Hi, I am trying to understand how DGL reconstructs the subgraph after sampling the neighborhood of target nodes. Let’s say we have target nodes = [0,1,2,3]. If we sample 3 neighbors for each target node, we have the following:
target node id: sampled nbrs
0: 2, 4, 5
1: 3, 6
2: 6, 4, 3
3: 0
Then, NeighborSampler.sample_blocks return seed_nodes, output_nodes, blocks, where seed_nodes is the unique list of sampled neighbors ([0,2,3,4,5,6] in this case), output_nodes is the original target nodes ([0,1,2,3]) and blocks should include the metadata ([Block(num_src_nodes=6, num_dst_nodes=4, num_edges=9)] in this case). Given these three outputs, I’m wondering how DGL is able to distinguish the neighbors from each target node. For example, 4 is the neighbor for both 0 and 2 but where is this information stored? I see in the training script the entire edges list is read into graph. Does that mean going to memory to fetch neighborhood information whenever the neighborhood information is needed?
Thank you in advance!