message passing for multiple feature arrays

[YichengDWu]

Hello and thank you for your excellent work!

If I have multiple feature arrays, how should I define a customized message function? I want to have something like

message(xi,xj,ui,uj) = xxx

This feature is available in torch_geometric. However, it seems like by design it is not possible in GraphNeuralNetworks.jl but we have to concatenate x and u?

[CarloLucibello]

You just put all your features inside a named tuple:

ndata = (a = rand(3, num_nodes), b = rand(3, num_nodes))

message(xi, xj, e)= xi.a .+ xj.b

propagate(message, g, +,  xi=ndata, xj=ndata)

Conversely, message functions can return named tuples as well.

[YichengDWu]

You just put all your features inside a named tuple:

ndata = (x = rand(3, num_nodes), b = rand(3, num_nodes))

message(xi, xj, e)= xi.x .+ xj.b

propagate(message, g, +,  x=ndata)

Conversely, message functions can return named tuples as well.

Sorry, I'm still confused about how it works. What is the keyword argument x here? I couldn't find its definiton in the source code.

[YichengDWu]

By the way, I found a typo in the doc:

xj: As xj, but to be materialized on edges' sources.

Should it be

xj: As xi, but to be materialized on edges' sources.

[CarloLucibello]

Sorry, I'm still confused about how it works. What is the keyword argument x here? I couldn't find its definiton in the source code.

sorry, I thought we had that shortcut. I edited the comment.

By the way, I found a typo in the doc

thanks, fixed now!

[YichengDWu]

I see. I honestly think it would be a good idea if we have

propagate(message, g, +,  x=ndata)==propagate(message, g, +,  xi=ndata, xj=ndata)

since quite often we indeed have x=xi=xj