|
1
|
|
|
# copyright: hyperactive developers, MIT License (see LICENSE file) |
|
2
|
|
|
|
|
3
|
|
|
from collections.abc import Callable |
|
4
|
|
|
from typing import Union |
|
5
|
|
|
|
|
6
|
|
|
from sklearn.base import BaseEstimator, clone |
|
7
|
|
|
from sklearn.utils.validation import indexable, _check_method_params |
|
8
|
|
|
|
|
9
|
|
|
from hyperactive.experiment.integrations.sklearn_cv import SklearnCvExperiment |
|
10
|
|
|
from hyperactive.integrations.sklearn.best_estimator import ( |
|
11
|
|
|
BestEstimator as _BestEstimator_ |
|
12
|
|
|
) |
|
13
|
|
|
from hyperactive.integrations.sklearn.checks import Checks |
|
14
|
|
|
|
|
15
|
|
|
|
|
16
|
|
|
class OptCV(BaseEstimator, _BestEstimator_, Checks): |
|
17
|
|
|
"""Tuning via any optimizer in the hyperactive API. |
|
18
|
|
|
|
|
19
|
|
|
Parameters |
|
20
|
|
|
---------- |
|
21
|
|
|
estimator : SklearnBaseEstimator |
|
22
|
|
|
The estimator to be tuned. |
|
23
|
|
|
optimizer : hyperactive BaseOptimizer |
|
24
|
|
|
The optimizer to be used for hyperparameter search. |
|
25
|
|
|
scoring : callable or str, default = accuracy_score or mean_squared_error |
|
26
|
|
|
sklearn scoring function or metric to evaluate the model's performance. |
|
27
|
|
|
Default is determined by the type of estimator: |
|
28
|
|
|
``accuracy_score`` for classifiers, and |
|
29
|
|
|
``mean_squared_error`` for regressors, as per sklearn convention |
|
30
|
|
|
through the default ``score`` method of the estimator. |
|
31
|
|
|
refit: bool, optional, default = True |
|
32
|
|
|
Whether to refit the best estimator with the entire dataset. |
|
33
|
|
|
If True, the best estimator is refit with the entire dataset after |
|
34
|
|
|
the optimization process. |
|
35
|
|
|
If False, does not refit, and predict is not available. |
|
36
|
|
|
cv : int or cross-validation generator, default = KFold(n_splits=3, shuffle=True) |
|
37
|
|
|
The number of folds or cross-validation strategy to be used. |
|
38
|
|
|
If int, the cross-validation used is KFold(n_splits=cv, shuffle=True). |
|
39
|
|
|
|
|
40
|
|
|
Example |
|
41
|
|
|
------- |
|
42
|
|
|
Tuning sklearn SVC via grid search |
|
43
|
|
|
|
|
44
|
|
|
1. defining the tuned estimator: |
|
45
|
|
|
>>> from sklearn.svm import SVC |
|
46
|
|
|
>>> from hyperactive.integrations.sklearn import OptCV |
|
47
|
|
|
>>> from hyperactive.opt import GridSearch |
|
48
|
|
|
>>> |
|
49
|
|
|
>>> param_grid = {"kernel": ["linear", "rbf"], "C": [1, 10]} |
|
50
|
|
|
>>> tuned_svc = OptCV(SVC(), GridSearch(param_grid)) |
|
51
|
|
|
|
|
52
|
|
|
2. fitting the tuned estimator: |
|
53
|
|
|
>>> from sklearn.datasets import load_iris |
|
54
|
|
|
>>> from sklearn.model_selection import train_test_split |
|
55
|
|
|
>>> X, y = load_iris(return_X_y=True) |
|
56
|
|
|
>>> X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2) |
|
57
|
|
|
>>> |
|
58
|
|
|
>>> tuned_svc.fit(X_train, y_train) |
|
59
|
|
|
OptCV(...) |
|
60
|
|
|
>>> y_pred = tuned_svc.predict(X_test) |
|
61
|
|
|
|
|
62
|
|
|
3. obtaining best parameters and best estimator |
|
63
|
|
|
>>> best_params = tuned_svc.best_params_ |
|
64
|
|
|
>>> best_estimator = tuned_svc.best_estimator_ |
|
65
|
|
|
""" |
|
66
|
|
|
|
|
67
|
|
|
_required_parameters = ["estimator", "optimizer"] |
|
68
|
|
|
|
|
69
|
|
|
def __init__( |
|
70
|
|
|
self, |
|
71
|
|
|
estimator, |
|
72
|
|
|
optimizer, |
|
73
|
|
|
*, |
|
74
|
|
|
scoring: Union[Callable, str, None] = None, |
|
75
|
|
|
refit: bool = True, |
|
76
|
|
|
cv=None, |
|
77
|
|
|
): |
|
78
|
|
|
super().__init__() |
|
79
|
|
|
|
|
80
|
|
|
self.estimator = estimator |
|
81
|
|
|
self.optimizer = optimizer |
|
82
|
|
|
self.scoring = scoring |
|
83
|
|
|
self.refit = refit |
|
84
|
|
|
self.cv = cv |
|
85
|
|
|
|
|
86
|
|
|
def _refit(self, X, y=None, **fit_params): |
|
87
|
|
|
self.best_estimator_ = clone(self.estimator).set_params( |
|
88
|
|
|
**clone(self.best_params_, safe=False) |
|
89
|
|
|
) |
|
90
|
|
|
|
|
91
|
|
|
self.best_estimator_.fit(X, y, **fit_params) |
|
92
|
|
|
return self |
|
93
|
|
|
|
|
94
|
|
|
def _check_data(self, X, y): |
|
95
|
|
|
X, y = indexable(X, y) |
|
96
|
|
|
if hasattr(self, "_validate_data"): |
|
97
|
|
|
validate_data = self._validate_data |
|
98
|
|
|
else: |
|
99
|
|
|
from sklearn.utils.validation import validate_data |
|
100
|
|
|
|
|
101
|
|
|
return validate_data(X, y) |
|
102
|
|
|
|
|
103
|
|
|
@Checks.verify_fit |
|
104
|
|
|
def fit(self, X, y, **fit_params): |
|
105
|
|
|
"""Fit the model. |
|
106
|
|
|
|
|
107
|
|
|
Parameters |
|
108
|
|
|
---------- |
|
109
|
|
|
X : {array-like, sparse matrix} of shape (n_samples, n_features) |
|
110
|
|
|
Training data. |
|
111
|
|
|
|
|
112
|
|
|
y : array-like of shape (n_samples,) or (n_samples, n_targets) |
|
113
|
|
|
Target values. Will be cast to X's dtype if necessary. |
|
114
|
|
|
|
|
115
|
|
|
Returns |
|
116
|
|
|
------- |
|
117
|
|
|
self : object |
|
118
|
|
|
Fitted Estimator. |
|
119
|
|
|
""" |
|
120
|
|
|
|
|
121
|
|
|
X, y = self._check_data(X, y) |
|
122
|
|
|
|
|
123
|
|
|
fit_params = _check_method_params(X, params=fit_params) |
|
124
|
|
|
|
|
125
|
|
|
experiment = SklearnCvExperiment( |
|
126
|
|
|
estimator=self.estimator, |
|
127
|
|
|
scoring=self.scoring, |
|
128
|
|
|
cv=self.cv, |
|
129
|
|
|
X=X, |
|
130
|
|
|
y=y, |
|
131
|
|
|
) |
|
132
|
|
|
self.scorer_ = experiment.scorer_ |
|
133
|
|
|
|
|
134
|
|
|
optimizer = self.optimizer.clone() |
|
135
|
|
|
optimizer.set_params(experiment=experiment) |
|
136
|
|
|
best_params = optimizer.run() |
|
137
|
|
|
|
|
138
|
|
|
self.best_params_ = best_params |
|
139
|
|
|
self.best_estimator_ = clone(self.estimator).set_params(**best_params) |
|
140
|
|
|
|
|
141
|
|
|
if self.refit: |
|
142
|
|
|
self._refit(X, y, **fit_params) |
|
143
|
|
|
|
|
144
|
|
|
return self |
|
145
|
|
|
|
|
146
|
|
|
def score(self, X, y=None, **params): |
|
147
|
|
|
"""Return the score on the given data, if the estimator has been refit. |
|
148
|
|
|
|
|
149
|
|
|
This uses the score defined by ``scoring`` where provided, and the |
|
150
|
|
|
``best_estimator_.score`` method otherwise. |
|
151
|
|
|
|
|
152
|
|
|
Parameters |
|
153
|
|
|
---------- |
|
154
|
|
|
X : array-like of shape (n_samples, n_features) |
|
155
|
|
|
Input data, where `n_samples` is the number of samples and |
|
156
|
|
|
`n_features` is the number of features. |
|
157
|
|
|
|
|
158
|
|
|
y : array-like of shape (n_samples, n_output) \ |
|
159
|
|
|
or (n_samples,), default=None |
|
160
|
|
|
Target relative to X for classification or regression; |
|
161
|
|
|
None for unsupervised learning. |
|
162
|
|
|
|
|
163
|
|
|
**params : dict |
|
164
|
|
|
Parameters to be passed to the underlying scorer(s). |
|
165
|
|
|
|
|
166
|
|
|
Returns |
|
167
|
|
|
------- |
|
168
|
|
|
score : float |
|
169
|
|
|
The score defined by ``scoring`` if provided, and the |
|
170
|
|
|
``best_estimator_.score`` method otherwise. |
|
171
|
|
|
""" |
|
172
|
|
|
return self.scorer_(self.best_estimator_, X, y, **params) |
|
173
|
|
|
|
|
174
|
|
|
@property |
|
175
|
|
|
def fit_successful(self): |
|
176
|
|
|
self._fit_successful |
|
177
|
|
|
|
SimonBlanke /
Hyperactive
