YOLOv3 on TensorRT in int8

I would like to run YOLOv3 on TensorRT in int8 mode. I’m on a Jetson AGX Xavier, TensorRT 5.0.6, CUDA 10.0.

I found the calibration table for yolov3 here: https://github.com/NVIDIA-AI-IOT/deepstream_reference_apps/blob/master/yolo/data/yolov3-calibration.table

How do I use the calibration table? I found the method below but I don’t know how to use it in my code. tensorrt.tensorrt.IInt8Calibrator.read_calibration_cache()

Alternatively, how can I convert for example “yolo_107: 3cb957a3” directly to the (min_float, max_float) for every layer?


You can use the calibration here: