(WIP) K3 Node: Multi-Backend Graph Neural Networks

Ported from Spektral, PyTorch Geometric and StellarGraph

Important

K3 Node only supports operations on sparse adjacency matrix with TensorFlow backend. Support for JAX or PyTorch backends will be added future releases.

What's Included:

Convolution Layers

AGNNConv: An Attention-based Graph Neural Network (AGNN)
APPNPConv PPNPPropagation: (Approximate) Personalized Propagation of Neural Predictions (PPNP/APPNP)
ARMAConv: An Auto-Regressive Moving Average convolutional layer (ARMA) from the paper
CrystalConv: Crystal Graph Convolutional Neural
GatedGraphConv: Gated Graph Sequence Neural Networks
GraphConvolution: Graph Convolutional Network (GCN)
GeneralConv: A general convolutional layer from the paper: Design Space for Graph Neural Networks
GINConv: A Graph Isomorphism Network (GIN) from the paper: How Powerful are Graph Neural Networks?
GraphAttention: Graph ATtention Network (GAT)
SAGEConv: A GraphSAGE layer from the paper: Inductive Representation Learning on Large Graphs

Aggregation Layers

SumAggregation
MaxAggregation
MeanAggregation
SoftmaxAggregation
PowerMeanAggregation

Normalization Layers

MeanSubtractionNorm
MessageNorm

Attention Layers:

Performer: The linear scaled attention mechanism from the "Rethinking Attention with Performers"

Examples

TensorFlow Backend

Model based on ARMAConv using features from spektral

PyTorch Backend

Model based on GatedGraphConv using features from torch_geometric

Contributions

Help us improve K3Node! Report issues, submit patches, or share your suggestions.

anas-rz / k3-node