CPOs don't work with runTaskMlr in Openml

[ja-thomas]

(Not sure if this has be solved in openml)

We cannot run (and upload) lrn with attached CPOs on OpenML Tasks

lrn = cpoFilterChiSquared(abs = 3) %>>%
   makeLearner("classif.rpart")
task = getOMLTask(3)
runTaskMlr(task, lrn)

Error in removeHyperPars.CPOLearner(lrn, ids = all.pars) : 
  CPO Parameters (chi.squared.perc, chi.squared.abs, chi.squared.threshold) can not be removed

@giuseppec Any hints?

[mb706]

CPO currently does not allow hyperparameters to be removed. A CPO with missing parameters could either be considered "incomplete" and would then throw an error when called (as when calling a function without giving all necessary arguments), or it could just use some kind of default values for the parameters instead. I made the decision to go with the first option, because I didn't want to have implicit behaviour in CPO parameters (suddenly you have to keep track of the parameters and of what their default values are). The "default" value of a parameter is therefore just the value it is initialized with:

> getHyperPars(cpoPca())
$pca.center
[1] TRUE

$pca.scale
[1] FALSE

CPOs with missing parameters give an error when invoked:

> iris.task %>>% cpoWrap()
Error in checkAllParams(cpo$par.vals, cpo$par.set, cpo$debug.name) :
  Parameter wrap.cpo of CPO wrap is missing
Either give it during construction, or with setHyperPars.

Removing hyperparameters from CPOLearners would just be a "make this not work anymore" switch, not a "reset to original state" switch as I think it is intended in that invocation at OpenML(?).

Why does OpenML do that? (Does it even work with ModelMultiplexer etc?)