Rockchip RKNN-Toolkit

Rockchip's model-conversion toolchain that turns neural networks into a format its NPUs can run.

RKNN-Toolkit is the development toolchain Rockchip provides for the NPU (neural processing unit) built into its own chips. It converts models trained in frameworks such as PyTorch, ONNX, or TensorFlow into the RKNN format, optionally applying INT8 quantization along the way, so they can then run on the NPU of chips such as the RK3588. The current generation of chips uses RKNN-Toolkit2; on-device Python inference uses RKNN-Toolkit-Lite2; a C interface is provided by the RKNPU runtime. Many low-cost robots and development boards use the RK3588 as their main computer, and running detection, segmentation, or small policy networks on it generally goes through this toolchain.

ExampleExport a trained YOLO model to ONNX, quantize and convert it to a .rknn file with RKNN-Toolkit2, then deploy it on an RK3588 development board for real-time object detection.

Also called

RKNN, RKNN-Toolkit2, RKNN-Toolkit-Lite2

Related

Rockchip RK3588 · On-Device / Edge Deployment · Post-Training Quantization · Open Neural Network Exchange (ONNX) · Neural Processing Unit (NPU) · Inference Deployment

Sources

airockchip/rknn-toolkit2 on GitHub