This package implements a randomized hyper parameter search for sklearn (similar to RandomizedSearchCV) but utilizes surrogate adaptive sampling from pySOT. Use this similarly to GridSearchCV with a few extra paramters.
pip install sklearn-surrogatesearchcv
The interface is unimaginative, stylistically similar to RandomizedSearchCV.
class SurrogateSearchCV(object):
"""Surrogate search with cross validation for hyper parameter tuning.
"""
def __init__(self, estimator, n_iter=10, param_def=None, refit=False,
**kwargs):
"""
:param estimator: estimator
:param n_iter: number of iterations to run (default 10)
:param param_def: list of dictionaries, e.g.
[
{
'name': 'alpha',
'integer': False,
'lb': 0.1,
'ub': 0.9,
},
{
'name': 'max_depth',
'integer': True,
'lb': 3,
'ub': 12,
}
]
:param **: every other parameter is the same as GridSearchCV
"""
The result can be found in the following properties of the class instance after running.
params_history_
score_history_
best_params_
best_score_
For a complete example, please refer to src/test/test_basic.py.
A slide about role of surrogate optimization in ml. link