pytorch/vision ? reverse-engineered prompt

Build me a Python package for computer vision that works with PyTorch.

I want it to make it easy to load popular public datasets, apply common image transforms, and use a set of ready to go pretrained models for image tasks. It should work with both tensor inputs and normal image files, and feel simple enough for someone doing training or inference in a notebook or script.

Please include clear install instructions, useful docs, a few example usages, and tests so I can trust it works. If there are any current best practices or compatibility details I should follow, look them up online and use the latest guidance.

Keep it practical and polished, with the kind of API people would actually want to use for everyday computer vision work.