from_networkx enhancement

From now on, we recommend using our discussion forum (https://github.com/rusty1s/pytorch_geometric/discussions) for proposing new features.

🚀 Feature

My proposal is to modify the from_networkx method to incorporate the option of creating data.x on the fly from node features (and similarly from edge features)

Motivation

The current from_networkx method does not convert the nodes features into property x (that is, data.x). I think enhancing this would be particularly useful for hand-crafted graph datasets that are saved and loaded using networkx methods.

Additional context

Some implementation details worth discussing:

• One idea is to leave numerical features as they are, and one-hot encode categorical ones. Is the one-hot encoding needed or should each feature contain at most one item? It may be worth the flexibility, we could for example one-hot encode the node feature values that are lists, and just convert to numbers the rest.
• Perhaps we should allow the user to define a dictionary of node features and possible values to allow coherent numericalization of many graphs if they form a “dataset” (not a Pytorch Geometric dataset, just a normal list of graphs)

[rusty1s]

I think we should ignore that for now, and extend functionality in later PRs if that’s okay for you.

[rusty1s]

One possible interface:

def from_networkx(G, group_node_attrs=['attr1', 'attr2', ...], group_edge_attrs=['attr3', 'attr4', ...]):
    pass

where attributes in group_node_attrs and group_edge_attrs get grouped/merged/concatenated into x and edge_attr, respectively. WDYT?