we may want fine tuning as a separate option since we need to check more stuff..

[alinelena]

          we may want fine tuning as a separate option since we need to check more stuff..

also will need a restart

Originally posted by @alinelena in #127 (comment)

[ElliottKasoar]

Should these checks not be handled by MACE's (or other MLIPs') training scripts?

We can check that foundation_model is passed, which triggers a few things including load_foundations from finetuning_utils.py, but otherwise would it not make more sense to contribute any more essential checks back to MACE, or are there things that would be specific to janus?

[alinelena]

for example will be useful to check that the foundation model exists, yes I agree the more fundamental stuff shall go to mace.

[ElliottKasoar]

for example will be useful to check that the foundation model exists, yes I agree the more fundamental stuff shall go to mace.

I'm not sure there's even a straight forward way of doing that, since the model loading is done within the training script e.g.

if args.foundation_model is not None:
    if args.foundation_model in ["small", "medium", "large"]:
        logging.info(
            f"Using foundation model mace-mp-0 {args.foundation_model} as initial checkpoint."
        )
        calc = mace_mp(
            model=args.foundation_model,
            device=args.device,
            default_dtype=args.default_dtype,
        )

Even if we loaded the model, we couldn't pass it explicitly, since we have to go via the config.

We could load it and pass the path to make use of the final else torch.load, but it does mean we miss any model-specific logic they include.

[ElliottKasoar]

We could load it and pass the path to make use of the final else torch.load, but it does mean we miss any model-specific logic they include.

I suppose we will need to do this, or something similar, for aiida anyway though.