plan for splitting the package

[CarloLucibello]

The codebase is currently undergoing a factorization whose final outcome is to have several packages hosted in the same monorepo (for developing convenience, see #433):

• GNNGraphs.jl: graph types and graph operations. Depends on Graphs.jl and on NNlib.jl.
• GNNlib.jl: functional implementation of graph convolutional layers. Depends on GNNGraphs.jl.
• GraphNeuralNetworks.jl or FluxGNN.jl: Stateful graph convolutional layers based on Flux.jl. This will be the fronted package for Flux users. Depends on GNNlib.jl, GNNGraphs.jl, and Flux.jl.
• LuxGNN.jl: Stateless graph convolutional layers satisfying Lux's interface. Depends on GNNlib.jl, GNNGraphs.jl, and Lux.jl.

This way we will introduce Lux support (#372) and the Flux and Lux frontends will be only tiny wrappers around a shared codebase given by GNNlib.jl and GNNGraphs.jl.

I will keep track here of the transition status:

• Create and register the package GNNlib.jl. GraphNeuralNetworks.jl will not depend on it yet. create GNNlib.jl #432
• Create and register the package GNNGraphs.jl. separate GNNGraphs from GNNlib #446, New package: GNNGraphs v0.1.0 JuliaRegistries/General#111443
• Setup multi-package CI (CI for GNNGraphs.jl #451)
• Make GraphNeuralNetworks.jl depend on GNNGraphs.jl make GraphNeuralNetworks.jl depend on GNNGraphs.jl #453
  • Make sure to include all the latest changes to the GNNGraphs module into the GNNGraphs package
• Make GNNlib.jl depend on GNNGraphs.jl
• Create GNNLux.jl depending on GNNlib.jl and GNNGraphs.jl. create GNNLux.jl package #460
• Make GraphNeuralNetworks.jl depend on NNlib.jl. use GNNlib in GNN.jl #464
• Setup multi-package documentation (consider using MultiDocumenter #456)
  • write developer docs