-
Notifications
You must be signed in to change notification settings - Fork 1
/
models.py
196 lines (184 loc) · 7.89 KB
/
models.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
import torch
from torch.nn import functional as F
from torch import nn
from torchvision.models import vgg16, VGG16_Weights, vgg19, VGG19_Weights
from torchvision import transforms
# based on https://nbviewer.org/github/amanchadha/coursera-gan-specialization/blob/main/C3%20-%20Apply%20Generative%20Adversarial%20Network%20(GAN)/Week%202/C3W2A_Assignment.ipynb
class ContractingStack(nn.Module):
"""
Class for contracting stack of the UNet (encoder). Repeats a sequence of
conv2d and leaky relu with alternative batch norm.
"""
def __init__(self, in_channels, out_channels, batch_norm=False, pre_activation=False):
super(ContractingStack, self).__init__()
self.conv1 = nn.Conv2d(in_channels, out_channels, 3, padding = "same")
self.conv2 = nn.Conv2d(out_channels, out_channels, 3, padding = "same")
self.activation = nn.ReLU()
if batch_norm:
if pre_activation:
self.bn1 = nn.BatchNorm2d(in_channels)
self.bn2 = nn.BatchNorm2d(out_channels)
else:
self.bn = nn.BatchNorm2d(out_channels)
self.batch_norm = batch_norm
self.pre_activation = pre_activation
def forward(self, x):
if self.pre_activation:
if self.batch_norm:
x = self.bn1(x)
x = self.activation(x)
x = self.conv1(x)
if self.batch_norm:
x = self.bn2(x)
x = self.activation(x)
x = self.conv2(x)
else:
x = self.conv1(x)
if self.batch_norm:
x = self.bn(x)
x = self.activation(x)
x = self.conv2(x)
if self.batch_norm:
x = self.bn(x)
x = self.activation(x)
return x
class MiddleStack(nn.Module):
"""
Class for middle stack of UNet. Repeats a sequence of conv2d and relu.
"""
def __init__(self, channels, batch_norm=False, pre_activation=False):
super(MiddleStack, self).__init__()
self.conv1 = nn.Conv2d(channels//2, channels, 3, padding = "same")
self.conv2 = nn.Conv2d(channels, channels//2, 3, padding = "same")
self.activation = nn.ReLU()
if batch_norm:
self.bn1 = nn.BatchNorm2d(channels//2)
self.bn2 = nn.BatchNorm2d(channels)
self.batch_norm = batch_norm
self.pre_activation = pre_activation
def forward(self, x):
if self.pre_activation:
if self.batch_norm:
x = self.bn1(x)
x = self.activation(x)
x = self.conv1(x)
if self.batch_norm:
x = self.bn2(x)
x = self.activation(x)
x = self.conv2(x)
else:
x = self.conv1(x)
if self.batch_norm:
x = self.bn2(x)
x = self.activation(x)
x = self.conv2(x)
if self.batch_norm:
x = self.bn1(x)
x = self.activation(x)
return x
class ExpandingStack(nn.Module):
"""
Class for expanding stack of the UNet (decoder). Repeats a squence of
upsampling, conv2d, and contacenation followed by conv2d and relu twice, with
alternative batch norm.
"""
def __init__(self, in_channels, out_channels, batch_norm=False, pre_activation=False):
super(ExpandingStack, self).__init__()
self.upsample = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)
self.conv1 = nn.Conv2d(2*in_channels, in_channels, 3, padding = "same")
self.conv2 = nn.Conv2d(in_channels, out_channels, 3, padding = "same")
self.activation = nn.ReLU()
if batch_norm:
if pre_activation:
self.bn1 = nn.BatchNorm2d(2*in_channels)
self.bn2 = nn.BatchNorm2d(in_channels)
else:
self.bn1 = nn.BatchNorm2d(in_channels)
self.bn2 = nn.BatchNorm2d(out_channels)
self.batch_norm = batch_norm
self.pre_activation = pre_activation
def forward(self, x, x_skip):
if self.pre_activation:
x = self.upsample(x)
x = torch.cat([x, x_skip],1)
if self.batch_norm:
x = self.bn1(x)
x = self.activation(x)
x = self.conv1(x)
if self.batch_norm:
x = self.bn2(x)
x = self.activation(x)
x = self.conv2(x)
else:
x = self.upsample(x)
x = torch.cat([x, x_skip],1)
x = self.conv1(x)
if self.batch_norm:
x = self.bn1(x)
x = self.activation(x)
x = self.conv2(x)
if self.batch_norm:
x = self.bn2(x)
x = self.activation(x)
return x
class UNet_alt(nn.Module):
'''
Class implementing UNet with 4x contracting stacks and corresponding
4x expanding stacks. Upconvolution has been replaced with upsampling.
Based on UNet used in Kim et al. 2019.
'''
def __init__(self, input_channels, hidden_channels=64, norm=False, skip=False, pre=False):
super(UNet_alt, self).__init__()
self.first = nn.Conv2d(input_channels, hidden_channels, 1, padding="same")
self.c1 = ContractingStack(hidden_channels,hidden_channels,batch_norm=norm,pre_activation=pre)
self.c2 = ContractingStack(hidden_channels,hidden_channels*2,batch_norm=norm,pre_activation=pre)
self.c3 = ContractingStack(hidden_channels*2,hidden_channels*4,batch_norm=norm,pre_activation=pre)
self.c4 = ContractingStack(hidden_channels*4,hidden_channels*8,batch_norm=norm,pre_activation=pre)
self.middle = MiddleStack(hidden_channels*16,batch_norm=norm,pre_activation=pre)
self.e1 = ExpandingStack(hidden_channels*8,hidden_channels*4,batch_norm=norm,pre_activation=pre)
self.e2 = ExpandingStack(hidden_channels*4,hidden_channels*2,batch_norm=norm,pre_activation=pre)
self.e3 = ExpandingStack(hidden_channels*2,hidden_channels,batch_norm=norm,pre_activation=pre)
self.e4 = ExpandingStack(hidden_channels,hidden_channels,batch_norm=norm,pre_activation=pre)
self.final = nn.Conv2d(hidden_channels, input_channels, 1, padding="same")
self.maxpool = nn.MaxPool2d(2)
self.skip = skip
def forward(self, x):
n = x.clone()
n0 = self.first(x)
n1 = self.c1(n0)
n2 = self.c2(self.maxpool(n1))
n3 = self.c3(self.maxpool(n2))
n4 = self.c4(self.maxpool(n3))
n5 = self.middle(self.maxpool(n4))
n6 = self.e1(n5,n4)
n7 = self.e2(n6,n3)
n8 = self.e3(n7,n2)
n9 = self.e4(n8,n1)
n10 = self.final(n9)
if self.skip:
return n+n10
else:
return n10
# Johnson et al. 2016:
# relu1_2 (4), relu2_2 (9), relu3_3 (16), relu4_3 (23)
# Kim et al. j = relu_j + 1
class VGG_Feature_Extractor_16(torch.nn.Module):
def __init__(self, layer=24, n_mat = 2, requires_grad=False):
super(VGG_Feature_Extractor_16, self).__init__()
vgg_pretrained_features = vgg16(weights=VGG16_Weights.IMAGENET1K_V1).features
self.n_mat = n_mat
self.slice = torch.nn.Sequential()
for x in range(layer):
self.slice.add_module(str(x), vgg_pretrained_features[x])
if not requires_grad:
for param in self.parameters():
param.requires_grad = False
def forward(self, X):
vgg_transform = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
if self.n_mat == 3:
out = self.slice(vgg_transform(X))
elif self.n_mat == 2:
out = torch.cat((self.slice(vgg_transform(X[:,0:1,:,:].repeat(1,3,1,1))),self.slice(vgg_transform(X[:,1:,:,:].repeat(1,3,1,1)))),dim=1)
else:
out = self.slice(vgg_transform(X[:,0:1,:,:].repeat(1,3,1,1)))
return out