Megvii-BaseDetection/YOLOX ? reverse-engineered prompt

Build me a YOLOX object detection project that I can use to train on COCO, run image and video demos, and test pretrained models from the command line.

I want it to have the standard YOLOX model choices like the small, medium, large, and extra large versions, plus the lightweight ones for faster use. It should let me point to a config or a model name, load weights, and save annotated results for images and videos. Also make sure the training flow is easy to run on one machine or multiple GPUs, with mixed precision and the usual options for batch size and image size.

If it makes sense, include the ability to export or work with common deployment formats like ONNX and TensorRT, since I may want to use it outside Python later. Keep the setup simple, include clear docs or example commands, and look up current docs online if you need to.