NequIP -> Allegro fine-tune training?

[turbosonics]

Hi,

Last year when I try to use NequIP and Allegro, I read from somewhere that, training from scratch using NequIP then continue the fine-tune training using Allegro from the NequIP model with the same training set & same hyperparameters can be helpful for Allegro model development. (Using initialize_from_state and initial_model_state)

But I cannot find that information now, or maybe this is something I read/remember wrong.

So, I hope to make sure about this again. Would NequIP->Allegro fine tuning (NequIP initial training from scratch then fine-tune using Allegro) be better than training just from Allegro from scratch? Or it doesn't matter that much?

Thanks