Inquiry about consumption of memory about seal

[muhanzhang]

The memory consumption increases exponentially with hops. If hop=2 OOMs and you still want to use hop=2, you may keep decreasing max_nodes_per_hop to a smaller value, or further reduce the max-train-num.

[joc32]

May I ask you @muhanzhang , what were the parameters of your computer setup when you were running SEAL with hops = 2 on the big datasets like (Arxiv, FB, Wikipedia, Blogpost) so that I know how much RAM I approximately need?

Seems like the building of the SEALDataset from TorchGeometric is the most demanding task. Even to build the PB dataset, it took me around 800 seconds to be done with hops = 1. I am using SEAL code from PTG Link Prediction demo.

Thanks,
Joe

[muhanzhang]

Pytorch Geometric default k-hop subgraph sampler is slow. I will release a new version (possibly merge to PyG) with much faster (~1000 times) speed up recently. Please stay tuned. I'll let you know when it is released.

[joc32]

@muhanzhang That would be much appreciated! Even if you could share it somehow here as a gist in some beta version. I need it for my MSc thesis and so far I cannot run Arxiv and Facebook with PTG SEAL as it is really slow.

[muhanzhang]

Hi @joc32 , the new version of SEAL is here. It runs on OGB benchmark datasets.

[joc32]

Thanks very much @muhanzhang , will look into it.