yolov8-onnx.cpp

YOLOv8 using onnx.cpp and stb image library, without ONNX Runtime or OpenCV. All the necessary files are included in the repository and you only need the basic build environment and protobuf library on your system.

Build and Run

By default, this repository contains test data files, which are not necessary for normal build and run. To clone the repository including the test data, just run:

git clone --recursive https://github.com/han-minhee/yolov8-onnx.cpp.git

However, if you want to clone the repository without the test data files:

mkdir yolov8-onnx.cpp && cd yolov8-onnx.cpp
git init
git remote add origin https://github.com/han-minhee/yolov8-onnx.cpp.git
git config core.sparseCheckout true
echo '/*' > .git/info/sparse-checkout
echo '!/tests/data/*' >> .git/info/sparse-checkout
git pull origin main
git submodule update --init --recursive

Sample Run

This will build and run the sample inference using /sample/yolov8n.onnx and /sample/people1.jpg, generate /sample/people1_output.jpg and /sample/people1_output.txt files.

mkdir build && cd build
cmake -G Ninja ..
cmake --build .
./yolov8

HIP Backend

You can give -DUSE_HIP=on flag to the cmake to use HIP Backend for GPU inference.

mkdir build && cd build
cmake -G Ninja .. -DUSE_HIP=on
cmake --build .
./yolov8

Full Test

This repo contains gtests and related data. If you want to run the test, you should clone the npy and onnx files too. After build, just run

ctest

The session_test is the comprehensive one that runs the whole session with a fixed input, and compare the results with the results from running the input using ONNX Runtime.

Python Scripts

The python scripts are used for building the onnx.cpp and this example.

• 0_download_model.py downloads the YOLOv8 model file and convert it to an ONNX file.
• 1_modify_onnx.py modifies the onnx file so that we can get the intermediate tensors.
• 2_export_npy.py runs the model on a demo input, exports the tensor values to npy files, to ./python_scripts/exports

These files are used to validate the implementation of each operators and the whole model inference result.

Credits

YOLOv8

stb image library