/
mv2_dilate_unet.py
178 lines (142 loc) · 5.65 KB
/
mv2_dilate_unet.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
'''
mobilenetv2_dilate_unet
Author: Zhengwei Li
Data: July 20 2018
'''
import torch
import torch.nn as nn
import torch.nn.functional as F
def conv_bn(inp, oup, stride):
return nn.Sequential(
nn.Conv2d(inp, oup, 3, stride, 1, bias=False),
nn.BatchNorm2d(oup),
nn.ReLU6(inplace=True)
)
class InvertedResidual(nn.Module):
def __init__(self, inp, oup, stride, expand_ratio):
super(InvertedResidual, self).__init__()
self.stride = stride
assert stride in [1, 2]
self.use_res_connect = self.stride == 1 and inp == oup
self.conv = nn.Sequential(
# pw
nn.Conv2d(inp, inp * expand_ratio, 1, 1, 0, bias=False),
nn.BatchNorm2d(inp * expand_ratio),
nn.ReLU6(inplace=True),
# dw
nn.Conv2d(inp * expand_ratio, inp * expand_ratio, 3, stride, 1, groups=inp * expand_ratio, bias=False),
nn.BatchNorm2d(inp * expand_ratio),
nn.ReLU6(inplace=True),
# pw-linear
nn.Conv2d(inp * expand_ratio, oup, 1, 1, 0, bias=False),
nn.BatchNorm2d(oup),
)
def forward(self, x):
if self.use_res_connect:
return x + self.conv(x)
else:
return self.conv(x)
#-------------------------------------------------------------------------------------------------
# MobileNet_v2_os_32_MFo
#--------------------
class MobileNet_v2_os_32_MFo(nn.Module):
def __init__(self, nInputChannels=3):
super(MobileNet_v2_os_32_MFo, self).__init__()
# 1/2
# 256 x 256
self.head_conv = conv_bn(nInputChannels, 32, 2)
# 1/2
# 256 x 256
self.block_1 = InvertedResidual(32, 16, 1, 1)
# 1/4 128 x 128
self.block_2 = nn.Sequential(
InvertedResidual(16, 24, 2, 6),
InvertedResidual(24, 24, 1, 6)
)
# 1/8 64 x 64
self.block_3 = nn.Sequential(
InvertedResidual(24, 32, 2, 6),
InvertedResidual(32, 32, 1, 6),
InvertedResidual(32, 32, 1, 6)
)
# 1/16 32 x 32
self.block_4 = nn.Sequential(
InvertedResidual(32, 64, 2, 6),
InvertedResidual(64, 64, 1, 6),
InvertedResidual(64, 64, 1, 6),
InvertedResidual(64, 64, 1, 6)
)
# 1/16 32 x 32
self.block_5 = nn.Sequential(
InvertedResidual(64, 96, 1, 6),
InvertedResidual(96, 96, 1, 6),
InvertedResidual(96, 96, 1, 6)
)
# 1/32 16 x 16
self.block_6 = nn.Sequential(
InvertedResidual(96, 160, 2, 6),
InvertedResidual(160, 160, 1, 6),
InvertedResidual(160, 160, 1, 6)
)
# 1/32 16 x 16
self.block_7 = InvertedResidual(160, 320, 1, 6)
def forward(self, x):
x = self.head_conv(x)
x1 = self.block_1(x)
x2 = self.block_2(x1)
x3 = self.block_3(x2)
x4 = self.block_4(x3)
x4 = self.block_5(x4)
x5 = self.block_6(x4)
x5 = self.block_7(x5)
return x1, x2, x3, x4, x5
# up amd concat and dilate
class UCD(nn.Module):
def __init__(self, inplanes, planes, dilation):
super(UCD, self).__init__()
self.up = nn.ConvTranspose2d(inplanes, planes, kernel_size=2, stride=2, padding=0)
self.aspp = nn.Sequential(nn.Conv2d(planes*2, planes*2, kernel_size=3,
stride=1, padding=dilation, dilation=dilation, bias=False),
nn.BatchNorm2d(planes*2))
def forward(self, e, x):
x = self.up(x)
x = torch.cat((x, e), dim=1)
x = self.aspp(x)
return x
#-------------------------------------------------------------------------------------------------
# mv2_dilate_unet
# feature exstractor : MobileNet_v2_os_32_MFo
#-----------------------------------------
class MobileNet_v2_Dilate_Unet(nn.Module):
def __init__(self, nInputChannels=3, n_classes=1):
super(MobileNet_v2_Dilate_Unet, self).__init__()
# mobilenetv2 feature
self.mobilenet_features = MobileNet_v2_os_32_MFo(nInputChannels)
self.up_concat_dilate_1 = UCD(320, 96, dilation = 2)
self.up_concat_dilate_2 = UCD(192, 32, dilation = 6)
self.up_concat_dilate_3 = UCD(64, 24, dilation = 12)
self.up_concat_dilate_4 = UCD(48, 16, dilation = 18)
self.last_conv = nn.Sequential(nn.Conv2d(32, 32, kernel_size=3, stride=1, padding=1, bias=False),
nn.BatchNorm2d(32),
nn.Conv2d(32, n_classes, kernel_size=1, stride=1))
# init weights
self._init_weight()
def _init_weight(self):
for m in self.modules():
if isinstance(m, nn.Conv2d) or isinstance(m, nn.Linear):
nn.init.xavier_normal_(m.weight)
if m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
def forward(self, x):
# x1 - x4 : 1/8 64 x 64
e1, e2, e3, e4, feature_map = self.mobilenet_features(x)
feature_map = self.up_concat_dilate_1(e4, feature_map)
feature_map = self.up_concat_dilate_2(e3, feature_map)
feature_map = self.up_concat_dilate_3(e2, feature_map)
feature_map = self.up_concat_dilate_4(e1, feature_map)
heat_map = self.last_conv(feature_map)
heat_map = F.upsample(heat_map, scale_factor=2, mode='bilinear', align_corners=True)
return heat_map