search

johmedr / GraphN

A convenient wrapper to develop graph neural networks with Keras. Currently under development with the objective of integrating Networkx, Owlready2 and oneM2M for cognitive IoT.
MIT License
2 stars 1 forks source link

Gated Graph Neural Network #1

Open jimanvlad opened 5 years ago

jimanvlad commented 5 years ago

Hi,

How would you extend your wrappers to build a GGNN layer (based on Microsoft's https://github.com/Microsoft/gated-graph-neural-network-samples)?

I am interested in refactoring https://github.com/CRIPAC-DIG/SR-GNN to Keras

johmedr commented 5 years ago

Hi,

Are you concerned about the non-squared shape of the adjacency matrix used in the paper? This would fail while checking the shape with the current wrapper.

However, it seems that their adjacency matrix will always be the graph adjacency matrix (outgoing edges) concatenated with its transposed (incoming edges). Hence an easy way to implement it would be to compute this matrix at each layer.

Otherwise, to avoid this overhead a solution would be to create a first layer “GatedGNNCell” that just implements the mechanisms of a layer, and then create a user-friendly “GatedGNN” that allows to build several connected “GatedGNNCell”s. It would take charge of the adjacency matrix transformation in input and output. Adding a “no_shape_check” attribute to GraphWrapper would help passing the non-square adjacency between GatedGNNCells.

Le 10 mars 2019 à 08:33, Vlad Jiman notifications@github.com a écrit :

Hi,

How would you extend your wrappers to build a GGNN layer (based on Microsoft's https://github.com/Microsoft/gated-graph-neural-networ)?

I am interested in reactoring https://github.com/CRIPAC-DIG/SR-GNN to Keras

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or mute the thread.