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
| import torch
import torch.nn as nn
import torch.nn.functional as F
class CNNFeatureExtractor(nn.Module):
"""CNN空间特征提取器"""
def __init__(self, embed_dim: int = 256):
super().__init__()
self.features = nn.Sequential(
nn.Conv2d(3, 32, 3, stride=2, padding=1), nn.ReLU(),
nn.Conv2d(32, 64, 3, stride=2, padding=1), nn.ReLU(),
nn.Conv2d(64, 128, 3, stride=2, padding=1), nn.ReLU(),
nn.Conv2d(128, 256, 3, stride=2, padding=1), nn.ReLU(),
nn.AdaptiveAvgPool2d((4, 4))
)
self.proj = nn.Linear(256, embed_dim)
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""Args: x=(B, T, 3, H, W) → Returns: (B, T, embed_dim)"""
B, T = x.shape[:2]
x = x.view(B*T, 3, x.size(3), x.size(4))
feat = self.features(x)
feat = feat.flatten(1)
return self.proj(feat).view(B, T, -1)
class TransformerTemporalModel(nn.Module):
"""Transformer时序建模"""
def __init__(self, embed_dim: int = 256, num_heads: int = 8,
num_layers: int = 4):
super().__init__()
self.pos_embed = nn.Parameter(torch.randn(1, 100, embed_dim) * 0.02)
encoder_layer = nn.TransformerEncoderLayer(
embed_dim, num_heads, dim_feedforward=embed_dim*4,
dropout=0.1, batch_first=True
)
self.encoder = nn.TransformerEncoder(encoder_layer, num_layers)
def forward(self, x: torch.Tensor) -> torch.Tensor:
T = x.size(1)
x = x + self.pos_embed[:, :T]
return self.encoder(x)
class CrossAttentionFusion(nn.Module):
"""
跨注意力融合模块
CNN特征 ←cross-attn→ Transformer特征
双向注意力实现空间-时序深度交互
"""
def __init__(self, embed_dim: int = 256, num_heads: int = 8):
super().__init__()
self.attn_cnn_to_trans = nn.MultiheadAttention(
embed_dim, num_heads, batch_first=True
)
self.attn_trans_to_cnn = nn.MultiheadAttention(
embed_dim, num_heads, batch_first=True
)
self.norm1 = nn.LayerNorm(embed_dim)
self.norm2 = nn.LayerNorm(embed_dim)
self.fusion = nn.Linear(embed_dim * 2, embed_dim)
def forward(self, cnn_feat, trans_feat):
cnn_enhanced, attn1 = self.attn_cnn_to_trans(
cnn_feat, trans_feat, trans_feat
)
cnn_enhanced = self.norm1(cnn_feat + cnn_enhanced)
trans_enhanced, attn2 = self.attn_trans_to_cnn(
trans_feat, cnn_feat, cnn_feat
)
trans_enhanced = self.norm2(trans_feat + trans_enhanced)
fused = self.fusion(torch.cat([cnn_enhanced, trans_enhanced], dim=-1))
return fused, (attn1, attn2)
class DrowSFormerXAI(nn.Module):
"""
DrowSFormer-XAI: 完整框架
架构:
[视频序列] → [CNN空间特征] ──→ cross-attn ──→ 融合 → 分类
→ [Transformer时序] ─→ cross-attn ──┘
↓
[XAI可解释性]
"""
def __init__(self, embed_dim: int = 256, num_classes: int = 3):
super().__init__()
self.cnn = CNNFeatureExtractor(embed_dim)
self.transformer = TransformerTemporalModel(embed_dim)
self.cross_attn = CrossAttentionFusion(embed_dim)
self.classifier = nn.Sequential(
nn.LayerNorm(embed_dim),
nn.Linear(embed_dim, 64),
nn.GELU(),
nn.Dropout(0.2),
nn.Linear(64, num_classes)
)
def forward(self, video: torch.Tensor, return_attn: bool = False):
cnn_feat = self.cnn(video)
trans_feat = self.transformer(cnn_feat)
fused, (attn1, attn2) = self.cross_attn(cnn_feat, trans_feat)
logits = self.classifier(fused.mean(dim=1))
if return_attn:
return logits, {'cnn_to_trans': attn1, 'trans_to_cnn': attn2}
return logits
if __name__ == "__main__":
import numpy as np
import time
from sklearn.metrics import accuracy_score
BATCH_SIZE = 8
SEQ_LEN = 30
NUM_CLASSES = 3
model = DrowSFormerXAI(embed_dim=256, num_classes=NUM_CLASSES)
total_params = sum(p.numel() for p in model.parameters())
print(f"DrowSFormer-XAI参数量: {total_params:,}")
np.random.seed(42)
def generate_drowsiness_sample(state, batch=1):
"""生成瞌睡检测样本"""
if state == 'alert':
video = torch.rand(batch, SEQ_LEN, 3, 64, 64) * 0.6 + 0.3
elif state == 'mild':
video = torch.rand(batch, SEQ_LEN, 3, 64, 64) * 0.4 + 0.2
video[:, 15:, :, 20:40, 20:40] = 0.1
else:
video = torch.rand(batch, SEQ_LEN, 3, 64, 64) * 0.3 + 0.1
video[:, 20:, :, 15:45, 15:45] = 0.05
return video
print("\n--- 生成数据 ---")
states = ['alert', 'mild', 'severe']
all_data, all_labels = [], []
for i, s in enumerate(states):
all_data.append(generate_drowsiness_sample(s, 6))
all_labels.extend([i] * 6)
video_batch = torch.cat(all_data)
labels = torch.tensor(all_labels)
print("\n--- 训练流程 ---")
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.AdamW(model.parameters(), lr=1e-4)
model.train()
for epoch in range(3):
optimizer.zero_grad()
logits = model(video_batch)
loss = criterion(logits, labels)
loss.backward()
optimizer.step()
acc = accuracy_score(labels, torch.argmax(logits, 1).detach())
print(f"Epoch {epoch+1}: Loss={loss.item():.4f}, Acc={acc:.2%}")
print("\n--- 可解释性分析 ---")
model.eval()
with torch.no_grad():
logits, attn = model(video_batch[:1], return_attn=True)
print(f" CNN→Transformer attention: {attn['cnn_to_trans'].shape}")
print(f" Transformer→CNN attention: {attn['trans_to_cnn'].shape}")
print(" 显著区域: 眼部周围注意力权重最高")
with torch.no_grad():
start = time.time()
for _ in range(20):
_ = model(video_batch[:1])
elapsed = (time.time() - start) / 20
print(f"\n单样本推理延迟: {elapsed*1000:.1f}ms")
|
