don't turn sparse matrix into dense matrix for glmnet prediction (#1210)

* don't turn sparse matrix into dense matrix for glmnet prediction

* remember to do subsetting

* don't add random newlines

* test glmnet predict doesn't remove sparseness

* regenerate snapshots to resolve merge conflicts

---------

Co-authored-by: ‘topepo’ <mxkuhn@gmail.com>

Author: EmilHvitfeldt

R/glmnet-engines.R

@@ -138,6 +138,16 @@ predict_raw._glmnetfit <- predict_raw_glmnet
   unname(x[, 1])
 }
 
+organize_glmnet_pre_pred <- function(x, object) {
+  x <- x[, rownames(object$fit$beta), drop = FALSE]
+  if (is_sparse_matrix(x)) {
+    return(x)
+  }
+  
+  as.matrix(x)
+}
+
+
 organize_glmnet_class <- function(x, object) {
   prob_to_class_2(x[, 1], object)
 }

R/linear_reg_data.R

@@ -250,7 +250,7 @@ set_pred(
     args =
       list(
         object = expr(object$fit),
-        newx = expr(as.matrix(new_data[, rownames(object$fit$beta), drop = FALSE])),
+        newx = expr(organize_glmnet_pre_pred(new_data, object)),
         type = "response",
         s = expr(object$spec$args$penalty)
       )

tests/testthat/_snaps/sparsevctrs.md

@@ -127,6 +127,14 @@
       Error in `maybe_sparse_matrix()`:
       ! no sparse vectors detected
 
+# we don't run as.matrix() on sparse matrix for glmnet pred #1210
+
+    Code
+      predict(lm_fit, hotel_data)
+    Condition
+      Error in `predict.elnet()`:
+      ! data is sparse
+
 # fit() errors if sparse matrix has no colnames
 
     Code

tests/testthat/test-sparsevctrs.R

@@ -314,6 +314,34 @@ test_that("maybe_sparse_matrix() is used correctly", {
   )
 })
 
+test_that("we don't run as.matrix() on sparse matrix for glmnet pred #1210", {
+  skip_if_not_installed("glmnet")
+
+  local_mocked_bindings(
+    predict.elnet = function(object, newx, ...) {
+      if (is_sparse_matrix(newx)) {
+        stop("data is sparse")
+      } else {
+        stop("data isn't sparse (should not happen)")
+      }
+    },
+    .package = "glmnet"
+  )
+
+  hotel_data <- sparse_hotel_rates()
+
+  spec <- linear_reg(penalty = 0) %>%
+    set_mode("regression") %>%
+    set_engine("glmnet")
+
+  lm_fit <- fit_xy(spec, x = hotel_data[, -1], y = hotel_data[, 1])
+
+  expect_snapshot(
+    error = TRUE,
+    predict(lm_fit, hotel_data)
+  )
+})
+
 test_that("fit() errors if sparse matrix has no colnames", {
   hotel_data <- sparse_hotel_rates()
   colnames(hotel_data) <- NULL