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
| """
DMS模型高通NPU部署工具链
PyTorch → ONNX → QNN → Hexagon NPU
"""
import torch
import onnx
import subprocess
class QualcommDeploy:
"""高通NPU部署工具"""
def __init__(self, target_device='QCS8295'):
self.target = target_device
self.qnn_sdk = '/opt/qnn-sdk'
def full_pipeline(self, pytorch_model, input_shape, output_dir):
"""完整部署流程"""
onnx_path = f'{output_dir}/model.onnx'
self.export_onnx(pytorch_model, input_shape, onnx_path)
qnn_path = f'{output_dir}/model_qnn.onnx'
self.quantize_onnx(onnx_path, qnn_path)
lib_path = f'{output_dir}/model.so'
self.compile_qnn(qnn_path, lib_path)
self.benchmark(lib_path)
return lib_path
def export_onnx(self, model, input_shape, output_path):
"""PyTorch → ONNX"""
model.eval()
dummy = torch.randn(*input_shape)
torch.onnx.export(
model, dummy, output_path,
input_names=['input'],
output_names=['output'],
opset_version=17,
dynamic_axes={'input': {0: 'batch'}, 'output': {0: 'batch'}}
)
print(f'✅ ONNX导出: {output_path}')
def quantize_onnx(self, onnx_path, output_path):
"""INT8量化"""
cmd = f'{self.qnn_sdk}/bin/qnn-quantize \
--input_list {onnx_path} \
--output {output_path} \
--quantization_mode int8 \
--target_hexagon'
subprocess.run(cmd, shell=True, check=True)
print(f'✅ INT8量化: {output_path}')
def compile_qnn(self, quantized_path, output_path):
"""QNN编译为Hexagon库"""
cmd = f'{self.qnn_sdk}/bin/qnn-net-run \
--model {quantized_path} \
--backend libQnnHtp.so \
--output_dir {output_path}'
subprocess.run(cmd, shell=True, check=True)
print(f'✅ Hexagon编译: {output_path}')
def benchmark(self, lib_path):
"""基准测试"""
cmd = f'{self.qnn_sdk}/bin/qnn-benchmark \
--model {lib_path} \
--device {self.target}'
result = subprocess.run(cmd, shell=True, capture_output=True, text=True)
print(f'基准测试结果:\n{result.stdout}')
def deploy_dms_pipeline():
"""DMS模型部署完整流程"""
deploy = QualcommDeploy(target_device='QCS8295')
print("DMS模型部署流程示例")
print("1. PyTorch模型 → ONNX导出")
print("2. ONNX → INT8量化 (QNN工具)")
print("3. QNN → Hexagon NPU编译")
print("4. 基准测试 (延迟/内存/功耗)")
class DMSInference:
"""DMS推理接口（Hexagon NPU）"""
def __init__(self, model_lib_path):
import libQnnHtp
self.context = libQnnHtp.create_context(model_lib_path)
def infer(self, frame):
"""
单帧推理
Args:
frame: (H, W, 3) BGR图像
Returns:
results: dict
"""
import cv2
import numpy as np
input = cv2.resize(frame, (224, 224))
input = input.transpose(2, 0, 1).astype(np.float32)
input = input / 255.0
input = input[np.newaxis, :]
output = self.context.execute(input)
return output
|
