Calculate experimental directed UMAP memberships from disk-backed KNN

Scans validated full-graph distances twice in bounded native buffers. The resulting .weights.f32 file shares the existing index file. These are directed memberships, not the symmetrized fuzzy UMAP graph, and are not yet accepted by full-data UMAP optimization.

Usage

massive_umap_memberships(
  x,
  checkpoint = FALSE,
  resume = FALSE,
  checkpoint_every = 100L,
  backend = "cpu",
  n.cores = 1L
)

Arguments

x

A distance-valued result of massive_knn_graph() or massive_full_knn_graph().

checkpoint

Save completed mean-scan and weight blocks for resume.

resume

Resume an interrupted checkpointed run.

checkpoint_every

Save progress after this many native blocks.

backend

This graph transformation currently runs on CPU only.

n.cores

CPU workers for independent per-row weights.

Value

A file-backed directed weighted-graph descriptor.