How to handle a batch of graph with similar structure

Models & Apps
#1

I try to use dgl.HeteroGraphConv as a feature extractor in reinforcement learning.
So, when I sample replay-buffer, it will return a batch of hetero graph.
Each graph has 9000 nodes, 26000 edges, and 1900*1 features for each nodes.

These graphs are come from a source graph with add or delete some edges (less than 5%), all graph has the same feats.
I try to use dgl.batch() but I find it is too slow.
So I try to using for-loop, but I find this approach does not make efficient use of GPU resources.

What should I do? should I reduce the batch size and keep using dgl.batch()?

#2

Did you profile and see where the bottleneck is? The bottleneck may not be in dgl.batch() itself but in other places. pyinstrument or line_profiler are two common choices.

Once you found the bottleneck we can optimize them.