Description
I am trying to set the hyperparameters of the base_estimator of an AdaBoostRegressor inside a MultioutputRegressor. I've defined a parameters dictionary as:
params = {'estimator__base_estimator': DecisionTreeRegressor(),
'estimator__loss': 'square',
'estimator__learning_rate': 0.05,
'estimator__n_estimators': 50,
'estimator__base_estimator__max_features': 'sqrt',
'estimator__base_estimator__max_leaf_nodes': 8
}
The values here aren't important, just the keys. When I call set_params(**params) it uses get_params with deep=True to find the parameters of the model and the sub-model. Great, except I also need to be able to tell the parameters of the sub-sub-model are valid and then set them.
I see that get_params calls itself but doesn't pass down deep=True, so I copied and used a version of my own where I did pass deep down. The trouble I ran in to then was that the sub-sub-model is a parameter of the sub-model and hasn't been set yet, so somewhere down the stack the call value.get_params(deep=True).items() fails when value is None, because value is the base_estimator of the AdaboostRegressor and hasn't been set yet.
I managed to hack a workaround in my code by only passing the parameters that have keys with less than two '__' substrings and then setting the subobject with the other subset of the params dictionary, but I would really like it if set_params() were truly completely generalized to arbitrarily deep structures. This will probably involve intelligently setting objects at each layer before going deeper.