Order of edge features in "sparse" mode

The documentation suggests that edge features (used by ECCConv, for instance) can be in sparse format to avoid padding large arrays with zeros. However, attempting to construct a Graph with e set to a scipy coo_matrix results in an error being raised by the following line:

https://github.com/danielegrattarola/spektral/blob/b6c3795cd12879c44e9591bd04b5d27acb511aab/spektral/data/graph.py#L81

It seems like only dense arrays are permitted for the edge features, unlike for the adjacency matrix (which also allows the type check to pass if sp.isspmatrix(a) is True). Is this intended, is there a workaround, or am I missing something?

[danielegrattarola]

You are correct, the documentation is imprecise in this regard and this is likely to cause issues to end-users. I will take some time today to:

1. Improve the data loaders so that the specific ordering of the adjacency matrix/edge feature is not something that the users need to worry about
2. Improve the documentation to clarify the meaning of “sparse format”.

Thanks for bringing this to my attention.

[danielegrattarola]

Ok so point 1 could be a little trickier than expected, but I’ve gone and added a significant section to the Getting Started tutorial that explains in detail how edge attributes should be represented in a Graph object. I’ve also fixed the wording of the Data Modes tutorial. I’ll update the online version of the docs soon.

I also made sure that, if edge attributes are sorted correctly, then all data loaders and layers will know how to deal with them. This still requires the users to correctly reorder the edge attributes, but I don’t see an easy way out for now.

I’ll change this issue into a discussion so that it remains part of the knowledge base of the library.

Thanks again