coreyolo 0.2.1

CoreYOLO
_{Object detection. Built for Apple.}

Train an independent detector in Python. Ship Core ML to Apple silicon (ReLU palettes on ANE; SiLU dense FP16 on GPU).
Code is MIT. Weights carry their own license, stated per checkpoint.

Website · Docs · Models · Deploy · iOS sample · Licenses

Roboflow data.yaml  →  train  →  .coreyolo  →  export  →  Core ML  →  iPhone / Mac

The public type is Detector (YOLO is a compatibility alias). App code loads CoreYOLO files only (.coreyolo, trained best.pt, or .mlpackage).

Install

pip install coreyolo
pip install "coreyolo[video]"   # webcam (`--source 0`)

from coreyolo import Detector

model = Detector("n")              # GELAN n; also "s" / "m" / "l", or family="e2e"
model.train(data="data.yaml", epochs=100, imgsz=640)
model.save("weights/app.coreyolo")
model.export(imgsz=640)            # ReLU → FP16 + palettes + ANE; SiLU convert → dense FP16 GPU

Designed for Apple silicon

ReLU prefers ANE (--device ane). Converted SiLU graphs prefer GPU.

Three graphs. One trainer.

n, s, m, and l

Same GELAN trainer. Pick the scale for the device.

coreyolo convert --weights v9-t.pt --out weights/coreyolo-n-coco.coreyolo
coreyolo convert --weights v9-s.pt --out weights/coreyolo-s-coco.coreyolo
coreyolo convert --weights v9-m.pt --out weights/coreyolo-m-coco.coreyolo
coreyolo convert --weights v9-c.pt --out weights/coreyolo-l-coco.coreyolo

Source: MultimediaTechLab/YOLO v1.0-alpha (copyright Kin-Yiu Wong and Hao-Tang Tsui). Same MIT files LibreYOLO converts. Ultralytics yolov9t.pt is a different dump (AGPL-3.0). Do not pass those pickles into Detector().

Zoo

Binaries live on GitHub Releases, not in git. mAP stays blank until you record COCO val. See weights/manifest.json and weights/LICENSE_NOTICE.txt. Choosing a model means choosing its license.

Python SDK

from coreyolo import Detector

model = Detector("n")
model.train(data="datasets/dummy/data.yaml", epochs=20, batch=8, imgsz=320)
model.save("weights/app.coreyolo")

model = Detector("weights/app.coreyolo")  # or a Core ML .mlpackage
results = model.predict("photo.jpg", classes="person", save=True)
results[0].save("out.jpg")
model.export(imgsz=320)

model.val(data="data.yaml") returns mAP. model.predict(0) opens the webcam.

From a clone:

python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

Apple silicon: PyTorch train / val / predict use MPS (GPU) when it works, then CUDA, then CPU. Core ML .mlpackage inference defaults to GPU (CPU_AND_GPU); --device all adds the Neural Engine, --device ane is ANE-only.

Labels

Roboflow Export → YOLO (detect) or YOLO-Seg (polygons):

dataset/
  data.yaml
  train/images/   train/labels/
  valid/images/   valid/labels/

Detect lines are normalized class xc yc w h. data.yaml may use list or dict names, val or valid.

coreyolo dummy-data --out datasets/dummy
coreyolo train --data datasets/dummy/data.yaml --model n --epochs 20 --imgsz 320
coreyolo export --weights runs/detect/train/weights/best.pt --imgsz 320

Segmentation: --task segment. Mosaic is off. Core ML writes three outputs; NMS and mask assembly stay on the host. iOS sample: examples/ios.

Train on COCO

coreyolo coco --out datasets/coco --download
coreyolo train --recipe coco-n-fast    # 100 epochs
coreyolo train --recipe coco-n         # 300 epochs, ReLU (ANE)
coreyolo train --recipe coco-n-e2e

CLI

coreyolo coco        --out datasets/coco --download [--max-images 512]
coreyolo convert     --weights v9-t.pt --out weights/coreyolo-n-coco.coreyolo
coreyolo train       --recipe coco-n-fast
coreyolo train       --data data.yaml --model n --family e2e --epochs 100 --device gpu
coreyolo val         --data data.yaml --weights best.pt --device gpu
coreyolo predict     --weights best.mlpackage --source photo.jpg --device gpu
coreyolo predict     --weights best.mlpackage --source 0 --device gpu   # webcam; Q quits
coreyolo export      --weights best.pt --imgsz 640 [--int8] [--no-palette]
coreyolo dummy-data  --out datasets/dummy --segment

License

CoreYOLO's code is MIT. It does not require you to open source your application, and it does not change if you sell what you build.

Pretrained weights are separate: each one carries the license of whoever trained it, stated per checkpoint. GELAN COCO files converted from MultimediaTechLab v9-*.pt are MIT (copyright Kin-Yiu Wong and Hao-Tang Tsui). Ultralytics yolov9*.pt converts stay AGPL-3.0. A model you train yourself is yours. Converting a file does not change its applicable terms.

Choosing a model means choosing its license. See docs/licenses.md, weights/LICENSE_NOTICE.txt, and NOTICE. This is a description of the licenses involved, not legal advice. YOLO is a trademark of its owners.