[Example] Dynamic Graph CNN on Point Cloud

[BarclayII]

Description

This PR implements the Dynamic Graph CNN paper, as well as the EdgeConv module and DynamicEdgeConv module in #748 . Namely, it provides

• EdgeConv in the said paper.
• NearestNeighborGraph: A module that constructs a graph from a feature matrix based on nearest neighbors.
• (Alternating between NearestNeighborGraph and EdgeConv yields DynamicEdgeConv.)

Hopefully this could be a starting point for issue #719.

This runs on ModelNet40 dataset and got a best validation accuracy of 93.5% and test accuracy of 91.8%.

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

[yzh119]

@BarclayII , could you please move the EdgeConv module to dgl.nn.*.conv? Plus, I recommend putting NearestNeighborGraph to dgl.transforms, thanks!

[BarclayII]

@BarclayII , could you please move the EdgeConv module to dgl.nn.*.conv? Plus, I recommend putting NearestNeighborGraph to dgl.transforms, thanks!

I'm actually thinking what the form of EdgeConv and NearestNeighborGraph should be. In particular, whether we want to support point cloud with different number of points.

[BarclayII]

Temporarily closing to save CI as I'm migrating code to dgl.nn.

[yzh119]

Great work! Go ahead and merge it.

[jermainewang]

@BarclayII could you rename nearest_neighbor_graph to just knn_graph and segmented_knn_graph? They are much shorter.