akanametov/yolo-face ? reverse-engineered prompt

Build me a simple PyTorch project for face detection based on YOLO, with ready to use pretrained weights for different model sizes and a clean way to run inference on images or video. I want it to be easy to install, easy to train on a face dataset like WIDERFace, and easy to export to ONNX for faster deployment.

Please include example commands and a couple of sample scripts so I can test it quickly on a face photo, then train my own model if I want. Make sure the project has good docs, works from the command line, and includes the usual stuff like tests, setup files, and a clear folder structure. If there are any current Ultralytics or YOLO docs that would help, look them up online and follow the modern approach. I’d also like the repo to support a few related pretrained detectors if that fits naturally, like person, football, parking, or drone detection.