Yolov8 on OrangePi 5

Configure PC for converting models to .rknn

1. Install requirements.

   # Download
   wget https://github.com/airockchip/rknn-toolkit2/raw/v2.1.0/rknn-toolkit2/packages/requirements_cp310-2.1.0.txt
   
   # Install
   pip install -r requirements_cp310-2.1.0.txt
   

2. Install whls for rknn-toolkit2.

   # Download
   wget https://github.com/airockchip/rknn-toolkit2/raw/v2.1.0/rknn-toolkit2/packages/rknn_toolkit2-2.1.0+708089d1-cp310-cp310-linux_x86_64.whl
   
   # Install
   pip install rknn_toolkit2-2.1.0+708089d1-cp310-cp310-linux_x86_64.whl
   

3. Install ultralytics_yolov8 special for converting pt -> onnx (optimized for rknn).

   # Clone repo
   git clone https://github.com/airockchip/ultralytics_yolov8
   
   # Go to cloned directory
   cd ultralytics_yolov8
   
   # Install as package
   python setup.py install
   

Convert model to .rknn

1. Convert pt to onnx.

   from ultralytics import YOLO
   model = YOLO("yolov8.pt")
   path = model.export(format="rknn")  # Internal method written by airockchip, don't be fooled by the format name
   

1.5 Check the input size when exporting the model. If necessary, change batch_size parameter in ultralytics/cfg/default.yaml to any value.

2. Convert onnx to rknn.

   # Clone repo
   git clone https://github.com/airockchip/rknn_model_zoo
   
   # Go to directory with converter
   cd rknn_model_zoo/examples/yolov8/python
   
   # Run converter
   python convert.py <path-to-onnx-model>/yolov8n.onnx rk3588 i8 ../model/yolov8n.rknn
   

2.5 If the model has issues or warnings in convertation process, you can change opset version from 12 to 17 or 19, depending on PyTorch version. Currently, RKNN==2.1.0 recommends opset 19.

3. Save and send it to Orange Pi.

Install OS

1. Download image:

   Ubuntu (OrangePi 5)	Ubuntu (OrangePi 5B)	Armbian (OrangePi 5/5B)

2. Burn it to SD card.

3. Plug SD card to Orange Pi.

Configure OrangePi for running models

1. Update librknnrt.so.

   # Download
   wget https://github.com/airockchip/rknn-toolkit2/raw/v2.1.0/rknpu2/runtime/Linux/librknn_api/aarch64/librknnrt.so
   
   # Move to /usr/lib
   sudo mv ./librknnrt.so /usr/lib
   

2. Install whls for rknn-toolkit-lite2.

   # Download
   wget https://github.com/airockchip/rknn-toolkit2/raw/v2.1.0/rknn-toolkit-lite2/packages/rknn_toolkit_lite2-2.1.0-cp310-cp310-linux_aarch64.whl
   
   # Install
   pip install rknn_toolkit_lite2-2.1.0-cp310-cp310-linux_aarch64.whl
   

3. Install opencv-python and other requirements (if necessary).

   pip install opencv-python
   

Run model

1. Move rknn model to models directory.

2. Change path to model in main.py.

3. Run main.py.

   python main.py