It would be useful to have save/load functionality for models, e.g., as a method with a primary preferred mode.

A way to implement this could be a save method that BaseEstimator has, similar to keras.save, which saves and/or returns (parameter dependent) a pickle.

"Ordinary" estimators would have the pickle default in save, which can be overridden by specific functionality, e.g., in BaseCNNClassifier, for neural networks.

The counterpart would be a load method - in each given module root, e.g., from sktime.classification import load.

Usage:

# session 1 - building/fitting an estimator
stuff
my_estimator.fit(data)
stuff
my_estimator.save(hd_location) # saves fitted estimator

# session 2 - retrieving/loading estimator
from sktime.classification import load

my_estimator = load(hd_location)