-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathsyncnet.py
168 lines (130 loc) · 6.53 KB
/
syncnet.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
import torch
from torch import nn
from torch.nn import functional as F
class SpatialAttention(nn.Module):
def __init__(self, kernel_size=7):
super(SpatialAttention, self).__init__()
self.conv1 = nn.Conv2d(2, 1, kernel_size, padding=kernel_size // 2, bias=False)
self.sigmoid = nn.Sigmoid()
def forward(self, x):
avg_out = torch.mean(x, dim=1, keepdim=True)
max_out, _ = torch.max(x, dim=1, keepdim=True)
x = torch.cat([avg_out, max_out], dim=1)
x = self.conv1(x)
return self.sigmoid(x)
class ChannelAttention(nn.Module):
def __init__(self, inplanes):
super(ChannelAttention, self).__init__()
self.max_pool = nn.MaxPool2d(1)
self.avg_pool = nn.AvgPool2d(1)
# 通道注意力，即两个全连接层连接
self.fc = nn.Sequential(
nn.Conv2d(in_channels=inplanes, out_channels=inplanes // 16, kernel_size=1, bias=False),
nn.ReLU(),
nn.Conv2d(in_channels=inplanes // 16, out_channels=inplanes, kernel_size=1, bias=False)
)
self.sigmoid = nn.Sigmoid()
def forward(self, x):
max_x = self.max_pool(x)
avg_x = self.avg_pool(x)
max_out = self.fc(max_x)
avg_out = self.fc(avg_x)
# 最后输出的注意力应该为非负
out = self.sigmoid(max_out + avg_out)
return out
class CBAM_Attention(nn.Module):
def __init__(self, in_channels):
super(CBAM_Attention, self).__init__()
self.channel_atten = ChannelAttention(in_channels)
self.spatial_atten = SpatialAttention()
def forward(self, x):
# CBAM attention
x = self.channel_atten(x) * x
x = self.spatial_atten(x) * x
return x
class Conv2dELU_BN_CBAM(nn.Module):
def __init__(self, cin, cout, kernel_size, stride, padding, residual=False):
super(Conv2dELU_BN_CBAM, self).__init__()
self.conv_block = nn.Sequential(
nn.Conv2d(cin, cout, kernel_size, stride, padding),
nn.BatchNorm2d(cout)
)
self.act = nn.ELU()
self.residual = residual
self.cbam_attention = CBAM_Attention(cout)
def forward(self, x):
out = self.conv_block(x)
# CBAM attention
out = self.cbam_attention(out)
if self.residual:
out += x
return self.act(out)
class SyncNet(nn.Module):
def __init__(self):
super(SyncNet, self).__init__()
self.face_encoder = nn.Sequential(
Conv2dELU_BN_CBAM(15, 64, kernel_size=(7, 8), stride=(1, 2), padding=3), # x/2
Conv2dELU_BN_CBAM(64, 64, kernel_size=7, stride=1, padding=3, residual=True),
Conv2dELU_BN_CBAM(64, 64, kernel_size=7, stride=1, padding=3, residual=True),
Conv2dELU_BN_CBAM(64, 128, kernel_size=6, stride=2, padding=2), # x/4
Conv2dELU_BN_CBAM(128, 128, kernel_size=5, stride=1, padding=2, residual=True),
Conv2dELU_BN_CBAM(128, 128, kernel_size=5, stride=1, padding=2, residual=True),
Conv2dELU_BN_CBAM(128, 256, kernel_size=4, stride=2, padding=1), # x/8
Conv2dELU_BN_CBAM(256, 256, kernel_size=3, stride=1, padding=1, residual=True),
Conv2dELU_BN_CBAM(256, 256, kernel_size=3, stride=1, padding=1, residual=True),
Conv2dELU_BN_CBAM(256, 256, kernel_size=3, stride=1, padding=1, residual=True),
Conv2dELU_BN_CBAM(256, 512, kernel_size=4, stride=2, padding=1), # x/16
Conv2dELU_BN_CBAM(512, 512, kernel_size=3, stride=1, padding=1, residual=True),
Conv2dELU_BN_CBAM(512, 512, kernel_size=3, stride=1, padding=1, residual=True),
Conv2dELU_BN_CBAM(512, 512, kernel_size=4, stride=2, padding=1), # x/32
Conv2dELU_BN_CBAM(512, 512, kernel_size=3, stride=1, padding=1, residual=True),
Conv2dELU_BN_CBAM(512, 512, kernel_size=3, stride=1, padding=1, residual=True),
Conv2dELU_BN_CBAM(512, 512, kernel_size=2, stride=2, padding=0), # x/64
Conv2dELU_BN_CBAM(512, 512, kernel_size=1, stride=1, padding=0),
nn.AdaptiveMaxPool2d(1),
)
self.audio_encoder = nn.Sequential(
Conv2dELU_BN_CBAM(1, 64, kernel_size=(9, 5), stride=(5, 1), padding=2), # 16 16
Conv2dELU_BN_CBAM(64, 64, kernel_size=5, stride=1, padding=2, residual=True),
Conv2dELU_BN_CBAM(64, 64, kernel_size=5, stride=1, padding=2, residual=True),
Conv2dELU_BN_CBAM(64, 128, kernel_size=4, stride=2, padding=1), # 8
Conv2dELU_BN_CBAM(128, 128, kernel_size=3, stride=1, padding=1, residual=True),
Conv2dELU_BN_CBAM(128, 128, kernel_size=3, stride=1, padding=1, residual=True),
Conv2dELU_BN_CBAM(128, 128, kernel_size=3, stride=1, padding=1, residual=True),
Conv2dELU_BN_CBAM(128, 256, kernel_size=4, stride=2, padding=1), # 4
Conv2dELU_BN_CBAM(256, 256, kernel_size=3, stride=1, padding=1, residual=True),
Conv2dELU_BN_CBAM(256, 256, kernel_size=3, stride=1, padding=1, residual=True),
Conv2dELU_BN_CBAM(256, 512, kernel_size=2, stride=2, padding=0), # 2
Conv2dELU_BN_CBAM(512, 512, kernel_size=1, stride=1, padding=0),
nn.AdaptiveMaxPool2d(1),
)
self.face_fc = nn.Sequential(
nn.Linear(512, 2048),
nn.BatchNorm1d(2048),
nn.ReLU(),
nn.Linear(2048, 2048),
nn.BatchNorm1d(2048),
nn.ReLU(),
nn.Linear(2048, 512),
nn.Sigmoid(),
)
self.audio_fc = nn.Sequential(
nn.Linear(512, 2048),
nn.BatchNorm1d(2048),
nn.ReLU(),
nn.Linear(2048, 2048),
nn.BatchNorm1d(2048),
nn.ReLU(),
nn.Linear(2048, 512),
nn.Sigmoid(),
)
def forward(self, audio_sequences, face_sequences, bs):
audio_embedding = self.audio_encoder(audio_sequences)
face_embedding = self.face_encoder(face_sequences)
audio_embedding = audio_embedding.reshape(bs, 512)
face_embedding = face_embedding.reshape(bs, 512)
audio_embedding = self.audio_fc(audio_embedding)
face_embedding = self.face_fc(face_embedding)
audio_embedding = F.normalize(audio_embedding, p=2, dim=1)
face_embedding = F.normalize(face_embedding, p=2, dim=1)
return audio_embedding, face_embedding