Incorporate global descriptors into MACE

[pfebrer]

Hi!

I have a global descriptor (a vector, the external electric field) that is very important for my predictions. What is the best way to incorporate this into MACE?

I have thought of incorporating the vector to all nodes in node_attrs, which would require a slight modification to accept node_attrs_irreps as an argument to initialize the model. I don't know if it's the best approach.

Is there some "established" way to do this?

Thanks!

[ilyes319]

Yes this is a good approach, just generalize to accept more than scalar node_attrs, it should be a small modification.

[pfebrer]

Cool, thanks for the quick response!

[pfebrer]

I'm doing some stupid test where the node_attrs are two times the one hot encoding. I.e. for a dataset of two species, a batch with one atom is [[0, 1, 0, 1]] instead of [[0, 1]].

It works until the forward call to EquivariantProductBasisBlock. There num_elements is 2, but now node_attrs irreps is 4x0e, and it fails with the message:

RuntimeError: contracted dimensions need to match, but first has size 4 in dim 1 and second has size 2 in dim 0

Is the num_elements attribute really meant to be num_elements or is it the size of node_attrs irreps?

[pfebrer]

Or should I pass only the one hot encoding to the node_attrs argument of EquivariantProductBasisBlock.forward ? 🤔

[ilyes319]

yes exactly the product basis will only accept one hot of elements.

[pfebrer]

What about the node_attrs argument of the interaction? Should it also be only the one hot of elements?

I see that different interaction classes use node_attrs differently.