torchcell.models.hetero_cell_bipartite_dango_gi_lazy.PairwiseGraphAggregation

class PairwiseGraphAggregation(hidden_dim: int, graph_names: list[str], dropout: float = 0.0, activation: str = 'relu', num_layers: int = 2, bottleneck_dim: int | None = None, norm: str | None = None)

Bases: Module

Pairwise interaction mechanism for aggregating multiple graph representations.

forward(graph_outputs: dict[str, Tensor]) → tuple[Tensor | None, Tensor | None]

Compute pairwise graph interactions and aggregate them.

Includes identity option (mean of individual graphs) alongside pairwise combinations.

Returns:

Aggregated features [num_nodes, hidden_dim] pair_weights: Learned importance weights for pairwise interactions + identity [num_nodes, num_interactions+1]

Return type:

aggregated