[MRG] MAINT add base class for voting and stacking #15084
Conversation
|
@thomasjpfan @ogrisel @rth @adrinjalali So the naming of the base class is terrible but I wanted to have a WIP PR such that we see what is in common and if it makes sense to merge code. NB: the tests will fail because I did not add support for WDYT? |
thomasjpfan
left a comment
There was a problem hiding this comment.
I like this refactoring. The _BaseEnsembleHeterogeneousEstimator class has have well defined boundaries.
|
Good to be reviewed. I will open a PR to deprecate |
|
Having the init make it explicit that the estimators parameter is the
common denominator for all inherited classes.
I am fine keeping it.
…On Wed, 2 Oct 2019 at 13:07, Nicolas Hug ***@***.***> wrote:
***@***.**** commented on this pull request.
------------------------------
In sklearn/ensemble/base.py
<#15084 (comment)>
:
> @@ -178,3 +182,76 @@ def _partition_estimators(n_estimators, n_jobs):
starts = np.cumsum(n_estimators_per_job)
return n_jobs, n_estimators_per_job.tolist(), [0] + starts.tolist()
+
+
+class _BaseHeterogeneousEnsemble(MetaEstimatorMixin, _BaseComposition,
+ metaclass=ABCMeta):
+ """Base class for ensemble learners based on heterogeneous estimators."""
+ _required_parameters = ['estimators']
+
+ @Property
+ def named_estimators(self):
+ return Bunch(**dict(self.estimators))
+
+ @AbstractMethod
+ def __init__(self, estimators):
+ self.estimators = estimators
i'm suggesting to not have the init method
—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
<#15084?email_source=notifications&email_token=ABY32P6RROAFMVM32WZOLMTQMR6FTA5CNFSM4I2CTMN2YY3PNVWWK3TUL52HS4DFWFIHK3DMKJSXC5LFON2FEZLWNFSXPKTDN5WW2ZLOORPWSZGOCGTRNOY#discussion_r330490319>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/ABY32PYGSAUSWSRVBIUX6RLQMR6FTANCNFSM4I2CTMNQ>
.
--
Guillaume Lemaitre
Scikit-learn @ Inria Foundation
https://glemaitre.github.io/
|
NicolasHug
left a comment
There was a problem hiding this comment.
I guess I'm becoming increasingly skeptical about the relevance of inheritance in some cases (like here where all it does is set a single attribute). Makes the code easy to write, but often harder to understand.
But LGTM anyway.
doc/whats_new/v0.22.rst
Outdated
| - |Fix| Stacking and Voting estimators now ensure that their underlying | ||
| estimators are either all classifiers or all regressors. | ||
| We introduced a new base class | ||
| :class:`ensemble.base._BaseHeterogeneousEnsemble` to raise consistent error |
There was a problem hiding this comment.
Should we include a private class in the whats new? This can be something like:
Stacking and Voting estimators now raise consistent error messages.
There was a problem hiding this comment.
I might be misunderstood what @NicolasHug meant by adding a link?
Did you mean mentioning the class or do you expect something else?
There was a problem hiding this comment.
Since we are not generating a new API doc for _BaseHeterogeneousEnsemble, there is nothing to link to: https://76528-843222-gh.circle-artifacts.com/0/doc/whats_new/v0.22.html#sklearn-ensemble
There was a problem hiding this comment.
The links referred to the Stacking and Voting estimators, sorry if that wasn't clear. I agree we shouldn't link a private class. (and I'm also fine not linking the estimators... it's just a nit)
doc/whats_new/v0.22.rst
Outdated
|
|
||
| - |Fix| Stacking and Voting estimators now ensure that their underlying | ||
| estimators are either all classifiers or all regressors. | ||
| We introduced a new base class |
There was a problem hiding this comment.
We do not need a "new base class" part?
:class:
ensemble.StackingClassifier, :class:ensemble.StackingRegressor, :class:ensemble.VotingClassifier, and :class:ensemeble.VotingRegressornow raise consistent error messages.
|
Thank you @glemaitre ! |
closes #15056
Create a base class for Voting* and Stacking*. They both are an ensemble of multiple learners type.
They could share the
get_params,set_paramsand validation ofestimators(as well as the fitted attributes then).This base class could be contrasted with the ensemble of single learner type such as boosting (adaboost, GBDT), RF and Bagging.