Carlo Lucibello

Support AMDGPU

1

Needs to implement gather/scatter kernels in NNlib

implement Graphormer

4

https://proceedings.neurips.cc/paper/2021/file/f1c1592588411002af340cbaedd6fc33-Paper.pdf

- [ ] DiffPool. paper: https://arxiv.org/abs/1806.08804 - [ ] MinCutPool. paper: https://arxiv.org/abs/1907.00481 - [ ] Graculus. paper: https://www.cs.utexas.edu/users/inderjit/public_papers/multilevel_pami.pdf / https://arxiv.org/pdf/1606.09375.pdf

One possibility in order to simplify the implementation is to incorporate the normalization into the edge weights.

Consider adding the definitions ```julia function Functors.functor(::Type{ getfield(m, f) for f in fieldnames(T))...) Tstripped = Base.typename(T).wrapper # remove all parameters. From https://discourse.julialang.org/t/stripping-parameter-from-parametric-types/8293/16 re = x -> Tstripped(x...) return childr, re...

find and fix invalidations

1

See https://timholy.github.io/SnoopCompile.jl/stable/snoopr/

Are there other asserts like the one in #259 hurting performances? One could be https://github.com/CarloLucibello/GraphNeuralNetworks.jl/blob/2912aeb5b9c7c33fc8eced6a203870ec61bf0346/src/layers/conv.jl#L1274

See https://timholy.github.io/SnoopCompile.jl/stable/snoop_pc/

support equivariant neural networks

6

- [ ] General forms using spherical harmonics https://www.nature.com/articles/s41467-022-29939-5 https://docs.e3nn.org/en/latest/guide/convolution.html - [x] Simpler equivariant layers https://proceedings.mlr.press/v139/satorras21a.html https://github.com/lucidrains/egnn-pytorch https://docs.dgl.ai/en/0.9.x/generated/dgl.nn.pytorch.conv.EGNNConv.html