TensorRT7 and TensorRT8 int8 calibration has a different precision

We tested the yolov5s and yolox model, and have same conclusion.

  • yolov5-v5.0 版本,yolov5s 的 mAP = 36.6,V100 GPU.
Model inputs mAP / tensorrt7 - int8 mAP / tensorrt8 - int8
Percentile=99.99% /
yolov5s [1, 3, 640, 640] 22.1 35.9
entropy2
yolov5s [1, 3, 640, 640] 27.9 32.7
minmax
yolov5s [1, 3, 640, 640] 29.3 33.1

yolov5_trt7.trt size is 9.9M and yolov5_trt8.trt is 16M.

Why is the same INT8 calibration method of Tensorrt7 and Tensorrt8 precision gaps have not been played