-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathAspect_Based_Fusion.py
228 lines (200 loc) · 11.6 KB
/
Aspect_Based_Fusion.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
import torch
import torch.nn as nn
import torch.nn.functional as F
bacth_size=1
cnn_kernal_size=96
landmark_init=73*2
landmark_final=cnn_kernal_size
aspect_number=20
aspect_dim=96
# from data_handler_boder import *
# temp_data,temp_label,temp_landmark=get_batch(20)
# from torchvision import models
# resnet=models.resnet18(pretrained=True)
#刚开始的时候尝试了resnet，失去了局部特性之后，对遮挡不鲁棒，且由于全脸建模，姿态多，致使数据稀疏，效果不好
# class ResidualBlock(nn.Module):
# def __init__(self, inchannel, outchannel, stride=1):
# super(ResidualBlock, self).__init__()
# self.left = nn.Sequential(
# nn.Conv2d(inchannel, outchannel, kernel_size=3, stride=stride, padding=1, bias=False),
# nn.BatchNorm2d(outchannel),
# nn.ReLU(inplace=True),
# nn.Conv2d(outchannel, outchannel, kernel_size=3, stride=1, padding=1, bias=False),
# nn.BatchNorm2d(outchannel)
# )
# self.shortcut = nn.Sequential()
# if stride != 1 or inchannel != outchannel:
# self.shortcut = nn.Sequential(
# nn.Conv2d(inchannel, outchannel, kernel_size=1, stride=stride, bias=False),
# nn.BatchNorm2d(outchannel)
# )
# def forward(self, x):
# out = self.left(x)
# out += self.shortcut(x)
# out = F.relu(out)
# return out
# class ResNet(nn.Module):
# def __init__(self, ResidualBlock=ResidualBlock, num_classes=10):
# super(ResNet, self).__init__()
# self.inchannel = 2
# self.layer1 = self.make_layer(ResidualBlock, 16, 2, stride=2)
# self.layer2 = self.make_layer(ResidualBlock, 32, 2, stride=2)
# self.layer3 = self.make_layer(ResidualBlock, 64, 2, stride=2)
# self.layer4 = self.make_layer(ResidualBlock, 128, 1, stride=2)
# self.layer5 = self.make_layer(ResidualBlock, 64, 1, stride=2)
# self.to_memory = nn.Sequential(nn.Linear(64, aspect_dim),nn.Tanh())
# def make_layer(self, block, channels, num_blocks, stride):
# strides = [stride] + [1] * (num_blocks - 1) #strides=[1,1]，单个block段内都是一样的stride。
# layers = []
# for stride in strides:
# layers.append(block(self.inchannel, channels, stride))
# self.inchannel = channels
# #收集参数的作用，就是将剩下的东西做成列表，Sequential的输入能够按照列表输入，因为平时的用法本来就是放到其收集参数当中去了
# return nn.Sequential(*layers)
# def forward(self, x):
# out = self.layer1(x)
# out = self.layer2(out)
# out = self.layer3(out)
# out = self.layer4(out)
# out = self.layer5(out)
# out = F.avg_pool2d(out, 4)
# out = out.view(1,out.size(0),-1)
# static_texture_memory = self.to_memory(out)
# return static_texture_memory
#实验结果显示，局部模型比全局cnn好，对au15，au20能预测出
class ResNet(nn.Module):
def __init__(self):
super(ResNet,self).__init__()
self.basic_layer=nn.Sequential(nn.Conv2d(in_channels=2,out_channels=16,kernel_size=3,stride=1),
# nn.BatchNorm2d(16),
nn.ELU(),
nn.MaxPool2d(kernel_size=2, stride=2),
nn.Conv2d(in_channels=16,out_channels=32,kernel_size=3,stride=1),
nn.ELU(),
nn.Conv2d(in_channels=32, out_channels=64, kernel_size=3, stride=1),
nn.BatchNorm2d(64),
nn.ELU(),
nn.MaxPool2d(kernel_size=2, stride=2),
# nn.Conv2d(in_channels=64, out_channels=128, kernel_size=3, stride=1),
nn.Conv2d(in_channels=64, out_channels=128, kernel_size=3, stride=1),
nn.ELU(),
nn.Conv2d(in_channels=128, out_channels=128, kernel_size=1, stride=1),
nn.BatchNorm2d(128),
nn.Sigmoid(),
nn.MaxPool2d(kernel_size=2, stride=2),
nn.Conv2d(in_channels=128, out_channels=cnn_kernal_size, kernel_size=1, stride=1),
nn.Sigmoid(),
nn.Conv2d(in_channels=cnn_kernal_size, out_channels=cnn_kernal_size, kernel_size=1, stride=1),
nn.BatchNorm2d(cnn_kernal_size),
nn.Sigmoid(),
)
self.cnn_serial=nn.GRU(cnn_kernal_size,aspect_dim,batch_first=True)
def forward(self, input):
pic=input
pic=self.basic_layer(pic).view([bacth_size, cnn_kernal_size, -1])
pic=torch.transpose(pic,1,2)
serial,h=self.cnn_serial(pic)
return h
class Landmark_diff_module(nn.Module):
def __init__(self,):
super(Landmark_diff_module,self).__init__()
self.shake_adapt=nn.Sequential(nn.Linear(landmark_init,landmark_init),)
self.to_memory=nn.Sequential(nn.Linear(landmark_init,landmark_init),
nn.Sigmoid(),
nn.Linear(landmark_init, landmark_init),
nn.BatchNorm1d(landmark_init),
nn.Sigmoid(),
nn.Linear(landmark_init,aspect_dim),
nn.Tanh())
def forward(self,diff):
adapt=self.shake_adapt(diff)
re_diff=diff+adapt
# temproal_feature,h=self.lstm_diff(re_diff)
temproal_landmark_memory=self.to_memory(re_diff)
temproal_landmark_memory = temproal_landmark_memory.view( 1,temproal_landmark_memory.size(0), -1)
return temproal_landmark_memory
class Aspect_Query(nn.Module):
def __init__(self):
super(Aspect_Query,self).__init__()
self.aspects_Embendding=nn.Embedding(aspect_number,aspect_dim)
self.au_Embendding = nn.Embedding(8, aspect_dim)
#如果用sequential包起来仅仅一个rnn，就会被报只能输入一个张量的错误
self.query1 = nn.GRU(aspect_dim, aspect_dim, num_layers=1, batch_first=True)
self.query2 = nn.GRU(aspect_dim, aspect_dim, num_layers=1, batch_first=True)
self.query_for_au = nn.GRU(aspect_dim, aspect_dim, num_layers=1, batch_first=True)
#关于西安bn还是先激活：实践中后BN，原始论文中先BN，本人先BN，
self.project2class_with_summmary=nn.Sequential(nn.BatchNorm1d(aspect_dim),
nn.Linear(aspect_dim,aspect_dim//4),
nn.BatchNorm1d(aspect_dim // 4),
nn.Sigmoid(),
nn.Linear(aspect_dim//4,aspect_dim//16),
nn.BatchNorm1d(aspect_dim // 16),
nn.Sigmoid(),
nn.Linear(aspect_dim//16,2),
nn.Softmax())
def forward(self,cnn_memory,landmark_memory):
aspect_embedding=self.aspects_Embendding(torch.LongTensor([list(range(aspect_number))]*bacth_size).cuda())
au_embedding=self.au_Embendding(torch.LongTensor([list(range(8))]*bacth_size).cuda())
final_aspect1,h1=self.query1(aspect_embedding,cnn_memory)
final_aspect2,h2=self.query2(final_aspect1,landmark_memory)
transposed_final_aspect=torch.transpose(final_aspect2,1,2)
alpha=torch.softmax(torch.bmm(final_aspect2,transposed_final_aspect),dim=-1)
self_attention=torch.sum(torch.bmm(alpha,final_aspect2),dim=-2).view(1,bacth_size,-1)
au_query,h=self.query_for_au(au_embedding,self_attention)
out=self.project2class_with_summmary(au_query.contiguous().view(bacth_size*8,-1))
return out
# class classifier_Semantic_correlation(nn.Module):
# def __init__(self):
# # out_put, attention_vector_a, max_pool_, temp_landmark
# super(classifier_Semantic_correlation,self).__init__()
# self.feature_dim=cnn_kernal_size+cnn_kernal_size
# self.toseq_image=nn.GRU(cnn_kernal_size,cnn_kernal_size,batch_first=True,bidirectional=False,num_layers=1)
# self.toseq_landmark = nn.GRU(cnn_kernal_size, cnn_kernal_size, batch_first=True, bidirectional=False, num_layers=1)
# self.Label_Embdding=nn.Embedding(8,cnn_kernal_size,)
# self.predictor=nn.Sequential(nn.Linear(cnn_kernal_size,64),
# nn.BatchNorm1d(64),
# nn.ELU(),
# nn.Linear(64,16),
# nn.ELU(),
# nn.Linear(16,2),
# nn.Softmax(dim=-1)
# )
# self.landmark_filter = nn.Sequential(nn.Linear(landmark_init,landmark_init//2),
# nn.ELU(),
# nn.Linear(landmark_init//2, landmark_init//2),
# nn.ELU(),
# nn.Linear(landmark_init//2,landmark_final),
# nn.ELU(),
# nn.Linear(landmark_final, landmark_final),
# nn.BatchNorm1d(landmark_final),
# nn.Sigmoid()
# )
# self.attention_transfer=nn.Sequential(nn.Linear(cnn_kernal_size,cnn_kernal_size),
# nn.Sigmoid(),
# nn.Linear(cnn_kernal_size,cnn_kernal_size),
# nn.Sigmoid())
# def forward(self,h_global_correlation_feature,cnn_hot_map,label_seq,temp_landmark):
# emb_label_seq=self.Label_Embdding(label_seq)
# out_put,h_global_correlation_feature_=self.toseq_image(emb_label_seq,h_global_correlation_feature)
#
# temp_landmark_hidden_state= self.landmark_filter(temp_landmark)
# temp_landmark_hidden_state = temp_landmark_hidden_state.unsqueeze(dim=0)
# out_put, temp_landmark_hidden_state=self.toseq_landmark(out_put, temp_landmark_hidden_state)
#
# emb_label_seq_a=self.attention_transfer(emb_label_seq)
# viewed_cnn_hot_map=cnn_hot_map.contiguous().view([bacth_size,cnn_kernal_size,-1])
# viewed_cnn_hot_map_t=torch.transpose(viewed_cnn_hot_map,1,2)
# attention_weight =torch.softmax(torch.bmm(emb_label_seq_a, viewed_cnn_hot_map),dim=2)
# attention_vector_a=torch.bmm(attention_weight,viewed_cnn_hot_map_t)
#
# # out_put = 0.5*out_put+0.5*attention_vector_a
# out_put = out_put+attention_vector_a
# out_put = out_put.view(out_put.size()[0] * 8, -1)
# out_put_p=self.predictor(out_put)
#
# return out_put_p
#
#
#
#
#