[MRG] Fast PDP for DecisionTreeRegressor and RandomForestRegressor #15864
Conversation
…rtial_dep_decision_tree
|
Codecov is red because of |
| @@ -351,19 +339,25 @@ def partial_dependence(estimator, X, features, response_method='auto', | |||
| if (isinstance(estimator, BaseGradientBoosting) and | |||
| estimator.init is None): | |||
| method = 'recursion' | |||
| elif isinstance(estimator, BaseHistGradientBoosting): | |||
| elif isinstance(estimator, (BaseHistGradientBoosting, | |||
There was a problem hiding this comment.
do you think that it would make sense to avoid this isinstance and check for hasattr(estimator, 'compute_partial_dependence') or a method that could be shared across all these estimators?
I could think that library as xgboost or lightgbm or tree base could expose the same method without us checking for type.
There was a problem hiding this comment.
that's a nice idea but that means we need to support a new public API, i.e. the interface of _compute_partial_dependence_recursion would now be fixed
We would also still need to hardcoded list of supported estimators for the error message.
Maybe we can keep that in mind for later?
There was a problem hiding this comment.
Yep we could check how other libraries are exposing this and put all these info in an issue
|
Does this have your +1 @glemaitre ? I think all comments were addressed |
|
thanks @NicolasHug |
This PR implements the fast 'recursion' method for DecisionTreeRegressor and RandomForestRegressor
Need to merge #16114 first
We're only exposing the method which already exists for the gradient boosting estimators.