RKNN Converter

RKNN Converter

About

This tool provides to convert YOLO models (or ONNX models) to RKNN format.

Important

This tool is based on rknn-toolkit2 version 1.6.0. You should be using rknn-toolkit-lite2 on an SBC of the same version.

Installation

System dependencies

System dependencies must be installed for a successful installation.

sudo apt install -y libgl1-mesa-dev

Cloning the repository and install

The following steps are for Python 3.11. If you have a different version, refer to this section first.

cd ~
git clone https://github.com/denisbondar/rknn-converter.git
cd rknn-converter
python3.11 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Conversion

Preparing

You will need a YOLO format model — usually the best.pt file. You will also need a dataset of 20 photos matching your model for the quantization process.

You don't have to place all these files inside the current rknn-converter directory, you can place them wherever you like. However, for simplicity and reliability, the instructions below will assume that the files are in the model subdirectory of the current rknn-converter directory:

Create the necessary directories:

mkdir -p model/images

And then copy your model file best.pt to the model directory.

Creating dataset

Create a model directory and navigate to it. In the model directory, create the images directory and copy your images there. While in the model directory, run the command that creates an image index, a text file that lists the paths to each of the image files.

cd model
ls ./images/* > images.txt
cd ..

The content of the images.txt file should be (you can see it with the cat model/images.txt command).

./images/01.jpg
./images/02.jpg
./images/03.jpg

... etc ...

In general, the file tree will be something like this:

.
├── model
│   ├── best.pt
│   ├── images
│   │   ├── 01.jpg
│   │   ├── 02.jpg
│   │   ├── 03.jpg
│   │   ├── . . .
│   │   └── 20.jpg
│   └── images.txt
├── pt2rknn.py
├── README.md
├── requirements.txt
└── rknn-toolkit2_pkgs

Executing

Make sure you are still in an active virtual environment - your command line prompt prefix should be (venv), and run the following command:

python3 pt2rknn.py -m model/best.pt -d model/images.txt

As a result, two new files should be created: model/best.onnx and model/best-rk3588.rknn.

The command also contains optional parameters such as image size (default 640x640) and target platform (default kr3588). You can run the command with the -h flag to see all available options.

Alternatively, you can specify a model of type ONNX as a source model and get a model in RKNN format.

$ python3 pt2rknn.py -h
usage: pt2rknn.py [-h] -m MODEL -d DATASET [-s IMGSIZE] [-p PLATFORM]

YOLOv8 to RKNN converter tool

options:
  -h, --help            show this help message and exit
  -m MODEL, --model MODEL
                        File mame of YOLO model (PyTorch format .pt)
  -d DATASET, --dataset DATASET
                        Path to dataset .txt file for quantization
  -s IMGSIZE, --imgsize IMGSIZE
                        Image size of the side of the square or H:W. (default: 640 or 640:640)
  -p PLATFORM, --platform PLATFORM
                        RKNN target platform (default: rk3588)

Other versions of Python

Python 3.10

If you are using Python version 3.10, before installing, edit the requirements.txt file by replacing the first line with

./rknn-toolkit2_pkgs/rknn_toolkit2-1.6.0+81f21f4d-cp311-cp311-linux_x86_64.whl

by replacing the first line

./rknn-toolkit2_pkgs/rknn_toolkit2-1.6.0+81f21f4d-cp310-cp310-linux_x86_64.whl

And only after that, run the pip install command.

Python 3.6 - 3.9

If your version of Python is between 3.6 and 3.9, you will first need to repackage the original rknn-toolkit2 package. To do this, add the desired version, such as "cp39", to the PYTHON_VERSIONS list of the rknn-toolkit2_pkgs/repack-whl.sh file. Then run this file, and it will create a package for your version of Python. Now modify requirements.txt as above with the name of your package.

You can now run the pip install ...command as shown in the installation section.

Python <3.6 or >3.11

rknn-toolkit2 v1.6.0 only supports Python in the 3.6 to 3.11 version range.

Support

Given the instability and persistent problems of the original rknn-toolkit package, there are likely to be future problems that simply did not exist when the converter was written.

Discussions and featurerequests

For questions and discussions, it is best to use discussions board so as not to clutter the issues with questions. If you're not sure, it's always best to start with discussions, as your issue may have already been or is currently being discussed.

Bugfixes

If you find a clear bug that can be fixed, see if there is an issue with a similar bug before creating a new one.

denisbondar/rknn-converter