-
Notifications
You must be signed in to change notification settings - Fork 0
/
multiresunet.py
190 lines (148 loc) · 7.27 KB
/
multiresunet.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
import torch
def Conv2dSame(in_channels, out_channels, kernel_size, use_bias=True, padding_layer=torch.nn.ReflectionPad2d):
ka = kernel_size // 2
kb = ka - 1 if kernel_size % 2 == 0 else ka
return [
padding_layer((ka, kb, ka, kb)),
torch.nn.Conv2d(in_channels, out_channels, kernel_size, bias=use_bias)
]
def conv2d_bn(in_channels, filters, kernel_size, padding='same', activation='relu'):
assert padding == 'same'
affine = False if activation == 'relu' or activation == 'sigmoid' else True
sequence = []
sequence += Conv2dSame(in_channels, filters, kernel_size, use_bias=False)
sequence += [torch.nn.BatchNorm2d(filters, affine=affine)]
if activation == "relu":
sequence += [torch.nn.ReLU()]
elif activation == "sigmoid":
sequence += [torch.nn.Sigmoid()]
elif activation == 'tanh':
sequence += [torch.nn.Tanh()]
return torch.nn.Sequential(*sequence)
class MultiResBlock(torch.nn.Module):
def __init__(self, in_channels, u, alpha=1.67, use_dropout=False):
super().__init__()
w = alpha * u
self.out_channel = int(w * 0.167) + int(w * 0.333) + int(w * 0.5)
self.conv2d_bn = conv2d_bn(in_channels, self.out_channel, 1, activation=None)
self.conv3x3 = conv2d_bn(in_channels, int(w * 0.167), 3, activation='relu')
self.conv5x5 = conv2d_bn(int(w * 0.167), int(w * 0.333), 3, activation='relu')
self.conv7x7 = conv2d_bn(int(w * 0.333), int(w * 0.5), 3, activation='relu')
self.bn_1 = torch.nn.BatchNorm2d(self.out_channel)
self.relu = torch.nn.ReLU()
self.bn_2 = torch.nn.BatchNorm2d(self.out_channel)
self.use_dropout = use_dropout
if use_dropout:
self.dropout = torch.nn.Dropout(0.5)
def forward(self, inp):
if self.use_dropout:
x = self.dropout(inp)
else:
x = inp
shortcut = self.conv2d_bn(x)
conv3x3 = self.conv3x3(x)
conv5x5 = self.conv5x5(conv3x3)
conv7x7 = self.conv7x7(conv5x5)
out = torch.cat([conv3x3, conv5x5, conv7x7], dim=1)
out = self.bn_1(out)
out = torch.add(shortcut, out)
out = self.relu(out)
out = self.bn_2(out)
return out
class ResPathBlock(torch.nn.Module):
def __init__(self, in_channels, filters):
super(ResPathBlock, self).__init__()
self.conv2d_bn1 = conv2d_bn(in_channels, filters, 1, activation=None)
self.conv2d_bn2 = conv2d_bn(in_channels, filters, 3, activation='relu')
self.relu = torch.nn.ReLU()
self.bn = torch.nn.BatchNorm2d(filters)
def forward(self, inp):
shortcut = self.conv2d_bn1(inp)
out = self.conv2d_bn2(inp)
out = torch.add(shortcut, out)
out = self.relu(out)
out = self.bn(out)
return out
class ResPath(torch.nn.Module):
def __init__(self, in_channels, filters, length):
super(ResPath, self).__init__()
self.first_block = ResPathBlock(in_channels, filters)
self.blocks = torch.nn.Sequential(*[ResPathBlock(filters, filters) for i in range(length - 1)])
def forward(self, inp):
out = self.first_block(inp)
out = self.blocks(out)
return out
class MultiResUnet(torch.nn.Module):
def __init__(self, in_channels, out_channels, nf=32, use_dropout=False):
super(MultiResUnet, self).__init__()
self.mres_block1 = MultiResBlock(in_channels, u=nf)
self.pool = torch.nn.MaxPool2d(kernel_size=2)
self.res_path1 = ResPath(self.mres_block1.out_channel, nf, 4)
self.mres_block2 = MultiResBlock(self.mres_block1.out_channel, u=nf * 2)
# self.pool2 = torch.nn.MaxPool2d(kernel_size=2)
self.res_path2 = ResPath(self.mres_block2.out_channel, nf * 2, 3)
self.mres_block3 = MultiResBlock(self.mres_block2.out_channel, u=nf * 4)
# self.pool3 = torch.nn.MaxPool2d(kernel_size=2)
self.res_path3 = ResPath(self.mres_block3.out_channel, nf * 4, 2)
self.mres_block4 = MultiResBlock(self.mres_block3.out_channel, u=nf * 8)
# self.pool4 = torch.nn.MaxPool2d(kernel_size=2)
self.res_path4 = ResPath(self.mres_block4.out_channel, nf * 8, 1)
self.mres_block5 = MultiResBlock(self.mres_block4.out_channel, u=nf * 16)
self.deconv1 = torch.nn.ConvTranspose2d(self.mres_block5.out_channel, nf * 8, (2, 2), (2, 2))
self.mres_block6 = MultiResBlock(nf * 8 + nf * 8, u=nf * 8, use_dropout=use_dropout)
# MultiResBlock(nf * 8 + self.mres_block4.out_channel, u=nf * 8)
self.deconv2 = torch.nn.ConvTranspose2d(self.mres_block6.out_channel, nf * 4, (2, 2), (2, 2))
self.mres_block7 = MultiResBlock(nf * 4 + nf * 4, u=nf * 4, use_dropout=use_dropout)
# MultiResBlock(nf * 4 + self.mres_block3.out_channel, u=nf * 4)
self.deconv3 = torch.nn.ConvTranspose2d(self.mres_block7.out_channel, nf * 2, (2, 2), (2, 2))
self.mres_block8 = MultiResBlock(nf * 2 + nf * 2, u=nf * 2, use_dropout=use_dropout)
# MultiResBlock(nf * 2 + self.mres_block2.out_channel, u=nf * 2)
self.deconv4 = torch.nn.ConvTranspose2d(self.mres_block8.out_channel, nf, (2, 2), (2, 2))
self.mres_block9 = MultiResBlock(nf + nf, u=nf)
# MultiResBlock(nf + self.mres_block1.out_channel, u=nf)
self.conv10 = conv2d_bn(self.mres_block9.out_channel, out_channels, 1, padding='same', activation='tanh')
def forward(self, inp):
mresblock1 = self.mres_block1(inp)
pool = self.pool(mresblock1)
mresblock1 = self.res_path1(mresblock1)
mresblock2 = self.mres_block2(pool)
pool = self.pool(mresblock2)
mresblock2 = self.res_path2(mresblock2)
mresblock3 = self.mres_block3(pool)
pool = self.pool(mresblock3)
mresblock3 = self.res_path3(mresblock3)
mresblock4 = self.mres_block4(pool)
pool = self.pool(mresblock4)
mresblock4 = self.res_path4(mresblock4)
mresblock = self.mres_block5(pool)
up = torch.cat([self.deconv1(mresblock), mresblock4], dim=1)
mresblock = self.mres_block6(up)
up = torch.cat([self.deconv2(mresblock), mresblock3], dim=1)
mresblock = self.mres_block7(up)
up = torch.cat([self.deconv3(mresblock), mresblock2], dim=1)
mresblock = self.mres_block8(up)
up = torch.cat([self.deconv4(mresblock), mresblock1], dim=1)
mresblock = self.mres_block9(up)
conv10 = self.conv10(mresblock)
return conv10
class MultiResUnetGenerator(torch.nn.Module):
def __init__(self, input_nc, output_nc, ngf=64, use_dropout=False, gpu_ids=[]):
super(MultiResUnetGenerator, self).__init__()
self.gpu_ids = gpu_ids
self.model = MultiResUnet(input_nc, output_nc, nf=ngf, use_dropout=use_dropout)
def forward(self, inp):
if self.gpu_ids and isinstance(inp.data, torch.cuda.FloatTensor):
return torch.nn.parallel.data_parallel(self.model, inp, self.gpu_ids)
else:
return self.model(inp)
def weights_init_uniform_rule(m):
classname = m.__class__.__name__
# for every Linear layer in a model..
if classname == 'Conv2d':
pass
print(classname)
# a = ResPath(10, 100,3)
# a.apply(weights_init_uniform_rule)
#a = MultiResUnet(512, 512, 3)
#x = torch.randn(2, 3, 512, 512)
#print(a(x).shape)