-
Notifications
You must be signed in to change notification settings - Fork 37
/
MoGA_A.py
191 lines (163 loc) · 5.87 KB
/
MoGA_A.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
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
def stem(inp, oup, stride):
return nn.Sequential(
nn.Conv2d(inp, oup, 3, stride, 1, bias=False),
nn.BatchNorm2d(oup),
# nn.ReLU6(inplace=True)
Hswish()
)
def separable_conv(inp, oup):
return nn.Sequential(
nn.Conv2d(inp, inp, 3, 1, 1, groups=inp, bias=False),
nn.BatchNorm2d(inp),
nn.ReLU(inplace=True),
nn.Conv2d(inp, oup, 1, 1, 0, bias=False),
nn.BatchNorm2d(oup),
)
def conv_before_pooling(inp, oup):
return nn.Sequential(
nn.Conv2d(inp, oup, 1, 1, 0, bias=False),
nn.BatchNorm2d(oup),
# nn.ReLU6(inplace=True)
Hswish()
)
def conv_head(inp, oup):
return nn.Sequential(
nn.Conv2d(inp, oup, 1, bias=False),
Hswish(inplace=True),
nn.Dropout2d(0.2)
)
def classifier(inp, nclass):
return nn.Linear(inp, nclass)
class Hswish(nn.Module):
def __init__(self, inplace=True):
super(Hswish, self).__init__()
self.inplace = inplace
def forward(self, x):
return x * F.relu6(x + 3., inplace=self.inplace) / 6.
class Hsigmoid(nn.Module):
def __init__(self, inplace=True):
super(Hsigmoid, self).__init__()
self.inplace = inplace
def forward(self, x):
return F.relu6(x + 3., inplace=self.inplace) / 6.
class SEModule(nn.Module):
def __init__(self, channel, act, reduction=4):
super(SEModule, self).__init__()
self.avg_pool = nn.AdaptiveAvgPool2d(1)
self.conv = nn.Sequential(
nn.Conv2d(channel, channel // reduction, 1, 1, 0, bias=True),
act
)
self.fc = nn.Sequential(
nn.Conv2d(channel // reduction, channel, 1, 1, 0, bias=True),
Hsigmoid()
)
def forward(self, x):
y = self.avg_pool(x)
y = self.conv(y)
y = self.fc(y)
return torch.mul(x, y)
class InvertedResidual(nn.Module):
def __init__(self, inp, oup, kernel_size, stride, expand_ratio, act, se):
super(InvertedResidual, self).__init__()
assert stride in [1, 2]
self.stride = stride
self.act = act
self.se = se
padding = kernel_size // 2
hidden_dim = round(inp * expand_ratio)
self.use_res_connect = self.stride == 1 and inp == oup
self.conv1 = nn.Conv2d(inp, hidden_dim, 1, 1, 0, bias=False)
self.bn1 = nn.BatchNorm2d(hidden_dim)
self.conv2 = nn.Conv2d(hidden_dim, hidden_dim, kernel_size, stride, padding, groups=hidden_dim, bias=False)
self.bn2 = nn.BatchNorm2d(hidden_dim)
if self.se:
self.mid_se = SEModule(hidden_dim, act)
self.conv3 = nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False)
self.bn3 = nn.BatchNorm2d(oup)
def forward(self, x):
inputs = x
x = self.conv1(x)
x = self.bn1(x)
x = self.act(x)
x = self.conv2(x)
x = self.bn2(x)
x = self.act(x)
if self.se:
x = self.mid_se(x)
x = self.conv3(x)
x = self.bn3(x)
if self.use_res_connect:
return inputs + x
else:
return x
class MoGaA(nn.Module):
def __init__(self, n_class=1000, input_size=224):
super(MoGaA, self).__init__()
assert input_size % 32 == 0
mb_config = [
# expansion, out_channel, kernel_size, stride, act(0 RE 1 Hs), se
[6, 24, 5, 2, 0, 0],
[6, 24, 7, 1, 0, 0],
[6, 40, 3, 2, 0, 0],
[6, 40, 3, 1, 0, 1],
[3, 40, 3, 1, 0, 1],
[6, 80, 3, 2, 1, 1],
[6, 80, 3, 1, 1, 0],
[6, 80, 7, 1, 1, 0],
[3, 80, 7, 1, 1, 1],
[6, 112, 7, 1, 1, 0],
[6, 112, 3, 1, 1, 0],
[6, 160, 3, 2, 1, 0],
[6, 160, 5, 1, 1, 1],
[6, 160, 5, 1, 1, 1],
]
first_filter = 16
second_filter = 16
second_last_filter = 960
last_channel = 1280
self.last_channel = last_channel
self.stem = stem(3, first_filter, 2)
self.separable_conv = separable_conv(first_filter, second_filter)
self.mb_module = list()
input_channel = second_filter
for t, c, k, s, a, se in mb_config:
output_channel = c
act = nn.ReLU(inplace=True) if a==0 else Hswish(inplace=True)
self.mb_module.append(InvertedResidual(input_channel, output_channel, k, s, expand_ratio=t, act=act, se=se!=0))
input_channel = output_channel
self.mb_module = nn.Sequential(*self.mb_module)
self.conv_before_pooling = conv_before_pooling(input_channel, second_last_filter)
self.global_pooling = nn.AdaptiveAvgPool2d(1)
self.conv_head = conv_head(second_last_filter, last_channel)
self.classifier = classifier(last_channel, n_class)
self._initialize_weights()
def forward(self, x):
x = self.stem(x)
x = self.separable_conv(x)
x = self.mb_module(x)
x = self.conv_before_pooling(x)
x = self.global_pooling(x)
x = self.conv_head(x)
x = x.view(x.size(0), -1)
x = self.classifier(x)
return x
def _initialize_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
m.weight.data.normal_(0, math.sqrt(2. / n))
if m.bias is not None:
m.bias.data.zero_()
elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(1)
m.bias.data.zero_()
elif isinstance(m, nn.Linear):
n = m.weight.size(0) # fan-out
init_range = 1.0 / math.sqrt(n)
m.weight.data.uniform_(-init_range, init_range)
m.bias.data.zero_()