Multi-n-graph best approach

[piyush-J]

Hi, Thank you for the great documentation! I am a student I want to implement an n-graph data object for my project. This would be an extension of the pair-data, where instead of 2 graphs, I want to create n graphs (n will also be an argument).

What's the most pythonic way to do this (especially for the __inc__ part)? Let's say the edge index will always have the naming convention (edge_index_1, edge_index_2, ..., edge_index_n) and similarly for x (x_1, x_2, ..., x_n).

from torch_geometric.data import Data

class NGraphData(Data):
    def __init__(self, n, **kwargs):
        super().__init__()
        ....

    def __inc__(self, key, value, *args, **kwargs):
        if ....
        else:
            return super().__inc__(key, value, *args, **kwargs)

[Padarn]

maybe simply

def __inc__(self, key, value, *args, **kwargs):
    if 'edge_index' in key:
        n = key.split("_")[-1]
        return getattr(self, f"x_{n}").size(0)

Could also use a regex if you feel its better

[piyush-J]

Thank you for your prompt reply. Will give it a try!

[rusty1s]

Batching of lists holding tensors should also be supported.

data.edge_index_list = [edge_index1, edge_index2, edge_index3]

although it does not support customizing __inc__ for each individual index currently.
For this, the approach from @Padarn might be a better alternative.