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
| import torch
import torch.nn as nn
from mamba_ssm import Mamba
class MS_Mamba(nn.Module):
"""
多光谱Mamba块
论文Section III-C:
1. 时间轴扫描：对每个光谱特征沿时间轴做双向Mamba
2. 通道轴扫描：沿通道轴双向扫描捕获跨光谱交互
"""
def __init__(self, d_model: int, d_state: int = 16,
d_conv: int = 4, expand: int = 2):
super().__init__()
self.d_model = d_model
self.temporal_mamba_rgb = Mamba(
d_model=d_model, d_state=d_state,
d_conv=d_conv, expand=expand, bidirectional=True
)
self.temporal_mamba_nir = Mamba(
d_model=d_model, d_state=d_state,
d_conv=d_conv, expand=expand, bidirectional=True
)
self.channel_mamba = Mamba(
d_model=d_model * 2,
d_state=d_state,
d_conv=d_conv, expand=expand, bidirectional=True
)
self.norm1 = nn.LayerNorm(d_model)
self.norm2 = nn.LayerNorm(d_model)
self.norm3 = nn.LayerNorm(d_model * 2)
self.proj = nn.Linear(d_model * 2, d_model)
def forward(self, rgb_feat: torch.Tensor, nir_feat: torch.Tensor):
"""
Args:
rgb_feat: [B, T, C] RGB时序特征
nir_feat: [B, T, C] NIR时序特征
Returns:
fused_feat: [B, T, C] 融合后特征
"""
B, T, C = rgb_feat.shape
rgb_temporal = self.temporal_mamba_rgb(
self.norm1(rgb_feat)
) + rgb_feat
nir_temporal = self.temporal_mamba_nir(
self.norm2(nir_feat)
) + nir_feat
concat_feat = torch.cat([rgb_temporal, nir_temporal], dim=-1)
concat_feat = self.norm3(concat_feat)
channel_input = concat_feat.transpose(1, 2)
channel_out = self.channel_mamba(channel_input)
channel_out = channel_out.transpose(1, 2)
fused = self.proj(channel_out + concat_feat)
return fused
class MS_rPPG(nn.Module):
"""
MS-rPPG完整模型
论文Section III: 多光谱rPPG估计框架
输入：RGB+NIR面部视频 → 输出：心率估计值
"""
def __init__(self, num_classes: int = 1):
super().__init__()
self.rgb_encoder = nn.Conv3d(3, 64, kernel_size=(3,3,3),
padding=(1,1,1))
self.nir_encoder = nn.Conv3d(1, 64, kernel_size=(3,3,3),
padding=(1,1,1))
self.cslm = CrossSpectralLinearModulation(64)
self.ms_mamba_blocks = nn.ModuleList([
MS_Mamba(d_model=64) for _ in range(3)
])
self.head = nn.Sequential(
nn.AdaptiveAvgPool3d(1),
nn.Flatten(),
nn.Linear(64, 32),
nn.GELU(),
nn.Linear(32, num_classes)
)
def forward(self, rgb: torch.Tensor, nir: torch.Tensor) -> torch.Tensor:
"""
Args:
rgb: [B, 3, T, H, W] RGB面部视频
nir: [B, 1, T, H, W] NIR面部视频
Returns:
hr: [B, 1] 估计心率值
"""
rgb_feat = self.rgb_encoder(rgb)
nir_feat = self.nir_encoder(nir)
fused = self.cslm(rgb_feat, nir_feat)
B, C, T, H, W = fused.shape
fused = fused.view(B, C, T, H * W).permute(0, 3, 2, 1)
fused = fused.mean(1)
for block in self.ms_mamba_blocks:
fused = block(fused, fused)
fused_4d = fused.transpose(1, 2).unsqueeze(-1).unsqueeze(-1)
hr = self.head(fused_4d)
return hr
if __name__ == "__main__":
model = MS_rPPG()
rgb_input = torch.randn(2, 3, 300, 64, 64)
nir_input = torch.randn(2, 1, 300, 64, 64)
hr_output = model(rgb_input, nir_input)
print(f"输入RGB shape: {rgb_input.shape}")
print(f"输入NIR shape: {nir_input.shape}")
print(f"输出心率估计: {hr_output.shape}")
print(f"模型参数量: {sum(p.numel() for p in model.parameters()):,}")
|