It provides a massive speed improvement which we definitely want in all methods where we use it, but due to mlr-org/mlr3pipelines#961 we have to anticipate it not being available for GraphLearners, which I use for simple things like po("removeconstants") or so in the benchmark comparing various methods on various tasks.
When doing performance benchmarking I also need to take into account whether predict_newdata_fast() was used or not.
As an aside: Currently I think there's technically nothing that prevents things like PFI or so to work with survival data, and since predict_newdata_fast() would need a dedicated implementation for survival learners, that's another edge case.
It provides a massive speed improvement which we definitely want in all methods where we use it, but due to mlr-org/mlr3pipelines#961 we have to anticipate it not being available for
GraphLearners, which I use for simple things likepo("removeconstants")or so in the benchmark comparing various methods on various tasks.When doing performance benchmarking I also need to take into account whether
predict_newdata_fast()was used or not.As an aside: Currently I think there's technically nothing that prevents things like
PFIor so to work with survival data, and sincepredict_newdata_fast()would need a dedicated implementation for survival learners, that's another edge case.