make set_tailor_type() play nice with infer_type() (closes #38)

Author: simonpcouch

R/tailor.R

@@ -209,8 +209,9 @@ predict.tailor <- function(object, new_data, ...) {
   new_data
 }
 
-set_tailor_type <- function(object, y) {
+set_tailor_type <- function(object, y, call = caller_env()) {
   if (object$type != "unknown") {
+    check_outcome_type(y, object$type, call = call)
     return(object)
   }
   if (is.factor(y)) {
@@ -223,7 +224,10 @@ set_tailor_type <- function(object, y) {
   } else if (is.numeric(y)) {
     object$type <- "regression"
   } else {
-    cli_abort("Only factor and numeric outcomes are currently supported.")
+    cli_abort(
+      "Only factor and numeric outcomes are currently supported.",
+      call = call
+    )
   }
   object
 }

R/utils.R

@@ -193,6 +193,27 @@ check_method <- function(method,
   method
 }
 
+# at `fit()` time, we check the type of the outcome vs the type
+# supported by the applied adjustments. where this is called currently,
+# we know already that `type` is not "unknown"
+check_outcome_type <- function(outcome, type, call) {
+  outcome_is_compatible <-
+    switch(
+      type,
+      regression = is.numeric(outcome),
+      binary = , multiclass = is.factor(outcome),
+      FALSE
+    )
+
+  if (!outcome_is_compatible) {
+    cli_abort(
+      "Tailors with {type} adjustments are not compatible
+       with {.cls {class(outcome)}} outcomes.",
+      call = call
+    )
+  }
+}
+
 check_selection <- function(selector, result, arg, call = caller_env()) {
   if (length(result) == 0) {
     cli_abort(

tests/testthat/_snaps/utils.md

@@ -6,3 +6,39 @@
       Error in `adjust_probability_threshold()`:
       ! `x` should be a <tailor> (`?tailor::tailor()`), not a string.
 
+# fit.tailor() errors informatively with incompatible outcome
+
+    Code
+      fit(tailor() %>% adjust_probability_threshold(0.1), two_class_example, outcome = c(
+        test_numeric), estimate = c(predicted), probabilities = c(Class1, Class2))
+    Condition
+      Error in `fit()`:
+      ! Tailors with binary adjustments are not compatible with <numeric> outcomes.
+
+---
+
+    Code
+      fit(tailor() %>% adjust_numeric_range(lower_limit = 0.1), two_class_example,
+      outcome = c(truth), estimate = c(Class1))
+    Condition
+      Error in `fit()`:
+      ! Tailors with regression adjustments are not compatible with <factor> outcomes.
+
+---
+
+    Code
+      fit(tailor() %>% adjust_probability_threshold(0.1), two_class_example, outcome = c(
+        test_date), estimate = c(predicted), probabilities = c(Class1, Class2))
+    Condition
+      Error in `fit()`:
+      ! Tailors with binary adjustments are not compatible with <POSIXct/POSIXt> outcomes.
+
+---
+
+    Code
+      fit(tailor() %>% adjust_predictions_custom(hey = "there"), two_class_example,
+      outcome = c(test_date), estimate = c(predicted), probabilities = c(Class1))
+    Condition
+      Error in `fit()`:
+      ! Only factor and numeric outcomes are currently supported.
+

tests/testthat/test-utils.R

@@ -73,3 +73,58 @@ test_that("tailor_requires_fit works", {
     )
   )
 })
+
+test_that("fit.tailor() errors informatively with incompatible outcome", {
+  skip_if_not_installed("modeldata")
+  library(modeldata)
+
+  two_class_example$test_numeric <- two_class_example$Class1 + 1
+  two_class_example$test_date <- as.POSIXct(two_class_example$Class1)
+
+  # supply a numeric outcome to a binary tailor
+  expect_snapshot(
+    error = TRUE,
+    fit(
+      tailor() %>% adjust_probability_threshold(.1),
+      two_class_example,
+      outcome = c(test_numeric),
+      estimate = c(predicted),
+      probabilities = c(Class1, Class2)
+    )
+  )
+
+  # supply a factor outcome to a regression tailor
+  expect_snapshot(
+    error = TRUE,
+    fit(
+      tailor() %>% adjust_numeric_range(lower_limit = .1),
+      two_class_example,
+      outcome = c(truth),
+      estimate = c(Class1)
+    )
+  )
+
+  # supply a totally wild outcome to a regression tailor
+  expect_snapshot(
+    error = TRUE,
+    fit(
+      tailor() %>% adjust_probability_threshold(.1),
+      two_class_example,
+      outcome = c(test_date),
+      estimate = c(predicted),
+      probabilities = c(Class1, Class2)
+    )
+  )
+
+  # supply a totally wild outcome to an unknown tailor
+  expect_snapshot(
+    error = TRUE,
+    fit(
+      tailor() %>% adjust_predictions_custom(hey = "there"),
+      two_class_example,
+      outcome = c(test_date),
+      estimate = c(predicted),
+      probabilities = c(Class1)
+    )
+  )
+})