---
license: other
license_name: upstream-terms-by-artifact
license_link: https://huggingface.co/HanzoHuang/YOLOv5-RKNN2/blob/main/LICENSES.md
pipeline_tag: image-classification
tags:
- rknn
- rknn2
- rockchip
- rk3576
- rk3588
- yolov5
---

# YOLOv5-RKNN2

> **Looking for Ultralytics RKNN exports?** Browse the [Ultralytics exports collection](https://huggingface.co/collections/HanzoHuang/ultralytics-rknn-exports-6af1d088ba04ccf2dcbf81c2).

YOLOv5 RKNN builds for **RK3576** and **RK3588**, including Rockchip Model Zoo builds and custom conversions.

## Choose a model

Models are organized under `models/v<toolkit-version>/<task>/<variant>/<chip>/<precision>/`. Download the model and metadata from the same directory. The [manifest](manifest.json) records output hashes.

- [Toolkit 2.3.2 builds](tree/main/models/v2.3.2) include classification, detection, and segmentation. The classification exports include n, s, and m variants in FP16 and INT8 for both chips.
- [Toolkit 2.5.0 builds](tree/main/models/v2.5.0) include detection and segmentation.
- Detection includes n, s, m, and s-ReLU variants in FP16 and INT8 for both chips. Check the conversion receipt for the source graph and compiler; Toolkit 2.3.2 builds were tested on their target board.

## Conversion and license

The classification weights came from the [YOLOv5 v7.0 release](https://github.com/ultralytics/yolov5/releases/tag/v7.0); their URLs and hashes are in [source](tree/main/source). The custom conversion exports ONNX and compiles it with RKNN Toolkit2. Run `python source/convert-extended.py --help` and use the [v7.0 source checkout](https://github.com/ultralytics/yolov5/tree/v7.0) with `--yolov5-source`. INT8 uses the COCO128 calibration list in `source/calibration.json`.

The classification distribution includes the [GPL-3.0 license](LICENSE) from the original release. See [source and license notices](LICENSES.md) for artifact-specific terms and provenance.