feat: preloading of machine learning models

[DawidPietrykowski]

In response to this discussion

Implements additional step of preloading select models at the start of the machine-learning process into memory permanently.

Models can be selected by changing the environment variable, for example:
MACHINE_LEARNING_PRELOAD="CLIP:ViT-B-32__openai,FACIAL_RECOGNITION:buffalo_s"

[mertalev]

Thanks for working on this! I think this can be simplified a lot. It should be good after the changes I mention in the comments.

[mertalev]

Great job! Just a few minor comments.

If you'd like, you can also change the env to the PRELOAD__CLIP=['model-name'] style based on this.

[DawidPietrykowski]

Based on the suggestions I implemented a pydantic config that parses:

MACHINE_LEARNING_PRELOAD__CLIP="ViT-B-32__openai"
MACHINE_LEARNING_PRELOAD__FACIAL_RECOGNITION="buffalo_s"

which seems like the least error-prone solution. As I mentioned in a comment above, this limits the env to storing a single model per type but it seems perfect for our use case as there shouldn't be more than one of each type as far as I know.

[mertalev]

Oh, that's a good point! I don't see a use-case for loading multiple of a model type.

[mertalev]

Awesome stuff :)

[DawidPietrykowski]

Thanks a lot for the detailed review and comments :)

[mertalev]

The formatting checks can be a bit strict haha. Running ruff app --fix should take care of it

[DawidPietrykowski]

Yeah, good lesson on running these types of checks before pushing though