Order of edge-gated graph convolutions?

[rees-c]

Hi,

Thanks for this great library.

I think I noticed a slight inconsistency in the code from the paper. The paper states that the ALIGNN layer first performs the edge-gated graph convolution on the line graph to update the pair and triplet features, and then the pair features are passed as edges to the atomistic/direct graph.

However, when I look at alignn.models.alignn.ALIGNNConv.forward, I see that the edge-gated graph convolution is actually applied on the atomistic/direct graph first, and then the updated pair features are passed as nodes to the line graph. Am I understanding this correctly?

[bdecost]

However, when I look at alignn.models.alignn.ALIGNNConv.forward, I see that the edge-gated graph convolution is actually applied on the atomistic/direct graph first, and then the updated pair features are passed as nodes to the line graph. Am I understanding this correctly?

I think so, yeah. Some of our experiments were done on an experimental branch where the order of these two operations was configurable (https://github.com/usnistgov/alignn/blob/dev/ray/alignn/models/alignn.py#L27). I thought we did a comparison study on that and found that the order didn't much matter in the end, but it looks like that didn't make it into the supplemental information of the paper

[rees-c]

Good to know, thanks! Would be cool to see that study at some point.