search

pyg-team / pytorch_geometric

Graph Neural Network Library for PyTorch
https://pyg.org
MIT License
20.91k stars 3.61k forks source link

Question about mini-batches training for hypergraph neural network #7365

Open ShuaiWang97 opened 1 year ago

ShuaiWang97 commented 1 year ago 
          This is not supported yet, sorry :( We currently only support mini-batch training for adjacency matrix representation but not for incidence matrix representation (although it is possible that we add such feature in the future).

_Originally posted by @rusty1s in https://github.com/pyg-team/pytorch_geometric/issues/2618#issuecomment-846538801_

ShuaiWang97 commented 1 year ago

Thanks for the answer! Is it possible to realize it using some graph mini-batches sampler (like GraphSAINT or GraphSAGE)?

I guess it is possible to update theData.__inc__in hypergraph by these built graph samplers like torch_geometric.sampler? I am very interested in this and wondering if you can give me some insights. Any suggestions would be helpful!

rusty1s commented 1 year ago

Yes, this is exactly what you would need to do (re-implementing Data.__inc__. We have a longer-term roadmap planed for this (see #7312). Feel free to join the discussion there, we are always looking for contributors :)

SauravMaheshkar commented 5 months ago

@CodeTal Is it safe to assume that the current implementation supports sampling using the various implementations in torch_geometric.sampler ?