-
Notifications
You must be signed in to change notification settings - Fork 1
/
model.py
300 lines (213 loc) · 11.8 KB
/
model.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
"""
Copyright 2019 Ilja Manakov
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated
documentation files (the "Software"), to deal in the Software without restriction, including without limitation the
rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit
persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or
substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE
WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
"""
import torch as pt
from torch.nn.functional import selu, relu
from torch.nn.parameter import Parameter
from autoencoder import model_parts as parts
from collections import namedtuple
class Generator(pt.nn.Module):
def __init__(self, scale_factor, n_residual=(6, 3), activation=relu, input_channels=1,
channel_factor=8, kernel_size=(3, 3), skip_conn='concat',
norm_func=pt.nn.InstanceNorm2d, up_conv=parts.ConvResize2d, pad=pt.nn.ZeroPad2d):
super().__init__()
self.skip_conn = skip_conn
self.downsampling = []
self.residuals = []
self.upsampling = []
# initial 7x7 convolution
self.initial_conv = parts.GeneralConvolution(input_channels, channel_factor, (7, 7), (1, 1), pt.nn.ZeroPad2d,
norm_func, activation, pt.nn.Conv2d, True)
# downsampling with strided convolutions, channels are double after each convolution
for i in range(scale_factor):
channel_factor *= 2
self.downsampling.append(
parts.GeneralConvolution(channel_factor // 2, channel_factor, kernel_size, (2, 2), pt.nn.ZeroPad2d,
norm_func, activation, pt.nn.Conv2d, True)
)
self.add_module(f'down_conv_{i}', self.downsampling[-1])
# add residual blocks
for i in range(n_residual[0]):
self.residuals.append(
parts.ResBlock2d(channel_factor, n_residual[1], kernel_size, pt.nn.ZeroPad2d, norm_func)
)
self.add_module(f'res_block_{i}', self.residuals[-1])
# upsampling
for i in range(scale_factor):
in_channels = channel_factor*2 if skip_conn == 'concat' else channel_factor
channel_factor = channel_factor // 2
self.upsampling.append(
up_conv(in_channels, channel_factor, kernel_size, (1, 1),
norm=norm_func, activation=activation, affine=True, padding=pad)
)
self.add_module(f'up_conv_{i}', self.upsampling[-1])
# final convolution
in_channels = channel_factor * 2 if skip_conn == 'concat' else channel_factor
self.final_conv = parts.GeneralConvolution(in_channels, input_channels, kernel_size, (1, 1), pt.nn.ZeroPad2d,
None, activation, pt.nn.Conv2d, True)
def forward(self, x):
out = x
out = self.initial_conv(out)
skips = []
for down_conv in self.downsampling:
skips.append(out)
out = down_conv(out)
skips.append(out)
for residual in self.residuals:
out = residual(out)
for up_conv in self.upsampling:
if self.skip_conn == 'concat':
out = pt.cat([out, skips.pop()], dim=1)
elif self.skip_conn == 'add':
out = out + skips.pop()
out = up_conv(out)
if self.skip_conn == 'concat':
out = pt.cat([out, skips.pop()], dim=1)
elif self.skip_conn == 'add':
out = out + skips.pop()
out = self.final_conv(out)
return out
class Discriminator(pt.nn.Module):
def __init__(self, n_out, channel_factor=2, n_layers=7, activation=relu, kernel_size=(4, 4),
n_residual=(0, 0), max_channels=1024, input_channels=1, affine=False, **kwargs):
super(Discriminator, self).__init__()
self.layers = []
current_channels = input_channels
for depth_index in range(n_layers):
out_channels = channel_factor * 2 ** depth_index
if out_channels > max_channels: out_channels = max_channels
for res_index in range(n_residual[0]):
self.layers.append(
parts.ResBlock2d(current_channels, n_residual[1], kernel_size, activation=activation, affine=affine, **kwargs))
self.add_module('r-block{}-{}'.format(depth_index + 1, res_index + 1), self.layers[-1])
self.layers.append(
parts.GeneralConvolution(current_channels, out_channels, kernel_size, (2, 2), activation=activation,
padding=pt.nn.ReflectionPad2d, norm=pt.nn.InstanceNorm2d,
convolution=pt.nn.Conv2d, affine=affine, **kwargs))
self.add_module('conv{}'.format(depth_index + 1), self.layers[-1])
current_channels = out_channels
self.layers.append(parts.GlobalAveragePooling2d())
self.add_module('average-pooling', self.layers[-1])
self.layers.append(parts.Flatten())
self.add_module('flatten', self.layers[-1])
self.layers.append(pt.nn.Linear(current_channels, n_out))
self.add_module('linear', self.layers[-1])
def forward(self, x):
out = x
for layer in self.layers:
out = layer(out)
out = pt.sigmoid(out) if out.shape[-1] == 1 else pt.softmax(out, -1)
return out
class ImagePool(pt.nn.Module):
def __init__(self, size, shape, write_probability=1):
super(ImagePool, self).__init__()
self.pool = Parameter(pt.rand(size, *shape), False)
self.write_probability = write_probability
def write(self, item):
if pt.rand(1) <= self.write_probability:
if item.shape[0] > 1:
item = item[self.random_index(len(item)), ...]
self.pool[self.random_index()] = item.detach()
def sample(self, batch_size):
samples = []
for _ in range(batch_size):
samples.append(self.pool[self.random_index()])
samples = pt.cat(samples)
return samples
def random_index(self, size=None):
size = size if size is not None else len(self.pool)
return (pt.rand(1)*size).long()
class CycleGAN(pt.nn.Module):
def __init__(self, generator, discriminator, input_size, pool_size=64, pool_write_probability=1):
super(CycleGAN, self).__init__()
self.discriminator = {'hn': discriminator[0](n_out=1, **discriminator[1]),
'ln': discriminator[0](n_out=1, **discriminator[1])}
self.generator = {'hn': generator[0](**generator[1]),
'ln': generator[0](**generator[1])}
self.add_module('discriminator_ln', self.discriminator['ln'])
self.add_module('discriminator_hn', self.discriminator['hn'])
self.add_module('generator_ln', self.generator['ln'])
self.add_module('generator_hn', self.generator['hn'])
self.pool_size = pool_size
self.pool_write_probability = pool_write_probability
self.pool = {'ln': ImagePool(self.pool_size, input_size, write_probability=self.pool_write_probability),
'hn': ImagePool(self.pool_size, input_size, write_probability=self.pool_write_probability)}
self.add_module('pool_ln', self.pool['ln'])
self.add_module('pool_hn', self.pool['hn'])
def generate(self, x, quality):
return self.generator[quality](x)
def discriminate(self, x, quality):
return self.discriminator[quality](x)
def cycle(self, x, start_quality):
other = 'hn' if start_quality == 'ln' else 'ln'
return self.generate(self.generate(x, other), start_quality)
def discriminate_from_pool(self, quality, batch_size):
return self.discriminate(self.pool[quality].sample(batch_size), quality)
def _forward(self, x, target_quality):
start_quality = 'hn' if target_quality == 'ln' else 'ln'
generated = self.generate(x, target_quality)
self.pool[target_quality].write(generated)
real = self.discriminate(x, start_quality)
fake = self.discriminate(generated, target_quality)
pool_fake = self.discriminate_from_pool(target_quality, len(generated))
# CAUTION: real score is for a sample from the other domain
scores = namedtuple('scores', ('real', 'fake', 'pool_fake'))
prediction = namedtuple(target_quality, ('generated', 'scores'))
return prediction(generated, scores(real, fake, pool_fake))
def forward(self, x):
hn, ln = x
generated_ln, prediction_ln = self._forward(hn, 'ln')
generated_hn, prediction_hn = self._forward(ln, 'hn')
cycled = self.generate(generated_ln, 'hn'), self.generate(generated_hn, 'ln')
result = namedtuple('Result', ('cycled', 'hn_scores', 'ln_scores'))
return result(cycled, prediction_hn, prediction_ln)
class HDCycleGAN(pt.nn.Module):
def __init__(self, generator, discriminator, input_size, pool_size=64, pool_write_probability=1):
super(HDCycleGAN, self).__init__()
self.discriminator = discriminator[0](n_out=3, **discriminator[1])
self.generator = {'hn': generator[0](**generator[1]),
'ln': generator[0](**generator[1])}
self.add_module('discriminator', self.discriminator)
self.add_module('generator_ln', self.generator['ln'])
self.add_module('generator_hn', self.generator['hn'])
self.pool_size = pool_size
self.pool_write_probability = pool_write_probability
self.pool = {'ln': ImagePool(self.pool_size, input_size, write_probability=self.pool_write_probability),
'hn': ImagePool(self.pool_size, input_size, write_probability=self.pool_write_probability)}
self.add_module('pool_ln', self.pool['ln'])
self.add_module('pool_hn', self.pool['hn'])
def generate(self, x, quality):
return self.generator[quality](x)
def discriminate(self, x):
return self.discriminator(x)
def cycle(self, x, start_quality):
other = 'hn' if start_quality == 'ln' else 'ln'
return self.generate(self.generate(x, other), start_quality)
def discriminate_from_pool(self, quality, batch_size):
return self.discriminate(self.pool[quality].sample(batch_size))
def _forward(self, x, target_quality):
generated = self.generate(x, target_quality)
self.pool[target_quality].write(generated)
real = self.discriminate(x)
fake = self.discriminate(generated)
pool_fake = self.discriminate_from_pool(target_quality, len(generated))
scores = namedtuple('scores', ('real', 'fake', 'pool_fake')) # caution: real score is for a sample from the other domain
prediction = namedtuple(target_quality, ('generated', 'scores'))
return prediction(generated, scores(real, fake, pool_fake))
def forward(self, x):
hn, ln = x
generated_ln, prediction_ln = self._forward(hn, 'ln')
generated_hn, prediction_hn = self._forward(ln, 'hn')
cycled = self.generate(generated_ln, 'hn'), self.generate(generated_hn, 'ln')
result = namedtuple('Result', ('cycled', 'hn_scores', 'ln_scores'))
return result(cycled, prediction_hn, prediction_ln)