Normalization constants for edge features during inference #13
Description
Hi @vxfung!
Hoping you are well!
Just wanted to bring to notice a minor issue I faced while trying to integrate MatDeepLearn with one of my custom packages. Currently, while processing the ASE structures, edge normalization happens as per the interatomic distance distribution of the input data features.
However, for the prediction or inference phase, this should not be the case. Especially, I found that for smaller datasets where crystal structures were similar, the normalization constants could be different from what were used for learning, and that resulted in inconsistent predictions. This is likely to happen only in cases where we have a set of crystal structures with interatomic distances less than the ones set in graph processing; and therefore, only when you are predicting properties for a small set of similar structures.
A couple of proposed fixes:
- Set the Min-Max normalization constants to be the same as what is set in the graph processing config
OR - Enforce it as a required field for inference/prediction jobs
Cheers!