How to set batch size when training AllSet?

[thu-wangz17]

Hi, this is a nice work. I read the code:

AllSet/src/train.py

Lines 325 to 327 in 0d0e399

     dataset = dataset_Hypergraph(name=dname,root = '../data/pyg_data/hypergraph_dataset_updated/', 
                                  p2raw = p2raw) 
 data = dataset.data 

AllSet/src/train.py

Line 345 in 0d0e399

 data = ExtractV2E(data) 

AllSet/src/train.py

Line 478 in 0d0e399

 out = model(data) 

It seems that you train the model with all the data in a batch. Thus I have a question about how to set batch size when I train the AllSet model. Thank you.

[Eli Chien]

Hi @sakuraiiiii ,

This is a great question! Unfortunately, mini-batch training in AllSet is currently an open problem, where we stated in our paper (Appendix A). We conjecture one can develop a GraphSAINT or ClusterGCN type of method to allow mini-batch training. There are also many open directions stated in Appendix A, which we conjecture that one can leverage the similar idea in the GNN literature.

Please let me know if you manage to develop mini-batch training! It would be a great improvement to AllSet and to general hypergraph neural networks!

Best,
Eli

[thu-wangz17]

Thank you very much. If I fix this question, I will give you feedback :)