qbiocode.apps.quvine.api.sgns module

SGNS embedding core: views -> walks -> corpus -> word2vec.

These functions are extracted verbatim from Pipeline._run_single_iteration and its private helpers so that both the Hydra pipeline and the quvine.embed() API share a single, behavior-preserving code path.

Reproducibility invariants that MUST be preserved (the cluster pipeline depends on them):

• Per-root RNG seed: cfg.experiment.base_seed + 10000 * it + idx where idx indexes sorted(roots).

• Embedding row order is list(graph.nodes) (never sorted).

• train_embeddings reads the top-level cfg.min_count.

• joblib Parallel(backend='loky', batch_size=1, prefer='processes').

Note on the kinds argument: when a subset of walk kinds is requested, only those walkers run, which changes the per-root RNG stream relative to a full multi-kind run. Calling run_sgns with kinds=list(cfg.walks.kinds) (as the pipeline does) reproduces the original behavior exactly.

Summary

Functions:

build_corpora	Run walks over every root and compile a per-walk-kind token corpus.
build_views	Build constrained views for a single root.
chunkify	Yield successive chunk_size-sized chunks of seq.
process_root	Build views + walks for a single root with deterministic seeding.
process_root_chunk	Process a batch of roots inside a single worker process.
run_sgns	One-shot SGNS embedding: build_corpora -> train_embeddings.
run_walks_for_root	Run every configured walk kind over each view for a single root.
train_embeddings	Train one SGNS (word2vec) embedding per walk kind.

Reference

chunkify(seq, chunk_size)

Yield successive chunk_size-sized chunks of seq.

build_views(cfg, graph, root, rng)

Build constrained views for a single root.

run_walks_for_root(cfg, graph, root, views, rng)

Run every configured walk kind over each view for a single root.

process_root(cfg, graph, root, node2idx, it)

Build views + walks for a single root with deterministic seeding.

process_root_chunk(cfg, graph, roots, node2idx, it)

Process a batch of roots inside a single worker process.

build_corpora(cfg, graph, it=0, *, n_jobs=1, chunk_size=30)

Run walks over every root and compile a per-walk-kind token corpus.

Returns a dict {walk_kind: corpus} where each corpus is a flat list of walks (each walk a list of node-id strings).

Return type:

Dict[str, List[List[str]]]

train_embeddings(cfg, graph, all_corpora)

Train one SGNS (word2vec) embedding per walk kind. Rows in node order.

run_sgns(cfg, graph, it=0, *, kinds=None, n_jobs=1, chunk_size=30)

One-shot SGNS embedding: build_corpora -> train_embeddings.

Parameters:

• cfg – OmegaConf config (reads walks.*, views.*, train.*, min_count, experiment.base_seed).

• graph (Graph) – NetworkX graph.

• it (int) – Iteration index used in the per-root seed.

• kinds (Optional[List[str]]) – Walk kinds to compute. None uses cfg.walks.kinds.

• n_jobs (int) – joblib parallelism over roots.

• chunk_size (int) – joblib parallelism over roots.

Return type:

Dict[str, ndarray]

Returns:

{walk_kind: embedding} with each embedding (n_nodes, dim) in list(graph.nodes) order.