Skip to contents

Estimate Runtime Before a Long Tuning or Scoring Run

Source: R/survival_metrics.R
estimate_runtime.Rd

Times a small sample of model fits (`n_sample`) and projects the total wall-clock time for the full grid and per machine.

Usage

estimate_runtime(
  fn,
  args,
  n_sample = 3,
  workers = NULL,
  n_machines = NULL,
  n_outer_folds = 1L,
  n_domains = 1L
)

Arguments

fn

Function to benchmark, such as [tune_alpha_by_split], [score_outer], [tune_over_alpha], or [glmnet_IBS] (or its character name).

args

Named list of arguments to pass to `fn`.

n_sample

Number of sample fits to time (default `3`).

workers

Number of parallel workers per machine. Defaults to [future::nbrOfWorkers()] when future is installed, or `1L`.

n_machines

Number of machines used for sharding (default `1L`, or inferred from `args$shard[2]` when present).

n_outer_folds

Optional multiplier for outer CV folds when `args` represents a single outer fold (default `1L`).

n_domains

Optional multiplier for domains when `args` represents a single domain (default `1L`).

Value

A tibble of class `cbe_runtime_estimate` with per-fit timing statistics (`mean_sec_per_fit`, `median_sec_per_fit`, `sd_sec_per_fit`), grid size (`n_tasks`, `n_total_fits`), and projected wall-clock times (`projected_sequential_sec`, `projected_total_sec`, `projected_per_machine_sec`, `projected_total_mins`, `projected_per_machine_mins`).

See also

[tune_alpha_by_split()], [score_outer()]