search

benedekrozemberczki / pytorch_geometric_temporal

PyTorch Geometric Temporal: Spatiotemporal Signal Processing with Neural Machine Learning Models (CIKM 2021)
MIT License
2.63k stars 370 forks source link

CUDA out of memory when the graph(edge_index and x) is big #247

Open jexterliangsufe opened 1 year ago

jexterliangsufe commented 1 year ago

My graph: tens of millions edges and 70000 nodes My model is something like GCNConv+GRU(GRU's reset, update and the other part is GCNConv)

Problem: cuda memory usage becomes bigger and bigger as time steps add.

How can I revise GNN part or RNNs part to solve this?