Extending Functionality for Custom Graph Generation

[thormtl]

First and foremost, I would like to express my admiration for the work all parties involved, have done with both DiGress and here MiDi. It is very impressive and I have already learned a lot from the work, particularly in the context of molecule generation. This is a new field for me and I appreciate the steep learning curve.

I would like to kindly ask about the potential adaptability of your codebase for a project I am currently working on (not molecules). My project involves generating graphs where nodes possess categorical features, like node types (which works with current DiGress) and specifically directional attributes represented categorically as 'north', 'south', 'east', and 'west'. Additionally, these nodes have a relative placement attribute that defines their position in relation to connected nodes that will also be expressed categorically. I believe all these features can be discretised (one-hot-encoded).

I am interested in understanding how feasible it would be to extend or modify your existing code to accommodate these features. Would the architecture of your system support such customisation without extensive overhauls? If so, could you provide any guidance or suggestions on where to start with the modifications?

I have currently looked into a DataModule implementation for my use-case and try to draw parallels to the features you define in your abstract_dataset.py. I am wondering if I draw parallels of e.g. my directional attributes to your molecule charges, that I can introduce these by only changing the input/output dimensions and changing the current Placeholder object being used.

I really hope this message reaches you and you have time to give some guidance 😊