control_last_fit() doesn't accept an allow_par argument, and sets it to TRUE by default when wrapping control_resamples. This causes issues after last_fit() for models that need native serialization methods to be passed between sessions.
I'd suggest we
- add an
allow_par argument to control_last_fit(), and
- set its default to
FALSE
I don't believe 2) would result in any slowdown for last_fit(), as there's only one fit occurring on the train fold of a train/test split. It would also allow users to last_fit() without needing to set allow_par for model engines that require native serialization.
Reprex and more discussion in tidymodels/bonsai#52 (comment).
control_last_fit()doesn't accept anallow_parargument, and sets it toTRUEby default when wrappingcontrol_resamples. This causes issues afterlast_fit()for models that need native serialization methods to be passed between sessions.I'd suggest we
allow_parargument tocontrol_last_fit(), andFALSEI don't believe 2) would result in any slowdown for
last_fit(), as there's only one fit occurring on the train fold of a train/test split. It would also allow users tolast_fit()without needing to setallow_parfor model engines that require native serialization.Reprex and more discussion in tidymodels/bonsai#52 (comment).