-
Notifications
You must be signed in to change notification settings - Fork 0
/
unet.py
231 lines (215 loc) · 8.07 KB
/
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
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
import torch
import torch.nn as nn
import torchvision
class UNetFactory(nn.Module):
"""
本质上就是一个U型的网络,先encode,后decode,中间可能有架bridge。
其中encoder需要输出skip到decode那边做concatenate,使得decode阶段能补充信息。
bridge不能存在下采样和上采样的操作。
"""
def __init__(self, encoder_blocks, decoder_blocks, bridge=None):
super(UNetFactory, self).__init__()
self.encoder = UNetEncoder(encoder_blocks)
self.bridge = bridge
self.decoder = UNetDecoder(decoder_blocks)
def forward(self, x):
res = self.encoder(x)
out, skips = res[0], res[1:]
if self.bridge is not None:
out = self.bridge(out)
out = self.decoder(out, skips)
return out
class UNetEncoder(nn.Module):
"""
encoder会有多次下采样,下采样前的feature map要作为skip缓存起来将来送到decoder用。
这里约定,以下采样为界线,将encoder分成多个block,其中第一个block无下采样操作,后面的每个block内都
含有一次下采样操作。
"""
def __init__(self, blocks):
super(UNetEncoder, self).__init__()
assert len(blocks) > 0
self.blocks = nn.ModuleList(blocks)
def forward(self, x):
skips = []
for i in range(len(self.blocks) - 1):
x = self.blocks[i](x)
skips.append(x)
res = [self.blocks[i+1](x)]
res += skips
return res # 只能以这种方式返回多个tensor
class UNetDecoder(nn.Module):
"""
decoder会有多次上采样,每次上采样后,要跟相应的skip做concatenate。
这里约定,以上采样为界线,将decoder分成多个block,其中最后一个block无上采样操作,其他block内
都含有一次上采样。如此一来,除第一个block以外,其他block都先做concatenate。
"""
def __init__(self, blocks):
super(UNetDecoder, self).__init__()
assert len(blocks) > 1
self.blocks = nn.ModuleList(blocks)
def _center_crop(self, skip, x):
"""
skip和x,谁比较大,就裁剪谁
"""
_, _, h1, w1 = skip.shape
_, _, h2, w2 = x.shape
ht, wt = min(h1, h2), min(w1, w2)
dh1 = (h1 - ht) // 2 if h1 > ht else 0
dw1 = (w1 - wt) // 2 if w1 > wt else 0
dh2 = (h2 - ht) // 2 if h2 > ht else 0
dw2 = (w2 - wt) // 2 if w2 > wt else 0
return skip[:, :, dh1: (dh1 + ht), dw1: (dw1 + wt)], \
x[:, :, dh2: (dh2 + ht), dw2: (dw2 + wt)]
def forward(self, x, skips, reverse_skips=True):
assert len(skips) == len(self.blocks) - 1
if reverse_skips:
skips = skips[::-1]
x = self.blocks[0](x)
for i in range(1, len(self.blocks)):
skip, x = self._center_crop(skips[i-1], x)
x = torch.cat([skip, x], dim=1)
x = self.blocks[i](x)
return x
def unet_convs(in_channels, out_channels, padding=0):
"""
unet论文里出现次数最多的2个conv3x3(non-padding)的结构
"""
return nn.Sequential(
nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=padding, bias=False),
nn.BatchNorm2d(out_channels),
nn.ReLU(inplace=True),
nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=padding, bias=False),
nn.BatchNorm2d(out_channels),
nn.ReLU(inplace=True),
)
def unet(in_channels, out_channels):
"""
构造跟论文一致的unet网络
https://arxiv.org/abs/1505.04597
"""
# encoder
encoder_blocks = [
# two conv3x3
unet_convs(in_channels, 64),
# max pool 2x2, two conv3x3
nn.Sequential(
nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True),
unet_convs(64, 128)
),
# max pool 2x2, two conv3x3
nn.Sequential(
nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True),
unet_convs(128, 256)
),
# max pool 2x2, two conv3x3
nn.Sequential(
nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True),
unet_convs(256, 512)
),
# max pool 2x2
nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True)
]
# bridge
bridge = nn.Sequential(
# two conv3x3
unet_convs(512, 1024)
)
# decoder
decoder_blocks = [
# up-conv2x2
nn.ConvTranspose2d(1024, 512, kernel_size=2, stride=2),
# two conv3x3, up-conv2x2
nn.Sequential(
unet_convs(1024, 512),
nn.ConvTranspose2d(512, 256, kernel_size=2, stride=2),
),
# two conv3x3, up-conv2x2
nn.Sequential(
unet_convs(512, 256),
nn.ConvTranspose2d(256, 128, kernel_size=2, stride=2),
),
# two conv3x3, up-conv2x2
nn.Sequential(
unet_convs(256, 128),
nn.ConvTranspose2d(128, 64, kernel_size=2, stride=2),
),
# two conv3x3, conv1x1
nn.Sequential(
unet_convs(128, 64),
nn.Conv2d(64, out_channels, kernel_size=1)
)
]
return UNetFactory(encoder_blocks, decoder_blocks, bridge)
def unet_resnet(resnet_type, in_channels, out_channels, pretrained=True):
"""
利用resnet作为encoder,相应地,decoder也做一些改动,使得输出的尺寸跟原始的一致
"""
if resnet_type == 'resnet18':
resnet = torchvision.models.resnet.resnet18(pretrained)
encoder_out_channels = [in_channels, 64, 64, 128, 256, 512] # encoder各个block的输出channel
elif resnet_type == 'resnet34':
resnet = torchvision.models.resnet.resnet34(pretrained)
encoder_out_channels = [in_channels, 64, 64, 128, 256, 512]
elif resnet_type == 'resnet50':
resnet = torchvision.models.resnet.resnet50(pretrained)
encoder_out_channels = [in_channels, 64, 256, 512, 1024, 2048]
elif resnet_type == 'resnet101':
resnet = torchvision.models.resnet.resnet101(pretrained)
encoder_out_channels = [in_channels, 64, 256, 512, 1024, 2048]
elif resnet_type == 'resnet152':
resnet = torchvision.models.resnet.resnet152(pretrained)
encoder_out_channels = [in_channels, 64, 256, 512, 1024, 2048]
elif resnet_type == 'resnext50_32x4d':
resnet = torchvision.models.resnet.resnext50_32x4d(pretrained)
encoder_out_channels = [in_channels, 64, 256, 512, 1024, 2048]
else:
raise ValueError("unexpected resnet_type")
# encoder
encoder_blocks = [
# org input
nn.Sequential(),
# conv1
nn.Sequential(
resnet.conv1,
resnet.bn1,
resnet.relu
),
# conv2_x
nn.Sequential(
resnet.maxpool,
resnet.layer1
),
# conv3_x
resnet.layer2,
# conv4_x
resnet.layer3,
# conv5_x
resnet.layer4
]
# bridge
bridge = None # 感觉并无必要
# decoder
decoder_blocks = []
in_ch = encoder_out_channels[-1]
out_ch = in_ch // 2
decoder_blocks.append(nn.ConvTranspose2d(in_ch, out_ch, kernel_size=2, stride=2)) # up-conv2x2
for i in range(1, len(encoder_blocks)-1):
in_ch = encoder_out_channels[-i-1] + out_ch # cat
decoder_blocks.append(nn.Sequential( # two conv3x3, up-conv2x2
unet_convs(in_ch, out_ch, padding=1),
nn.ConvTranspose2d(out_ch, out_ch//2, kernel_size=2, stride=2),
))
out_ch = out_ch // 2
in_ch = encoder_out_channels[0] + out_ch # cat
decoder_blocks.append(nn.Sequential( # two conv3x3, conv1x1
unet_convs(in_ch, out_ch, padding=1),
nn.Conv2d(out_ch, out_channels, kernel_size=1)
))
return UNetFactory(encoder_blocks, decoder_blocks, bridge)
if __name__ == "__main__":
from torchsummary import summary
# net = unet(3, 3).cuda()
# summary(net, (3, 572, 572), device='cuda')
# net = unet_resnet('resnet50', 3, 3, False).cuda()
net = unet_resnet('resnext50_32x4d', 3, 3, False).cuda()
summary(net, (3, 224, 224), device='cuda')