Construct experimental file-backed t-SNE affinities — massive_tsne

Streams every row of a validated full KNN graph, computes the same conditional probabilities as native float32 t-SNE, externally sorts directed pairs, and writes symmetric normalized CSR affinities. The complete graph is never materialized in memory. A content-hashed manifest permits massive_affinity_graph() to reload the result. The saved graph can feed the experimental CPU massive_tsne_optimize() route.

Usage

massive_tsne_affinities(
  x,
  perplexity,
  output,
  memory_limit = "8GB",
  backend = "cpu",
  n.cores = 1L,
  checkpoint = FALSE,
  resume = FALSE,
  checkpoint_every = 100L
)

Arguments

x: A distance-valued massive_knn_graph() descriptor.
perplexity: Positive perplexity no larger than the KNN width.
output: New CSR output prefix.
memory_limit: RAM budget for external sorting, at least 64 MB.
backend: Only "cpu" is supported for graph construction.
n.cores: CPU workers for independent conditional-probability rows.
checkpoint: Save completed sorted-pair runs for resume.
resume: Continue a matching interrupted graph sort.
checkpoint_every: Save progress after this many input blocks.

Value

An experimental symmetric fastEmbedR_massive_graph descriptor.