kernel_size
|
473 |
+
self.demodulate = demodulate
|
474 |
+
self.sample_mode = sample_mode
|
475 |
+
self.eps = eps
|
476 |
+
|
477 |
+
# modulation inside each modulated conv
|
478 |
+
self.modulation = nn.Linear(num_style_feat, in_channels, bias=True)
|
479 |
+
# initialization
|
480 |
+
default_init_weights(self.modulation, scale=1, bias_fill=1, a=0, mode='fan_in', nonlinearity='linear')
|
481 |
+
|
482 |
+
self.weight = nn.Parameter(
|
483 |
+
torch.randn(1, out_channels, in_channels, kernel_size, kernel_size) /
|
484 |
+
math.sqrt(in_channels * kernel_size**2))
|
485 |
+
self.padding = kernel_size // 2
|
486 |
+
|
487 |
+
def forward(self, x, style):
|
488 |
+
b, c, h, w = x.shape
|
489 |
+
style = self.modulation(style).view(b, 1, c, 1, 1)
|
490 |
+
weight = self.weight * style
|
491 |
+
|
492 |
+
if self.demodulate:
|
493 |
+
demod = torch.rsqrt(weight.pow(2).sum([2, 3, 4]) + self.eps)
|
494 |
+
weight = weight * demod.view(b, self.out_channels, 1, 1, 1)
|
495 |
+
|
496 |
+
weight = weight.view(b * self.out_channels, c, self.kernel_size, self.kernel_size)
|
497 |
+
|
498 |
+
# upsample or downsample if necessary
|
499 |
+
if self.sample_mode == 'upsample':
|
500 |
+
x = F.interpolate(x, scale_factor=2, mode='bilinear', align_corners=False)
|
501 |
+
elif self.sample_mode == 'downsample':
|
502 |
+
x = F.interpolate(x, scale_factor=0.5, mode='bilinear', align_corners=False)
|
503 |
+
|
504 |
+
b, c, h, w = x.shape
|
505 |
+
x = x.view(1, b * c, h, w)
|
506 |
+
out = F.conv2d(x, weight, padding=self.padding, groups=b)
|
507 |
+
out = out.view(b, self.out_channels, *out.shape[2:4])
|
508 |
+
return out
|
509 |
+
|
510 |
+
def __repr__(self):
|
511 |
+
return (f'{self.__class__.__name__}(in_channels={self.in_channels}, out_channels={self.out_channels}, '
|
512 |
+
f'kernel_size={self.kernel_size}, demodulate={self.demodulate}, sample_mode={self.sample_mode})')
|
513 |
+
|
514 |
+
|
515 |
+
class StyleConv(nn.Module):
|
516 |
+
def __init__(self, in_channels, out_channels, kernel_size, num_style_feat, demodulate=True, sample_mode=None):
|
517 |
+
super(StyleConv, self).__init__()
|
518 |
+
self.modulated_conv = ModulatedConv2d(
|
519 |
+
in_channels, out_channels, kernel_size, num_style_feat, demodulate=demodulate, sample_mode=sample_mode)
|
520 |
+
self.weight = nn.Parameter(torch.zeros(1)) # for noise injection
|
521 |
+
self.bias = nn.Parameter(torch.zeros(1, out_channels, 1, 1))
|
522 |
+
self.activate = nn.LeakyReLU(negative_slope=0.2, inplace=True)
|
523 |
+
|
524 |
+
def forward(self, x, style, noise=None):
|
525 |
+
# modulate
|
526 |
+
out = self.modulated_conv(x, style) * 2**0.5 # for conversion
|
527 |
+
# noise injection
|
528 |
+
if noise is None:
|
529 |
+
b, _, h, w = out.shape
|
530 |
+
noise = out.new_empty(b, 1, h, w).normal_()
|
531 |
+
out = out + self.weight * noise
|
532 |
+
# add bias
|
533 |
+
out = out + self.bias
|
534 |
+
# activation
|
535 |
+
out = self.activate(out)
|
536 |
+
return out
|
537 |
+
|
538 |
+
|
539 |
+
class ToRGB(nn.Module):
|
540 |
+
def __init__(self, in_channels, num_style_feat, upsample=True):
|
541 |
+
super(ToRGB, self).__init__()
|
542 |
+
self.upsample = upsample
|
543 |
+
self.modulated_conv = ModulatedConv2d(
|
544 |
+
in_channels, 3, kernel_size=1, num_style_feat=num_style_feat, demodulate=False, sample_mode=None)
|
545 |
+
self.bias = nn.Parameter(torch.zeros(1, 3, 1, 1))
|
546 |
+
|
547 |
+
def forward(self, x, style, skip=None):
|
548 |
+
out = self.modulated_conv(x, style)
|
549 |
+
out = out + self.bias
|
550 |
+
if skip is not None:
|
551 |
+
if self.upsample:
|
552 |
+
skip = F.interpolate(skip, scale_factor=2, mode='bilinear', align_corners=False)
|
553 |
+
out = out + skip
|
554 |
return out
|