-
Notifications
You must be signed in to change notification settings - Fork 375
/
segnet.py
245 lines (201 loc) · 9.18 KB
/
segnet.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
231
232
233
234
235
236
237
238
239
240
241
242
243
from base import BaseModel
import torch
import torch.nn as nn
import torch.nn.functional as F
from torchvision import models
from itertools import chain
from math import ceil
class SegNet(BaseModel):
def __init__(self, num_classes, in_channels=3, pretrained=True, freeze_bn=False, **_):
super(SegNet, self).__init__()
vgg_bn = models.vgg16_bn(pretrained= pretrained)
encoder = list(vgg_bn.features.children())
# Adjust the input size
if in_channels != 3:
encoder[0] = nn.Conv2d(in_channels, 64, kernel_size=3, stride=1, padding=1)
# Encoder, VGG without any maxpooling
self.stage1_encoder = nn.Sequential(*encoder[:6])
self.stage2_encoder = nn.Sequential(*encoder[7:13])
self.stage3_encoder = nn.Sequential(*encoder[14:23])
self.stage4_encoder = nn.Sequential(*encoder[24:33])
self.stage5_encoder = nn.Sequential(*encoder[34:-1])
self.pool = nn.MaxPool2d(kernel_size=2, stride=2, return_indices=True)
# Decoder, same as the encoder but reversed, maxpool will not be used
decoder = encoder
decoder = [i for i in list(reversed(decoder)) if not isinstance(i, nn.MaxPool2d)]
# Replace the last conv layer
decoder[-1] = nn.Conv2d(64, 64, kernel_size=3, stride=1, padding=1)
# When reversing, we also reversed conv->batchN->relu, correct it
decoder = [item for i in range(0, len(decoder), 3) for item in decoder[i:i+3][::-1]]
# Replace some conv layers & batchN after them
for i, module in enumerate(decoder):
if isinstance(module, nn.Conv2d):
if module.in_channels != module.out_channels:
decoder[i+1] = nn.BatchNorm2d(module.in_channels)
decoder[i] = nn.Conv2d(module.out_channels, module.in_channels, kernel_size=3, stride=1, padding=1)
self.stage1_decoder = nn.Sequential(*decoder[0:9])
self.stage2_decoder = nn.Sequential(*decoder[9:18])
self.stage3_decoder = nn.Sequential(*decoder[18:27])
self.stage4_decoder = nn.Sequential(*decoder[27:33])
self.stage5_decoder = nn.Sequential(*decoder[33:],
nn.Conv2d(64, num_classes, kernel_size=3, stride=1, padding=1)
)
self.unpool = nn.MaxUnpool2d(kernel_size=2, stride=2)
self._initialize_weights(self.stage1_decoder, self.stage2_decoder, self.stage3_decoder,
self.stage4_decoder, self.stage5_decoder)
if freeze_bn: self.freeze_bn()
def _initialize_weights(self, *stages):
for modules in stages:
for module in modules.modules():
if isinstance(module, nn.Conv2d):
nn.init.kaiming_normal_(module.weight)
if module.bias is not None:
module.bias.data.zero_()
elif isinstance(module, nn.BatchNorm2d):
module.weight.data.fill_(1)
module.bias.data.zero_()
def forward(self, x):
# Encoder
x = self.stage1_encoder(x)
x1_size = x.size()
x, indices1 = self.pool(x)
x = self.stage2_encoder(x)
x2_size = x.size()
x, indices2 = self.pool(x)
x = self.stage3_encoder(x)
x3_size = x.size()
x, indices3 = self.pool(x)
x = self.stage4_encoder(x)
x4_size = x.size()
x, indices4 = self.pool(x)
x = self.stage5_encoder(x)
x5_size = x.size()
x, indices5 = self.pool(x)
# Decoder
x = self.unpool(x, indices=indices5, output_size=x5_size)
x = self.stage1_decoder(x)
x = self.unpool(x, indices=indices4, output_size=x4_size)
x = self.stage2_decoder(x)
x = self.unpool(x, indices=indices3, output_size=x3_size)
x = self.stage3_decoder(x)
x = self.unpool(x, indices=indices2, output_size=x2_size)
x = self.stage4_decoder(x)
x = self.unpool(x, indices=indices1, output_size=x1_size)
x = self.stage5_decoder(x)
return x
def get_backbone_params(self):
return []
def get_decoder_params(self):
return self.parameters()
def freeze_bn(self):
for module in self.modules():
if isinstance(module, nn.BatchNorm2d): module.eval()
class DecoderBottleneck(nn.Module):
def __init__(self, inchannels):
super(DecoderBottleneck, self).__init__()
self.conv1 = nn.Conv2d(inchannels, inchannels//4, kernel_size=1, bias=False)
self.bn1 = nn.BatchNorm2d(inchannels//4)
self.conv2 = nn.ConvTranspose2d(inchannels//4, inchannels//4, kernel_size=2, stride=2, bias=False)
self.bn2 = nn.BatchNorm2d(inchannels//4)
self.conv3 = nn.Conv2d(inchannels//4, inchannels//2, 1, bias=False)
self.bn3 = nn.BatchNorm2d(inchannels//2)
self.relu = nn.ReLU(inplace=True)
self.downsample = nn.Sequential(
nn.ConvTranspose2d(inchannels, inchannels//2, kernel_size=2, stride=2, bias=False),
nn.BatchNorm2d(inchannels//2))
def forward(self, x):
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
out = self.conv2(out)
out = self.bn2(out)
out = self.relu(out)
out = self.conv3(out)
out = self.bn3(out)
identity = self.downsample(x)
out += identity
out = self.relu(out)
return out
class LastBottleneck(nn.Module):
def __init__(self, inchannels):
super(LastBottleneck, self).__init__()
self.conv1 = nn.Conv2d(inchannels, inchannels//4, kernel_size=1, bias=False)
self.bn1 = nn.BatchNorm2d(inchannels//4)
self.conv2 = nn.Conv2d(inchannels//4, inchannels//4, kernel_size=3, padding=1, bias=False)
self.bn2 = nn.BatchNorm2d(inchannels//4)
self.conv3 = nn.Conv2d(inchannels//4, inchannels//4, 1, bias=False)
self.bn3 = nn.BatchNorm2d(inchannels//4)
self.relu = nn.ReLU(inplace=True)
self.downsample = nn.Sequential(
nn.Conv2d(inchannels, inchannels//4, kernel_size=1, bias=False),
nn.BatchNorm2d(inchannels//4))
def forward(self, x):
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
out = self.conv2(out)
out = self.bn2(out)
out = self.relu(out)
out = self.conv3(out)
out = self.bn3(out)
identity = self.downsample(x)
out += identity
out = self.relu(out)
return out
class SegResNet(BaseModel):
def __init__(self, num_classes, in_channels=3, pretrained=True, freeze_bn=False, **_):
super(SegResNet, self).__init__()
resnet50 = models.resnet50(pretrained=pretrained)
encoder = list(resnet50.children())
if in_channels != 3:
encoder[0] = nn.Conv2d(in_channels, 64, kernel_size=3, stride=1, padding=1)
encoder[3].return_indices = True
# Encoder
self.first_conv = nn.Sequential(*encoder[:4])
resnet50_blocks = list(resnet50.children())[4:-2]
self.encoder = nn.Sequential(*resnet50_blocks)
# Decoder
resnet50_untrained = models.resnet50(pretrained=False)
resnet50_blocks = list(resnet50_untrained.children())[4:-2][::-1]
decoder = []
channels = (2048, 1024, 512)
for i, block in enumerate(resnet50_blocks[:-1]):
new_block = list(block.children())[::-1][:-1]
decoder.append(nn.Sequential(*new_block, DecoderBottleneck(channels[i])))
new_block = list(resnet50_blocks[-1].children())[::-1][:-1]
decoder.append(nn.Sequential(*new_block, LastBottleneck(256)))
self.decoder = nn.Sequential(*decoder)
self.last_conv = nn.Sequential(
nn.ConvTranspose2d(64, 64, kernel_size=2, stride=2, bias=False),
nn.Conv2d(64, num_classes, kernel_size=3, stride=1, padding=1)
)
if freeze_bn: self.freeze_bn()
def forward(self, x):
inputsize = x.size()
# Encoder
x, indices = self.first_conv(x)
x = self.encoder(x)
# Decoder
x = self.decoder(x)
h_diff = ceil((x.size()[2] - indices.size()[2]) / 2)
w_diff = ceil((x.size()[3] - indices.size()[3]) / 2)
if indices.size()[2] % 2 == 1:
x = x[:, :, h_diff:x.size()[2]-(h_diff-1), w_diff: x.size()[3]-(w_diff-1)]
else:
x = x[:, :, h_diff:x.size()[2]-h_diff, w_diff: x.size()[3]-w_diff]
x = F.max_unpool2d(x, indices, kernel_size=2, stride=2)
x = self.last_conv(x)
if inputsize != x.size():
h_diff = (x.size()[2] - inputsize[2]) // 2
w_diff = (x.size()[3] - inputsize[3]) // 2
x = x[:, :, h_diff:x.size()[2]-h_diff, w_diff: x.size()[3]-w_diff]
if h_diff % 2 != 0: x = x[:, :, :-1, :]
if w_diff % 2 != 0: x = x[:, :, :, :-1]
return x
def get_backbone_params(self):
return chain(self.first_conv.parameters(), self.encoder.parameters())
def get_decoder_params(self):
return chain(self.decoder.parameters(), self.last_conv.parameters())
def freeze_bn(self):
for module in self.modules():
if isinstance(module, nn.BatchNorm2d): module.eval()