Merge pull request #1167 from tidymodels/sparse-matrix-predict

Make sure sparse matrices can be used with `predict()`

Author: EmilHvitfeldt

NEWS.md

@@ -4,6 +4,8 @@
 
 * `fit()` and `fit_xy()` can now take sparse tibbles as data values (#1165).
 
+* `predict()` can now take dgCMatrix and sparse tibble input for `new_data` argument, and error informatively when model doesn't support it (#1167).
+
 * Transitioned package errors and warnings to use cli (#1147 and #1148 by
   @shum461, #1153 by @RobLBaker and @wright13, #1154 by @JamesHWade, #1160, 
   #1161, #1081).

R/fit.R

@@ -444,6 +444,10 @@ check_xy_interface <- function(x, y, cl, model) {
 }
 
 allow_sparse <- function(x) {
+  if (inherits(x, "model_fit")) {
+    x <- x$spec
+  }
+
   res <- get_from_env(paste0(class(x)[1], "_encoding"))
   all(res$allow_sparse_x[res$engine == x$engine])
 }

R/predict.R

@@ -160,6 +160,8 @@ predict.model_fit <- function(object, new_data, type = NULL, opts = list(), ...)
   }
   check_pred_type_dots(object, type, ...)
 
+  new_data <- to_sparse_data_frame(new_data, object)
+
   res <- switch(
     type,
     numeric     = predict_numeric(object = object, new_data = new_data, ...),

R/sparsevctrs.R

@@ -3,6 +3,10 @@ to_sparse_data_frame <- function(x, object) {
     if (allow_sparse(object)) {
       x <- sparsevctrs::coerce_to_sparse_data_frame(x)
     } else {
+      if (inherits(object, "model_fit")) {
+        object <- object$spec
+      }
+    
       cli::cli_abort(
       "{.arg x} is a sparse matrix, but {.fn {class(object)[1]}} with
        engine {.code {object$engine}} doesn't accept that.")
@@ -19,6 +23,10 @@ is_sparse_tibble <- function(x) {
 
 materialize_sparse_tibble <- function(x, object, input) {
   if (is_sparse_tibble(x) && (!allow_sparse(object))) {
+    if (inherits(object, "model_fit")) {
+      object <- object$spec
+    }
+    
     cli::cli_warn(
       "{.arg {input}} is a sparse tibble, but {.fn {class(object)[1]}} with
       engine {.code {object$engine}} doesn't accept that. Converting to 

tests/testthat/_snaps/sparsevctrs.md

@@ -22,6 +22,22 @@
       Error in `to_sparse_data_frame()`:
       ! `x` is a sparse matrix, but `linear_reg()` with engine `lm` doesn't accept that.
 
+# sparse tibble can be passed to `predict()
+
+    Code
+      preds <- predict(lm_fit, sparse_mtcars)
+    Condition
+      Warning:
+      `x` is a sparse tibble, but `linear_reg()` with engine `lm` doesn't accept that. Converting to non-sparse.
+
+# sparse matrices can be passed to `predict()
+
+    Code
+      predict(lm_fit, sparse_mtcars)
+    Condition
+      Error in `to_sparse_data_frame()`:
+      ! `x` is a sparse matrix, but `linear_reg()` with engine `lm` doesn't accept that.
+
 # to_sparse_data_frame() is used correctly
 
     Code

tests/testthat/test-sparsevctrs.R

@@ -67,6 +67,66 @@ test_that("sparse matrices can be passed to `fit_xy()", {
   )
 })
 
+test_that("sparse tibble can be passed to `predict()", {
+  skip_if_not_installed("ranger")
+
+  hotel_data <- sparse_hotel_rates()
+  hotel_data <- sparsevctrs::coerce_to_sparse_tibble(hotel_data)
+
+  spec <- rand_forest(trees = 10) %>%
+    set_mode("regression") %>%
+    set_engine("ranger")
+
+  tree_fit <- fit_xy(spec, x = hotel_data[, -1], y = hotel_data[, 1])
+
+  expect_no_error(
+    predict(tree_fit, hotel_data)
+  )
+
+  spec <- linear_reg() %>%
+    set_mode("regression") %>%
+    set_engine("lm")
+
+  lm_fit <- fit(spec, mpg ~ ., data = mtcars)
+
+  sparse_mtcars <- mtcars %>%
+    sparsevctrs::coerce_to_sparse_matrix() %>%
+    sparsevctrs::coerce_to_sparse_tibble()
+
+  expect_snapshot(
+    preds <- predict(lm_fit, sparse_mtcars)
+  )
+})
+
+test_that("sparse matrices can be passed to `predict()", {
+  skip_if_not_installed("ranger")
+
+  hotel_data <- sparse_hotel_rates()
+
+  spec <- rand_forest(trees = 10) %>%
+    set_mode("regression") %>%
+    set_engine("ranger")
+
+  tree_fit <- fit_xy(spec, x = hotel_data[, -1], y = hotel_data[, 1])
+
+  expect_no_error(
+    predict(tree_fit, hotel_data)
+  )
+
+  spec <- linear_reg() %>%
+    set_mode("regression") %>%
+    set_engine("lm")
+
+  lm_fit <- fit(spec, mpg ~ ., data = mtcars)
+
+  sparse_mtcars <- sparsevctrs::coerce_to_sparse_matrix(mtcars)
+
+  expect_snapshot(
+    error = TRUE,
+    predict(lm_fit, sparse_mtcars)
+  )
+})
+
 test_that("to_sparse_data_frame() is used correctly", {
   skip_if_not_installed("xgboost")