mlfoundations/open_clip ? reverse-engineered prompt

I want to build an open-source Python library that's an implementation of OpenAI's CLIP model. The main goal is to connect images and text. It should be able to take an image and a few text descriptions, and then tell me which description best fits the image by providing a probability score.

This should be built with PyTorch. It's really important that the library makes it easy to load different pre-trained models, especially the ones trained on large datasets like LAION. I'd like a simple way to list all the available pre-trained models and then instantiate one by name.

Please package this up so it can be installed with pip. The most important part is a straightforward example I can run. Show me how to load a model, preprocess an image from a file, tokenize a few text labels, and then compare them to get the final classification probabilities.