PyTorch Geometric Adapter

PyG conversion helpers from annnet.adapters.pyg_adapter.

annnet.adapters.pyg_adapter

AnnNet-PyTorch Geometric adapter for AnnNet.

Provides: to_pyg(G) -> torch_geometric.data.HeteroData

PyTorch Geometric represents graph data as tensors. This adapter exports AnnNet vertices and edges into a heterogeneous graph structure suitable for downstream GNN workflows.

AnnNet-specific structures such as slices, multilayer metadata, hyperedge semantics, and rich attribute tables are only exported where they can be mapped to tensor-compatible node, edge, or graph-level fields.

Classes

Functions

to_pyg

to_pyg(
    graph,
    node_features=None,
    edge_features=None,
    slice_id=None,
    hyperedge_mode="reify",
    device="cpu",
)

Export AnnNet → torch_geometric.data.HeteroData.

Builds a heterogeneous graph: vertices are grouped into node types by their kind attribute and edges into relation types by their endpoint kinds, so AnnNet's entity/edge typing is preserved. Selected vertex/edge attribute columns become node/edge feature tensors, and slice membership is carried as boolean masks. Hyperedges have no native PyG equivalent and are handled per hyperedge_mode.

Parameters:

Name	Type	Description	Default
`graph`	`AnnNet`		required
`node_features`	`dict[str, list[str]]`	Per node-kind, the vertex-attribute columns to stack into the node feature tensor `x`. Kinds absent from the mapping get no features.	`None`
`edge_features`	`dict[tuple[str, str, str], list[str]]`	Per relation triple `(src_kind, relation, dst_kind)`, the edge-attribute columns to stack into `edge_attr`.	`None`
`slice_id`	`str`	Slice to export; defaults to the graph's active slice (`graph.slices.active`).	`None`
`hyperedge_mode`	`('reify', 'expand', 'skip')`	`"reify"`: add a node per hyperedge plus membership edges carrying per-endpoint roles/coefficients. `"expand"`: replace each hyperedge with binary edges over its endpoints. `"skip"`: drop hyperedges (kept only in the manifest).	`"reify"`
`device`	`str`	Torch device the returned tensors are allocated on.	`"cpu"`

Returns:

Type	Description
`HeteroData`	Heterogeneous graph with per-type `x` / `edge_index` / `edge_attr` and slice masks.