[WIP] variable num_neighbors for different hops & add self loop in the sampled subgraph

[ZiyueHuang]

Description

Different hops might require different num_neighbors to be sampled (For example, in GraphSage, num_neigh = 25 for the first layer and num_neigh = 10 for the second layer). test_node_flow.py passed.

TODO: For GraphSage, self loop is needed in the sampled subgraph, which is required to implement an equivalent version in DGL.

Checklist

Please feel free to remove inapplicable items for your PR.

• The PR title starts with [$CATEGORY] (such as [Model], [Doc], [Feature]])
• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage
• Code is well-documented
• To the my best knowledge, examples are either not affected by this change,
  or have been fixed to be compatible with this change
• Related issue is referred in this PR

Changes

[zheng-da]

How do you add self-loop in the sampled subgraph?

[ZiyueHuang]

I tried skip-connection (second-order proximity from Adaptive Sampling) in gcn_sc.py, which has amazing performance. On Cora, it achieves 80% test accuracy after ~30 epochs and finally ~84% test accuracy (if without skip-connection, it achieves 80% test accuracy after ~110 epochs and finally ~81% test accuracy). For neighbor sampling (num_neighbors=2), it stably and quickly achieves ~84% test accuracy (if without skip-connection, it achieves ~60% test accuracy unstably).

DGLBACKEND=mxnet python3 gcn_sc.py --dataset cora --num-neighbors NUM_NEIGHBORS  --batch-size 10000 --self-loop

DGLBACKEND=mxnet python3 gcn_ns.py --dataset cora --num-neighbors NUM_NEIGHBORS  --batch-size 10000 --self-loop