refactor(ml): model downloading

[mertalev]

Description

This PR distinguishes between downloading and loading models. Specifically, it adds a new download method that can be used independently from the load method for each model type.

There are several useful benefits to this:

• Lazy loading of models without compromising on the ability to download models at startup
• Simplified startup logic
  • Model classes can be instantiated freely since they can handle eager/lazy loading internally
• Faster startup times
  • Insightface's logic for loading models is very inefficient: it loads every model in the folder and inspects their inputs and outputs, then unloads the ones that aren't relevant.
  • The new implementation only loads relevant models, and in fact only extracts these models from the zip file on download.

How Has This Been Tested?

• No errors on startup
• Deleting the model cache and restarting downloads each model once
• After restarting the container, it does not download any models
• Object tagging, CLIP and facial recognition jobs run as normal

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

1 Ignored Deployment

Name	Status	Preview	Comments	Updated (UTC)
immich	⬜️ Ignored (Inspect)	Visit Preview		Aug 5, 2023 10:25pm

[mertalev]

I should probably update the tests...

[jrasm91]

Nice work!