regnet.py
4.44 KB
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
'''RegNet in PyTorch.
Paper: "Designing Network Design Spaces".
Reference: https://github.com/keras-team/keras-applications/blob/master/keras_applications/efficientnet.py
'''
import torch
import torch.nn as nn
import torch.nn.functional as F
class SE(nn.Module):
'''Squeeze-and-Excitation block.'''
def __init__(self, in_planes, se_planes):
super(SE, self).__init__()
self.se1 = nn.Conv2d(in_planes, se_planes, kernel_size=1, bias=True)
self.se2 = nn.Conv2d(se_planes, in_planes, kernel_size=1, bias=True)
def forward(self, x):
out = F.adaptive_avg_pool2d(x, (1, 1))
out = F.relu(self.se1(out))
out = self.se2(out).sigmoid()
out = x * out
return out
class Block(nn.Module):
def __init__(self, w_in, w_out, stride, group_width, bottleneck_ratio, se_ratio):
super(Block, self).__init__()
# 1x1
w_b = int(round(w_out * bottleneck_ratio))
self.conv1 = nn.Conv2d(w_in, w_b, kernel_size=1, bias=False)
self.bn1 = nn.BatchNorm2d(w_b)
# 3x3
num_groups = w_b // group_width
self.conv2 = nn.Conv2d(w_b, w_b, kernel_size=3,
stride=stride, padding=1, groups=num_groups, bias=False)
self.bn2 = nn.BatchNorm2d(w_b)
# se
self.with_se = se_ratio > 0
if self.with_se:
w_se = int(round(w_in * se_ratio))
self.se = SE(w_b, w_se)
# 1x1
self.conv3 = nn.Conv2d(w_b, w_out, kernel_size=1, bias=False)
self.bn3 = nn.BatchNorm2d(w_out)
self.shortcut = nn.Sequential()
if stride != 1 or w_in != w_out:
self.shortcut = nn.Sequential(
nn.Conv2d(w_in, w_out,
kernel_size=1, stride=stride, bias=False),
nn.BatchNorm2d(w_out)
)
def forward(self, x):
out = F.relu(self.bn1(self.conv1(x)))
out = F.relu(self.bn2(self.conv2(out)))
if self.with_se:
out = self.se(out)
out = self.bn3(self.conv3(out))
out += self.shortcut(x)
out = F.relu(out)
return out
class RegNet(nn.Module):
def __init__(self, cfg, num_classes=10):
super(RegNet, self).__init__()
self.cfg = cfg
self.in_planes = 64
self.conv1 = nn.Conv2d(3, 64, kernel_size=3,
stride=1, padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(64)
self.layer1 = self._make_layer(0)
self.layer2 = self._make_layer(1)
self.layer3 = self._make_layer(2)
self.layer4 = self._make_layer(3)
self.linear = nn.Linear(self.cfg['widths'][-1], num_classes)
def _make_layer(self, idx):
depth = self.cfg['depths'][idx]
width = self.cfg['widths'][idx]
stride = self.cfg['strides'][idx]
group_width = self.cfg['group_width']
bottleneck_ratio = self.cfg['bottleneck_ratio']
se_ratio = self.cfg['se_ratio']
layers = []
for i in range(depth):
s = stride if i == 0 else 1
layers.append(Block(self.in_planes, width,
s, group_width, bottleneck_ratio, se_ratio))
self.in_planes = width
return nn.Sequential(*layers)
def forward(self, x):
out = F.relu(self.bn1(self.conv1(x)))
out = self.layer1(out)
out = self.layer2(out)
out = self.layer3(out)
out = self.layer4(out)
out = F.adaptive_avg_pool2d(out, (1, 1))
out = out.view(out.size(0), -1)
out = self.linear(out)
return out
def RegNetX_200MF():
cfg = {
'depths': [1, 1, 4, 7],
'widths': [24, 56, 152, 368],
'strides': [1, 1, 2, 2],
'group_width': 8,
'bottleneck_ratio': 1,
'se_ratio': 0,
}
return RegNet(cfg)
def RegNetX_400MF():
cfg = {
'depths': [1, 2, 7, 12],
'widths': [32, 64, 160, 384],
'strides': [1, 1, 2, 2],
'group_width': 16,
'bottleneck_ratio': 1,
'se_ratio': 0,
}
return RegNet(cfg)
def RegNetY_400MF():
cfg = {
'depths': [1, 2, 7, 12],
'widths': [32, 64, 160, 384],
'strides': [1, 1, 2, 2],
'group_width': 16,
'bottleneck_ratio': 1,
'se_ratio': 0.25,
}
return RegNet(cfg)
def test():
net = RegNetX_200MF()
print(net)
x = torch.randn(2, 3, 32, 32)
y = net(x)
print(y.shape)
if __name__ == '__main__':
test()