Spatial Graph Convolutional Networks

[msainsburydale]

Hi, many thanks for your work on this amazing package.

There is a paper called "Spatial Graph Convolutional Networks", with accompanying PyTorch code. You may be familiar with this work already but, from my understanding, the premise of this architecture is that it is a proper generalisation of CNNs from gridded spatial data to irregular spatial data; see Figure 3 and Theorem 1 in the paper.

Do you think it would be worthwhile to include this architecture in your package, perhaps as one of the convolutional layers?

[CarloLucibello]

we don't have layers taking in spatial features, I think it would be a nice addition to the package. Please file a PR if you have time.

Right now due to the monorepo structure setting up the environment for developing the package has become more complicated. I have to improve the developers docs but meanwhile feel free to ask if you need any help with that.

[msainsburydale]

I have a more specialised version of a spatial graph convolution already implemented in Julia (see here), and I think it would be fairly easy to adapt this code to the formulation given in the paper linked above.

However, I implemented this specifically to be used in my package NeuralEstimators.jl, and I think there would be a little bit of work needed to adapt this code for use in GraphNeuralNetworks.jl. It shouldn’t take too much work, but I’ll need to wait until I have more time to take a closer look.