Add upet and sevennet models#591
Conversation
| trained_on_dispersion=cfg.get("trained_on_dispersion", False), | ||
| dispersion_kwargs=cfg.get("dispersion_kwargs", {}), | ||
| ) | ||
| case "UPETCalculator": |
There was a problem hiding this comment.
Why do we not want to replace PETMADCalculator?
There was a problem hiding this comment.
can we still use the old model?
There was a problem hiding this comment.
Should be able to, yep, it's the same repo, just renamed. Models available are here: https://github.com/lab-cosmo/upet/blob/main/src/upet/_version.py#L1. You set the version e.g.
calculator = UPETCalculator(model="pet-mad-s", version="1.5.0", device="cuda") but v1.0.2 should also be available based on https://lab-cosmo.github.io/upet/latest/models.html
| case "SevenNetCalculator": | ||
| kwargs = cfg.get("kwargs", {}) | ||
| loaded_models[name] = SevenNetCalc( | ||
| model=kwargs["model"], |
There was a problem hiding this comment.
Why do we need to pull this out when the key word is model anyway?
| device: Device | None = None | ||
| kwargs: dict = dataclasses.field(default_factory=dict) | ||
|
|
||
| def get_calculator(self, **kwargs) -> Calculator: |
There was a problem hiding this comment.
Does this work when we set precision? I don't think they've implemented anything yet (MDIL-SNU/SevenNet#138), and then precision will get passed to SevenNetCalculator
| uma = [ | ||
| "fairchem-core == 2.19.0; python_version >= '3.11'", | ||
| ] | ||
| upet = [ |
There was a problem hiding this comment.
I think we just want to replace pet-mad? It's essentially the same repo
| "upet", | ||
| ] | ||
| sevenn = [ | ||
| "sevenn", |
There was a problem hiding this comment.
I would pin these so we know exactly what version we are using, and can rely on their interfaces
| kwargs: | ||
| model: "7net-l3i5" | ||
|
|
||
| sevennet-omni-i12-mpa: |
There was a problem hiding this comment.
Why this one? From their list, the standard omni is recommended?
There was a problem hiding this comment.
this is the one they submitted recently to matbench discovery so i assume its their model of choice. strange that one is reccomended. hmm we can discuss
|
Also just to note, the pre-commit it failing |
2 participants
Pre-review checklist for PR author
PR author must check the checkboxes below when creating the PR.
Summary
adds:
Linked issue
Resolves #
Testing