search

FluxML / NNlib.jl

Neural Network primitives with multiple backends
Other
198 stars 121 forks source link

spconv ... sparse convolution #493

Open dorn-gerhard opened 1 year ago

dorn-gerhard commented 1 year ago

Motivation and description

For many applications (graph neural networks, 3d point clouds, ...) sparse convolution would be a cool feature

Possible Implementation

A guide to implementation is given here (with some references): https://towardsdatascience.com/how-does-sparse-convolution-work-3257a0a8fd1

ToucheSir commented 1 year ago

I doubt anyone is going to step up to write custom CUDA kernels for this like in that post unless they also need the feature, so it's in solid "PRs welcome" status for now. You may want to see if anything in GeometricFlux, GraphNeuralNetworks.jl or Flux3D helps with your use case. Those three should cover GNNs and point clouds between them.