Merge pull request #273 from DavisVaughan/last-fit-randomness

Construct a "manual" rset for usage in `last_fit()`

DESCRIPTION

@@ -25,7 +25,7 @@ Imports:
     cli (>= 2.0.0),
     crayon,
     yardstick,
-    rsample,
+    rsample (>= 0.0.7.9000),
     tidyr,
     GPfit,
     foreach,
@@ -48,3 +48,5 @@ LazyData: true
 Roxygen: list(markdown = TRUE)
 RoxygenNote: 7.1.1
 Language: en-US
+Remotes:
+    tidymodels/rsample

NEWS.md

@@ -2,6 +2,8 @@
 
 ## Bug Fixes
 
+* `last_fit()` no longer accidentally adjusts the random seed (#264).
+
 * Fixed two bugs in the acquisition function calculations.
 
 # tune 0.1.1

R/last_fit.R

@@ -122,29 +122,22 @@ last_fit.workflow <- function(object, split, ..., metrics = NULL) {
   last_fit_workflow(object, split, metrics)
 }
 
-split_to_rset <- function(x) {
-  prop <- length(x$in_id)/nrow(x$data)
-  res <- rsample::mc_cv(x$data, times = 1, prop = prop)
-  res$splits[[1]] <- x
-  res
-}
-
 last_fit_workflow <- function(object, split, metrics) {
   extr <- function(x)
     x
   ctrl <- control_resamples(save_pred = TRUE, extract = extr)
+  splits <- list(split)
+  resamples <- rsample::manual_rset(splits, ids = "train/test split")
   res <-
     fit_resamples(
       object,
-      resamples = split_to_rset(split),
+      resamples = resamples,
       metrics = metrics,
       control = ctrl
     )
-  res$id[[1]] <- "train/test split"
   res$.workflow <- res$.extracts[[1]][[1]]
   res$.extracts <- NULL
   class(res) <- c("last_fit", class(res))
   class(res) <- unique(class(res))
   res
 }
-

tests/testthat/test-last-fit.R

@@ -29,20 +29,6 @@ test_that("recipe method", {
   expect_equal(res$.predictions[[1]]$.pred, unname(test_pred))
 })
 
-test_that("split_to_rset", {
-
-  res <- tune:::split_to_rset(split)
-  expect_true(inherits(res, "mc_cv"))
-  expect_true(nrow(res) == 1)
-  expect_true(nrow(res) == 1)
-
-  res <- linear_reg() %>% set_engine("lm") %>% last_fit(f, split)
-  expect_true(is.list(res$.workflow))
-  expect_true(inherits(res$.workflow[[1]], "workflow"))
-  expect_true(is.list(res$.predictions))
-  expect_true(inherits(res$.predictions[[1]], "tbl_df"))
-})
-
 test_that("collect metrics of last fit", {
 
   res <- linear_reg() %>% set_engine("lm") %>% last_fit(f, split)