cuda 10.2
lastest onnx2trt from this repo: GitHub - onnx/onnx-tensorrt: ONNX-TensorRT: TensorRT backend for ONNX
I want to convert my onnx model to tensorrt engine int8 not fp16 when running my deepstream app. How can I do that? Thank you.
The wiki have details about build and run, did you meet any issue? GitHub - onnx/onnx-tensorrt: ONNX-TensorRT: TensorRT backend for ONNX
besides, we support onnx model parsing in deepstream, you can set your model in infer configuration
onnx-file = your onnx model
yes, I built it successfully. But the engine model was generated was in fp16, I really want my engine model built with int8 option. How can I configure the config file?
Do we need a calibration file for int8 mode?
Thank you.
Yes, we need INT8 calibration file only in INT8, if engine file not specified.
int8-calib-file
1 Like
I am using Yolov5 model. How can I create a calib file for that model?
For INT8 calibration, you may refer to TRT documentation,
Developer Guide :: NVIDIA Deep Learning TensorRT Documentation