Merge pull request #225 from patr1ckm/nnet-pred

Standardize nnet class prediction column names

Author: topepo

NEWS.md

@@ -6,6 +6,9 @@
 
 * A bug was fixed related to the column names generated by `multi_predict()`. The top-level tibble will always have a column named `.pred` and this list column contains tibbles across sub-models. The column names for these sub-model tibbles will have names consistent with `predict()` (which was previously incorrect). See [43c15db](https://github.com/tidymodels/parsnip/commit/43c15db377ea9ef27483ff209f6bd0e98cb830d2).
 
+# [A bug](https://github.com/tidymodels/parsnip/issues/174) was fixed
+standardizing the column names of `nnet` class probability predictions.
+
 # parsnip 0.0.3.1
 
 Test case update due to CRAN running extra tests [(#202)](https://github.com/tidymodels/parsnip/issues/202)

R/mlp.R

@@ -381,7 +381,7 @@ nnet_softmax <- function(results, object) {
 
   results <- apply(results, 1, function(x) exp(x)/sum(exp(x)))
   results <- t(results)
-  names(results) <- paste0(".pred_", object$lvl)
+  colnames(results) <- object$lvl
   results <- as_tibble(results)
   results
 }

tests/testthat/test_mlp.R

@@ -157,3 +157,11 @@ test_that('bad input', {
   expect_error(translate(mlp(mode = "regression", formula = y ~ x) %>% set_engine()))
 })
 
+test_that("nnet_softmax", {
+  obj <- mlp(mode = 'classification')
+  obj$lvls <- c("a", "b")
+  res <- nnet_softmax(matrix(c(.8, .2)), obj)
+  expect_equal(names(res), obj$lvls)
+  expect_equal(res$b, 1 - res$a)
+})
+