GeometricFlux.jl
GeometricFlux.jl copied to clipboard
FluxML
Geometric Deep Learning for Flux
> Ref. [Hierarchical Graph Representation Learning with Differentiable Pooling](https://arxiv.org/abs/1806.08804)
> Ref. [Deep Graph Infomax](https://arxiv.org/abs/1809.10341)
- [ ] [phanein/deepwalk](https://github.com/phanein/deepwalk) > Ref. [DeepWalk: online learning of social representations](https://dl.acm.org/doi/10.1145/2623330.2623732)
- [ ] `Interact` layer - [ ] `InteractNet` example > Ref. [Interaction Networks for Learning about Objects, Relations and Physics](https://arxiv.org/abs/1612.00222)
- [ ] `GeodConv` layer - [ ] `GeodCNN` example > Ref. [Geodesic Convolutional Neural Networks on Riemannian Manifolds](https://www.cv-foundation.org/openaccess/content_iccv_2015_workshops/w22/papers/Masci_Geodesic_Convolutional_Neural_ICCV_2015_paper.pdf)
This model is a spectral-based graph convolution. It requires operations over a real-valued function with complex coefficient. - [ ] `CayleyConv` layer - [ ] `CayleyNet` example > Ref. [CayleyNets:...
> Ref: [Weisfeiler and Leman Go Neural: Higher-order Graph Neural Networks](https://arxiv.org/abs/1810.02244)
When training on multiple small graphs, typically one batches several graphs together into a larger graph for efficiency. This operation is called [blockdiag](https://juliagraphs.org/LightGraphs.jl/v1.0/operators.html#SparseArrays.blockdiag-Union{Tuple{T},%20Tuple{T,T}}%20where%20T%3C:AbstractGraph) in SparseArrays and LightGraphs.jl. For `FeaturedGraph`s, node...
Exposed in #214 ```julia julia> A = [0 1 0 1 0 0 0 0 0]; julia> fg = FeaturedGraph(A, nf=rand(2,3)); julia> m = GraphConv(2 => 2); julia> m(fg) ERROR:...
The current design where layers store graph references is a bit odd, since the two should be decoupled. Other GNN libraries such as DGL and pytorch geometric pass the graph...