Skip to contents

Optimize experimental disk-backed t-SNE affinities

Source: R/massive_tsne_optimizer.R
massive_tsne_optimize.Rd

Streams a normalized symmetric CSR affinity graph on every iteration. The CPU optimizer retains only its float32 coordinates and update state in RAM, reuses the package's 2D or 3D FFT repulsion, and writes the final coordinates to a file. The CUDA route streams the same affinities through its 2D FFT optimizer and retains coordinates on the GPU. This is a full-graph fit, not landmark projection. Optimization uses fixed iteration counts, without automatic KL stopping. Disk reads can be substantial. Both routes can checkpoint optimizer state.

Usage

massive_tsne_optimize(
  graph,
  init,
  output,
  early_exaggeration_iter = 250L,
  n_iter = 500L,
  early_exaggeration = 12,
  exaggeration = 1,
  learning_rate = "auto",
  initial_momentum = 0.8,
  final_momentum = 0.8,
  min_gain = 0.01,
  max_step_norm = Inf,
  memory_limit = "8GB",
  backend = "cpu",
  n.cores = 1L,
  checkpoint = FALSE,
  resume = FALSE,
  checkpoint_every = 100L
)

Arguments

graph

A result of massive_tsne_affinities().

init

File-backed .f32 initial coordinates with two or three columns.

output

New .f32 output path.

early_exaggeration_iter, n_iter

Iterations in each phase.

early_exaggeration, exaggeration

Phase affinity multipliers.

learning_rate

Positive value or "auto" for n / exaggeration.

initial_momentum, final_momentum

Phase momentum values.

min_gain

Minimum per-coordinate adaptive gain.

max_step_norm

Maximum update norm; Inf disables clipping.

memory_limit

RAM budget, including coordinates, optimizer state, and a conservative FFT workspace allowance.

backend

"cpu" or "cuda"; CUDA currently requires 2D output.

n.cores

CPU worker count for FFT and coordinate updates.

checkpoint

Save coordinates and optimizer state periodically.

resume

Continue a matching interrupted fit.

checkpoint_every

Iterations between state snapshots.

Value

An experimental file-backed fastEmbedR_massive_matrix. CUDA results include host-call stage times. attraction_sync can include previously queued FFT work.