rk3588's yolov8 model conversion from pt to rknn

Posted on | Edited on | In | Comments:

Background

Yolov8’s original model includes post-processing. Some shapes exceed the matrix calculation limit of 3588, so some cropping of the output layer is required.

pt to onnx

Clone ultralytics_yolov8 repository and pull docker

1
2
3
4
docker pull kaylor/rk3588_pt2onnx
git clone https://github.com/airockchip/ultralytics_yolov8.git
cd ultralytics_yolov8
git checkout 5b7ddd8f821c8f6edb389aa30cfbc88bd903867b

Download the newest model files from Yolov8 github repository.
For example, I download the model named yolov8n.pt

1
wget https://github.com/ultralytics/assets/releases/download/v8.1.0/yolov8n.pt

Edit ./ultralytics/cfg/default.yaml, replace “yolov8m-seg.pt” with “yolov8n.pt”

1
model: yolov8n.pt # (str, optional) path to model file, i.e. yolov8n.pt, yolov8n.yaml

Convert pt to onnx…

1
2
3
4
5
6
7
8
# run the command in your host 
docker run -it -v ${PWD}:/root/ws kaylor/rk3588_pt2onnx bash
----------------------------------
# run commnads in your container
cd /root/ws
export PYTHONPATH=./ 
python ./ultralytics/engine/exporter.py
exit

onnx to rknn

Clone rk3588-convert-to-rknn repository and pull docker

1
2
3
4
5
6
cd ../
docker pull kaylor/rk3588_onnx2rknn # for yolov8 
docker pull kaylor/rk3588_onnx2rknn:beta # for yolov10
git clone https://github.com/kaylorchen/rk3588-convert-to-rknn.git
cp ultralytics_yolov8/yolov8n.onnx rk3588-convert-to-rknn
cd rk3588-convert-to-rknn

Convert onnx to rknn

1
2
3
4
5
6
7
8
9
# run the command in your host 
docker run -it -v ${PWD}:/root/ws kaylor/rk3588_onnx2rknn bash # for yolov8
docker run -it -v ${PWD}:/root/ws kaylor/rk3588_onnx2rknn:beta bash # for yolov10
docker run -it -v ${PWD}:/root/ws kaylor/rk3588_onnx2rknn:2.3.0 bash # for all the yolos
----------------------------------
# run commnads in your container
cd /root/ws
python convert.py yolov8n.onnx rk3588 i8 yolov8n.rknn
exit

Enjoy ~~

0%