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
| """
Wave2Body 推理管道测试
论文：arXiv:2607.18875
"""
import torch
import numpy as np
def test_wave2body():
"""Wave2Body 完整推理测试"""
print("=" * 60)
print("Wave2Body 推理测试")
print("=" * 60)
B, N = 2, 200
point_dim = 7
model = Wave2Body(
point_dim=point_dim,
hidden=256,
num_radar_tokens=64,
num_body_tokens=8,
codebook_size=512,
pose_dim=66
)
np.random.seed(42)
point_cloud = torch.randn(B, N, point_dim) * 5
point_cloud[:, :, :3] += torch.tensor([2.0, 0.0, 1.0])
pose, root, tokens = model(point_cloud)
print(f"输入: {point_cloud.shape} (B, N, dim)")
print(f"输出姿态: {pose.shape} (B, J*3)")
print(f"骨盆位置: {root.shape} (B, 3)")
print(f"Body tokens: {tokens.shape} (B, num_tokens)")
print()
total = sum(p.numel() for p in model.parameters())
trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
frozen = total - trainable
print(f"总参数量: {total / 1e6:.2f}M")
print(f"可训练参数: {trainable / 1e6:.2f}M ({trainable/total*100:.1f}%)")
print(f"冻结参数: {frozen / 1e6:.2f}M ({frozen/total*100:.1f}%)")
print()
import time
model.eval()
with torch.no_grad():
for _ in range(10):
_ = model(point_cloud)
times = []
for _ in range(50):
start = time.perf_counter()
_ = model(point_cloud)
times.append((time.perf_counter() - start) * 1000)
print(f"推理延迟: {np.mean(times):.2f} ms (±{np.std(times):.2f})")
print(f"理论帧率: {1000 / np.mean(times):.0f} fps")
return model
def test_cross_domain():
"""跨域泛化测试"""
print("\n" + "=" * 60)
print("跨域泛化测试")
print("=" * 60)
model = Wave2Body()
model.eval()
indoor_points = torch.randn(1, 200, 7) * 3
indoor_points[:, :, :3] += torch.tensor([1.5, 0, 0.8])
outdoor_points = torch.randn(1, 150, 7) * 8
outdoor_points[:, :, :3] += torch.tensor([5.0, 0, 2.0])
occluded_points = torch.randn(1, 80, 7) * 4
occluded_points[:, :, :3] += torch.tensor([2.0, 0, 1.0])
with torch.no_grad():
for name, points in [("室内近距离", indoor_points),
("室外远距离", outdoor_points),
("遮挡场景", occluded_points)]:
pose, root, _ = model(points)
print(f"{name}: 骨盆位置 = [{root[0,0]:.2f}, {root[0,1]:.2f}, {root[0,2]:.2f}], "
f"姿态范围 = [{pose.min():.2f}, {pose.max():.2f}]")
def test_body_token_analysis():
"""人体 Token 空间分析"""
print("\n" + "=" * 60)
print("人体 Token 空间分析")
print("=" * 60)
model = Wave2Body()
codebook = model.body_tokenizer.codebook.weight
print(f"码本大小: {codebook.shape}")
print(f"码本距离统计: 均值={torch.cdist(codebook, codebook).mean():.4f}, "
f"最大={torch.cdist(codebook, codebook).max():.4f}")
poses = []
for _ in range(5):
pc = torch.randn(1, 200, 7) * 5
with torch.no_grad():
_, _, tokens = model(pc)
poses.append(tokens)
all_tokens = torch.cat(poses, dim=0)
print(f"\nToken 多样性分析:")
print(f" 唯一 token 数: {len(torch.unique(all_tokens))}")
print(f" Token 分布:")
for t in all_tokens.flatten():
print(f" Token {t.item()}", end="")
print()
if __name__ == "__main__":
model = test_wave2body()
test_cross_domain()
test_body_token_analysis()
|
