-
Notifications
You must be signed in to change notification settings - Fork 341
/
mobilenetv3.py
215 lines (173 loc) · 7.63 KB
/
mobilenetv3.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
'''MobileNetV3 in PyTorch.
See the paper "Inverted Residuals and Linear Bottlenecks:
Mobile Networks for Classification, Detection and Segmentation" for more details.
'''
import torch.nn as nn
import torch.nn.functional as F
from torch.nn import init
class hswish(nn.Module):
def forward(self, x):
out = x * F.relu6(x + 3, inplace=True) / 6
return out
class hsigmoid(nn.Module):
def forward(self, x):
out = F.relu6(x + 3, inplace=True) / 6
return out
class SeModule(nn.Module):
def __init__(self, in_size, reduction=4):
super(SeModule, self).__init__()
expand_size = max(in_size // reduction, 8)
self.se = nn.Sequential(
nn.AdaptiveAvgPool2d(1),
nn.Conv2d(in_size, expand_size, kernel_size=1, bias=False),
nn.BatchNorm2d(expand_size),
nn.ReLU(inplace=True),
nn.Conv2d(expand_size, in_size, kernel_size=1, bias=False),
nn.Hardsigmoid()
)
def forward(self, x):
return x * self.se(x)
class Block(nn.Module):
'''expand + depthwise + pointwise'''
def __init__(self, kernel_size, in_size, expand_size, out_size, act, se, stride):
super(Block, self).__init__()
self.stride = stride
self.conv1 = nn.Conv2d(in_size, expand_size, kernel_size=1, bias=False)
self.bn1 = nn.BatchNorm2d(expand_size)
self.act1 = act(inplace=True)
self.conv2 = nn.Conv2d(expand_size, expand_size, kernel_size=kernel_size, stride=stride, padding=kernel_size//2, groups=expand_size, bias=False)
self.bn2 = nn.BatchNorm2d(expand_size)
self.act2 = act(inplace=True)
self.se = SeModule(expand_size) if se else nn.Identity()
self.conv3 = nn.Conv2d(expand_size, out_size, kernel_size=1, bias=False)
self.bn3 = nn.BatchNorm2d(out_size)
self.act3 = act(inplace=True)
self.skip = None
if stride == 1 and in_size != out_size:
self.skip = nn.Sequential(
nn.Conv2d(in_size, out_size, kernel_size=1, bias=False),
nn.BatchNorm2d(out_size)
)
if stride == 2 and in_size != out_size:
self.skip = nn.Sequential(
nn.Conv2d(in_channels=in_size, out_channels=in_size, kernel_size=3, groups=in_size, stride=2, padding=1, bias=False),
nn.BatchNorm2d(in_size),
nn.Conv2d(in_size, out_size, kernel_size=1, bias=True),
nn.BatchNorm2d(out_size)
)
if stride == 2 and in_size == out_size:
self.skip = nn.Sequential(
nn.Conv2d(in_channels=in_size, out_channels=out_size, kernel_size=3, groups=in_size, stride=2, padding=1, bias=False),
nn.BatchNorm2d(out_size)
)
def forward(self, x):
skip = x
out = self.act1(self.bn1(self.conv1(x)))
out = self.act2(self.bn2(self.conv2(out)))
out = self.se(out)
out = self.bn3(self.conv3(out))
if self.skip is not None:
skip = self.skip(skip)
return self.act3(out + skip)
class MobileNetV3_Small(nn.Module):
def __init__(self, num_classes=1000, act=nn.Hardswish):
super(MobileNetV3_Small, self).__init__()
self.conv1 = nn.Conv2d(3, 16, kernel_size=3, stride=2, padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(16)
self.hs1 = act(inplace=True)
self.bneck = nn.Sequential(
Block(3, 16, 16, 16, nn.ReLU, True, 2),
Block(3, 16, 72, 24, nn.ReLU, False, 2),
Block(3, 24, 88, 24, nn.ReLU, False, 1),
Block(5, 24, 96, 40, act, True, 2),
Block(5, 40, 240, 40, act, True, 1),
Block(5, 40, 240, 40, act, True, 1),
Block(5, 40, 120, 48, act, True, 1),
Block(5, 48, 144, 48, act, True, 1),
Block(5, 48, 288, 96, act, True, 2),
Block(5, 96, 576, 96, act, True, 1),
Block(5, 96, 576, 96, act, True, 1),
)
self.conv2 = nn.Conv2d(96, 576, kernel_size=1, stride=1, padding=0, bias=False)
self.bn2 = nn.BatchNorm2d(576)
self.hs2 = act(inplace=True)
self.gap = nn.AdaptiveAvgPool2d(1)
self.linear3 = nn.Linear(576, 1280, bias=False)
self.bn3 = nn.BatchNorm1d(1280)
self.hs3 = act(inplace=True)
self.drop = nn.Dropout(0.2)
self.linear4 = nn.Linear(1280, num_classes)
self.init_params()
def init_params(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
init.kaiming_normal_(m.weight, mode='fan_out')
if m.bias is not None:
init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
init.constant_(m.weight, 1)
init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
init.normal_(m.weight, std=0.001)
if m.bias is not None:
init.constant_(m.bias, 0)
def forward(self, x):
out = self.hs1(self.bn1(self.conv1(x)))
out = self.bneck(out)
out = self.hs2(self.bn2(self.conv2(out)))
out = self.gap(out).flatten(1)
out = self.drop(self.hs3(self.bn3(self.linear3(out))))
return self.linear4(out)
class MobileNetV3_Large(nn.Module):
def __init__(self, num_classes=1000, act=nn.Hardswish):
super(MobileNetV3_Large, self).__init__()
self.conv1 = nn.Conv2d(3, 16, kernel_size=3, stride=2, padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(16)
self.hs1 = act(inplace=True)
self.bneck = nn.Sequential(
Block(3, 16, 16, 16, nn.ReLU, False, 1),
Block(3, 16, 64, 24, nn.ReLU, False, 2),
Block(3, 24, 72, 24, nn.ReLU, False, 1),
Block(5, 24, 72, 40, nn.ReLU, True, 2),
Block(5, 40, 120, 40, nn.ReLU, True, 1),
Block(5, 40, 120, 40, nn.ReLU, True, 1),
Block(3, 40, 240, 80, act, False, 2),
Block(3, 80, 200, 80, act, False, 1),
Block(3, 80, 184, 80, act, False, 1),
Block(3, 80, 184, 80, act, False, 1),
Block(3, 80, 480, 112, act, True, 1),
Block(3, 112, 672, 112, act, True, 1),
Block(5, 112, 672, 160, act, True, 2),
Block(5, 160, 672, 160, act, True, 1),
Block(5, 160, 960, 160, act, True, 1),
)
self.conv2 = nn.Conv2d(160, 960, kernel_size=1, stride=1, padding=0, bias=False)
self.bn2 = nn.BatchNorm2d(960)
self.hs2 = act(inplace=True)
self.gap = nn.AdaptiveAvgPool2d(1)
self.linear3 = nn.Linear(960, 1280, bias=False)
self.bn3 = nn.BatchNorm1d(1280)
self.hs3 = act(inplace=True)
self.drop = nn.Dropout(0.2)
self.linear4 = nn.Linear(1280, num_classes)
self.init_params()
def init_params(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
init.kaiming_normal_(m.weight, mode='fan_out')
if m.bias is not None:
init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
init.constant_(m.weight, 1)
init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
init.normal_(m.weight, std=0.001)
if m.bias is not None:
init.constant_(m.bias, 0)
def forward(self, x):
out = self.hs1(self.bn1(self.conv1(x)))
out = self.bneck(out)
out = self.hs2(self.bn2(self.conv2(out)))
out = self.gap(out).flatten(1)
out = self.drop(self.hs3(self.bn3(self.linear3(out))))
return self.linear4(out)