Inspect the e-means training recipe — emeans

Resolves batch size and initial step size from the fitted scaling laws of the e-means paper, exactly as emeans() does.

Usage

emeans_recipe(
  n,
  k,
  d,
  nepochs = 5,
  batch_law = c("auto", "throughput", "quality"),
  batch_size = NULL,
  alpha0 = NULL,
  unused_threshold = NULL
)

Arguments

n: Number of training rows.
k: Number of clusters.
d: Dimension (selects the "auto" batch law and revival default).
nepochs: Epoch budget; fractions allowed.
batch_law: "auto" (quality law when d <= 16, else throughput), "throughput", or "quality" (batch about 8 times smaller).
batch_size, alpha0: Optional overrides. Giving batch_size alone keeps the step law for alpha0.
unused_threshold: Revival threshold tau: clusters with mass below tau * mean(mass) are revived. NULL gives 0.1, or 0 (off) when d <= 2.

Value

A list with batch_size, alpha0, epoch_rows, epoch_steps, total_steps, batch_law, and unused_threshold.

Examples

emeans_recipe(n = 1e6, k = 1024, d = 128, nepochs = 3)
#> $batch_size
#> [1] 32768
#> 
#> $alpha0
#> [1] 0.6955212
#> 
#> $epoch_rows
#> [1] 1e+06 1e+06 1e+06
#> 
#> $epoch_steps
#> [1] 31 31 31
#> 
#> $total_steps
#> [1] 93
#> 
#> $batch_law
#> [1] "throughput"
#> 
#> $unused_threshold
#> [1] 0.1
#> 
#> $nepochs
#> [1] 3
#>