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
| """
论文：BiCFlow-MER: Conditional Transport for Multimodal Emotion Recognition
arXiv: 2609.27615
核心方法：双向条件流融合判别和生成
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
class BiCFlowMER(nn.Module):
"""
BiCFlow-MER: 双向条件流多模态情感识别
1. 从音频-文本中解耦情感证据(与说话风格、词汇内容分离)
2. 构建冲突感知情感条件
3. 通过双向整流流传输到显式情感空间
4. 自适应原型云评分 + 反向一致性验证
"""
def __init__(self,
audio_dim: int = 256,
text_dim: int = 384,
emotion_dim: int = 64,
num_emotions: int = 7):
super().__init__()
self.emotion_encoder = nn.Sequential(
nn.Linear(audio_dim + text_dim, 256),
nn.GELU(),
nn.Linear(256, emotion_dim)
)
self.style_encoder = nn.Sequential(
nn.Linear(audio_dim, 128),
nn.GELU(),
nn.Linear(128, 64)
)
self.content_encoder = nn.Sequential(
nn.Linear(text_dim, 128),
nn.GELU(),
nn.Linear(128, 64)
)
self.condition_builder = ConditionBuilder(
emotion_dim, style_dim=64, content_dim=64
)
self.flow = BidirectionalRectifiedFlow(emotion_dim)
self.prototype_cloud = PrototypeCloud(num_emotions, emotion_dim)
self.backward_consistency = BackwardConsistency(
emotion_dim, style_dim=64, content_dim=64
)
def forward(self, audio_feat, text_feat):
"""
Args:
audio_feat: (B, audio_dim) 音频特征
text_feat: (B, text_dim) 文本特征
Returns:
emotion_logits: (B, num_emotions)
"""
emotion_evidence = self.emotion_encoder(
torch.cat([audio_feat, text_feat], dim=-1)
)
style = self.style_encoder(audio_feat)
content = self.content_encoder(text_feat)
condition = self.condition_builder(emotion_evidence, style, content)
transported = self.flow(emotion_evidence, condition)
logits = self.prototype_cloud(transported)
return logits
class ConditionBuilder(nn.Module):
"""构建冲突感知情感条件"""
def __init__(self, emotion_dim, style_dim, content_dim):
super().__init__()
self.fusion = nn.Sequential(
nn.Linear(emotion_dim + style_dim + content_dim, 256),
nn.GELU(),
nn.Linear(256, emotion_dim),
nn.LayerNorm(emotion_dim)
)
def forward(self, emotion, style, content):
concat = torch.cat([emotion, style, content], dim=-1)
return self.fusion(concat)
class BidirectionalRectifiedFlow(nn.Module):
"""双向整 rectified flow"""
def __init__(self, dim, num_steps=10):
super().__init__()
self.num_steps = num_steps
self.velocity_net = nn.Sequential(
nn.Linear(dim * 2, 256),
nn.GELU(),
nn.Linear(256, 256),
nn.GELU(),
nn.Linear(256, dim)
)
def forward(self, x0, condition):
"""从x0传输到情感空间端点"""
x = x0
dt = 1.0 / self.num_steps
for _ in range(self.num_steps):
v = self.velocity_net(torch.cat([x, condition], dim=-1))
x = x + v * dt
return x
class PrototypeCloud(nn.Module):
"""自适应原型云评分"""
def __init__(self, num_emotions, dim):
super().__init__()
self.prototypes = nn.Parameter(torch.randn(num_emotions, dim))
self.temperature = nn.Parameter(torch.ones(1))
def forward(self, x):
dist = torch.cdist(x, self.prototypes)
logits = -dist / self.temperature
return F.softmax(logits, dim=-1)
class BackwardConsistency(nn.Module):
"""反向类别到条件一致性验证"""
def __init__(self, emotion_dim, style_dim, content_dim):
super().__init__()
self.decoder = nn.Sequential(
nn.Linear(emotion_dim + style_dim + content_dim, 256),
nn.GELU(),
nn.Linear(256, style_dim + content_dim)
)
def forward(self, emotion, style, content):
"""验证传输后的情感是否与原始条件一致"""
concat = torch.cat([emotion, style, content], dim=-1)
reconstructed = self.decoder(concat)
return reconstructed
if __name__ == "__main__":
model = BiCFlowMER()
audio = torch.randn(8, 256)
text = torch.randn(8, 384)
logits = model(audio, text)
print(f"输入: audio={audio.shape}, text={text.shape}")
print(f"输出: {logits.shape}")
print(f"\n=== 论文结果 ===")
print("IEMOCAP: 优于所有对比方法")
print("MELD: 优于所有对比方法")
print("CASE (零样本): 优于所有对比方法")
|
