Hello, I would like to ask about accelerating graph sampling with CUDA. It seems relatively easy to perform sampling with a specified fanout, but for models that require full neighborhood sampling like GCN, how does DGL handle it? Also, for CUDA programs, don’t we need to know the size of the output variables in advance? I apologize that I couldn’t find the specific part I was looking for in the code. Could you provide some information or give a brief description?
I don’t think people really use GCN + neighbor sampling. If you really want to perform sampling-based training of GCN, it might be more appropriate to perform subgraph sampling instead.
What do you mean by “Also, for CUDA programs, don’t we need to know the size of the output variables in advance?”
This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.