Merge pull request #317 from FluxML/new_features

New features: positional encoding

Author: yuehhua

Project.toml

@@ -31,7 +31,7 @@ DataStructures = "0.18"
 FillArrays = "0.13"
 Flux = "0.12 - 0.13"
 GraphMLDatasets = "0.1"
-GraphSignals = "0.4 - 0.5"
+GraphSignals = "0.6"
 Graphs = "1"
 NNlib = "0.8"
 NNlibCUDA = "0.2"

docs/Project.toml

@@ -1,6 +1,7 @@
 [deps]
 Documenter = "e30172f5-a6a5-5a46-863b-614d45cd2de4"
 DocumenterCitations = "daee34ce-89f3-4625-b898-19384cb65244"
+Flux = "587475ba-b771-5e3f-ad9e-33799f191a9c"
 
 [compat]
 Documenter = "0.27"

docs/bibliography.bib

@@ -184,6 +184,29 @@ @inproceedings{Hamilton2017
    year = {2017},
 }
 
+@inproceedings{Satorras2021,
+   abstract = {This paper introduces a new model to learn graph neural networks equivariant to rotations, translations, reflections and permutations called E(n)-Equivariant Graph Neural Networks (EGNNs). In contrast with existing methods, our work does not require computationally expensive higher-order representations in intermediate layers while it still achieves competitive or better performance. In addition, whereas existing methods are limited to equivariance on 3 dimensional spaces, our model is easily scaled to higher-dimensional spaces. We demonstrate the effectiveness of our method on dynamical systems modelling, representation learning in graph autoencoders and predicting molecular properties.},
+   author = {Victor Garcia Satorras and Emiel Hoogeboom and Max Welling},
+   editor = {Marina Meila and Tong Zhang},
+   booktitle = {Proceedings of the 38th International Conference on Machine Learning},
+   month = {2},
+   pages = {9323-9332},
+   publisher = {PMLR},
+   title = {E(n) Equivariant Graph Neural Networks},
+   volume = {139},
+   url = {http://arxiv.org/abs/2102.09844},
+   year = {2021},
+}
+
+@article{Dwivedi2021,
+   abstract = {Graph neural networks (GNNs) have become the standard learning architectures for graphs. GNNs have been applied to numerous domains ranging from quantum chemistry, recommender systems to knowledge graphs and natural language processing. A major issue with arbitrary graphs is the absence of canonical positional information of nodes, which decreases the representation power of GNNs to distinguish e.g. isomorphic nodes and other graph symmetries. An approach to tackle this issue is to introduce Positional Encoding (PE) of nodes, and inject it into the input layer, like in Transformers. Possible graph PE are Laplacian eigenvectors. In this work, we propose to decouple structural and positional representations to make easy for the network to learn these two essential properties. We introduce a novel generic architecture which we call LSPE (Learnable Structural and Positional Encodings). We investigate several sparse and fully-connected (Transformer-like) GNNs, and observe a performance increase for molecular datasets, from 2.87% up to 64.14% when considering learnable PE for both GNN classes.},
+   author = {Vijay Prakash Dwivedi and Anh Tuan Luu and Thomas Laurent and Yoshua Bengio and Xavier Bresson},
+   month = {10},
+   title = {Graph Neural Networks with Learnable Structural and Positional Representations},
+   url = {http://arxiv.org/abs/2110.07875},
+   year = {2021},
+}
+
 @article{Battaglia2018,
    abstract = {Artificial intelligence (AI) has undergone a renaissance recently, making major progress in key domains such as vision, language, control, and decision-making. This has been due, in part, to cheap data and cheap compute resources, which have fit the natural strengths of deep learning. However, many defining characteristics of human intelligence, which developed under much different pressures, remain out of reach for current approaches. In particular, generalizing beyond one's experiences--a hallmark of human intelligence from infancy--remains a formidable challenge for modern AI. The following is part position paper, part review, and part unification. We argue that combinatorial generalization must be a top priority for AI to achieve human-like abilities, and that structured representations and computations are key to realizing this objective. Just as biology uses nature and nurture cooperatively, we reject the false choice between "hand-engineering" and "end-to-end" learning, and instead advocate for an approach which benefits from their complementary strengths. We explore how using relational inductive biases within deep learning architectures can facilitate learning about entities, relations, and rules for composing them. We present a new building block for the AI toolkit with a strong relational inductive bias--the graph network--which generalizes and extends various approaches for neural networks that operate on graphs, and provides a straightforward interface for manipulating structured knowledge and producing structured behaviors. We discuss how graph networks can support relational reasoning and combinatorial generalization, laying the foundation for more sophisticated, interpretable, and flexible patterns of reasoning. As a companion to this paper, we have released an open-source software library for building graph networks, with demonstrations of how to use them in practice.},
    author = {Peter W. Battaglia and Jessica B. Hamrick and Victor Bapst and Alvaro Sanchez-Gonzalez and Vinicius Zambaldi and Mateusz Malinowski and Andrea Tacchetti and David Raposo and Adam Santoro and Ryan Faulkner and Caglar Gulcehre and Francis Song and Andrew Ballard and Justin Gilmer and George Dahl and Ashish Vaswani and Kelsey Allen and Charles Nash and Victoria Langston and Chris Dyer and Nicolas Heess and Daan Wierstra and Pushmeet Kohli and Matt Botvinick and Oriol Vinyals and Yujia Li and Razvan Pascanu},

docs/make.jl

@@ -42,8 +42,10 @@ makedocs(
              "Dynamic Graph Update" => "dynamicgraph.md",
              "Manual" => [
                "FeaturedGraph" => "manual/featuredgraph.md",
-               "Graph Convolutional Layers" => "manual/conv.md",
+               "Graph Convolutional Layers" => "manual/graph_conv.md",
                "Graph Pooling Layers" => "manual/pool.md",
+               "Group Convolutional Layers" => "manual/group_conv.md",
+               "Positional Encoding Layers" => "manual/positional.md",
                "Embeddings" => "manual/embedding.md",
                "Models" => "manual/models.md",
                "Linear Algebra" => "manual/linalg.md",

docs/src/manual/featuredgraph.md

@@ -9,6 +9,8 @@ GraphSignals.edge_feature
 GraphSignals.has_edge_feature
 GraphSignals.global_feature
 GraphSignals.has_global_feature
+GraphSignals.positional_feature
+GraphSignals.has_positional_feature
 GraphSignals.subgraph
 GraphSignals.ConcreteFeaturedGraph
 ```

docs/src/manual/conv.md renamed to docs/src/manual/graph_conv.md

@@ -0,0 +1,19 @@
+# Group Convolutional Layers
+
+## ``E(n)``-equivariant Convolutional Layer
+
+It employs message-passing scheme and can be defined by following functions:
+
+- message function (Eq. 3 from paper): ``m_{ij} = \phi_e(h_i^l, h_j^l, ||x_i^l - x_j^l||^2, a_{ij})``
+- aggregate (Eq. 5 from paper): ``m_i = \sum_j m_{ij}``
+- update function (Eq. 6 from paper): ``h_i^{l+1} = \phi_h(h_i^l, m_i)``
+
+where ``h_i^l`` and ``h_j^l`` denotes the node feature for node ``i`` and ``j``, respectively, in ``l``-th layer, as well as ``x_i^l`` and ``x_j^l`` denote the positional feature for node ``i`` and ``j``, respectively, in ``l``-th layer. ``a_{ij}`` is the edge feature for edge ``(i,j)``. ``\phi_e`` and ``\phi_h`` are neural network for edges and nodes.
+
+```@docs
+EEquivGraphConv
+```
+
+Reference: [Satorras2021](@cite)
+
+---

docs/src/manual/group_conv.md

@@ -0,0 +1,43 @@
+# Positional Encoding
+
+## Positional Encoding Methods
+
+```@docs
+AbstractPositionalEncoding
+RandomWalkPE
+LaplacianPE
+positional_encode
+```
+
+## Positional Encoding Layers
+
+### ``E(n)``-equivariant Positional Encoding Layer
+
+It employs message-passing scheme and can be defined by following functions:
+
+- message function: ``y_{ij}^l = (x_i^l - x_j^l)\phi_x(m_{ij})``
+- aggregate: ``y_i^l = \frac{1}{M} \sum_{j \in \mathcal{N}(i)} y_{ij}^l``
+- update function: ``x_i^{l+1} = x_i^l + y_i^l``
+
+where ``x_i^l`` and ``x_j^l`` denote the positional feature for node ``i`` and ``j``, respectively, in ``l``-th layer, ``\phi_x`` is the neural network for positional encoding and ``m_{ij}`` is the edge feature for edge ``(i,j)``. ``y_{ij}^l`` and ``y_i^l`` represent the encoded positional feature and aggregated positional feature, respectively, and ``M`` denotes number of neighbors of node ``i``.
+
+```@docs
+EEquivGraphPE
+```
+
+Reference: [Satorras2021](@cite)
+
+---
+
+### Learnable Structural Positional Encoding layer
+
+(WIP)
+
+```@docs
+LSPE
+```
+
+Reference: [Dwivedi2021](@cite)
+
+---
+

docs/src/manual/positional.md

@@ -30,7 +30,15 @@ export
     # layers/msgpass
     MessagePassing,
 
-    # layers/conv
+    # layers/positional
+    AbstractPositionalEncoding,
+    RandomWalkPE,
+    LaplacianPE,
+    positional_encode,
+    EEquivGraphPE,
+    LSPE,
+
+    # layers/graph_conv
     GCNConv,
     ChebConv,
     GraphConv,
@@ -44,6 +52,9 @@ export
     MeanAggregator, MeanPoolAggregator, MaxPoolAggregator,
     LSTMAggregator,
 
+    # layers/group_conv
+    EEquivGraphConv,
+
     # layer/pool
     GlobalPool,
     LocalPool,
@@ -71,7 +82,9 @@ include("layers/graphlayers.jl")
 include("layers/gn.jl")
 include("layers/msgpass.jl")
 
-include("layers/conv.jl")
+include("layers/positional.jl")
+include("layers/graph_conv.jl")
+include("layers/group_conv.jl")
 include("layers/pool.jl")
 include("models.jl")
 

src/GeometricFlux.jl

@@ -165,7 +165,7 @@ aggregate_vertices(::GraphNet, aggr, V) = aggregate(aggr, V)
 @inline aggregate_vertices(::GraphNet, ::Nothing, V) = nothing
 
 function propagate(gn::GraphNet, sg::SparseGraph, E, V, u, naggr, eaggr, vaggr)
-    el = to_namedtuple(sg)
+    el = GraphSignals.to_namedtuple(sg)
     return propagate(gn, el, E, V, u, naggr, eaggr, vaggr)
 end
 
@@ -179,14 +179,11 @@ function propagate(gn::GraphNet, el::NamedTuple, E, V, u, naggr, eaggr, vaggr)
     return E, V, u
 end
 
-WithGraph(fg::AbstractFeaturedGraph, gn::GraphNet) = WithGraph(to_namedtuple(fg), gn)
-WithGraph(gn::GraphNet; dynamic=nothing) = WithGraph(DynamicGraph(dynamic), gn)
+WithGraph(fg::AbstractFeaturedGraph, gn::GraphNet) =
+    WithGraph(GraphSignals.to_namedtuple(fg), gn, positional_feature(fg))
 
-to_namedtuple(fg::AbstractFeaturedGraph) = to_namedtuple(graph(fg))
+WithGraph(gn::GraphNet; dynamic=nothing) =
+    WithGraph(DynamicGraph(dynamic), gn, GraphSignals.NullDomain())
 
-function to_namedtuple(sg::SparseGraph)
-    es, nbrs, xs = collect(edges(sg))
-    return (N=nv(sg), E=ne(sg), es=es, nbrs=nbrs, xs=xs)
-end
-
-@non_differentiable to_namedtuple(x...)
+WithGraph(fg::AbstractFeaturedGraph, gn::GraphNet, pos::GraphSignals.AbstractGraphDomain) =
+    WithGraph(GraphSignals.to_namedtuple(fg), gn, pos)