Running TODOs

[jermainewang]

Hi all,

Let's use this thread to book-keep all the major TODOs before our first major milestone.

High priority Tasks that have a major contributor to be in charge of.

• [x] C <=> Python interface (probably borrow from TVM).
• [ ] Lowering graph related logic to C++
  • [x] Light weight & efficient C++ graph data structure
  • [x] Graph construction and graph batching
  • [x] Graph query logic (the current cached graph)
  • [ ] Graph traversal logic
  • [x] Fast generation of adj or incident matrix
• [ ] generic-SPMV in C++
  • [ ] Operator interface and implementation
  • [ ] Plugin to Pytorch
• [x] Improve end-to-end training time of tree-lstm to be 2x faster than DyNet.
• [ ] Improve GAT-spmv implementation to be faster than TF.
• [ ] Document & tutorial plans
  • [x] Setup sphinx doc

Low priority Tasks that anyone can contribute when they don't have high priority tasks at the moment.

• [x] Setup CI
  • [x] Basic CI (#30 )
  • [x] GPU CI
• [x] Setup lint check
• [x] Scaling tests using generated graphs (#32 )
• [ ] Scale to multi-card

Applications

• [ ] GCN-like
  • [x] GCN (#35 )
  • [x] GAT (#37)
  • [ ] MGCN
  • [ ] relational GCN
  • [ ] GeniePath
  • [ ] PageRank
  • [ ] MPNN
  • [ ] GraphSAGE with different aggregators
• [x] TreeLSTM
• [x] LineGraph
• [ ] Graph generation model
  • [x] DGMG (#14 )
  • [ ] Junction Tree VAE
• [ ] DeepWalk/node2vec
• [ ] Transformer (?)
• [ ] CapsuleNet (?)

[zzhang-cn]

Some thoughts: [] Extend networkx's dict-of-dict to store tensors [] Cache adjacency matrix (for the reduce function) [] Establish scale-testing [] Think about how to scale out to multiple cards

[jermainewang]

Incorporated ZZ's advices.

[BarclayII]

I have gone through the nvGraph API and I doubt if we can use the SrSPMV function in the reducers. Their graph structure and SrSPMV only operates on scalars.

Also for keeping track of TODOs I recommend Trello. It can set deadlines, assign people and send reminders to mailboxes.

[zzhang-cn]

Also let's re-org the folds and make a more complete listing of examples. We have:

• PageRank
• TreeLSTM
• GAT, GCN, MGCN, GeniePath,
• Line-graph
• Dynamic graph (in the generative_model branch)

I might have missed some. Let's put a readme in the example folder, with links to the appropriate papers.

[jermainewang]

Close as we are using trello to keep track of progress.

[kitaev-chen]

I think "Dynamic graph (in the generative_model branch)" will be a real competitive feature compared with other graph libraries since graph based generative model / gaussian process will be very prevalent .

[mufeili]

Thanks a lot for the suggestion @kitaev-chen . I'm personally a big fan of generative models and yes we are likely to incorporate some more examples on generative models in the future. As this issue is closed now, you may want to instead post your suggestion at #450 . Also let us know if there are some models that you are particularly interested.