-
Notifications
You must be signed in to change notification settings - Fork 1
/
resnet.py
132 lines (108 loc) · 3.71 KB
/
resnet.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
import torch
from torch import nn
import torch.nn.functional as F
import numpy as np
def get_resnet_structure(n_layers=50):
n_layer_structs = {
18: [2, 2, 2, 2],
34: [3, 4, 6, 3],
50: [3, 4, 14, 3],
101: [3, 13, 30, 3],
152: [3, 8, 36, 3],
}
struct = n_layer_structs[n_layers]
return [
{
"in_channels": 64,
"out_channels": 64,
"num_units": struct[0]
},
{
"in_channels": 64,
"out_channels": 128,
"num_units": struct[1]
},
{
"in_channels": 128,
"out_channels": 256,
"num_units": struct[2]
},
{
"in_channels": 256,
"out_channels": 512,
"num_units": struct[3]
},
]
def create_resnet_block(in_channels, out_channels, num_units, stride=2):
modules = [ResnetBlock(in_channels, out_channels, stride)]
for i in range(num_units - 1):
modules.append(ResnetBlock(out_channels, out_channels, 1))
return nn.Sequential(*modules)
class ResnetBlock(nn.Module):
"""
Essentially a ResNet block
"""
def __init__(self, in_channels, out_channels, stride):
super().__init__()
if in_channels == out_channels:
self.shortcut = nn.MaxPool2d(1, stride)
else:
self.shortcut = nn.Sequential(
nn.Conv2d(in_channels, out_channels, 1, stride, bias=False),
nn.BatchNorm2d(out_channels)
)
self.res_layer = nn.Sequential(
nn.BatchNorm2d(in_channels),
nn.Conv2d(in_channels, out_channels, 3, 1, 1, bias=False),
nn.PReLU(out_channels),
nn.Conv2d(out_channels, out_channels, 3, stride, 1, bias=False),
nn.BatchNorm2d(out_channels),
SqueezeExcitation(out_channels, 16)
)
def forward(self, x):
return self.shortcut(x) + self.res_layer(x)
class SqueezeExcitation(nn.Module):
"""
Calculates a value for each channel in the input between [0, 1]
and multiplies every value in the channel by the amount
"""
def __init__(self, channels, reduction):
super().__init__()
self.avg_pool = nn.AdaptiveAvgPool2d(1)
self.fc1 = nn.Conv2d(channels, channels //
reduction, 1, padding=0, bias=False)
self.fc2 = nn.Conv2d(channels // reduction,
channels, 1, padding=0, bias=False)
self.relu = nn.ReLU(inplace=True)
self.sigmoid = nn.Sigmoid()
def forward(self, x):
identity = x
x = self.relu(self.fc1(self.avg_pool(x)))
x = self.sigmoid(self.fc2(x))
return identity * x
class ResnetModel(nn.Module):
def __init__(self, input_nc, output_nc, activation=None, n_layers=50, res=(256, 256)):
super().__init__()
self.activation = activation
self.input = nn.Sequential(
nn.Conv2d(input_nc, 64, 3, 1, 1, bias=False),
nn.BatchNorm2d(64),
nn.PReLU(64)
)
self.body = nn.Sequential(*[create_resnet_block(**block)
for block in get_resnet_structure(n_layers=n_layers)])
self.fc1 = nn.Linear(self.get_size(input_nc, res), output_nc)
def get_size(self, input_nc, resolution):
x = torch.zeros(1, input_nc, resolution[0], resolution[1])
x = self.input(x)
x = self.body(x)
x = x.view(x.size(0), -1)
return x.size(1)
def forward(self, x):
x = self.input(x)
x = self.body(x)
x = x.view(x.size(0), -1)
x = self.fc1(x)
if self.activation:
x = self.activation(x)
return x