Sound support of variable graphs

[yuehhua]

There are more and more applications in GNN. In training, people demand model to be trained on distinct graphs. Variable graphs are needed to customize their training process.
So far, I implement the FeaturedGraph to equip graph structure on features, so one can input feature with a graph. However, it is experimentally implemented to support GCNConv. To support all layers, there are more work to do.

[ilancoulon]

First, thank you for your work on GeometricFlux.jl 😃

Are you currently working on that issue?
If not, could I try implemeting some other layers?
Finally, is the FeaturedGraph struct going to last? Should we use MetaGraphs.jl instead or is it better to keep that structure? (as discussed in #6)

[yuehhua]

I am welcome and happy to receive contributions from others.
I will maintain FeaturedGraph in the mainstream. As for the MetaGraph support, I will keep the conversion to FeaturedGraph available.
If one wants to design new layer, please take FeaturedGraph as first choice.

[yuehhua]

FeaturedGraph is supported by graph net and message passing scheme in #63

[tdfalc1]

Hi, thank you for the awesome package! Im curious how the GCNConv layer incorporates edge features from the FeaturedGraph type? And what structure the the edge features should take, i.e. can the features for each edge be greater than 1 dim, or are edge features just treated as a weighted adjaceny matrix?

[yuehhua]

@tdfalc Thank you for your question. Would you please open another issue for this? I will give you a reply in that issue.