use QuantFormat.QOperator by default to avoid fake quantization (#88)
Browse files
tools/quantize/quantize-ort.py
CHANGED
@@ -12,7 +12,7 @@ import cv2 as cv
|
|
12 |
import onnx
|
13 |
from onnx import version_converter
|
14 |
import onnxruntime
|
15 |
-
from onnxruntime.quantization import quantize_static, CalibrationDataReader, QuantType
|
16 |
|
17 |
from transform import Compose, Resize, CenterCrop, Normalize, ColorConvert
|
18 |
|
@@ -70,6 +70,7 @@ class Quantize:
|
|
70 |
self.check_opset()
|
71 |
output_name = '{}-act_{}-wt_{}-quantized.onnx'.format(self.model_path[:-5], self.act_type, self.wt_type)
|
72 |
quantize_static(self.model_path, output_name, self.dr,
|
|
|
73 |
per_channel=self.per_channel,
|
74 |
weight_type=self.type_dict[self.wt_type],
|
75 |
activation_type=self.type_dict[self.act_type])
|
|
|
12 |
import onnx
|
13 |
from onnx import version_converter
|
14 |
import onnxruntime
|
15 |
+
from onnxruntime.quantization import quantize_static, CalibrationDataReader, QuantType, QuantFormat
|
16 |
|
17 |
from transform import Compose, Resize, CenterCrop, Normalize, ColorConvert
|
18 |
|
|
|
70 |
self.check_opset()
|
71 |
output_name = '{}-act_{}-wt_{}-quantized.onnx'.format(self.model_path[:-5], self.act_type, self.wt_type)
|
72 |
quantize_static(self.model_path, output_name, self.dr,
|
73 |
+
quant_format=QuantFormat.QOperator, # start from onnxruntime==1.11.0, quant_format is set to QuantFormat.QDQ by default, which performs fake quantization
|
74 |
per_channel=self.per_channel,
|
75 |
weight_type=self.type_dict[self.wt_type],
|
76 |
activation_type=self.type_dict[self.act_type])
|