Jraph

The GraphsTuple data structure supports nested features for nodes, edges, and globals, allowing flexible representation of complex graph data without heavy overhead, as shown in the quick start examples.

Built on JAX, it enables automatic differentiation and JIT compilation for high-performance training on GPUs/TPUs, with utilities for batching variable-sized graphs via padding and masking.

Provides reference implementations like GraphNet that users can easily adapt and extend, serving as practical starting points for custom GNN development, as detailed in the model zoo section.

Allows storing multiple feature sets per graph element with different types and meanings, enabling complex data handling without rigid structural constraints.