The rknn-toolkit2 is used for compiling your inference models into RKNN format. Rockchip also provides a Model Zoo with scripts to assist this conversion.

Compiling to RKNN format is done on a x86 workstation and the Dockerfile included in this directory has the rknn-toolkit2 and the Model Zoo files prepared with python dependencies installed.

Along with the compile-models.sh script the original ONNX model files used in the examples can be easily compiled to RKNN format for your target platform: rk3562, rk3566, rk3568, rk3576, or rk3588.

Note that rk3566 and rk3568 share the same RKNN compiled model, as does rk3582 and rk3588 are the same.

Compile Example Models

Models compiled to RKNN format used in the examples are already available at example/data/models/, however these are generated as follows on a x86 workstation.

# run from project toolkit directory 
cd toolkit/

docker run --rm \
  -v ../example/data/models:/opt/rkmodels \
  -v "$(pwd)/compile-models.sh:/compile-models.sh" \
  swdee/rknn-toolkit:latest \
  bash -c "chmod +x /compile-models.sh && /compile-models.sh <platform>"

An explanation of each parameter in the docker command is as follows;

Parameter	Description
-v ../example/data/models:/opt/rkmodels	Mount the directory ../example/data/models from your x86 workstation into the container as /opt/rkmodels
-v "$(pwd)/compile-models.sh:/compile-models.sh"	Copy the compile-models.sh script into the container
swdee/rknn-toolkit:latest	Use the prebuilt rknn-toolkit docker image
bash -c "chmod +x /compile-models.sh && /compile-models.sh "	Make the compile-models.sh script executable and compile models for

Make sure you replace <platform> in the above docker command with the Rockchip platform options of rk3562|rk3566|rk3568|rk3576|rk3588. The parameter all can be used to compile models for all platforms.

Compile Custom Model

If you have trained your own model you can use the Docker image to compile to RKNN format.

Take a custom trained YOLOv8 model for example with ONNX file located on your x86 workstation at /tmp/my-model/yolov8.onnx, we can use the Model Zoo convert.py script to compile to RKNN format for the rk3588 platform.

# run from project toolkit directory
cd toolkit/

docker run --rm \
  -v /tmp/my-model:/tmp/my-model \
  swdee/rknn-toolkit:latest \
  bash -c "cd /opt/rknn_model_zoo/examples/yolov8/python && python convert.py /tmp/my-model/yolov8.onnx rk3588 i8 /tmp/my-model/yolov8-rk3588.rknn"

After this command successfully runs the compiled RKNN model will be on your x86 workstation at /tmp/my-model/yolov8-rk3588.rknn. You can then copy this file to your Rockchip based SBC for running on the NPU.