Min-Batch for HAN

Hi,everyone!
How to implement HAN by using neighbor sampling and min-batch?

same question,How to use the new sampler API?Is there any Doc about this??? :pray:

up! please
up!!
up!!
up!!
up!!
up!!

The bulk of HAN is essentially GATs on a set of homogeneous graphs coalesced from metapaths.

Implementing GAT with neighbor sampling and minibatch should be straightforward if we start from the GraphSAGE + neighbor sampling example in https://github.com/dmlc/dgl/blob/master/examples/pytorch/graphsage/train_sampling.py. Simply replacing SAGEConv with GATConv should be everything. If you find the code hard to understand, a step-by-step walkthrough on implementing GraphSAGE with minibatch and neighbor sampling is explained in this upcoming WebConf 2020 tutorial.

Now once you have multiple homogeneous graphs, you can perform neighbor sampling on each graph separately, obtain the output from the corresponding GATs, and combine them with semantic attention.

This should be straightforward but currently I can’t promise when I will add the minibatch implementation of HAN to our collection. You are very welcome to give it a try if possible.

Please feel free to follow up.