Open Graph Benchmark integration

[DoktorMike]

There are quite a few useful datasets, benchmarks and leaderboards in ogb. The paper is here.

From what I can see it plays nicely with PyTorch Geometric and Deep Graph Library both of which are Python packages.

I would think that having access to these resources through GraphNeuralNetworks.jl or another julia package could ease the attraction of new users. I haven't used many of these datasets myself so I don't know more than this.

[CarloLucibello]

They are already available in MLDatasets.jl. It should be quite easy to create a custom function converting MLDatasets' type to GNN.jl's types.

Once JuliaML/MLDatasets.jl#114 is merged, the interface of MLDatasets' datasets will be streamlined and I will be able to implement conversion utilities here without having to depend on MLDatasets directly.

[CarloLucibello]

With #164 we have

julia> using MLDatasets, GraphNeuralNetworks

julia> dataset = OGBDataset("ogbn-arxiv")
dataset OGBDataset:
  name        =>    ogbn-arxiv
  metadata    =>    Dict{String, Any} with 17 entries
  graphs      =>    1-element Vector{MLDatasets.Graph}
  graph_data  =>    nothing

julia> mldataset2gnngraph(dataset)
GNNGraph:
    num_nodes = 169343
    num_edges = 1166243
    ndata:
        val_mask => 169343-element BitVector
        test_mask => 169343-element BitVector
        year => 169343-element Vector{Int64}
        features => 128×169343 Matrix{Float32}
        label => 169343-element Vector{Int64}
        train_mask => 169343-element BitVector