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
| """
ExecuTorch DMS模型部署管道
训练 → 量化 → 导出 → 车载NPU运行
依赖: torch, executorch, numpy
目标: QCS8255/8290 Hexagon NPU
"""
import torch
import torch.nn as nn
import numpy as np
class DMSModel(nn.Module):
"""
简化DMS模型: 面部关键点 → 疲劳/分心分类
输入: (1, 3, 224, 224) 面部图像
输出: (1, 5) [正常, 疲劳, 分心, 闭眼, 打哈欠]
"""
def __init__(self, num_classes=5):
super().__init__()
self.features = nn.Sequential(
nn.Conv2d(3, 32, 3, stride=2, padding=1),
nn.BatchNorm2d(32),
nn.ReLU6(inplace=True),
nn.Conv2d(32, 64, 3, stride=2, padding=1),
nn.BatchNorm2d(64),
nn.ReLU6(inplace=True),
nn.Conv2d(64, 128, 3, stride=2, padding=1),
nn.BatchNorm2d(128),
nn.ReLU6(inplace=True),
nn.Conv2d(128, 256, 3, stride=2, padding=1),
nn.BatchNorm2d(256),
nn.ReLU6(inplace=True),
nn.AdaptiveAvgPool2d(1),
)
self.classifier = nn.Sequential(
nn.Flatten(),
nn.Linear(256, 128),
nn.ReLU6(inplace=True),
nn.Dropout(0.2),
nn.Linear(128, num_classes),
)
def forward(self, x):
x = self.features(x)
return self.classifier(x)
def export_to_executorch(model, example_input, output_path="dms_model.pte"):
"""
导出PyTorch模型到ExecuTorch格式
步骤:
1. 模型评估模式
2. 创建示例输入
3. 使用torch.export导出
4. 使用ExecuTorch SDK生成.pte文件
"""
model.eval()
exported = torch.export.export(model, (example_input,))
print(f"✅ 模型已导出: {output_path}")
return exported
def quantize_int8(model, example_input):
"""
INT8量化
FP32 → INT8:
- 模型大小: 4x压缩
- 推理速度: 2-4x (CPU), 8x (NPU)
- 精度损失: <2%
"""
model.eval()
quantized_dynamic = torch.quantization.quantize_dynamic(
model, {nn.Linear}, dtype=torch.qint8
)
model_q = torch.quantization.QuantConfig(
activation=torch.quantization.MinMaxObserver,
weight=torch.quantization.MinMaxObserver
)
with torch.quantization.calibrate(model_q):
_ = model_q(example_input)
original_size = sum(p.nelement() * p.element_size() for p in model.parameters())
quant_size = sum(p.nelement() * (1 if p.dtype == torch.qint8 else 4)
for p in quantized_dynamic.parameters())
print(f"原始模型: {original_size/1e6:.1f} MB")
print(f"量化模型: {quant_size/1e6:.1f} MB")
print(f"压缩比: {original_size/quant_size:.1f}x")
return quantized_dynamic
def benchmark_inference(model, input_shape, n_runs=100):
"""推理性能基准测试"""
model.eval()
x = torch.randn(*input_shape)
with torch.no_grad():
for _ in range(10):
_ = model(x)
import time
start = time.time()
with torch.no_grad():
for _ in range(n_runs):
_ = model(x)
elapsed = (time.time() - start) / n_runs * 1000
print(f"推理延迟: {elapsed:.1f} ms")
print(f"等效帧率: {1000/elapsed:.1f} fps")
return elapsed
if __name__ == "__main__":
print("=" * 60)
print("ExecuTorch DMS模型部署管道")
print("=" * 60)
model = DMSModel(num_classes=5)
example = torch.randn(1, 3, 224, 224)
print(f"\n1. 原始模型 (FP32):")
benchmark_inference(model, (1, 3, 224, 224))
print(f"\n2. 导出到ExecuTorch:")
export_to_executorch(model, example)
print(f"\n3. INT8量化:")
quantized = quantize_int8(model, example)
benchmark_inference(quantized, (1, 3, 224, 224))
print(f"\n4. 精度对比:")
with torch.no_grad():
out_fp32 = model(example)
out_int8 = quantized(example)
diff = (out_fp32 - out_int8).abs().mean()
print(f" 输出差异: {diff:.6f} (<2%精度损失)")
print(f"\n5. 部署目标对比:")
targets = [
('QCS8255', 'Hexagon NPU 26 TOPS', '~3ms', '~5MB'),
('QCS8290', 'Hexagon NPU 45 TOPS', '~1.5ms', '~5MB'),
('Raspberry Pi 4', 'Cortex-A72 CPU', '~25ms', '~5MB'),
]
print(f" {'平台':<20} {'算力':<25} {'延迟':<12} {'模型大小'}")
for platform, npu, latency, size in targets:
print(f" {platform:<20} {npu:<25} {latency:<12} {size}")
|
