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
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
| import torch
import torch.nn as nn
import torch.quantization as quant
from torch.quantization import get_default_qconfig, prepare_qat, convert
class DMSModel(nn.Module):
"""
DMS模型示例
包含：
- 人脸检测
- 眼睛检测
- 疲劳判断
"""
def __init__(self):
super().__init__()
self.backbone = nn.Sequential(
nn.Conv2d(3, 32, 3, 2, 1),
nn.BatchNorm2d(32),
nn.ReLU(),
nn.Conv2d(32, 64, 3, 2, 1),
nn.BatchNorm2d(64),
nn.ReLU(),
nn.Conv2d(64, 128, 3, 2, 1),
nn.BatchNorm2d(128),
nn.ReLU(),
)
self.head = nn.Sequential(
nn.AdaptiveAvgPool2d(1),
nn.Flatten(),
nn.Linear(128, 64),
nn.ReLU(),
nn.Linear(64, 3)
)
def forward(self, x):
feat = self.backbone(x)
out = self.head(feat)
return out
def quantize_model_ptq(model, calibration_loader):
"""
训练后量化（PTQ）
Args:
model: FP32模型
calibration_loader: 校准数据集
Returns:
quantized_model: INT8量化模型
"""
model.qconfig = get_default_qconfig('fbgemm')
model_prepared = quant.prepare(model, inplace=True)
print("开始校准...")
with torch.no_grad():
for batch in calibration_loader:
model_prepared(batch)
model_quantized = quant.convert(model_prepared)
print("量化完成！")
return model_quantized
def quantize_model_qat(model, train_loader, epochs=10):
"""
量化感知训练（QAT）
优势：精度损失更小，适合精度敏感场景
Args:
model: FP32模型
train_loader: 训练数据集
epochs: 训练轮数
Returns:
quantized_model: INT8量化模型
"""
model.qconfig = get_default_qconfig('fbgemm')
model_prepared = prepare_qat(model, inplace=True)
optimizer = torch.optim.SGD(model_prepared.parameters(), lr=0.001)
criterion = nn.CrossEntropyLoss()
print("开始QAT训练...")
for epoch in range(epochs):
for batch, labels in train_loader:
optimizer.zero_grad()
outputs = model_prepared(batch)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
print(f"Epoch {epoch+1}/{epochs}, Loss: {loss.item():.4f}")
model_quantized = convert(model_prepared)
print("QAT完成！")
return model_quantized
def export_to_onnx(model, output_path, input_shape=(1, 3, 224, 224)):
"""
导出为ONNX格式
Args:
model: 量化模型
output_path: 输出路径
input_shape: 输入尺寸
"""
model.eval()
dummy_input = torch.randn(input_shape)
torch.onnx.export(
model,
dummy_input,
output_path,
export_params=True,
opset_version=13,
do_constant_folding=True,
input_names=['input'],
output_names=['output'],
dynamic_axes={
'input': {0: 'batch_size'},
'output': {0: 'batch_size'}
}
)
print(f"ONNX模型已导出：{output_path}")
def benchmark_quantization(model_fp32, model_int8, test_loader, device='cpu'):
"""
对比FP32和INT8模型性能
Returns:
results: 性能对比结果
"""
import time
results = {
'fp32': {'accuracy': 0, 'latency': 0, 'size_mb': 0},
'int8': {'accuracy': 0, 'latency': 0, 'size_mb': 0}
}
model_fp32.eval()
model_fp32.to(device)
correct = 0
total = 0
latencies = []
with torch.no_grad():
for batch, labels in test_loader:
batch = batch.to(device)
labels = labels.to(device)
start = time.time()
outputs = model_fp32(batch)
latency = (time.time() - start) * 1000
latencies.append(latency)
_, predicted = torch.max(outputs, 1)
total += labels.size(0)
correct += (predicted == labels).sum().item()
results['fp32']['accuracy'] = 100 * correct / total
results['fp32']['latency'] = sum(latencies) / len(latencies)
results['fp32']['size_mb'] = sum(p.numel() * 4 for p in model_fp32.parameters()) / 1024 / 1024
model_int8.eval()
correct = 0
total = 0
latencies = []
with torch.no_grad():
for batch, labels in test_loader:
start = time.time()
outputs = model_int8(batch)
latency = (time.time() - start) * 1000
latencies.append(latency)
_, predicted = torch.max(outputs, 1)
total += labels.size(0)
correct += (predicted == labels).sum().item()
results['int8']['accuracy'] = 100 * correct / total
results['int8']['latency'] = sum(latencies) / len(latencies)
results['int8']['size_mb'] = sum(p.numel() for p in model_int8.parameters()) / 1024 / 1024
print("\n性能对比：")
print(f"{'指标':<15} {'FP32':<15} {'INT8':<15} {'提升':<15}")
print("-" * 60)
print(f"{'准确率 (%)':<15} {results['fp32']['accuracy']:<15.2f} {results['int8']['accuracy']:<15.2f} {results['int8']['accuracy'] - results['fp32']['accuracy']:+.2f}")
print(f"{'延迟 (ms)':<15} {results['fp32']['latency']:<15.2f} {results['int8']['latency']:<15.2f} {(1 - results['int8']['latency']/results['fp32']['latency'])*100:+.1f}%")
print(f"{'大小 (MB)':<15} {results['fp32']['size_mb']:<15.2f} {results['int8']['size_mb']:<15.2f} {(1 - results['int8']['size_mb']/results['fp32']['size_mb'])*100:+.1f}%")
return results
if __name__ == '__main__':
model = DMSModel()
|