Introduction

CheetahInfer is a pure C++ inference SDK based on TensorRT, which supports fast inference of CNNs based computer vision model.

Features

• Efficient

  With the help of TensorRT's optimization to CNNs and the pure C++ implmentation of preprocessing and postprocessing, CheetahInfer is really efficent. If you are interested in flexibleness, you can refer to FlexInfer.

Prerequisites

CheetahInfer has several dependencies:

• OpenCV
  • tested on version 4.3.0
• CUDA
  • tested on version 10.2
• TensorRT
  • tested on version 7.1.3.4
• cuDNN
  • optional
  • tested on version 8.0.0
• GCC
  • tested on version 5.4.0

After the installation of above dependencies, we need modify the TENSORRT_INSTALL_DIR and OPENCV_INSTALL_DIR in file Makefile.config and the environment variable LD_LIBRARY_PATH and PATH in .bashrc file accordingly like the following.

export LD_LIBRARY_PATH="${LD_LIBRARY_PATH}:/home/yichaoxiong/opt/lib/tensorrt:/home/yichaoxiong/opt/lib/opencv"
export PATH="${PATH}:/usr/local/cuda-10.2/bin"

Preparation for model and data

• Prepare the ONNX file
  • If your model has a PyTorch format, you can use vedadep to convert PyTorch model to ONNX model.
• Modify the ONNX file path
  • Some related configurations in main.cpp in classifier folder also need be corrected accordingly.
• Get some images for testing

Compilation and running

cd classifier
make -j4
./build/main --imgfp /path/to/image

If you want speficy which GPU to use, you can make it by setting the environment variable CUDA_VISIBLE_DEVICES.

Contact

This repository is currently maintained by Hongxiang Cai (@hxcai), Yichao Xiong (@mileistone).

Credits

We got some code from TensorRT and retinanet-examples.