aorsf - engine: model fit fails if mtry is specified

[MasterLuke84]

Hi,

the model fit fails if mtry is specified for the aorsf-engine. If it is not specified, it works with the default engine values.

library(bonsai)
#> Loading required package: parsnip

# This works with default mtry value
rf_mod <- 
  rand_forest() %>%
  set_engine(engine = "aorsf") %>%
  set_mode(mode = "regression") %>% 
  set_args(min_n = 1, trees = 2, importance = "permute") %>% 
  fit(
    formula = mpg  ~ . , 
    data = mtcars 
  )


rf_mod
#> parsnip model object
#> 
#> ---------- Oblique random regression forest
#> 
#>      Linear combinations: Accelerated Linear regression
#>           N observations: 32
#>                  N trees: 2
#>       N predictors total: 10
#>    N predictors per node: 4
#>  Average leaves per tree: 7.5
#> Min observations in leaf: 1
#>           OOB stat value: 0.27
#>            OOB stat type: RSQ
#>      Variable importance: permute
#> 
#> -----------------------------------------



# Error occurs...
rf_mod_w_mtry <- 
  rand_forest() %>%
  set_engine(engine = "aorsf") %>%
  set_mode(mode = "regression") %>% 
  set_args(mtry = 3, min_n = 1, trees = 2, importance = "permute") %>% 
  fit(
    formula = mpg  ~ . , 
    data = mtcars 
  )
#> Error in ncol(source): object 'x' not found

^{Created on 2024-08-08 with reprex v2.0.2}

Thank you in advance and best regards

[simonpcouch]

Thanks for the issue! Just confirming that I can reproduce this and 1) it does seem to be aorsf-specific (i.e. xgboost is not an issue) and 2) it doesn't seem to be due to any changes in parsnip (issue persists with parsnip v1.0.0). min_cols() seems to be evaluated in a different environment than its usual.

[bcjaeger]

Just saw this and thought I'd check it out. The issue appears to occur in parsnip:::make_form_call:

If we modify:

  if (object$engine == "spark") {
    env$x <- env$data
  }

to

  if (object$engine %in% c("spark", "aorsf")) {
    env$x <- env$data
  }

the x object will be recognized when you specify mtry.

@simonpcouch - would this solution be too hacky? A more general approach might change the call to aorsf::orsf, making it so that instead of mtry = min_cols(~3, x) we use mtry = min_cols(~3, data)

[EmilHvitfeldt]

This is what happens:

For the ranger engine, you end up in form_xy() where x is passed in the evaluation environment

parsnip/R/fit_helpers.R

Lines 148 to 157 in 6d4c684

	env$x <- data_obj$x
	env$y <- data_obj$y
	res <- xy_xy(
	object = object,
	env = env, #weights!
	control = control,
	target = target,
	call = call
	)

On the other hand, with the aorsf engine, you end up in form_form().

parsnip/R/fit_helpers.R

Lines 40 to 55 in 6d4c684

	fit_call <- make_form_call(object, env = env)
	res <- list(
	lvl = y_levels$lvl,
	ordered = y_levels$ordered,
	spec = object
	)
	time <- proc.time()
	res$fit <- eval_mod(
	fit_call,
	capture = control$verbosity == 0,
	catch = control$catch,
	envir = env,
	...
	)

The created call is

aorsf::orsf(formula = mpg ~ ., data = data, mtry = min_cols(~3, 
    x), n_tree = ~2, leaf_min_obs = ~1, n_thread = 1, verbose_progress = FALSE)

and the env has the elements "formula", "data", and "weights".

Leading to the error because we try to evaluate x in min_cols(~3, x) while it isn't available anywhere.

We could either do what @bcjaeger suggests, OR add a env$x <- env$data in form_form().