Sor far the whole model class wrapping sklearn estimators is serialized in a pickle, including estimator classes and the parent class base_estimator. Pickle files contain references to libraries, so that when the code in base_estimator changes, the pickle changes also making flame to break with old models.
New strategy will serialize only the variable self.estimator containing solely the sklearn estimator, estimators inheriting from sklearn, or aggregated sklearn models used in conformal prediction.
The latest version implements Jose Carlos proposed strategy. Moreover, the pickls which store the model are now labelled with a version number which allow further compatibility control
Sor far the whole model class wrapping sklearn estimators is serialized in a pickle, including estimator classes and the parent class base_estimator. Pickle files contain references to libraries, so that when the code in base_estimator changes, the pickle changes also making flame to break with old models.
New strategy will serialize only the variable self.estimator containing solely the sklearn estimator, estimators inheriting from sklearn, or aggregated sklearn models used in conformal prediction.