from evalml.exceptions import EnsembleMissingPipelinesError
from evalml.model_family import ModelFamily
from evalml.pipelines.components import Estimator
from evalml.pipelines.components.utils import scikit_learn_wrapped_estimator
from evalml.utils import classproperty
_nonstackable_model_families = [ModelFamily.BASELINE, ModelFamily.NONE]
[docs]class StackedEnsembleBase(Estimator):
"""Stacked Ensemble Base Class.
Arguments:
input_pipelines (list(PipelineBase or subclass obj)): List of pipeline instances to use as the base estimators.
This must not be None or an empty list or else EnsembleMissingPipelinesError will be raised.
final_estimator (Estimator or subclass): The estimator used to combine the base estimators.
cv (int, cross-validation generator or an iterable): Determines the cross-validation splitting strategy used to train final_estimator.
For int/None inputs, if the estimator is a classifier and y is either binary or multiclass, StratifiedKFold is used. In all other cases, KFold is used.
Possible inputs for cv are:
- None: 5-fold cross validation
- int: the number of folds in a (Stratified) KFold
- An scikit-learn cross-validation generator object
- An iterable yielding (train, test) splits
n_jobs (int or None): Non-negative integer describing level of parallelism used for pipelines.
None and 1 are equivalent. If set to -1, all CPUs are used. For n_jobs below -1, (n_cpus + 1 + n_jobs) are used.
Defaults to -1.
- Note: there could be some multi-process errors thrown for values of `n_jobs != 1`. If this is the case, please use `n_jobs = 1`.
random_seed (int): Seed for the random number generator. Defaults to 0.
"""
model_family = ModelFamily.ENSEMBLE
"""ModelFamily.ENSEMBLE"""
_stacking_estimator_class = None
_default_final_estimator = None
_default_cv = None
def __init__(
self,
input_pipelines=None,
final_estimator=None,
cv=None,
n_jobs=-1,
random_seed=0,
**kwargs,
):
if not input_pipelines:
raise EnsembleMissingPipelinesError(
"`input_pipelines` must not be None or an empty list."
)
if [
pipeline
for pipeline in input_pipelines
if pipeline.model_family in _nonstackable_model_families
]:
raise ValueError(
"Pipelines with any of the following model families cannot be used as base pipelines: {}".format(
_nonstackable_model_families
)
)
parameters = {
"input_pipelines": input_pipelines,
"final_estimator": final_estimator,
"cv": cv,
"n_jobs": n_jobs,
}
parameters.update(kwargs)
if len(set([pipeline.problem_type for pipeline in input_pipelines])) > 1:
raise ValueError("All pipelines must have the same problem type.")
cv = cv or self._default_cv(n_splits=3, random_state=random_seed, shuffle=True)
estimators = [
scikit_learn_wrapped_estimator(pipeline) for pipeline in input_pipelines
]
final_estimator = scikit_learn_wrapped_estimator(
final_estimator or self._default_final_estimator()
)
sklearn_parameters = {
"estimators": [
(f"({idx})", estimator) for idx, estimator in enumerate(estimators)
],
"final_estimator": final_estimator,
"cv": cv,
"n_jobs": n_jobs,
}
sklearn_parameters.update(kwargs)
super().__init__(
parameters=parameters,
component_obj=self._stacking_estimator_class(**sklearn_parameters),
random_seed=random_seed,
)
@property
def feature_importance(self):
"""Not implemented for StackedEnsembleClassifier and StackedEnsembleRegressor"""
raise NotImplementedError(
"feature_importance is not implemented for StackedEnsembleClassifier and StackedEnsembleRegressor"
)
@classproperty
def default_parameters(cls):
"""Returns the default parameters for stacked ensemble classes.
Returns:
dict: default parameters for this component.
"""
return {
"final_estimator": None,
"cv": None,
"n_jobs": -1,
}