Node-, edge-, and graph- wise attributes

[yliu1240]

Hi, all!
I'm currently learning MatGL and I would like to follow up and know more about the current progress for a recent update, log #125 . Generally, there are node- (can also be called atom- or site-) wise quantities, edge- (or bond-) wise quantities, and graph- (or structure/molecule-) wise quantities that can be integrated into graph-based models, as either inputs (or attributes) or outputs (prediction targets). I'm wondering which of these features are available in current version of MatGL?
.
1) I guess node-wise attributes and node-wise predictions are partially implemented. May I presume the 'state_attr' refers to node-wise attributes which will be concatenated with atomic embeddings in training process? Then, the 'include_state', 'ntypes_state', 'dim_state_feats', and 'dim_state_embedding' variables control whether the model (MEGNet, M3GNet, and CHGNet) is going to be trained on these attributes, right?
.
2) When adding the data of node-wise attributes, do I need to add all these attributes to the Site objects in Structure/Molecule? Then, shall I use the 'state_attr' to select which of the attributes to include in the training process (or I misunderstand)?
.
3) If I intend to customize some of the features (predict node- and edge-wise targets, and train node-, edge-, and graph-wise attributes). Would you like to suggest which of the functions do I need to customize? I guess collate_fn will surely be one.
.
Thank you very much for your time!

Best regards,