WongKinYiu/yolov7 ? reverse-engineered prompt

Build me a YOLOv7 object detection project that can train on COCO or on a custom dataset, then run inference on images, videos, or a webcam and draw clean bounding boxes with labels and confidence scores.

I want the usual workflow to just work, so I can download a pretrained model, test it, fine tune it on my own data, and export it if needed. Make sure there are simple scripts for training, validation, detection, and model export, and that the project is set up so it can run on one GPU or multiple GPUs. If there are any current setup details or missing steps in the docs, look them up online and fill them in.

Please keep it practical and ready to run, with sensible defaults, example commands, and a straightforward way to point it at my own dataset without having to understand the internals.