[WIP] Adapt Makers to use pydantic models instead of yamls

[naik-aakash]

Closes #305 , #329, #324 and #330

Changes

• providing default config file arg seems redundant (Maybe remove)
• check if MLIPFitMaker also needs this change
• Add type checks before updating kwargs of config (use pydantic to achieve this automatically)
• add test to ensure RssConfig custom model work as intended
• add MLIP hypers pydantic models
• use pydantic models instead of load yamls for defaults (adapt fitting jobs / flows)
• Move database_dir arg back to Make see RssMaker does not work as expected #329 (comment)
• Extend MLIP hyper parameter models where possible (Maybe MACE for now > also no typechecking for fine-tuning hypers for MACE at this point in models)
• remove use defaults arg of MLIPFitMaker/ machine learning fit method
• Check if nequip_fitting function can be cleaned up
• Add hypeparameters arg to machine_learning_fit function
• Think if value Error should be raised or a warning when unexpected kwarg gets passed for fits.
• Figure out missing args in RssMaker doc-strings and placement of args (need some input from @YuanbinLiu)
• Update fine-tuning docs for MACE
• Update RssMaker documentation to reflect changes
• Update AIRSS install instructions

[JaGeo]

Shouldn't we rather do something similar to atomate2 settings?

[JaGeo]

Or this approach? https://github.com/materialsproject/pymatgen/blob/da607e86f9ce8aec942067c6c1a4fda6e04915dd/src/pymatgen/io/vasp/sets.py#L86

[naik-aakash]

Hi @JaGeo , i think both above approach we can use to load the default yamls, which I plan to do next : similar to atomate2 settings

But we would still need postinit in the maker if we expect user to provide modified yaml path (so it gets read when the flow is created) and we don't want to copy that yamls to remote cluster. Due to nature of jobflow delayed execution

[JaGeo]

@naik-aakash okay! Likely pymatgen approach is th closest to what we want for defaults

[JaGeo]

Another thought: post init will not work if you initialize the flow during run time

[QuantumChemist]

Hi @JaGeo , @QuantumChemist , I have started a bit to replace use of default jsons with pydantic models. (Just done for GAP and J-ACE) If you think the way it is done is fine, I will do the same for all other methods. If you have any comments or suggestions let me know, happy to address it before I make lot of changes.

We now have type checks already as part of Models with update_fields method (handy when updating defaults with user supplied kwargs)

It looks really good to me!
The only thing I wonder is if we want still keep the possibility that users can also pass the hyperparas via a file? Let's say someone has a certain default setup and wants to submit a lot of wfs with theses same settings? Putting the hyperparas in the submission script every time seems rather inconvenient in this case. What do you think?

[naik-aakash]

Hi @JaGeo , @QuantumChemist , I have started a bit to replace use of default jsons with pydantic models. (Just done for GAP and J-ACE) If you think the way it is done is fine, I will do the same for all other methods. If you have any comments or suggestions let me know, happy to address it before I make lot of changes.
We now have type checks already as part of Models with update_fields method (handy when updating defaults with user supplied kwargs)

It looks really good to me! The only thing I wonder is if we want still keep the possibility that users can also pass the hyperparas via a file? Let's say someone has a certain default setup and wants to submit a lot of wfs with theses same settings? Putting the hyperparas in the submission script every time seems rather inconvenient in this case. What do you think?

I think files are be problematic when using remote submissions, user then needs to copy the file, I feel it is much more convenient to rather create MlLIP_HYPERS object from a custom hyper parameter args and simply pass this instance to the workflow. I still have to test this though with a remote submission if it really works as I think it would.

Btw, Option to create this object from file is on my todo

[JaGeo]

Thanks for fixing the bug well @naik-aakash 😉.

[QuantumChemist]

why don't we just do run_locally(...)?

[naik-aakash]

Tests can be made better in future by accessing the responses and assertions.

I can remove it yes. That's true

[QuantumChemist]

Up to you, just wondered about it 😄