-
Notifications
You must be signed in to change notification settings - Fork 0
/
pix2pixHD.py
403 lines (352 loc) · 14.6 KB
/
pix2pixHD.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
import argparse
from typing import List, Optional, Tuple, Union
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch import Tensor
from firewood import layers, utils
from firewood.common.types import INT
from firewood.models.gan.pix2pix import PatchGAN
class ResBlock(nn.Module):
def __init__(
self,
channels: int,
kernel_size: int = 3,
stride: int = 1,
bias: bool = True,
padding_mode="reflect",
normalization="in",
activation="relu",
) -> None:
super().__init__()
# fmt: off
conv_kwargs = dict(kernel_size=kernel_size, stride=stride, padding="same", bias=bias,
padding_mode=padding_mode, normalization=normalization)
self.layers = nn.ModuleList([
layers.Conv2dBlock(channels, channels, activation=activation, **conv_kwargs),
layers.Conv2dBlock(channels, channels, **conv_kwargs),
])
# fmt: on
def forward(self, input: Tensor) -> Tensor:
# input -> conv -> norm -> act -> conv -> norm -> output + input
output = input
for layer in self.layers:
output = layer(output)
return input + output
class GlobalGenerator(nn.Module):
"""
GlobalGenerator of Pix2PixHD
High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs
https://arxiv.org/abs/1711.11585
"""
def __init__(
self,
in_channels: int,
out_channels: Optional[int] = None,
n_filters: int = 64,
n_down_blocks: int = 4,
n_res_blocks: int = 9,
padding_mode: str = "reflect",
normalization: str = "in",
activation: str = "relu",
max_filters: int = 2**10,
) -> None:
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels or in_channels
n_filters = min(max_filters, n_filters)
# fmt: off
kwargs = dict(bias=True, padding_mode=padding_mode, normalization=normalization, activation=activation)
# Downsample Blocks
self.downsamples = nn.ModuleList([
layers.Conv2dBlock(in_channels, n_filters, 7, 1, 0,
normalization=normalization, activation=activation)
])
for i in range(n_down_blocks):
in_channels = min(max_filters, n_filters * 2**i)
out_channels = min(max_filters, in_channels * 2)
self.downsamples.append(
layers.Conv2dBlock(in_channels, out_channels, 3, 2, 1, **kwargs)
)
# Residual Blocks
self.residuals = nn.ModuleList()
res_channels = min(max_filters, n_filters * 2**n_down_blocks)
for i in range(n_res_blocks):
self.residuals.append(ResBlock(res_channels, 3, 1, **kwargs))
# Upsample Blocks
self.upsamples = nn.ModuleList()
for i in range(n_down_blocks):
in_channels = n_filters * 2**(n_down_blocks - i)
out_channels = in_channels // 2
if in_channels > max_filters:
in_channels = max_filters
out_channels = max_filters
self.upsamples.append(
layers.ConvTranspose2dBlock(in_channels, out_channels, 3, 2, 1, output_padding=1,
bias=True, normalization=normalization, activation=activation)
)
self.upsamples.append(layers.Conv2dBlock(n_filters, self.out_channels, 7, 1, "same", bias=True,
padding_mode=padding_mode, activation="tanh"))
# fmt: on
def forward(self, input: Tensor) -> Tensor:
output = input
for downsample in self.downsamples:
output = downsample(output)
for residual in self.residuals:
output = residual(output)
for upsample in self.upsamples:
output = upsample(output)
return output
class LocalEnhancer(nn.Module):
"""
LocalEnhancer of Pix2PixHD
High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs
https://arxiv.org/abs/1711.11585
Operation Graph:
AvgPool(input)**n -> GlobalGenerator = output_0\n
AvgPool(input)**(n-1) -> Downsample -> Add(output_0) -> Upsample -> output_1\n
...\n
AvgPool(input)**1 -> Downsample -> Add(output_(n-2)) -> Upsample -> output_(n-1)\n
input -> Downsample -> Add(output_(n-1)) -> Upsample -> output_n\n
OutputLayer(output_n) -> output
"""
def __init__(
self,
in_channels: int,
out_channels: Optional[int] = None,
n_local_filters: int = 32,
n_local_enhancers: int = 1,
n_down_blocks: int = 4,
n_local_res_blocks: int = 3,
n_global_res_blocks: int = 9,
padding_mode: str = "reflect",
normalization: str = "in",
activation: str = "relu",
) -> None:
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels or in_channels
self.n_local_filters = n_local_filters
self.n_local_enhancers = n_local_enhancers
self.global_generator = GlobalGenerator(
in_channels=self.in_channels,
out_channels=self.out_channels,
n_filters=self.n_local_filters * 2**self.n_local_enhancers,
n_down_blocks=n_down_blocks,
n_res_blocks=n_global_res_blocks,
padding_mode=padding_mode,
normalization=normalization,
activation=activation,
max_filters=2**10,
)
# Remove last convolution layer of GlobalGenerator to extract features
self.global_generator.upsamples = self.global_generator.upsamples[:-1]
for param in self.global_generator.parameters():
param.requires_grad = False
# fmt: off
self.downsamples = nn.ModuleList()
self.residuals = nn.ModuleList()
self.upsamples = nn.ModuleList()
downsample_kwargs = dict(bias=True, padding_mode=padding_mode, normalization=normalization, activation=activation)
residual_kwargs = dict(bias=True, padding_mode=padding_mode, normalization=normalization, activation=activation)
upsample_kwargs = dict(bias=True, normalization=normalization, activation=activation)
for i in range(self.n_local_enhancers):
global_filters = self.n_local_filters * 2**i
self.downsamples.append(nn.ModuleList([
layers.Conv2dBlock(self.in_channels, global_filters, 7, 1, 3, **downsample_kwargs),
layers.Conv2dBlock(global_filters, 2 * global_filters, 3, 2, 1, **downsample_kwargs)
]))
self.residuals.append(nn.ModuleList([
ResBlock(2 * global_filters, 3, 1, **residual_kwargs)
for _ in range(n_local_res_blocks)
]))
self.upsamples.append(nn.ModuleList([
layers.ConvTranspose2dBlock(2 * global_filters, global_filters, 3, 2, 1, output_padding=1, **upsample_kwargs),
]))
self.output_layer = layers.Conv2dBlock(self.n_local_filters, self.out_channels, 7, 1, "same", bias=True,
padding_mode=padding_mode, activation="tanh")
# fmt: on
def forward(self, input: Tensor) -> Tensor:
inputs = [input]
for _ in range(self.n_local_enhancers):
downsampled: Tensor = F.avg_pool2d(
inputs[-1], 3, stride=2, padding=1, count_include_pad=False
)
inputs.append(downsampled)
# GlobalGenerator features
self.global_generator.eval()
output: Tensor = self.global_generator(inputs[-1])
for i in range(self.n_local_enhancers - 1, -1, -1):
downsampled = inputs[i]
for downsample in self.downsamples[i]:
downsampled = downsample(downsampled)
# Add GlobalGenerator features to local enhancer
output += downsampled
for residual in self.residuals[i]:
output = residual(output)
for upsample in self.upsamples[i]:
output = upsample(output)
return self.output_layer(output)
class Encoder(nn.Module):
"""
Encoder of Pix2PixHD for instance-wise embedding
High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs
https://arxiv.org/abs/1711.11585
"""
def __init__(
self,
in_channels: int,
out_channels: int = 3,
n_filters: int = 16,
n_down_blocks: int = 4,
padding_mode: str = "reflect",
normalization: str = "in",
activation: str = "relu",
) -> None:
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
# fmt: off
kwargs = dict(bias=True, padding_mode=padding_mode, normalization=normalization, activation=activation)
self.layers = nn.ModuleList(
[layers.Conv2dBlock(self.in_channels, n_filters, 7, 1, 3, **kwargs)]
)
for i in range(n_down_blocks):
in_channels = n_filters * 2**i
out_channels = in_channels * 2
self.layers.append(
layers.Conv2dBlock(in_channels, out_channels, 3, 2, 1, **kwargs)
)
for i in range(n_down_blocks, 0, -1):
in_channels = n_filters * 2**i
out_channels = in_channels // 2
self.layers.append(
layers.ConvTranspose2dBlock(in_channels, out_channels, 3, 2, 1, output_padding=1,
bias=True, normalization=normalization, activation=activation)
)
self.layers.append(
layers.Conv2dBlock(n_filters, self.out_channels, 7, 1, 3,
**utils.updated_dict(kwargs, activation="tanh"))
)
# fmt: on
def forward(
self, input: Tensor, instance_input: Tensor, normalize: bool = False
) -> Tensor:
"""
input: (N, C, H, W), range [-1, 1]
instance_input: (N, 1, H, W), range [0, 255]
normalize: normalize instance_input to [0, 255] if range in [-1, 1]
"""
output = input
for layer in self.layers:
output: Tensor = layer(output)
# instance-wise average pooling
mean_output = output.clone()
if normalize:
instance_input = instance_input.mul(127.5).add(128).clamp_(0, 255)
instance_input = instance_input.byte()
instance_list = instance_input.unique()
for instance in instance_list:
for b in range(output.size(0)):
indices = (instance_input[b : b + 1] == instance).nonzero()
for j in range(self.out_channels):
output_instance = output[
indices[:, 0] + b,
indices[:, 1] + j,
indices[:, 2],
indices[:, 3],
]
mean_feature = torch.mean(output_instance).expand_as(
output_instance
)
mean_output[
indices[:, 0] + b,
indices[:, 1] + j,
indices[:, 2],
indices[:, 3],
] = mean_feature
return mean_output
class MultiScalePatchGAN(nn.Module):
"""
Multi-Scale Discriminator of Pix2PixHD
High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs
https://arxiv.org/abs/1711.11585
"""
def __init__(
self,
in_channels: int = 6,
n_layers: int = 4,
n_discriminators: int = 3,
normalization: str = "in",
activation: str = "lrelu",
) -> None:
super().__init__()
self.n_discriminators = n_discriminators
# fmt: off
self.discriminators = nn.ModuleList()
for _ in range(self.n_discriminators):
self.discriminators.append(
PatchGAN(in_channels, n_layers=n_layers, normalization=normalization, activation=activation)
)
# fmt: on
def avg_pool(self, *inputs: Tensor) -> Tuple[Tensor, ...]:
outputs: List[Tensor] = []
for input in inputs:
outputs.append(
F.avg_pool2d(input, 3, 2, 1, count_include_pad=False)
)
return tuple(outputs)
def forward(
self, input: Tensor, label: Tensor, extract_features: bool = False
) -> Union[List[Tensor], List[List[Tensor]]]:
inputs = [(input, label)]
for _ in range(self.n_discriminators - 1):
inputs.append(self.avg_pool(*inputs[-1]))
outputs: Union[List[Tensor], List[List[Tensor]]] = []
for i in range(self.n_discriminators):
outputs.append(self.discriminators[i](*inputs[i], extract_features))
return outputs
def main() -> None:
import pytorch_lightning as pl
from pytorch_lightning.utilities.model_summary import summarize
parser = argparse.ArgumentParser()
parser.add_argument("--depth", "-d", type=int, default=3)
args = vars(parser.parse_args())
class Summary(pl.LightningModule):
def __init__(
self,
in_channels: int = 1,
out_channels: int = 3,
instance_out_channels: int = 3,
resolution: INT = (1024, 512),
) -> None:
super().__init__()
self.encoder = Encoder(out_channels, instance_out_channels)
self.local_enhancer = LocalEnhancer(
in_channels + instance_out_channels,
out_channels,
n_local_enhancers=2,
)
self.discriminator = MultiScalePatchGAN(
in_channels + out_channels, n_discriminators=2
)
self.example_input_array = (
torch.empty(2, in_channels, *utils._pair(resolution)),
torch.empty(2, out_channels, *utils._pair(resolution)),
True,
)
def forward(
self, source: Tensor, target: Tensor, extract_features: bool = False
) -> List[Tensor]: # type: ignore
feature_map = self.encoder(target, target[:, :1], True)
input = torch.cat((source, feature_map), dim=1)
generated_image: Tensor = self.local_enhancer(input)
score: List[Tensor] = self.discriminator(
source, generated_image, extract_features
)
return score
summary = Summary()
print(summary)
print(summarize(summary, max_depth=args["depth"]))
if __name__ == "__main__":
main()