-
Notifications
You must be signed in to change notification settings - Fork 2
/
model_n2n.py
228 lines (208 loc) · 9.31 KB
/
model_n2n.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
import torch
import torch.nn as nn
def conv3x3(in_channels, out_channels):
return nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1, bias=False)
class UpsamplingNearest2d(nn.Module):
def __init__(self, scale_factor=2):
super().__init__()
self.scale_factor = scale_factor
def forward(self, x):
return nn.functional.interpolate(x, scale_factor=self.scale_factor, mode='nearest')
class N2N(nn.Module):
"""https://github.com/NVlabs/noise2noise
Lehtinen, Jaakko, et al. "Noise2Noise: Learning Image Restoration without
Clean Data." arXiv preprint arXiv:1803.04189 (2018).
"""
def __init__(self, in_channels, out_channels):
super(N2N, self).__init__()
self.enc_conv0 = conv3x3(in_channels, 48)
self.enc_relu0 = nn.LeakyReLU(0.1)
self.enc_conv1 = conv3x3(48, 48)
self.enc_relu1 = nn.LeakyReLU(0.1)
self.pool1 = nn.MaxPool2d(kernel_size=2)
# 128
self.enc_conv2 = conv3x3(48, 48)
self.enc_relu2 = nn.LeakyReLU(0.1)
self.pool2 = nn.MaxPool2d(kernel_size=2)
# 64
self.enc_conv3 = conv3x3(48, 48)
self.enc_relu3 = nn.LeakyReLU(0.1)
self.pool3 = nn.MaxPool2d(kernel_size=2)
# 32
self.enc_conv4 = conv3x3(48, 48)
self.enc_relu4 = nn.LeakyReLU(0.1)
self.pool4 = nn.MaxPool2d(kernel_size=2)
# 16
self.enc_conv5 = conv3x3(48, 48)
self.enc_relu5 = nn.LeakyReLU(0.1)
self.pool5 = nn.MaxPool2d(kernel_size=2)
# 8
self.enc_conv6 = conv3x3(48, 48)
self.enc_relu6 = nn.LeakyReLU(0.1)
self.upsample5 = UpsamplingNearest2d(scale_factor=2)
# 16
self.dec_conv5a = conv3x3(96, 96)
self.dec_relu5a = nn.LeakyReLU(0.1)
self.dec_conv5b = conv3x3(96, 96)
self.dec_relu5b = nn.LeakyReLU(0.1)
self.upsample4 = UpsamplingNearest2d(scale_factor=2)
# 32
self.dec_conv4a = conv3x3(144, 96)
self.dec_relu4a = nn.LeakyReLU(0.1)
self.dec_conv4b = conv3x3(96, 96)
self.dec_relu4b = nn.LeakyReLU(0.1)
self.upsample3 = UpsamplingNearest2d(scale_factor=2)
# 64
self.dec_conv3a = conv3x3(144, 96)
self.dec_relu3a = nn.LeakyReLU(0.1)
self.dec_conv3b = conv3x3(96, 96)
self.dec_relu3b = nn.LeakyReLU(0.1)
self.upsample2 = UpsamplingNearest2d(scale_factor=2)
# 128
self.dec_conv2a = conv3x3(144, 96)
self.dec_relu2a = nn.LeakyReLU(0.1)
self.dec_conv2b = conv3x3(96, 96)
self.dec_relu2b = nn.LeakyReLU(0.1)
self.upsample1 = UpsamplingNearest2d(scale_factor=2)
# 256
self.dec_conv1a = conv3x3(96 + in_channels, 64)
self.dec_relu1a = nn.LeakyReLU(0.1)
self.dec_conv1b = conv3x3(64, 32)
self.dec_relu1b = nn.LeakyReLU(0.1)
self.dec_conv1c = conv3x3(32, out_channels)
def forward(self, x):
out_pool1 = self.pool1(self.enc_relu1(self.enc_conv1(self.enc_relu0(self.enc_conv0(x)))))
out_pool2 = self.pool2(self.enc_relu2(self.enc_conv2(out_pool1)))
out_pool3 = self.pool3(self.enc_relu3(self.enc_conv3(out_pool2)))
out_pool4 = self.pool4(self.enc_relu4(self.enc_conv4(out_pool3)))
out_pool5 = self.pool5(self.enc_relu5(self.enc_conv5(out_pool4)))
out = self.upsample5(self.enc_relu6(self.enc_conv6(out_pool5)))
out = self.upsample4(
self.dec_relu5b(self.dec_conv5b(self.dec_relu5a(self.dec_conv5a(torch.cat((out, out_pool4), 1))))))
out = self.upsample3(
self.dec_relu4b(self.dec_conv4b(self.dec_relu4a(self.dec_conv4a(torch.cat((out, out_pool3), 1))))))
out = self.upsample2(
self.dec_relu3b(self.dec_conv3b(self.dec_relu3a(self.dec_conv3a(torch.cat((out, out_pool2), 1))))))
out = self.upsample1(
self.dec_relu2b(self.dec_conv2b(self.dec_relu2a(self.dec_conv2a(torch.cat((out, out_pool1), 1))))))
out = self.dec_conv1c(
self.dec_relu1b(self.dec_conv1b(self.dec_relu1a(self.dec_conv1a(torch.cat((out, x), 1))))))
return out
class UnetN2Nv2(nn.Module):
"""Add BatchNorm and Tanh
Lehtinen, Jaakko, et al. "Noise2Noise: Learning Image Restoration without
Clean Data." arXiv preprint arXiv:1803.04189 (2018).
Add BatchNorm and Tanh out activation
"""
def __init__(self, in_channels, out_channels):
super(UnetN2Nv2, self).__init__()
self.enc_conv0 = conv3x3(in_channels, 48)
self.enc_relu0 = nn.LeakyReLU(0.1)
self.enc_conv1 = conv3x3(48, 48)
self.enc_bn1 = nn.BatchNorm2d(48)
self.enc_relu1 = nn.LeakyReLU(0.1)
self.pool1 = nn.MaxPool2d(kernel_size=2)
# 128
self.enc_conv2 = conv3x3(48, 48)
self.enc_bn2 = nn.BatchNorm2d(48)
self.enc_relu2 = nn.LeakyReLU(0.1)
self.pool2 = nn.MaxPool2d(kernel_size=2)
# 64
self.enc_conv3 = conv3x3(48, 48)
self.enc_bn3 = nn.BatchNorm2d(48)
self.enc_relu3 = nn.LeakyReLU(0.1)
self.pool3 = nn.MaxPool2d(kernel_size=2)
# 32
self.enc_conv4 = conv3x3(48, 48)
self.enc_bn4 = nn.BatchNorm2d(48)
self.enc_relu4 = nn.LeakyReLU(0.1)
self.pool4 = nn.MaxPool2d(kernel_size=2)
# 16
self.enc_conv5 = conv3x3(48, 48)
self.enc_bn5 = nn.BatchNorm2d(48)
self.enc_relu5 = nn.LeakyReLU(0.1)
self.pool5 = nn.MaxPool2d(kernel_size=2)
# 8
self.enc_conv6 = conv3x3(48, 48)
self.enc_bn6 = nn.BatchNorm2d(48)
self.enc_relu6 = nn.LeakyReLU(0.1)
self.upsample5 = UpsamplingNearest2d(scale_factor=2)
# 16
self.dec_conv5a = conv3x3(96, 96)
self.dec_bn5a = nn.BatchNorm2d(96)
self.dec_relu5a = nn.LeakyReLU(0.1)
self.dec_conv5b = conv3x3(96, 96)
self.dec_bn5b = nn.BatchNorm2d(96)
self.dec_relu5b = nn.LeakyReLU(0.1)
self.upsample4 = UpsamplingNearest2d(scale_factor=2)
# 32
self.dec_conv4a = conv3x3(144, 96)
self.dec_bn4a = nn.BatchNorm2d(96)
self.dec_relu4a = nn.LeakyReLU(0.1)
self.dec_conv4b = conv3x3(96, 96)
self.dec_bn4b = nn.BatchNorm2d(96)
self.dec_relu4b = nn.LeakyReLU(0.1)
self.upsample3 = UpsamplingNearest2d(scale_factor=2)
# 64
self.dec_conv3a = conv3x3(144, 96)
self.dec_bn3a = nn.BatchNorm2d(96)
self.dec_relu3a = nn.LeakyReLU(0.1)
self.dec_conv3b = conv3x3(96, 96)
self.dec_bn3b = nn.BatchNorm2d(96)
self.dec_relu3b = nn.LeakyReLU(0.1)
self.upsample2 = UpsamplingNearest2d(scale_factor=2)
# 128
self.dec_conv2a = conv3x3(144, 96)
self.dec_bn2a = nn.BatchNorm2d(96)
self.dec_relu2a = nn.LeakyReLU(0.1)
self.dec_conv2b = conv3x3(96, 96)
self.dec_bn2b = nn.BatchNorm2d(96)
self.dec_relu2b = nn.LeakyReLU(0.1)
self.upsample1 = UpsamplingNearest2d(scale_factor=2)
# 256
self.dec_conv1a = conv3x3(96 + in_channels, 64)
self.dec_bn1a = nn.BatchNorm2d(64)
self.dec_relu1a = nn.LeakyReLU(0.1)
self.dec_conv1b = conv3x3(64, 32)
self.dec_bn1b = nn.BatchNorm2d(32)
self.dec_relu1b = nn.LeakyReLU(0.1)
self.dec_conv1c = conv3x3(32, out_channels)
self.dec_act = nn.Tanh()
def forward(self, x):
out_pool1 = self.pool1(self.enc_relu1(self.enc_bn1(self.enc_conv1(self.enc_relu0(self.enc_conv0(x))))))
out_pool2 = self.pool2(self.enc_relu2(self.enc_bn2(self.enc_conv2(out_pool1))))
out_pool3 = self.pool3(self.enc_relu3(self.enc_bn3(self.enc_conv3(out_pool2))))
out_pool4 = self.pool4(self.enc_relu4(self.enc_bn4(self.enc_conv4(out_pool3))))
out_pool5 = self.pool5(self.enc_relu5(self.enc_bn5(self.enc_conv5(out_pool4))))
out = self.upsample5(self.enc_relu6(self.enc_bn6(self.enc_conv6(out_pool5))))
out = self.upsample4(self.dec_relu5b(self.dec_bn5b(
self.dec_conv5b(self.dec_relu5a(self.dec_bn5a(self.dec_conv5a(torch.cat((out, out_pool4), 1))))))))
out = self.upsample3(self.dec_relu4b(self.dec_bn4b(
self.dec_conv4b(self.dec_relu4a(self.dec_bn4a(self.dec_conv4a(torch.cat((out, out_pool3), 1))))))))
out = self.upsample2(self.dec_relu3b(self.dec_bn3b(
self.dec_conv3b(self.dec_relu3a(self.dec_bn3a(self.dec_conv3a(torch.cat((out, out_pool2), 1))))))))
out = self.upsample1(self.dec_relu2b(self.dec_bn2b(
self.dec_conv2b(self.dec_relu2a(self.dec_bn2a(self.dec_conv2a(torch.cat((out, out_pool1), 1))))))))
out = self.dec_conv1c(self.dec_relu1b(
self.dec_bn1b(self.dec_conv1b(self.dec_relu1a(self.dec_bn1a(self.dec_conv1a(torch.cat((out, x), 1))))))))
out = self.dec_act(out)
return out
@property
def model_size(self):
return self._model_size()
def _model_size(self):
n_params, n_conv_layers = 0, 0
for param in self.parameters():
n_params += param.numel()
for module in self.modules():
if 'Conv' in module.__class__.__name__ \
or 'conv' in module.__class__.__name__:
n_conv_layers += 1
return n_params, n_conv_layers
if __name__ == '__main__':
unet = N2N(3, 3).to('cuda:2')
print(unet)
print(unet.model_size)
x = torch.randn(16, 3, 256, 256).to('cuda:2')
y = unet(x)
print(y.shape)