Fork me on GitHub Top

Tuning a scikit-learn estimator with skopt

Gilles Louppe, July 2016
Katie Malone, August 2016

%matplotlib inline
import numpy as np
import matplotlib.pyplot as plt

Problem statement

Tuning the hyper-parameters of a machine learning model is often carried out using an exhaustive exploration of (a subset of) the space all hyper-parameter configurations (e.g., using sklearn.model_selection.GridSearchCV), which often results in a very time consuming operation.

In this notebook, we illustrate how to couple gp_minimize with sklearn's estimators to tune hyper-parameters using sequential model-based optimisation, hopefully resulting in equivalent or better solutions, but within less evaluations.

Note: scikit-optimize provides a dedicated interface for estimator tuning via BayesSearchCV class which has a similar interface to those of GridSearchCV. This class uses functions of skopt to perform hyperparameter search efficiently. For example usage of this class, see "Scikit Learn HPO Wrapper" example notebook.


The first step is to define the objective function we want to minimize, in this case the cross-validation mean absolute error of a gradient boosting regressor over the Boston dataset, as a function of its hyper-parameters:

from sklearn.datasets import load_boston
from sklearn.ensemble import GradientBoostingRegressor
from sklearn.model_selection import cross_val_score

boston = load_boston()
X, y =,
n_features = X.shape[1]

reg = GradientBoostingRegressor(n_estimators=50, random_state=0)

def objective(params):
    max_depth, learning_rate, max_features, min_samples_split, min_samples_leaf = params


    return -np.mean(cross_val_score(reg, X, y, cv=5, n_jobs=-1,

Next, we need to define the bounds of the dimensions of the search space we want to explore:

space  = [(1, 5),                           # max_depth
          (10**-5, 10**0, "log-uniform"),   # learning_rate
          (1, n_features),                  # max_features
          (2, 100),                         # min_samples_split
          (1, 100)]                         # min_samples_leaf

Optimize all the things!

With these two pieces, we are now ready for sequential model-based optimisation. Here we use gaussian process-based optimisation.

from skopt import gp_minimize
res_gp = gp_minimize(objective, space, n_calls=100, random_state=0)

"Best score=%.4f" %
'Best score=2.7903'
print("""Best parameters:
- max_depth=%d
- learning_rate=%.6f
- max_features=%d
- min_samples_split=%d
- min_samples_leaf=%d""" % (res_gp.x[0], res_gp.x[1], 
                            res_gp.x[2], res_gp.x[3], 
Best parameters:
- max_depth=3
- learning_rate=0.130851
- max_features=6
- min_samples_split=2
- min_samples_leaf=1

Convergence plot

from skopt.plots import plot_convergence