[ANN] GeometricFlux.jl - Geometric Deep Learning for Flux

Hi, Julians!

I am pleased to announce GeometricFlux.jl v0.1.0, which is an extension toolbox enabled geometric deep learning or graph neural networks in Flux.

This package aims to integrate existing JuliaGraphs ecosystem (i.e. SimpleGraph, WrightedGraph etc.) with Flux and accelerate by CuArrays and SparseArray. Zygote is used as AD engine. Several mainstream layers are implemented. An simple working example is available on CPU.

However, this package is still under fast development period. CUDA support is not tested. Suggestions and pull requests are welcome.

25 Likes

This looks very interesting! Thanks a lot for your effort.
I am particularly interested in the Message Passing layers. :+1:

Really cool. This is a space where Julia can really (and ecosystem) can really shine. I’ll be forking this project as soon as I’m home!