Details on integrating Loki model into OpenCLIP training pipeline

[biofocker]

Dear Authors,

Thanks for the clarification in previous issues that the training leverages the OpenCLIP codebase.

I am planning to train a Loki-like model from scratch (or finetune) using my own custom dataset. However, I am not sure how to properly integrate the Loki model definition into the standard OpenCLIP pipeline.

Could you please share some guidance on:

1. The Configuration: Is there a specific config file (.json) or a set of command-line arguments (hyperparameters like learning rate, scheduler, etc.) that you recommend for Loki?
2. Model Registration: Since Loki involves specific architectural changes, do we need to manually register the model in OpenCLIP's registry? Or is there a specific script in this repo that acts as an entry point?

Sharing a sample training command would be extremely helpful for those of us trying to apply Loki to new data.

Thank you very much!