PipeOpTargetMutate breaks internal validation #1401
Description
The mlr3 book shows how to set the internal_valid_task property of a task for internal validation, including with graph learners. But the code breaks when using PipeOpTargetMutate.
It looks like PipeOpTargetMutate calls convert_task, which doesn't preserve internal_valid_task. Then the base learner raises an error because it can't find the validation data. Other target-related pipeops also have this issue.
It's not clear to me whether this should be considered a bug in PipeOpTargetMutate or in convert_task.
tsk_mtcars = tsk("mtcars")
tsk_mtcars$internal_valid_task = sample(tsk_mtcars$nrow, 10)
convert_task(tsk_mtcars) # validation task is droppedExample:
tsk_mtcars = tsk("mtcars")
tsk_mtcars$internal_valid_task = sample(tsk_mtcars$nrow, 10)
lrn_xgb = lrn("regr.xgboost")
# working graph learner
glrn = as_learner(po("pca") %>>% lrn_xgb)
set_validate(glrn, validate = "predefined")
glrn$train(tsk_mtcars) # it works
# failing graph learner
ttscalerange = ppl("targettrafo", trafo_pipeop = PipeOpTargetTrafoScaleRange$new(),
graph = PipeOpLearner$new(lrn_xgb))
glrn2 = as_learner(ttscalerange)
set_validate(glrn2, validate = "predefined")
glrn2$train(tsk_mtcars) # errorIt raises this error:
Error in create_internal_valid_task(validate, task, test_row_ids, prev_valid, :
Parameter 'validate' is set to 'predefined' but no internal validation task is present. This commonly happens in GraphLearners and can be avoided by configuring the validation data for the GraphLearner via `set_validate(<glrn>, validate = <value>)`. See https://mlr3book.mlr-org.com/chapters/chapter15/predsets_valid_inttune.html for more information.
This happened in PipeOp regr.xgboost's $train()