Skip to contents

Optimize an experimental disk-backed UMAP graph

Source: R/massive_umap_optimizer.R
massive_umap_optimize.Rd

Streams a symmetric fuzzy CSR graph for each epoch. CPU coordinates use float32 RAM or an explicitly requested writable memory-mapped file. CUDA streams bounded edge batches. Its default layout must fit VRAM; explicit managed storage allows CUDA paging from host RAM. Sampling times are derived from edge weights instead of storing a per-edge schedule. CPU supports file-backed 2D or 3D initialization; Completed CPU or CUDA epochs can be checkpointed and resumed without rescanning earlier epochs.

Usage

massive_umap_optimize(
  graph,
  init,
  output,
  n_epochs = 200L,
  negative_sample_rate = 5L,
  learning_rate = 1,
  min_dist = 0.1,
  repulsion_strength = 1,
  seed = 1L,
  chunk_rows = 8192L,
  memory_limit = "16GB",
  backend = "cpu",
  n.cores = 1L,
  checkpoint = FALSE,
  resume = FALSE,
  checkpoint_every = 10L,
  layout_storage = "memory"
)

Arguments

graph

A result of massive_umap_fuzzy_graph().

init

A file-backed massive_matrix() of initial coordinates.

output

New .f32 file for the resulting coordinates.

n_epochs

Number of optimization epochs, at least two because the first epoch initializes the edge-sampling schedule.

negative_sample_rate

Negative samples per positive edge.

learning_rate

Initial learning rate.

min_dist

UMAP minimum-distance curve parameter.

repulsion_strength

Repulsive-force multiplier.

seed

Random seed for negative sampling.

chunk_rows

Maximum graph rows per sequential read.

memory_limit

Host working-buffer budget. The full layout counts in CPU RAM mode; OS page cache for a mapped layout is not capped by this value. CUDA preflights device allocations against free device memory; managed CUDA coordinates count against the host budget.

backend

"cpu" or "cuda"; CUDA has no CPU fallback.

n.cores

Only one worker is currently supported.

checkpoint

Save a full coordinate snapshot at completed epochs.

resume

Continue a matching interrupted fit on the same backend.

checkpoint_every

Number of epochs between snapshots.

layout_storage

"memory" (default), "mmap" for CPU writable mapping, or "managed" for explicit CUDA unified memory. Managed storage may page between host and GPU and requires a supporting GPU.

Value

An experimental file-backed fastEmbedR_massive_matrix.