mixle.represent.graph module

Structured / graph encoding – message passing over a structure’s elements, so edges actually matter.

A molecule, a knowledge graph, a fossil section’s co-occurrence network, a protein contact map: these are structures, not bags. SetSegmenter gives their elements’ features but throws the edges away. GraphEmbedding keeps them – a small message-passing net (GCN-style) embeds each element and then mixes it with its neighbours over layers rounds, so an atom’s vector reflects its bonds and a node’s reflects its graph context. The output is one (n_elements, dim) block in the same shared space as every other modality, so a molecule composes with text/images/signals in one HeterogeneousEncoder.

A graph-modality record value is (node_features (n, f), adjacency (n, n)); GraphEncoder is a ModalityEncoder you drop into the registry via register_encoder. Trains end to end like any encoder.

class GraphEmbedding(in_features, dim, *, layers=2, name=None)

Bases: object

A message-passing embedding: (node_features, adjacency) -> (n_nodes, dim) with layers GCN rounds.

Parameters:

• in_features (int)

• dim (int)

• layers (int)

• name (str | None)

module()

Return type:

Any

class GraphEncoder(embedding)

Bases: ModalityEncoder

A structure modality: raw = (node_features (n, f), adjacency (n, n)) -> per-node vectors in shared R^dim.

Parameters:

embedding (GraphEmbedding)

encode(raw)

Torch tensor (n_units, dim) – gradients flow to the embedding (and, via it, train the encoder).

Parameters:

raw (Any)

Return type:

Any

encode_numpy(raw)

Detached (n_units, dim) array – for eval, quantization, or feeding a non-torch model.

Parameters:

raw (Any)

Return type:

ndarray