Hello everyone. Based on the documentation and the post from the end of July, seems like DL4J has currently no out-of-the-box support for Graph Neural Networks. Is there any support planned in the future releases? Are there maybe some basic examples of GNN implementation using SameDiff ?
I think there will have no deeplearning4j gnn, I’m using tf gnn by python4j.
Eg: Is there a manual about GCN(Graph Convolutional Networks) using DL4J - #12 by xiumuzidiao
It seems like some folks make it work. I think what @SidneyLann wanted was something higher level. If you’re willing to (sort of like what you did with your other models) I’d be happy to work with you on what is/isn’t supported.
@agibsonccc , right now I’m looking for some basic stuff, nothing fancy. I posted this question because I didn’t want to re-invent the wheel in case it’s already there. But even if it’s not - 2-d convolution and FF layers will be the starting point and I’ll go on with them for now. If then I’ll need some support - I’ll definitely let you know. Thanks a lot !
@SidneyLann , thanks for a tip! Personally I’m not fond of using TF imported models because it brings more efforts to work with them in case I need significant modifications. I think using SameDiff directly will be a quicker solution for my case. But it’s good to know that there’s an already working model which could be imported!
@partarstu I think he has some fair points. Some specific abstractions would be needed for most people. I do think the building blocks are there.We built this library for deepwalk (graph embeddings):
The thread I linked I think has most of the basics you’d need. I think you could roll your own blocks that achieve the same results.
@agibsonccc thanks a lot! I’ll take a look at it !