TensorRT-Image-Classification icon indicating copy to clipboard operation
TensorRT-Image-Classification copied to clipboard

Published 20 hours ago • verified publisher icon ivder

About Int8

Open guods opened this issue 6 years ago • 3 comments

Did you set nvinfer1::DataType as Int8 ? I set nvinfer1::DataType as Int8, did not create engine file? I think that it should use the calibtration data such as detector with TensorRT when set nvinfer1::DataType as Int8.

guods avatar Sep 18 '19 10:09 guods

@guods Hi is there any error message or just the engine is not created?

Also, what's your purpose of changing the nvinfer1::DataType as Int8? To calibrate the data? What happens if you set builder->setInt8Mode(true); ?

ivder avatar Sep 19 '19 00:09 ivder

Run on KINT8, error message "nvinfer1::Network::addScale::163, condition:scale.type == shift.type && shift.type == power.type, error parsing layer type Scale index 3.". Run on KFLOAT mode, the program is correct.

guods avatar Sep 19 '19 01:09 guods

Run on INT8 mode, the speed of classification is more faster; But in order to ensure recognition accuracy when the parameter is from fp32 to int8, it is necessary to calibrate the create engine by calibration data. NVIDIA official website has corresponding instructions.

guods avatar Sep 19 '19 01:09 guods