Adding AQSOL Dataset (#240)

* AQSOL

* aqsol working

* remove redundant file

* remove redundant file

* Update graphs.jl

* Update graphs.jl: Fixing tests

* Update AQSOL.jl: docs

* Update AQSOL.jl: better docstring

* Update MLDatasets.jl

* Update AQSOL.jl: fix edge_index

* Update graphs.jl: update tests

* Update graphs.jl: check only random graphs

* Update graphs.md: add to docs

Author: rbSparky

docs/src/datasets/graphs.md

@@ -20,6 +20,7 @@ MLDatasets.HeteroGraph
 ```
 
 ```@docs
+AQSOL
 ChickenPox
 CiteSeer
 Cora

src/MLDatasets.jl

@@ -119,6 +119,8 @@ include("datasets/graphs/citeseer.jl")
 export CiteSeer
 include("datasets/graphs/karateclub.jl")
 export KarateClub
+include("datasets/graphs/AQSOL.jl")
+export AQSOL
 include("datasets/graphs/movielens.jl")
 export MovieLens
 include("datasets/graphs/ogbdataset.jl")
@@ -151,6 +153,7 @@ function __init__()
     # TODO automatically find and execute __init__xxx functions
 
     # graph
+    __init__aqsol()    
     __init__chickenpox()
     __init__citeseer()
     __init__cora()
@@ -166,7 +169,6 @@ function __init__()
     __init__temporalbrains()
     __init__windmillenergy()
 
-
     # misc
     __init__iris()
     __init__mutagenesis()

src/datasets/graphs/AQSOL.jl

@@ -0,0 +1,100 @@
+function __init__aqsol()
+    DEPNAME = "AQSOL"
+    LINK = "https://www.dropbox.com/s/lzu9lmukwov12kt/aqsol_graph_raw.zip?dl=1"
+    register(DataDep(DEPNAME,
+                     """
+                     Dataset: The AQSOL dataset.
+                     Website: http://arxiv.org/abs/2003.00982
+                     """,
+                     LINK,
+                     post_fetch_method = unpack))
+end
+
+struct AQSOL <: AbstractDataset
+    split::Symbol
+    metadata::Dict{String,Any}
+    graphs::Vector{Graph}
+end
+
+"""
+    AQSOL(; split=:train, dir=nothing)
+
+The AQSOL (Aqueous Solubility) dataset from the paper 
+[Graph Neural Network for Predicting Aqueous Solubility of Organic Molecules](http://arxiv.org/abs/2003.00982).
+
+The dataset contains 9,882 graphs representing small organic molecules. Each graph represents a molecule, where nodes correspond to atoms and edges to bonds. The node features represent the atomic number, and the edge features represent the bond type. The target is the aqueous solubility of the molecule, measured in mol/L.
+
+# Arguments
+
+- `split`: Which split of the dataset to load. Can be one of `:train`, `:val`, or `:test`. Defaults to `:train`.
+- `dir`: Directory in which the dataset is in.
+
+# Examples
+
+```julia-repl
+julia> using MLDatasets
+
+julia> data = AQSOL()
+dataset AQSOL:
+  split     =>    :train
+  metadata  =>    Dict{String, Any} with 1 entry
+  graphs    =>    7985-element Vector{MLDatasets.Graph}
+
+julia> length(data)
+7985
+
+julia> g = data[1]
+Graph:
+  num_nodes   =>    23
+  num_edges   =>    42
+  edge_index  =>    ("42-element Vector{Int64}", "42-element Vector{Int64}")
+  node_data   =>    (features = "23-element Vector{Int64}",)
+  edge_data   =>    (features = "42-element Vector{Int64}",)
+
+julia> g.num_nodes
+23
+
+julia> g.node_data.features
+23-element Vector{Int64}:
+ 0
+ 1
+ 1
+ ⋮
+ 1
+ 1
+ 1
+
+julia> g.edge_index
+([2, 3, 3, 4, 4, 5, 5, 6, 6, 7  …  18, 19, 19, 20, 20, 21, 20, 22, 20, 23], [3, 2, 4, 3, 5, 4, 6, 5, 7, 6  …  19, 18, 20, 19, 21, 20, 22, 20, 23, 20])
+```
+"""
+function AQSOL(;split=:train, dir=nothing)
+    @assert split ∈ [:train, :val, :test]
+    DEPNAME = "AQSOL"
+    path = datafile(DEPNAME, "asqol_graph_raw/$(split).pickle", dir)
+    graphs = Pickle.npyload(path)
+    g = [create_aqsol_graph(g...) for g in graphs]
+    metadata = Dict{String, Any}("n_observations" => length(g))
+    return AQSOL(split, metadata, g)
+end
+
+function create_aqsol_graph(x, edge_attr, edge_index, y)
+    x = Int.(x)
+    edge_attr = Int.(edge_attr)
+    edge_index = Int.(edge_index .+ 1)
+    
+    if size(edge_index, 2) == 0
+        s, t = Int[], Int[] 
+    else
+        s, t = edge_index[1, :], edge_index[2, :]
+    end
+
+    return Graph(; num_nodes = length(x),
+                  edge_index = (s, t),
+                  node_data = (features = x,),
+                  edge_data = (features = edge_attr,))
+end
+
+Base.length(d::AQSOL) = length(d.graphs)
+Base.getindex(d::AQSOL, ::Colon) = d.graphs
+Base.getindex(d::AQSOL, i) = getindex(d.graphs, i)

src/datasets/graphs/movielens.jl

@@ -549,4 +549,4 @@ Base.length(data::MovieLens) = length(data.graphs)
 function Base.getindex(data::MovieLens, ::Colon)
     length(data.graphs) == 1 ? data.graphs[1] : data.graphs
 end
-Base.getindex(data::MovieLens, i) = getobs(data.graphs, i)
+Base.getindex(data::MovieLens, i) = getobs(data.graphs, i)

test/datasets/graphs.jl

@@ -278,3 +278,23 @@ end
     @test size(g.edge_data.features) == (2, g.num_edges)
     @test size(g.edge_data.targets) == (g.num_edges,)
 end
+
+@testset "AQSOL" begin
+    split_counts = Dict(:train => 7985, :val => 998, :test => 999)
+    for split in [:train, :val, :test]
+        data = AQSOL(split=split)
+        @test data isa AbstractDataset
+        @test data.split == split
+        @test length(data) == data.metadata["n_observations"]
+        @test length(data.graphs) == split_counts[split]
+
+        i = rand(1:length(data))
+        g = data[i]
+        @test g isa MLDatasets.Graph
+        s, t = g.edge_index
+        @test all(1 .<= s .<= g.num_nodes)
+        @test all(1 .<= t .<= g.num_nodes)
+        @test length(s) == g.num_edges
+        @test length(t) == g.num_edges
+    end
+end