Skip to content

Latest commit

 

History

History
72 lines (53 loc) · 1.72 KB

README.md

File metadata and controls

72 lines (53 loc) · 1.72 KB

YOLOv8 deploy in RK3588

1. Prepare your environment

1.1 On X86 PC

Suggest to use anaconda to create a virtual environment.

conda create -n rknn python=3.8
conda activate rknn

Install yolov8:

git clone https://github.com/triple-Mu/yolov8.git -b triplemu/model-only
# uninstall ultralytics first
pip uninstall ultralytics
# install yolov8
cd yolov8
pip install -r requirements.txt
pip install .

Convert pt to onnx:

git clone https://github.com/triple-Mu/AI-on-Board.git
cd AI-on-Board/Rockchip/python/yolov8
# modify the export.py: pt_path to your own first
python export.py

Convert onnx to rknn:

rknn_toolkit2-1.5.0+1fa95b5c-cp38-cp38-linux_x86_64.whl is in packages

pip install rknn_toolkit2-1.5.0+1fa95b5c-cp38-cp38-linux_x86_64.whl
# modify the onnx2rknn.py: ONNX_MODEL RKNN_MODEL IMG_PATH DATASET IMG_SIZE
python onnx2rknn.py

1.2 On ARM RK3588

Copy this repo to your board.

Install rknn-lite and triplemu tools: rknn_toolkit_lite2-1.5.0-cp38-cp38-linux_aarch64.whl and triplemu-0.0.1-cp38-cp38-linux_aarch64.whl is in packages

cd AI-on-Board/Rockchip/python/yolov8
# install rknn_toolkit_lite and triplemu tools on RK3588
pip install rknn_toolkit_lite2-1.5.0-cp38-cp38-linux_aarch64.whl
pip install triplemu-0.0.1-cp38-cp38-linux_aarch64.whl

2. Run

python rknn_infer.py --input zidane.jpg --rknn yolov8s.rknn --show

Description of all arguments

  • --input : The image path or images dir or mp4 path.
  • --rknn : The rknn model path.
  • --show : Whether to show results.
  • --output : The output dir path for saving results.
  • --iou-thres : IoU threshold for NMS algorithm.
  • --conf-thres : Confidence threshold for NMS algorithm.