I’m porting onnx model to tensorrt engine. I’ve tried onnx2trt and trtexec to generate fp32 and fp16 model. When it comes to int8, it seems onnx2trt does not support int8 quantization. After I set --int8 flag when converting onnx model to tensorrt, without providing the calib file, the inference result from the int8 engine differs a lot from the fp32 one.
So I’d like to try with some calibration image and then do the int8 quantization but no idea how to generate the calib file.
Any sample code or tools is welcome. Thanks.
Freshly installed Jetpack 4.4 DP