GraphSAINT test

[asd10asd]

Hi,

I have a large graph(about 50,000 nodes,10,000,000 edges) and try to train a GNN model . But it seems that the graph is too large to train. I try to use GraphSAINT and read the example graph_saint.py. But when test on the whole graph, it can't work beacuse of limited memory. How can I test the whole graph? Or maybe there is a more convenient way to train and test on my data?

Thanks!

[rusty1s]

50k nodes isn't that large, so I suggest to look into our memory-efficient aggregation options.

[asd10asd]

Thank you for your reply. I have changed my code to use SparseTensor to save memory. But now I ran into another problem. My initialized feature matrix X is an identity matrix. This matrix is 50000* 50000 and need about 12G memory. Can I also use sparse matrix multiplication for X？

[rusty1s]

Yes, this is possible for some, but not all GNN operators. I guess a more elegant solution is to utilize torch.nn.Embedding for learning node features rather than relying on identity matrices (which basically pushes the task of learning node features to the first GNN op).

[asd10asd]

Thank you very much. Your answer helped me very well. I have solved this problem now.