forked from foamliu/Autoencoder
-
Notifications
You must be signed in to change notification settings - Fork 0
/
models.py
177 lines (141 loc) · 6.14 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
import torch.nn as nn
from torchsummary import summary
from config import device, imsize
class conv2DBatchNormRelu(nn.Module):
def __init__(
self,
in_channels,
n_filters,
k_size,
stride,
padding,
bias=True,
dilation=1,
with_bn=True,
):
super(conv2DBatchNormRelu, self).__init__()
conv_mod = nn.Conv2d(int(in_channels),
int(n_filters),
kernel_size=k_size,
padding=padding,
stride=stride,
bias=bias,
dilation=dilation, )
if with_bn:
self.cbr_unit = nn.Sequential(conv_mod,
nn.BatchNorm2d(int(n_filters)),
nn.ReLU(inplace=True))
else:
self.cbr_unit = nn.Sequential(conv_mod, nn.ReLU(inplace=True))
def forward(self, inputs):
outputs = self.cbr_unit(inputs)
return outputs
class segnetDown2(nn.Module):
def __init__(self, in_size, out_size):
super(segnetDown2, self).__init__()
self.conv1 = conv2DBatchNormRelu(in_size, out_size, 3, 1, 1)
self.conv2 = conv2DBatchNormRelu(out_size, out_size, 3, 1, 1)
self.maxpool_with_argmax = nn.MaxPool2d(2, 2, return_indices=True)
def forward(self, inputs):
outputs = self.conv1(inputs)
outputs = self.conv2(outputs)
unpooled_shape = outputs.size()
outputs, indices = self.maxpool_with_argmax(outputs)
return outputs, indices, unpooled_shape
class segnetDown3(nn.Module):
def __init__(self, in_size, out_size):
super(segnetDown3, self).__init__()
self.conv1 = conv2DBatchNormRelu(in_size, out_size, 3, 1, 1)
self.conv2 = conv2DBatchNormRelu(out_size, out_size, 3, 1, 1)
self.conv3 = conv2DBatchNormRelu(out_size, out_size, 3, 1, 1)
self.maxpool_with_argmax = nn.MaxPool2d(2, 2, return_indices=True)
def forward(self, inputs):
outputs = self.conv1(inputs)
outputs = self.conv2(outputs)
outputs = self.conv3(outputs)
unpooled_shape = outputs.size()
outputs, indices = self.maxpool_with_argmax(outputs)
return outputs, indices, unpooled_shape
class segnetUp2(nn.Module):
def __init__(self, in_size, out_size):
super(segnetUp2, self).__init__()
self.unpool = nn.MaxUnpool2d(2, 2)
self.conv1 = conv2DBatchNormRelu(in_size, in_size, 3, 1, 1)
self.conv2 = conv2DBatchNormRelu(in_size, out_size, 3, 1, 1)
def forward(self, inputs, indices, output_shape):
outputs = self.unpool(input=inputs, indices=indices, output_size=output_shape)
outputs = self.conv1(outputs)
outputs = self.conv2(outputs)
return outputs
class segnetUp3(nn.Module):
def __init__(self, in_size, out_size):
super(segnetUp3, self).__init__()
self.unpool = nn.MaxUnpool2d(2, 2)
self.conv1 = conv2DBatchNormRelu(in_size, in_size, 3, 1, 1)
self.conv2 = conv2DBatchNormRelu(in_size, in_size, 3, 1, 1)
self.conv3 = conv2DBatchNormRelu(in_size, out_size, 3, 1, 1)
def forward(self, inputs, indices, output_shape):
outputs = self.unpool(input=inputs, indices=indices, output_size=output_shape)
outputs = self.conv1(outputs)
outputs = self.conv2(outputs)
outputs = self.conv3(outputs)
return outputs
class SegNet(nn.Module):
def __init__(self, n_classes=3, in_channels=3, is_unpooling=True):
super(SegNet, self).__init__()
self.in_channels = in_channels
self.is_unpooling = is_unpooling
self.down1 = segnetDown2(self.in_channels, 64)
self.down2 = segnetDown2(64, 128)
self.down3 = segnetDown3(128, 256)
self.down4 = segnetDown3(256, 512)
self.down5 = segnetDown3(512, 512)
self.up5 = segnetUp3(512, 512)
self.up4 = segnetUp3(512, 256)
self.up3 = segnetUp3(256, 128)
self.up2 = segnetUp2(128, 64)
self.up1 = segnetUp2(64, n_classes)
def forward(self, inputs):
down1, indices_1, unpool_shape1 = self.down1(inputs)
down2, indices_2, unpool_shape2 = self.down2(down1)
down3, indices_3, unpool_shape3 = self.down3(down2)
down4, indices_4, unpool_shape4 = self.down4(down3)
down5, indices_5, unpool_shape5 = self.down5(down4)
up5 = self.up5(down5, indices_5, unpool_shape5)
up4 = self.up4(up5, indices_4, unpool_shape4)
up3 = self.up3(up4, indices_3, unpool_shape3)
up2 = self.up2(up3, indices_2, unpool_shape2)
up1 = self.up1(up2, indices_1, unpool_shape1)
return up1
def init_vgg16_params(self, vgg16):
blocks = [self.down1, self.down2, self.down3, self.down4, self.down5]
ranges = [[0, 4], [5, 9], [10, 16], [17, 23], [24, 29]]
features = list(vgg16.features.children())
vgg_layers = []
for _layer in features:
if isinstance(_layer, nn.Conv2d):
vgg_layers.append(_layer)
merged_layers = []
for idx, conv_block in enumerate(blocks):
if idx < 2:
units = [conv_block.conv1.cbr_unit, conv_block.conv2.cbr_unit]
else:
units = [
conv_block.conv1.cbr_unit,
conv_block.conv2.cbr_unit,
conv_block.conv3.cbr_unit,
]
for _unit in units:
for _layer in _unit:
if isinstance(_layer, nn.Conv2d):
merged_layers.append(_layer)
assert len(vgg_layers) == len(merged_layers)
for l1, l2 in zip(vgg_layers, merged_layers):
if isinstance(l1, nn.Conv2d) and isinstance(l2, nn.Conv2d):
assert l1.weight.size() == l2.weight.size()
assert l1.bias.size() == l2.bias.size()
l2.weight.data = l1.weight.data
l2.bias.data = l1.bias.data
if __name__ == '__main__':
model = SegNet().to(device)
summary(model, (3, imsize, imsize))