-
Notifications
You must be signed in to change notification settings - Fork 60
/
mobileFacenet_32_PReLU.py
235 lines (190 loc) · 8.71 KB
/
mobileFacenet_32_PReLU.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
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
import torch.nn as nn
import math
import torch
import torch.nn.functional as F
def conv_bn(inp, oup, stride, k_size=3):
return nn.Sequential(
nn.Conv2d(inp, oup, k_size, stride, 1, bias=False),
nn.BatchNorm2d(oup),
nn.PReLU()
)
def conv_1x1_bn(inp, oup):
return nn.Sequential(
nn.Conv2d(inp, oup, 1, 1, 0, bias=False),
nn.BatchNorm2d(oup),
nn.PReLU()
)
class DWC(nn.Module):
def __init__(self, in_channels, out_channels):
super(DWC, self).__init__()
#self.depthwise = nn.Conv2d(in_channels=in_channels, out_channels=in_channels, kernel_size=(7,6),
#stride=1, padding=0, groups=in_channels, bias=False)
self.batch_norm_in = nn.BatchNorm2d(in_channels)
self.depthwise = nn.AvgPool2d((7, 6), stride=1, padding=0)
self.pointwise = nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=1,
stride=1, padding=0, bias=False)
def forward(self, x):
x = self.depthwise(x)
#x = self.batch_norm_in(x)
x = self.pointwise(x)
return x
class Max_AvgPool(nn.Module):
def __init__(self, kernel_size=(3,3), stride=2, padding=1, dim=128):
super(Max_AvgPool, self).__init__()
self.Maxpool = nn.MaxPool2d(kernel_size=kernel_size, stride=stride, padding=padding)
self.Avgpool = nn.AvgPool2d(kernel_size=kernel_size, stride=stride, padding=padding)
def forward(self, x):
x = self.Maxpool(x) + self.Avgpool(x) # add some channelwise gating?
return x
class Max_AvgPool(nn.Module):
def __init__(self, kernel_size=(3,3), stride=2, padding=1, dim=128):
super(Max_AvgPool, self).__init__()
self.Maxpool = nn.MaxPool2d(kernel_size=kernel_size, stride=stride, padding=padding)
self.Avgpool = nn.AvgPool2d(kernel_size=kernel_size, stride=stride, padding=padding)
def forward(self, x):
x = self.Maxpool(x) + self.Avgpool(x) # add some channelwise gating?
return x
class gated_conv1x1(nn.Module):
def __init__(self, inc=128, outc=128):
super(gated_conv1x1, self).__init__()
self.inp = int(inc/2)
self.oup = int(outc/2)
self.conv1x1_1 = nn.Conv2d(self.inp, self.oup, 1, 1, 0, bias=False)
self.gate_1 = nn.Conv2d(self.inp, self.oup, 1, 1, 0, bias=True)
self.conv1x1_2 = nn.Conv2d(self.inp, self.oup, 1, 1, 0, bias=False)
self.gate_2 = nn.Conv2d(self.inp, self.oup, 1, 1, 0, bias=True)
def forward(self, x):
x_1 = x[:, :self.inp, :, :]
x_2 = x[:, self.inp:, :, :]
a_1 = self.conv1x1_1(x_1)
g_1 = F.sigmoid(self.gate_1(x_1))
a_2 = self.conv1x1_2(x_2)
g_2 = F.sigmoid(self.gate_2(x_2))
ret = torch.cat((a_1*g_1, a_2*g_2), 1)
return ret
class InvertedResidual_dwc(nn.Module):
def __init__(self, inp, oup, stride, expand_ratio):
super(InvertedResidual_dwc, self).__init__()
self.stride = stride
assert stride in [1, 2]
hidden_dim = int(round(inp * expand_ratio))
self.use_res_connect = self.stride == 1 and inp == oup
self.conv = []
if expand_ratio == 1:
self.conv.append(nn.Conv2d(inp, hidden_dim, kernel_size=(3, 3), stride=stride, padding=1, groups=hidden_dim))
self.conv.append(nn.BatchNorm2d(hidden_dim))
self.conv.append(nn.PReLU())
#self.conv.append(nn.MaxPool2d(kernel_size=(3, 3), stride=stride, padding=1))
#self.conv.append(gated_conv1x1(inc=hidden_dim,outc=oup))
self.conv.append(nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False))
self.conv.append(nn.BatchNorm2d(oup))
self.conv.append(nn.PReLU())
else:
#self.conv.append(gated_conv1x1(inc=inp,outc=hidden_dim))
self.conv.append(nn.Conv2d(inp, hidden_dim, 1, 1, 0, bias=False))
self.conv.append(nn.BatchNorm2d(hidden_dim))
self.conv.append(nn.PReLU())
self.conv.append(nn.Conv2d(hidden_dim, hidden_dim, kernel_size=(3, 3), stride=stride, padding=1, groups=hidden_dim))
self.conv.append(nn.BatchNorm2d(hidden_dim))
self.conv.append(nn.PReLU())
#self.conv.append(gated_conv1x1(inc=hidden_dim,outc=oup))
self.conv.append(nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False))
self.conv.append(nn.BatchNorm2d(oup))
self.conv.append(nn.PReLU())
self.conv = nn.Sequential(*self.conv)
def forward(self, x):
if self.use_res_connect:
return x + self.conv(x)
else:
return self.conv(x)
class InvertedResidual(nn.Module):
def __init__(self, inp, oup, stride, expand_ratio):
super(InvertedResidual, self).__init__()
self.stride = stride
assert stride in [1, 2]
hidden_dim = int(round(inp * expand_ratio))
self.use_res_connect = self.stride == 1 and inp == oup
self.conv = []
if expand_ratio == 1:
self.conv.append(nn.MaxPool2d(kernel_size=(3, 3), stride=stride, padding=1))
#self.conv.append(gated_conv1x1(inc=hidden_dim,outc=oup))
self.conv.append(nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False))
self.conv.append(nn.BatchNorm2d(oup))
else:
#self.conv.append(gated_conv1x1(inc=inp,outc=hidden_dim))
self.conv.append(nn.Conv2d(inp, hidden_dim, 1, 1, 0, bias=False))
self.conv.append(nn.BatchNorm2d(hidden_dim))
self.conv.append(nn.PReLU())
self.conv.append(nn.MaxPool2d(kernel_size=(3, 3), stride=stride, padding=1))
#self.conv.append(gated_conv1x1(inc=hidden_dim,outc=oup))
self.conv.append(nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False))
self.conv.append(nn.BatchNorm2d(oup))
self.conv = nn.Sequential(*self.conv)
def forward(self, x):
if self.use_res_connect:
return x + self.conv(x)
else:
return self.conv(x)
class Net(nn.Module): #mobileNet v2
def __init__(self, embedding_size=128, input_size=224, width_mult=1.):
super(Net, self).__init__()
block = InvertedResidual
block_dwc = InvertedResidual_dwc
input_channel = 64
last_channel = 256
interverted_residual_setting = [
# t, c, n, s
[1, 32, 1, 1], # depthwise conv for first row
[2, 32, 2, 1],
[4, 32, 2, 1],
[2, 32, 2, 2],
[4, 32, 5, 1],
[2, 32, 2, 2],
[2, 32, 6, 2],
]
# building first layer
input_channel = int(input_channel * width_mult)
self.last_channel = int(last_channel * width_mult) if width_mult > 1.0 else last_channel
self.features = [conv_bn(3, input_channel, 2)]
# building inverted residual
cnt = 0
for t, c, n, s in interverted_residual_setting:
output_channel = int(c * width_mult)
for i in range(n):
if cnt>1:
if i == n - 1: # reduce the featuremap in the last.
self.features.append(block_dwc(input_channel, output_channel, s, expand_ratio=t))
else:
self.features.append(block_dwc(input_channel, output_channel, 1, expand_ratio=t))
input_channel = output_channel
else:
if i == n - 1: # reduce the featuremap in the last.
self.features.append(block_dwc(input_channel, output_channel, s, expand_ratio=t))
else:
self.features.append(block_dwc(input_channel, output_channel, 1, expand_ratio=t))
input_channel = output_channel
cnt+=1
# building last several layers
self.features.append(gated_conv1x1(input_channel, self.last_channel))
# make it nn.Sequential
self.features_sequential = nn.Sequential(*self.features)
# Global depthwise conv
#self.GDCconv = DWC(self.last_channel, embedding_size)
self._initialize_weights()
def forward(self, x):
x = self.features_sequential(x).view(-1, 256*4)
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(1)
m.weight.data.normal_(0, 0.01)
m.bias.data.zero_()