If a heterogeneous graph generates many subgraphs, and each subgraph requires independent graph convolution, how can we speed up the convolution of so many subgraphs?
At present, I cycle a graph to generate multiple subgraphs, and then convolute each subgraph. It takes a lot of time.
You could use dgl.batch
to batch them into a bigger graph and then do computation on that big graph.
Later on you can split it back with dgl.unbatch
, or using readout functions such as dgl.sum_nodes
to get a representation for each individual graph.
Thank you for your reply.
This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.