K3Node: Multi-Backend Graph Neural Networks

High-performance GNN library achieving 100% public API parity with PyTorch Geometric across TensorFlow, PyTorch, and JAX with 700+ tests per backend.

K3Node: Multi-Backend Graph Neural Networks for Keras 3

Stack: Keras 3, PyTorch, TensorFlow, JAX, Python
Code: github.com/anas-rz/k3-node


Overview

K3Node is a next-generation Graph Neural Network (GNN) library designed to achieve 100% public API parity with PyTorch Geometric (PyG), fully executable across TensorFlow, PyTorch, and JAX with native GPU and TPU acceleration through Keras 3.


Architectural Highlights

  • 65+ Convolution Layers: Implemented core and advanced message-passing operators including GCNConv, GATv2Conv, TransformerConv, GPSConv, PNAConv, SchNet, DimeNetPlusPlus, and ViSNet.
  • 26 Aggregation & 31 Pooling Operators: Advanced graph pooling schemes including hierarchical coarsening (TopK, SAG, ASAP, EdgePooling) and 3D spatial voxel/k-NN pooling, paired with flexible aggregations (SetTransformer, GraphMultisetTransformer, Set2Set, DeepSets, LSTMAggregation).
  • 5 Pre-Trained Foundation Models: Complete weight-converter pipelines for pre-trained foundation models:
    • GraphMAE2 (masked graph autoencoder)
    • Graphormer (2D & 3D molecular transformer)
    • GraphGPS (general powerful spatial/spectral architecture)
    • GROVER & Mole-BERT (molecular representation models)
  • Knowledge Graph Embeddings: Implemented 5 KG embedding models (TransE, RotatE, DistMult, ComplEx, KGTripletLoader), 46 high-level models, and 26 mini-batch samplers (NeighborLoader, ClusterLoader, GraphSAINTSampler).
  • Rigorous Parity Testing: Engineered 700+ unit tests per backend, checking gradient flow, weight learning, and numerical parity against PyG reference checkpoints; authored 90+ Colab example notebooks.