GNN Core

class GNN(*args, **kwargs)[source]

Bases: Module

Simple Graph Neural Network (GNN) model designed to handle Hyper Heterogeneous Multi Graphs (H2MGs).

The model consists of a normalization step, an encoding step, a coupling step, and a decoding step. The decoder can either be invariant or equivariant, depending on the task requirements.

Parameters:

  • normalizer (Normalizer) – Maps the input features to a learning-compatible range.

  • encoder (Encoder) – Embeds hyper-edge set features into a latent space.

  • coupler (Coupler) – Outputs latent coordinates for each address present in the input graph.

  • decoder (Decoder) – Maps latent coordinates and encoded graph to a meaningful output.

Return type:

Any

GNN.__call__

Processes a given graph through a sequence of steps: normalization, encoding, coupling, and decoding.

GNN.forward_batch

Applies the model to a batch of graphs.