-
Notifications
You must be signed in to change notification settings - Fork 9
/
shufflenet_v2.py
171 lines (141 loc) · 5.47 KB
/
shufflenet_v2.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
import os
import sys
import torch
import torch.nn as nn
from torch.autograd import Variable
import math
import numpy as np
def conv3x3(in_channels, out_channels, stride, padding=1, groups=1):
"""3x3 convolution"""
return nn.Conv2d(in_channels, out_channels,
kernel_size=3, stride=stride, padding=padding,
groups=groups,
bias=False)
def conv1x1(in_channels, out_channels, stride=1):
"""1x1 convolution"""
return nn.Conv2d(in_channels, out_channels,
kernel_size=1, stride=stride,padding=0,
bias=False)
class ShufflenetUnit(nn.Module):
def __init__(self, inplanes, planes, stride=1, downsample=None):
super(ShufflenetUnit, self).__init__()
self.downsample = downsample
if not self.downsample: #---if not downsample, then channel split, so the channel become half
inplanes = inplanes // 2
planes = planes // 2
self.conv1x1_1 = conv1x1(in_channels=inplanes, out_channels=planes)
self.conv1x1_1_bn = nn.BatchNorm2d(planes)
self.dwconv3x3 = conv3x3(in_channels=planes, out_channels=planes, stride=stride, groups=planes)
self.dwconv3x3_bn= nn.BatchNorm2d(planes)
self.conv1x1_2 = conv1x1(in_channels=planes, out_channels=planes)
self.conv1x1_2_bn = nn.BatchNorm2d(planes)
self.relu = nn.ReLU(inplace=True)
def _channel_split(self, features, ratio=0.5):
"""
ratio: c'/c, default value is 0.5
"""
size = features.size()[1]
split_idx = int(size * ratio)
return features[:,:split_idx], features[:,split_idx:]
def _channel_shuffle(self, features, g=2):
channels = features.size()[1]
index = torch.from_numpy(np.asarray([i for i in range(channels)]))
index = index.view(-1, g).t().contiguous()
index = index.view(-1).cuda()
features = features[:, index]
return features
def forward(self, x):
if self.downsample:
#x1 = x.clone() #----deep copy x, so where x2 is modified, x1 not be affected
x1 = x
x2 = x
else:
x1, x2 = self._channel_split(x)
#----right branch-----
x2 = self.conv1x1_1(x2)
x2 = self.conv1x1_1_bn(x2)
x2 = self.relu(x2)
x2 = self.dwconv3x3(x2)
x2 = self.dwconv3x3_bn(x2)
x2 = self.conv1x1_2(x2)
x2 = self.conv1x1_2_bn(x2)
x2 = self.relu(x2)
#---left branch-------
if self.downsample:
x1 = self.downsample(x1)
x = torch.cat([x1, x2], 1)
x = self._channel_shuffle(x)
return x
class ShuffleNet(nn.Module):
def __init__(self, feature_dim, layers_num, num_classes=1000):
super(ShuffleNet, self).__init__()
dim1, dim2, dim3, dim4, dim5 = feature_dim
self.conv1 = conv3x3(in_channels=3, out_channels=dim1,
stride=2, padding=1)
self.bn1 = nn.BatchNorm2d(dim1)
self.relu = nn.ReLU(inplace=True)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
self.stage2 = self._make_layer(dim1, dim2, layers_num[0])
self.stage3 = self._make_layer(dim2, dim3, layers_num[1])
self.stage4 = self._make_layer(dim3, dim4, layers_num[2])
self.conv5 = conv1x1(in_channels=dim4, out_channels=dim5)
self.globalpool = nn.AvgPool2d(kernel_size=7, stride=1)
self.fc = nn.Linear(dim5, num_classes)
"""
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))
elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(1)
m.bias.data.zero_()
"""
def _make_layer(self, dim1, dim2, blocks_num):
half_channel = dim2 // 2
downsample = nn.Sequential(
conv3x3(in_channels=dim1, out_channels=dim1, stride=2, padding=1, groups=dim1),
nn.BatchNorm2d(dim1),
conv1x1(in_channels=dim1, out_channels=half_channel),
nn.BatchNorm2d(half_channel),
nn.ReLU(inplace=True)
)
layers = []
layers.append(ShufflenetUnit(dim1, half_channel, stride=2, downsample=downsample))
for i in range(blocks_num):
layers.append(ShufflenetUnit(dim2, dim2, stride=1))
return nn.Sequential(*layers)
def forward(self,x):
x = self.conv1(x)
x = self.bn1(x)
x = self.relu(x)
#print("x0.size:\t", x.size())
x = self.maxpool(x)
#print("x1.size:\t", x.size())
x = self.stage2(x)
#print("x2.size:\t", x.size())
x = self.stage3(x)
#print("x3.size:\t", x.size())
x = self.stage4(x)
#print("x4.size:\t", x.size())
x = self.conv5(x)
#print("x5.size:\t", x.size())
x = self.globalpool(x)
#print("x6.size:\t", x.size())
x = x.view(-1, 1024)
x = self.fc(x)
return x
features = {
"0.5x":[24, 48, 96, 192, 1024],
"1x":[24, 116, 232, 464, 1024],
"1.5x":[24, 176, 352, 704, 1024],
"2x":[24, 244, 488, 976, 2048]
}
def shufflenet():
model = ShuffleNet(features["1x"], [3, 7, 3])
return model
if __name__=="__main__":
model = shufflenet().cuda()
print(model)
x = torch.rand((1,3,224,224))
x = Variable(x).cuda()
x = model(x)