feat: add NeighborLoader (#497)

Author: askorupka

GNNGraphs/src/sampling.jl

@@ -177,6 +177,8 @@ function Graphs.induced_subgraph(graph::GNNGraph, nodes::Vector{Int})
 
     node_map = Dict(node => i for (i, node) in enumerate(nodes))
 
+    edge_list = [collect(t) for t in zip(edge_index(graph)[1],edge_index(graph)[2])]
+
     # Collect edges to add
     source = Int[]
     target = Int[]
@@ -187,8 +189,7 @@ function Graphs.induced_subgraph(graph::GNNGraph, nodes::Vector{Int})
             if neighbor in keys(node_map)
                 push!(target, node_map[node])
                 push!(source, node_map[neighbor])
-
-                eindex = findfirst(x -> x == [neighbor, node], edge_index(graph))
+                eindex = findfirst(x -> x == [neighbor, node], edge_list)
                 push!(eindices, eindex)
             end
         end

GraphNeuralNetworks/Project.toml

@@ -9,6 +9,7 @@ Flux = "587475ba-b771-5e3f-ad9e-33799f191a9c"
 Functors = "d9f16b24-f501-4c13-a1f2-28368ffc5196"
 GNNGraphs = "aed8fd31-079b-4b5a-b342-a13352159b8c"
 GNNlib = "a6a84749-d869-43f8-aacc-be26a1996e48"
+Graphs = "86223c79-3864-5bf0-83f7-82e725a168b6"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 MLUtils = "f1d291b0-491e-4a28-83b9-f70985020b54"
 MacroTools = "1914dd2f-81c6-5fcd-8719-6d5c9610ff09"
@@ -28,6 +29,7 @@ CUDA = "4, 5"
 ChainRulesCore = "1"
 Flux = "0.14"
 Functors = "0.4.1"
+Graphs = "1.12"
 GNNGraphs = "1.0"
 GNNlib = "0.2"
 LinearAlgebra = "1"

GraphNeuralNetworks/docs/make.jl

@@ -46,6 +46,7 @@ makedocs(;
                  "Message Passing" => "api/messagepassing.md",
                  "Heterogeneous Graphs" => "api/heterograph.md",
                  "Temporal Graphs" => "api/temporalgraph.md",
+                 "Samplers" => "api/samplers.md",
                  "Utils" => "api/utils.md",
              ],
              "Developer Notes" => "dev.md",

GraphNeuralNetworks/docs/src/api/samplers.md

@@ -0,0 +1,14 @@
+```@meta
+CurrentModule = GraphNeuralNetworks
+```
+
+# Samplers
+
+
+## Docs
+
+```@autodocs
+Modules = [GraphNeuralNetworks]
+Pages   = ["samplers.jl"]
+Private = false
+```

GraphNeuralNetworks/src/GraphNeuralNetworks.jl

@@ -10,6 +10,7 @@ using NNlib: scatter, gather
 using ChainRulesCore
 using Reexport
 using MLUtils: zeros_like
+using Graphs: Graphs
 
 using GNNGraphs:  COO_T, ADJMAT_T, SPARSE_T,
                   check_num_nodes, check_num_edges,
@@ -66,4 +67,7 @@ export GlobalPool,
 
 include("deprecations.jl")
 
+include("samplers.jl")
+export NeighborLoader
+
 end

GraphNeuralNetworks/src/samplers.jl

@@ -0,0 +1,103 @@
+"""
+    NeighborLoader(graph; num_neighbors, input_nodes, num_layers, [batch_size])
+
+A data structure for sampling neighbors from a graph for training Graph Neural Networks (GNNs). 
+It supports multi-layer sampling of neighbors for a batch of input nodes, useful for mini-batch training
+originally introduced in "Inductive Representation Learning on Large Graphs" paper.
+[see https://arxiv.org/abs/1706.02216]
+
+# Fields
+- `graph::GNNGraph`: The input graph.
+- `num_neighbors::Vector{Int}`: A vector specifying the number of neighbors to sample per node at each GNN layer.
+- `input_nodes::Vector{Int}`: A vector containing the starting nodes for neighbor sampling.
+- `num_layers::Int`: The number of layers for neighborhood expansion (how far to sample neighbors).
+- `batch_size::Union{Int, Nothing}`: The size of the batch. If not specified, it defaults to the number of `input_nodes`.
+
+# Usage
+```jldoctest
+julia>  loader = NeighborLoader(graph; num_neighbors=[10, 5], input_nodes=[1, 2, 3], num_layers=2)
+
+julia> batch_counter = 0
+julia> for mini_batch_gnn in loader
+            batch_counter += 1
+            println("Batch ", batch_counter, ": Nodes in mini-batch graph: ", nv(mini_batch_gnn))
+```
+"""
+struct NeighborLoader
+    graph::GNNGraph             # The input GNNGraph (graph + features from GraphNeuralNetworks.jl)
+    num_neighbors::Vector{Int}  # Number of neighbors to sample per node, for each layer
+    input_nodes::Vector{Int}    # Set of input nodes (starting nodes for sampling)
+    num_layers::Int             # Number of layers for neighborhood expansion
+    batch_size::Union{Int, Nothing}  # Optional batch size, defaults to the length of input_nodes if not given
+    neighbors_cache::Dict{Int, Vector{Int}}  # Cache neighbors to avoid recomputation
+end
+
+function NeighborLoader(graph::GNNGraph; num_neighbors::Vector{Int}, input_nodes::Vector{Int}=nothing, 
+                        num_layers::Int, batch_size::Union{Int, Nothing}=nothing)
+    return NeighborLoader(graph, num_neighbors, input_nodes === nothing ? collect(1:graph.num_nodes) : input_nodes, num_layers, 
+                            batch_size === nothing ? length(input_nodes) : batch_size, Dict{Int, Vector{Int}}())
+end
+
+# Function to get cached neighbors or compute them
+function get_neighbors(loader::NeighborLoader, node::Int)
+    if haskey(loader.neighbors_cache, node)
+        return loader.neighbors_cache[node]
+    else
+        neighbors = Graphs.neighbors(loader.graph, node, dir = :in)  # Get neighbors from graph
+        loader.neighbors_cache[node] = neighbors
+        return neighbors
+    end
+end
+
+# Function to sample neighbors for a given node at a specific layer
+function sample_nbrs(loader::NeighborLoader, node::Int, layer::Int)
+    neighbors = get_neighbors(loader, node)
+    if isempty(neighbors)
+        return Int[]
+    else
+        num_samples = min(loader.num_neighbors[layer], length(neighbors))  # Limit to required samples for this layer
+        return rand(neighbors, num_samples)  # Randomly sample neighbors
+    end
+end
+
+# Iterator protocol for NeighborLoader with lazy batch loading
+function Base.iterate(loader::NeighborLoader, state=1)
+    if state > length(loader.input_nodes) 
+        return nothing  # End of iteration if batches are exhausted (state larger than amount of input nodes or current batch no >= batch number)
+    end
+
+    # Determine the size of the current batch
+    batch_size = min(loader.batch_size, length(loader.input_nodes) - state + 1) # Conditional in case there is not enough nodes to fill the last batch
+    batch_nodes = loader.input_nodes[state:state + batch_size - 1] # Each mini-batch uses different set of input nodes 
+
+    # Set for tracking the subgraph nodes
+    subgraph_nodes = Set(batch_nodes)
+
+    for node in batch_nodes
+        # Initialize current layer of nodes (starting with the node itself)
+        sampled_neighbors = Set([node])
+
+        # For each GNN layer, sample the neighborhood
+        for layer in 1:loader.num_layers
+            new_neighbors = Set{Int}()
+            for n in sampled_neighbors
+                neighbors = sample_nbrs(loader, n, layer)  # Sample neighbors of the node for this layer
+                new_neighbors = union(new_neighbors, neighbors)  # Avoid duplicates in the neighbor set
+            end
+            sampled_neighbors = new_neighbors
+            subgraph_nodes = union(subgraph_nodes, sampled_neighbors)  # Expand the subgraph with the new neighbors
+        end
+    end
+
+    # Collect subgraph nodes and their features
+    subgraph_node_list = collect(subgraph_nodes)
+
+    if isempty(subgraph_node_list)
+        return GNNGraph(), state + batch_size
+    end
+
+    mini_batch_gnn = Graphs.induced_subgraph(loader.graph, subgraph_node_list)  # Create a subgraph of the nodes
+
+    # Continue iteration for the next batch
+    return mini_batch_gnn, state + batch_size
+end

GraphNeuralNetworks/test/runtests.jl

@@ -30,6 +30,7 @@ tests = [
     "layers/temporalconv",
     "layers/pool",
     "examples/node_classification_cora",
+    "samplers"
 ]
 
 !CUDA.functional() && @warn("CUDA unavailable, not testing GPU support")

GraphNeuralNetworks/test/samplers.jl

@@ -0,0 +1,125 @@
+# Helper function to create a simple graph with node features using GNNGraph
+function create_test_graph()
+    source = [1, 2, 3, 4]  # Define source nodes of edges
+    target = [2, 3, 4, 5]  # Define target nodes of edges
+    node_features = rand(Float32, 5, 5)  # Create random node features (5 features for 5 nodes)
+
+    return GNNGraph(source, target, ndata = node_features)  # Create a GNNGraph with edges and features
+end
+
+# Tests for NeighborLoader structure and its functionalities
+@testset "NeighborLoader tests" begin
+
+    # 1. Basic functionality: Check neighbor sampling and subgraph creation
+    @testset "Basic functionality" begin
+        g = create_test_graph()
+
+        # Define NeighborLoader with 2 neighbors per layer, 2 layers, batch size 2
+        loader = NeighborLoader(g; num_neighbors=[2, 2], input_nodes=[1, 2], num_layers=2, batch_size=2)
+
+        mini_batch_gnn, next_state = iterate(loader)
+
+        # Test if the mini-batch graph is not empty
+        @test !isempty(mini_batch_gnn.graph)
+
+        num_sampled_nodes = mini_batch_gnn.num_nodes
+        println("Number of nodes in mini-batch: ", num_sampled_nodes)
+
+        @test num_sampled_nodes == 2
+
+        # Test if there are edges in the subgraph
+        @test mini_batch_gnn.num_edges > 0
+    end
+
+    # 2. Edge case: Single node with no neighbors
+    @testset "Single node with no neighbors" begin
+        g = SimpleDiGraph(1)  # A graph with a single node and no edges
+        node_features = rand(Float32, 5, 1)
+        graph = GNNGraph(g, ndata = node_features)
+
+        loader = NeighborLoader(graph; num_neighbors=[2], input_nodes=[1], num_layers=1)
+
+        mini_batch_gnn, next_state = iterate(loader)
+
+        # Test if the mini-batch graph contains only one node
+        @test size(mini_batch_gnn.x, 2) == 1
+    end
+
+    # 3. Edge case: A node with no outgoing edges (isolated node)
+    @testset "Node with no outgoing edges" begin
+        g = SimpleDiGraph(2)  # Graph with 2 nodes, no edges
+        node_features = rand(Float32, 5, 2)
+        graph = GNNGraph(g, ndata = node_features)
+
+        loader = NeighborLoader(graph; num_neighbors=[1], input_nodes=[1, 2], num_layers=1)
+
+        mini_batch_gnn, next_state = iterate(loader)
+
+        # Test if the mini-batch graph contains the input nodes only (as no neighbors can be sampled)
+        @test size(mini_batch_gnn.x, 2) == 2  # Only two isolated nodes
+    end
+
+    # 4. Edge case: A fully connected graph
+    @testset "Fully connected graph" begin
+        g = SimpleDiGraph(3)
+        add_edge!(g, 1, 2)
+        add_edge!(g, 2, 3)
+        add_edge!(g, 3, 1)
+        node_features = rand(Float32, 5, 3)
+        graph = GNNGraph(g, ndata = node_features)
+
+        loader = NeighborLoader(graph; num_neighbors=[2, 2], input_nodes=[1], num_layers=2)
+
+        mini_batch_gnn, next_state = iterate(loader)
+
+        # Test if all nodes are included in the mini-batch since it's fully connected
+        @test size(mini_batch_gnn.x, 2) == 3  # All nodes should be included
+    end
+
+    # 5. Edge case: More layers than the number of neighbors
+    @testset "More layers than available neighbors" begin
+        g = SimpleDiGraph(3)
+        add_edge!(g, 1, 2)
+        add_edge!(g, 2, 3)
+        node_features = rand(Float32, 5, 3)
+        graph = GNNGraph(g, ndata = node_features)
+
+        # Test with 3 layers but only enough connections for 2 layers
+        loader = NeighborLoader(graph; num_neighbors=[1, 1, 1], input_nodes=[1], num_layers=3)
+
+        mini_batch_gnn, next_state = iterate(loader)
+
+        # Test if the mini-batch graph contains all available nodes
+        @test size(mini_batch_gnn.x, 2) == 1
+    end
+
+    # 6. Edge case: Large batch size greater than the number of input nodes
+    @testset "Large batch size" begin
+        g = create_test_graph()
+
+        # Define NeighborLoader with a larger batch size than input nodes
+        loader = NeighborLoader(g; num_neighbors=[2], input_nodes=[1, 2], num_layers=1, batch_size=10)
+
+        mini_batch_gnn, next_state = iterate(loader)
+
+        # Test if the mini-batch graph is not empty
+        @test !isempty(mini_batch_gnn.graph)
+
+        # Test if the correct number of nodes are sampled
+        @test size(mini_batch_gnn.x, 2) == length(unique([1, 2]))  # Nodes [1, 2] are expected
+    end
+
+    # 7. Edge case: No neighbors sampled (num_neighbors = [0]) and 1 layer
+    @testset "No neighbors sampled" begin
+        g = create_test_graph()
+
+        # Define NeighborLoader with 0 neighbors per layer, 1 layer, batch size 2
+        loader = NeighborLoader(g; num_neighbors=[0], input_nodes=[1, 2], num_layers=1, batch_size=2)
+
+        mini_batch_gnn, next_state = iterate(loader)
+
+        # Test if the mini-batch graph contains only the input nodes
+        @test size(mini_batch_gnn.x, 2) == 2  # No neighbors should be sampled, only nodes 1 and 2 should be in the graph
+    end
+
+end