Loading models from other platforms like Keras, HuggingFace, etc.

[AdityaKane2001]

@DavidLandup0

Great work on the vision of this library! This is a huge step in unifying the frameworks, workflows, etc.

Most of the community uses models from renowned platforms like Keras, TF Model Hub, HuggingFace Hub, etc. It would be a monumental addition if we have an API which can read an load models from these platforms. I think this would be a big step towards making a truly unified platform.

[DavidLandup0]

Thank you @AdityaKane2001!

I've been thinking about hubs (HF, TF, PT, etc.) and we definitely want to support them as much as we can, and not limit users to the implementations in a single library. There are two ways we can go about this:

• The easiest way to do this would be to make DeepVision utilities conform to HF/TF/PT hub models to make it easy to use it with them.
• On the other hand, it'd be nice to have DeepVision utilities that automatically does this for users as well.

Exposing both levels would be ideal. If someone wants to use the hubs manually and plug the models in DV, they can, and if they want to leave it up to DV, they can.

In a way, we'd then have DeepVision models (TF/PT, same API, same implementations, automatic weight conversions), and community models from hubs. Is this what you meant?

If so, it's definitely on the radar, but the road to it will require some refinement and a bit more thought (though, the sooner the better). Thank you for bringing this up!

[AdityaKane2001]

The easiest way to do this would be to make DeepVision utilities conform to HF/TF/PT hub models to make it easy to use it with them.

Not really sure what do you mean by "conform" in this case.

My very first thought on this matter was to simply add a source="..." argument to the model initialization and handle the rest in the library.