support mlr3

[simonpcouch]

Some work on the mlr3 branch already.

[simonpcouch]

After spending more time with this, I think our best approach here will to wait for more machinery from mlr3 folks before implementing anything more principled. Prediction in a new session works fine, so this shouldn't be a pain point!

library(mlr3)

task <- tsk("mtcars")
fit <- lrn("regr.rpart")
  
# train a model of this learner for a subset of the task
fit$train(task, row_ids = 1:26)

callr::r(
  function(fit) {
    library(mlr3)
    
    predict(fit, mtcars[27:32,])
  },
  args = list(fit = fit)
)
#> [1] 27.71429 16.87368 16.87368 16.87368 16.87368 16.87368

^{Created on 2022-07-18 by the reprex package (v2.0.1)}

[simonpcouch]

Look for the fitted model object slot, these should wrap like caret or parsnip.

[sebffischer]

I have started a discussion in mlr3 about serialization here: https://github.com/mlr-org/mlr3/issues/891 If we implement something like this, adding a bundle method for mlr3 Learners should be straightforward! :)