Don't coerce sparse data to non-sparse during predict()

[EmilHvitfeldt]

This bug was first reported here: https://community.rstudio.com/t/predict-not-working-with-ranger-model-when-using-sparse-data/163352/7

When you call predict() you eventually get to run prepare_data(), which didn't know about the alllow_sparse_x encoding so it would try to turn it into matrices/data.frames. This was bad for 2 reasons. First we are losing some performance. Secondly, in the case of the ranger method, it wants the data as a data.frame which a sparse data can't be turned into, yielding the error seen below.

Main

library(parsnip)

data(agaricus.train, package = 'xgboost')
train <- agaricus.train

rf_model <- parsnip::rand_forest(trees = 100) %>% 
  set_engine("ranger") %>% 
  set_mode("classification")

rf_fit <- fit_xy(rf_model, x = train$data, y = factor(train$label))

predict(rf_fit, train$data)
#> Error in as.data.frame.default(new_data): cannot coerce class 'structure("dgCMatrix", package = "Matrix")' to a data.frame

This PR

library(parsnip)

data(agaricus.train, package = 'xgboost')
train <- agaricus.train

rf_model <- parsnip::rand_forest(trees = 100) %>% 
  set_engine("ranger") %>% 
  set_mode("classification")

rf_fit <- fit_xy(rf_model, x = train$data, y = factor(train$label))

predict(rf_fit, train$data)
#> # A tibble: 6,513 × 1
#>    .pred_class
#>    <fct>      
#>  1 1          
#>  2 0          
#>  3 0          
#>  4 1          
#>  5 0          
#>  6 0          
#>  7 0          
#>  8 1          
#>  9 0          
#> 10 0          
#> # ℹ 6,503 more rows

[EmilHvitfeldt]

Small speedup benchmark. (this is likely to vary majorly for different data, but it should almost always be in the right direction

Code used for the following reprexes

library(parsnip)

data(agaricus.train, package = 'xgboost')
train <- agaricus.train

xg_model <- parsnip::boost_tree(trees = 100) %>% 
  set_engine("xgboost") %>% 
  set_mode("classification")

xg_fit <- fit_xy(xg_model, x = train$data, y = factor(train$label))

Main

bench::mark(
  old = predict(xg_fit, train$data)
)
#> # A tibble: 1 × 6
#>   expression      min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr> <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 old          8.96ms   9.55ms      105.     7.2MB     60.4

This PR

bench::mark(
  new = predict(xg_fit, train$data)
)
#> # A tibble: 1 × 6
#>   expression      min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr> <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 new          6.24ms   6.65ms      150.    1.13MB     4.17

[simonpcouch]

Love it. :)

Closes #694. Does this also address #690?

Could we also clarify in these docs that allow_sparse_x now applies to predict()ion, too? (I think?) Are there any engines that would allow sparsity at fit() but not predict() time?

parsnip/R/aaa_models.R

Lines 553 to 555 in 51b0cd7

	#' Finally, `allow_sparse_x` specifies whether the model function can natively
	#' accommodate a sparse matrix representation for predictors during fitting
	#' and tuning.

[github-actions]

This pull request has been automatically locked. If you believe you have found a related problem, please file a new issue (with a reprex: https://reprex.tidyverse.org) and link to this issue.