This issue describes how, once the data are given to a workflow, prediction always requires it as-is.
This will be increasingly important for supervised feature selection.
Look at my initial work on colander (still in private repo), there is an api called reset_columns() that edits the workflow's data objects. That's not great since we always try to remake the model or recipe when the data changes.
Alternatively, it might be better to have some sort of ignore_predictors() api that leaves the data intact but updates what the fit() and predict() methods for workflows do.
This issue describes how, once the data are given to a workflow, prediction always requires it as-is.
This will be increasingly important for supervised feature selection.
Look at my initial work on colander (still in private repo), there is an api called
reset_columns()that edits the workflow's data objects. That's not great since we always try to remake the model or recipe when the data changes.Alternatively, it might be better to have some sort of
ignore_predictors()api that leaves the data intact but updates what thefit()andpredict()methods for workflows do.