parsnip_speedup.R

# benchmarking the new parsnip release 1.0.3 vs previous 1.0.2
#
# forked from:  @simonpcouch
# https://gist.github.com/simonpcouch/651d0ea4d968b455ded8194578dabf52
# gist name:simonpcouch/parsnip_speedup.R
# 2022-11-22
#
library(tidymodels)

# with v1.0.2 ------------------------------------------------------------
pak::pkg_install("tidymodels/parsnip@v1.0.2")

num_samples <- 10^(3:7)
num_resamples <- c(5, 10, 20)
nrow <- length(num_samples) * length(num_resamples)

res_v102 <-
  tibble(
    version = character(nrow),
    num_samples = numeric(nrow),
    num_resamples = numeric(nrow),
    time_to_fit = numeric(nrow)
  )

set.seed(1)

for (i in seq_along(num_samples)) {
  dat <- tibble(x = rnorm(num_samples[i]), y = x + rnorm(num_samples[i], sd = .2))

  for (j in seq_along(num_resamples)) {
    folds <- vfold_cv(dat, v = num_resamples[j])

    timing <-
      system.time({
        fit_resamples(linear_reg(), y ~ x, folds)
      })

    idx <- (length(num_samples) * (j - 1)) + i
    res_v102[idx,] <- list("v1.0.2", num_samples[i], num_resamples[j], timing[["elapsed"]])
  }
}

# with v1.0.3 ------------------------------------------------------------
pak::pkg_install("tidymodels/parsnip@v1.0.3")

res_v103 <-
  tibble(
    version = character(nrow),
    num_samples = numeric(nrow),
    num_resamples = numeric(nrow),
    time_to_fit = numeric(nrow)
  )

set.seed(1)

for (i in seq_along(num_samples)) {
  dat <- tibble(x = rnorm(num_samples[i]), y = x + rnorm(num_samples[i], sd = .2))

  for (j in seq_along(num_resamples)) {
    folds <- vfold_cv(dat, v = num_resamples[j])

    timing <-
      system.time({
        fit_resamples(linear_reg(), y ~ x, folds)
      })

    idx <- (length(num_samples) * (j - 1)) + i
    res_v103[idx,] <- list("v1.0.3", num_samples[i], num_resamples[j], timing[["elapsed"]])
  }
}

# plotting ---------------------------------------------------------------
res <- bind_rows(res_v102, res_v103)

res_plot <-
  res %>%
  pivot_wider(
    id_cols = c(num_samples, num_resamples),
    names_from = version,
    values_from = time_to_fit
  ) %>%
  mutate(speedup = `v1.0.2` / `v1.0.3`) %>%
  select(-starts_with("v")) %>%
  mutate(num_resamples = factor(num_resamples, levels = c("20", "10", "5"), ordered = TRUE)) %>%
  ggplot() +
  aes(x = num_samples, y = speedup, col = num_resamples) +
  geom_line() +
  scale_x_log10() +
  labs(
    x = "Number of Rows in Training Data",
    y = "Speedup (v1.0.2 / v1.0.3)",
    col = "Number \nof Folds",
    title = "The 'parsnip' R Package got a bit faster",
    subtitle = "The new release of parsnip contributed speedup for model fitting."
  ) +
  scale_color_viridis_d(end = .8) +
  theme(plot.subtitle = element_text(face = "italic"))

res_plot
ggsave("parsnip-102vs103-speedup.svg", res_plot, width = 6, height = 4, dpi = 400)
ggsave("parsnip-102vs103-speedup.png", res_plot, width = 6, height = 4, dpi = 400)

This is the speedup I got on my machine. R 4.2.2, i686 Haswell i4770K from 2012, 16 GB RAM,