I use TensorRT to accelerate the inception v1 in onnx format, and get top1-accuracy 67.5% in fp32 format/67.5% in fp16 format, while get 0.1% in int8 after calibration.
The image preprocessing of the model is in bgr format, with mean subtraction [103.939, 116.779, 123.680]. Since tensorrt is not opensourced, I’ve no idea what’s going on inside the calibration tools. The images fed into the calibration tools should be the same format with the ones for inference, right?
Was there anything wrong when I was using the calibration or inference? Or this type of unnormalized image format not friendly as input?
I attached my script , onnx weight and calibration cache below.
Could you help to inspect it ? thanks.
Actually the environment is the Flashed environment by Jetpack 4.4 on Jetson AGX Xavier.
TensorRT Version 7: 7.1 (Flashed by Jetpack4.4)
GPU Type: GPU of Jetson Xavier
Nvidia Driver Version:
CUDA Version: 10.2
Operating System + Version: Ubuntu 18.04
Python Version (if applicable): 3.6
TensorFlow Version (if applicable):
PyTorch Version (if applicable): 1.2 aarch64, downloaded from this forum
Baremetal or Container (if container which image + tag):