i’m new to using transformers. i’m trying to use the equiformer model docs here to train a score function on atomic coordinate data. the atomic coordinate data is an Nx3 array, where N is the number of atoms and 3 because of 3 spatial dimensions. likewise, i also have an Nx118 array of one-hot encoded labels for each atom type (118 because of 118 atoms in periodic table).

however, i am confused on how to train this model. i would normally provide some code for this but ive tried so many things that it seems pretty frivolous to include anything.

its a very simple short script that im looking for, and would like to ask anyone of you who might be experienced with this type of model.

just assume the loss function is an empty return statement.

thanks

i tried inputting my Nx118 atom labels into the “feats” parameter but it gave me a ton of random errors, notably that my dim must be equal to 2.