## ----setup, include = FALSE--------------------------------------------------- knitr::opts_chunk$set(collapse = TRUE, comment = "#>") library(functionals) set.seed(1) has_rsample <- requireNamespace("rsample", quietly = TRUE) ## ----------------------------------------------------------------------------- by_species <- fmapg(iris, function(d) { fit <- lm(Sepal.Length ~ Petal.Length, data = d) c(slope = unname(coef(fit)[2]), r2 = summary(fit)$r.squared) }, by = "Species") do.call(rbind, by_species) ## ----------------------------------------------------------------------------- specs <- data.frame( response = "mpg", predictor = c("wt", "hp", "wt + hp", "wt * hp"), stringsAsFactors = FALSE ) aic_tbl <- fmapr(specs, function(row) { f <- reformulate(row$predictor, row$response) fit <- lm(f, data = mtcars) data.frame(model = deparse(f), aic = AIC(fit), adj_r2 = summary(fit)$adj.r.squared) }) do.call(rbind, aic_tbl) ## ----------------------------------------------------------------------------- grid <- expand.grid(df = c(2, 3, 4, 5), span = NA) fits <- fmapn(list(df = grid$df), function(df) { fit <- smooth.spline(mtcars$wt, mtcars$mpg, df = df) c(df = df, gcv = fit$cv.crit) }) do.call(rbind, fits) ## ----eval = has_rsample------------------------------------------------------- set.seed(42) folds <- rsample::vfold_cv(mtcars, v = 5) cv_rows <- fcv(folds$splits, function(split) { tr <- rsample::analysis(split) te <- rsample::assessment(split) fit <- lm(mpg ~ wt + hp, data = tr) pred <- predict(fit, newdata = te) data.frame(rmse = sqrt(mean((te$mpg - pred)^2))) }) cv <- do.call(rbind, cv_rows) c(mean_rmse = mean(cv$rmse), sd_rmse = sd(cv$rmse)) ## ----eval = FALSE------------------------------------------------------------- # cv_rows <- fcv(folds$splits, evaluate_one, ncores = 5, pb = TRUE) ## ----eval = has_rsample------------------------------------------------------- safe_rows <- fcv(folds$splits, function(split) { tr <- rsample::analysis(split) if (nrow(tr) < 3) stop("degenerate fold") data.frame(rmse = sd(rsample::assessment(split)$mpg)) }, .on_error = "pass") ok <- !vapply(safe_rows, inherits, logical(1), "functionals_error") do.call(rbind, safe_rows[ok]) ## ----------------------------------------------------------------------------- boot_coef <- frepeat(times = 200, .seed = 2024, simplify = TRUE, expr = function() { idx <- sample(nrow(mtcars), replace = TRUE) coef(lm(mpg ~ wt, data = mtcars[idx, ]))[["wt"]] }) c( estimate = coef(lm(mpg ~ wt, data = mtcars))[["wt"]], boot_se = sd(boot_coef), ci_lo = unname(quantile(boot_coef, 0.025)), ci_hi = unname(quantile(boot_coef, 0.975)) ) ## ----------------------------------------------------------------------------- # same result, more cores; .seed is invariant to ncores b2 <- frepeat(times = 200, .seed = 2024, simplify = TRUE, ncores = 2, expr = function() { idx <- sample(nrow(mtcars), replace = TRUE) coef(lm(mpg ~ wt, data = mtcars[idx, ]))[["wt"]] }) identical(round(sd(boot_coef), 10), round(sd(b2), 10)) ## ----eval = has_rsample------------------------------------------------------- learners <- list( lm_wt = function(tr) lm(mpg ~ wt, data = tr), lm_wt_hp = function(tr) lm(mpg ~ wt + hp, data = tr), lm_full = function(tr) lm(mpg ~ ., data = tr) ) cv_rmse <- function(fit_fun) { rows <- fcv(folds$splits, function(split) { tr <- rsample::analysis(split); te <- rsample::assessment(split) p <- predict(fit_fun(tr), newdata = te) sqrt(mean((te$mpg - p)^2)) }) mean(unlist(rows)) } bench_tbl <- fmap(learners, cv_rmse, .on_error = "fill", .fill = NA_real_) data.frame(learner = names(bench_tbl), cv_rmse = round(unlist(bench_tbl), 3)) .