Using fairchem pre-trained models on custom datasets.

[nevinngyt]

What would you like to report?

Hello, I'm trying to use the fairchem pre-trained models and evaluate their performances on my custom datasets (xyz files). The headers in my xyz file look like:
Lattice="5.494219689819475 0.0 0.0 2.7471098449097373 4.758133825356323 0.0 0.0 0.0 27.21502897900848" Properties=species:S:1:pos:R:3:forces:R:3 energy=-136.82442956 pbc="T T T"

I converted my xyz files into atoms objects and geometric graph objects with ASE for use of fairchem pre-trained models. However, I ran into some errors and would need help navigating the documentation to solve the issues. Here's the code I used and the errors I encountered:

from fairchem.core.models.model_registry import model_name_to_local_file

checkpoint_path = model_name_to_local_file('EquiformerV2-31M-S2EF-OC20-All+MD', local_cache='/tmp/fairchem_checkpoints/')

Error:
WARNING:root:No seed has been set in modelcheckpoint or OCPCalculator! Results may not be reproducible on re-run

Attempt to calculate eenrgies and forces with atoms object:
calc.calculate(atoms, properties=['energy', 'forces'], system_changes=False)

Result
None

Could anyone advise on what I did wrong here and how I could best use fairchems models for my own xyz dataset? Would greatly appreciate it!! Thanks!

[lbluque]

Hello @nevinngyt 👋

If you are trying to run predictions using the ASE calculator interface the recommended way to do so is something like this:

from ase.io import read
from fairchem.core import OCPCalculator

atoms = read("path_to_your_structure.xyz")

# set cpu=False to use GPU, set the seed if you want exactly reproducible results and avoid the Warning.
calc = OCPCalculator(model_name="EquiformerV2-31M-S2EF-OC20-All+MD", cpu=True, seed=0) 

atoms.calc = calc

# calculate 
energy = atoms.get_potential_energy()
forces = atoms.get_forces()

[nevinngyt]

@lbluque Ahh thank you so much!! What about if I hope to play around with freezing some layers of the models and the parameters and re-training the models on my custom datasets?