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
| import torch
import torch.nn as nn
import torch.nn.functional as F
class RAFMSerPlusPlus(nn.Module):
"""
RAFM_SER++: 轻量多模态语音情绪识别
架构:
1. 语音分支: HuBERT-Base (冻结) + 适配层
2. 文本分支: BERT (冻结) + 适配层
3. RAFM: 非对称残差注意力融合 (语音 → 文本 单向注入)
4. BYOL 对齐: 跨模态表示对齐
5. 注意力引导池化
6. 分类头
"""
def __init__(self, speech_dim=768, text_dim=768,
hidden_dim=256, num_classes=4):
super().__init__()
self.speech_adapter = nn.Sequential(
nn.Linear(speech_dim, hidden_dim),
nn.LayerNorm(hidden_dim),
nn.GELU()
)
self.text_adapter = nn.Sequential(
nn.Linear(text_dim, hidden_dim),
nn.LayerNorm(hidden_dim),
nn.GELU()
)
self.rafm = AsymmetricResidualAttentionFusion(
dim=hidden_dim, num_heads=4
)
self.attn_pool = AttentionPooling(hidden_dim)
self.classifier = nn.Sequential(
nn.Dropout(0.2),
nn.Linear(hidden_dim, num_classes)
)
def forward(self, speech_feat, text_feat):
"""
Args:
speech_feat: (B, L_s, 768) HuBERT 输出
text_feat: (B, L_t, 768) BERT 输出
Returns:
logits: (B, num_classes)
"""
speech = self.speech_adapter(speech_feat)
text = self.text_adapter(text_feat)
fused = self.rafm(speech, text)
pooled = self.attn_pool(fused)
logits = self.classifier(pooled)
return logits
class AsymmetricResidualAttentionFusion(nn.Module):
"""
RAFM: 非对称残差注意力融合
核心思想: 语音情感线索单向注入文本语义表示
- 不用双向交互 (太贵)
- 语音特征做 Query, 文本特征做 K/V
- 残差连接保持文本表示
"""
def __init__(self, dim=256, num_heads=4):
super().__init__()
self.cross_attn = nn.MultiheadAttention(
dim, num_heads, batch_first=True, dropout=0.1
)
self.norm1 = nn.LayerNorm(dim)
self.norm2 = nn.LayerNorm(dim)
self.ffn = nn.Sequential(
nn.Linear(dim, dim * 2),
nn.GELU(),
nn.Dropout(0.1),
nn.Linear(dim * 2, dim)
)
def forward(self, speech, text):
"""
Args:
speech: (B, L_s, D) 语音特征
text: (B, L_t, D) 文本特征
Returns:
fused: (B, L_t, D) 语音增强的文本特征
"""
normed_q = self.norm1(speech)
normed_kv = self.norm1(text)
attn_out, _ = self.cross_attn(
normed_q,
normed_kv,
normed_kv
)
fused = text + attn_out[:, :text.shape[1], :]
fused = fused + self.ffn(self.norm2(fused))
return fused
class AttentionPooling(nn.Module):
"""注意力引导池化"""
def __init__(self, dim=256):
super().__init__()
self.attn = nn.Linear(dim, 1)
def forward(self, x):
"""
Args:
x: (B, L, D)
Returns:
pooled: (B, D)
"""
weights = F.softmax(self.attn(x), dim=1)
pooled = (x * weights).sum(dim=1)
return pooled
class BYOLAlignment(nn.Module):
"""
BYOL 跨模态对齐
将语音和文本表示拉到同一空间
无需负样本 (vs SimCLR)
"""
def __init__(self, dim=256, projection_dim=128):
super().__init__()
self.speech_proj = nn.Sequential(
nn.Linear(dim, dim), nn.GELU(),
nn.Linear(dim, projection_dim)
)
self.text_proj = nn.Sequential(
nn.Linear(dim, dim), nn.GELU(),
nn.Linear(dim, projection_dim)
)
def forward(self, speech_repr, text_repr):
s = self.speech_proj(speech_repr)
t = self.text_proj(text_repr)
loss = 2 - 2 * F.cosine_similarity(
s, t.detach(), dim=-1
).mean()
return loss
if __name__ == "__main__":
model = RAFMSerPlusPlus(num_classes=4)
speech = torch.randn(2, 100, 768)
text = torch.randn(2, 32, 768)
logits = model(speech, text)
print(f"Speech: {speech.shape}")
print(f"Text: {text.shape}")
print(f"Logits: {logits.shape}")
print(f"Params: {sum(p.numel() for p in model.parameters()):,}")
import time
model.eval()
with torch.no_grad():
start = time.time()
for _ in range(100):
_ = model(speech, text)
elapsed = time.time() - start
print(f"Speed: {100/elapsed:.1f} it/s")
|
