| --- |
| license: agpl-3.0 |
| tags: |
| - rknn |
| - rknn2 |
| - rockchip |
| - rk3576 |
| - rk3588 |
| --- |
| |
| # YOLO26-RKNN2-Ultralytics |
|
|
| > **Looking for the Rockchip Model Zoo and custom exports?** See [YOLO26-RKNN2](https://huggingface.co/HanzoHuang/YOLO26-RKNN2). |
|
|
| 140 FP16 and INT8 YOLO26 exports for **RK3576** and **RK3588**, produced with the Ultralytics export workflow and task-specific RKNN conversion tools. Supported tasks include classification, depth, detection, oriented bounding boxes, pose, instance segmentation, and semantic segmentation. |
|
|
| ## Choose a model |
|
|
| Models are organized by toolkit version, task, variant, target chip, and precision: |
|
|
| ```text |
| models/v<toolkit-version>/<task>/<variant>/<chip>/<precision>/ |
| ``` |
|
|
| Download `model.rknn` and `metadata.yaml` from the same directory. The [manifest](manifest.json) lists the available models and checksums. |
|
|
| ## INT8 detection |
|
|
| Use the `metadata.yaml` from the same model directory. INT8 detection outputs normalized box coordinates and requires `coordinates: normalized`. Import both files together in the warehouse, or keep them in the same directory for the CLI. Treating normalized coordinates as pixels can result in no boxes. |
|
|
| ## Reproduce the conversion |
|
|
| The [source directory](tree/main/source) contains the original checkpoints, conversion scripts, and environment and calibration records. Conversion used [RKNN Toolkit2 2.3.2](https://github.com/airockchip/rknn-toolkit2); upstream Ultralytics source is available on [PyPI](https://pypi.org/project/ultralytics/#history). |
|
|
| ```bash |
| python source/convert-model.py --help |
| ``` |
|
|
| For INT8 detection, pass a COCO128 dataset YAML. For other INT8 tasks, pass a COCO128 image list. Use `--calibration-limit 128` and the input size listed in the manifest. |
|
|
| ## Source and license |
|
|
| The inputs are official Ultralytics checkpoints. Conversion changes the format and quantizes the INT8 graphs. The RKNN files and original weights are distributed under AGPL-3.0; see [LICENSE](LICENSE) and the [checkpoint URLs and hashes](source/weights.json). |
|
|
