This directory contains tools for working with RKNN models, specifically for converting YOLO models to the RKNN format for use with Rockchip NPU hardware.

The model converter tool allows you to convert YOLO models (in ONNX, PyTorch, or TFLite format) to RKNN format for deployment on Rockchip NPU devices.

• YOLOv5
• YOLOv8
• YOLOv10
• YOLOv11
• YOLOX
• YOLOv5-Segmentation
• YOLOv8-Segmentation
• YOLOv8-Pose
• YOLOv8-OBB (Oriented Bounding Box)

• RK3566
• RK3568
• RK3588
• RK3588S

go run main.go convert --input <input_model> --output <output_model> [options]

• --input string: Path to input YOLO model (.onnx, .pt, .pth, .tflite)
• --output string: Path for output RKNN model (.rknn)
• --type string: YOLO model type (v5, v8, v10, v11, x, v5seg, v8seg, v8pose, v8obb) (default "v5")
• --target string: Target platform (rk3566, rk3568, rk3588, rk3588s) (default "rk3588")
• --quant string: Quantization type (int8, float) (default "int8")
• --size string: Input size in format widthxheight (default "640x640")
• --verbose: Enable verbose logging

go run main.go convert --input yolov5s.onnx --output yolov5s-rk3588.rknn --type v5 --target rk3588 --quant int8 --size 640x640

To build the tool as a standalone executable:

cd tools
go build -o rknn-convert

Then you can use it directly:

./rknn-convert convert --input yolov5s.onnx --output yolov5s-rk3588.rknn

• This tool currently provides a simulation of the conversion process. In a real implementation, it would use the RKNN SDK to perform the actual conversion.
• The conversion process requires the RKNN Toolkit to be installed on your system.
• For optimal performance, it's recommended to use the int8 quantization option, which provides faster inference with minimal accuracy loss.
• Make sure to specify the correct input size that matches your model's expected input dimensions.