|
1
|
|
|
"""Experiment adapter for sklearn cross-validation experiments.""" |
|
2
|
|
|
# copyright: hyperactive developers, MIT License (see LICENSE file) |
|
3
|
|
|
|
|
4
|
|
|
from sklearn import clone |
|
5
|
|
|
from sklearn.metrics import check_scoring |
|
6
|
|
|
from sklearn.model_selection import cross_validate |
|
7
|
|
|
from sklearn.utils.validation import _num_samples |
|
8
|
|
|
|
|
9
|
|
|
from hyperactive.base import BaseExperiment |
|
10
|
|
|
|
|
11
|
|
|
class SklearnCvExperiment(BaseExperiment): |
|
12
|
|
|
"""Experiment adapter for sklearn cross-validation experiments. |
|
13
|
|
|
|
|
14
|
|
|
This class is used to perform cross-validation experiments using a given |
|
15
|
|
|
sklearn estimator. It allows for hyperparameter tuning and evaluation of |
|
16
|
|
|
the model's performance using cross-validation. |
|
17
|
|
|
|
|
18
|
|
|
The score returned is the mean of the cross-validation scores, |
|
19
|
|
|
of applying cross-validation to ``estimator`` with the parameters given in |
|
20
|
|
|
``score`` ``params``. |
|
21
|
|
|
|
|
22
|
|
|
The cross-validation performed is specified by the ``cv`` parameter, |
|
23
|
|
|
and the scoring metric is specified by the ``scoring`` parameter. |
|
24
|
|
|
The ``X`` and ``y`` parameters are the input data and target values, |
|
25
|
|
|
which are used in fit/predict cross-validation. |
|
26
|
|
|
|
|
27
|
|
|
Parameters |
|
28
|
|
|
---------- |
|
29
|
|
|
estimator : sklearn estimator |
|
30
|
|
|
The estimator to be used for the experiment. |
|
31
|
|
|
X : array-like, shape (n_samples, n_features) |
|
32
|
|
|
The input data for the model. |
|
33
|
|
|
y : array-like, shape (n_samples,) or (n_samples, n_outputs) |
|
34
|
|
|
The target values for the model. |
|
35
|
|
|
cv : int or cross-validation generator, default = KFold(n_splits=3, shuffle=True) |
|
36
|
|
|
The number of folds or cross-validation strategy to be used. |
|
37
|
|
|
If int, the cross-validation used is KFold(n_splits=cv, shuffle=True). |
|
38
|
|
|
scoring : callable or str, default = accuracy_score or mean_squared_error |
|
39
|
|
|
sklearn scoring function or metric to evaluate the model's performance. |
|
40
|
|
|
Default is determined by the type of estimator: |
|
41
|
|
|
``accuracy_score`` for classifiers, and |
|
42
|
|
|
``mean_squared_error`` for regressors, as per sklearn convention |
|
43
|
|
|
through the default ``score`` method of the estimator. |
|
44
|
|
|
|
|
45
|
|
|
Example |
|
46
|
|
|
------- |
|
47
|
|
|
>>> from hyperactive.experiment.integrations import SklearnCvExperiment |
|
48
|
|
|
>>> from sklearn.datasets import load_iris |
|
49
|
|
|
>>> from sklearn.svm import SVC |
|
50
|
|
|
>>> from sklearn.metrics import accuracy_score |
|
51
|
|
|
>>> from sklearn.model_selection import KFold |
|
52
|
|
|
>>> |
|
53
|
|
|
>>> X, y = load_iris(return_X_y=True) |
|
54
|
|
|
>>> |
|
55
|
|
|
>>> sklearn_exp = SklearnCvExperiment( |
|
56
|
|
|
... estimator=SVC(), |
|
57
|
|
|
... scoring=accuracy_score, |
|
58
|
|
|
... cv=KFold(n_splits=3, shuffle=True), |
|
59
|
|
|
... X=X, |
|
60
|
|
|
... y=y, |
|
61
|
|
|
... ) |
|
62
|
|
|
>>> params = {"C": 1.0, "kernel": "linear"} |
|
63
|
|
|
>>> score, add_info = sklearn_exp.score(params) |
|
64
|
|
|
|
|
65
|
|
|
For default choices of ``scoring`` and ``cv``: |
|
66
|
|
|
>>> sklearn_exp = SklearnCvExperiment( |
|
67
|
|
|
... estimator=SVC(), |
|
68
|
|
|
... X=X, |
|
69
|
|
|
... y=y, |
|
70
|
|
|
... ) |
|
71
|
|
|
>>> params = {"C": 1.0, "kernel": "linear"} |
|
72
|
|
|
>>> score, add_info = sklearn_exp.score(params) |
|
73
|
|
|
|
|
74
|
|
|
Quick call without metadata return or dictionary: |
|
75
|
|
|
>>> score = sklearn_exp(C=1.0, kernel="linear") |
|
76
|
|
|
""" |
|
77
|
|
|
|
|
78
|
|
|
def __init__(self, estimator, X, y, scoring=None, cv=None): |
|
79
|
|
|
self.estimator = estimator |
|
80
|
|
|
self.X = X |
|
81
|
|
|
self.y = y |
|
82
|
|
|
self.scoring = scoring |
|
83
|
|
|
self.cv = cv |
|
84
|
|
|
|
|
85
|
|
|
super().__init__() |
|
86
|
|
|
|
|
87
|
|
|
if cv is None: |
|
88
|
|
|
from sklearn.model_selection import KFold |
|
89
|
|
|
|
|
90
|
|
|
self._cv = KFold(n_splits=3, shuffle=True) |
|
91
|
|
|
elif isinstance(cv, int): |
|
92
|
|
|
from sklearn.model_selection import KFold |
|
93
|
|
|
|
|
94
|
|
|
self._cv = KFold(n_splits=cv, shuffle=True) |
|
95
|
|
|
else: |
|
96
|
|
|
self._cv = cv |
|
97
|
|
|
|
|
98
|
|
|
# check if scoring is a scorer by checking for "estimator" in signature |
|
99
|
|
|
if scoring is None: |
|
100
|
|
|
self._scoring = check_scoring(self.estimator) |
|
101
|
|
|
# check using inspect.signature for "estimator" in signature |
|
102
|
|
|
elif callable(scoring): |
|
103
|
|
|
from inspect import signature |
|
104
|
|
|
|
|
105
|
|
|
if "estimator" in signature(scoring).parameters: |
|
106
|
|
|
self._scoring = scoring |
|
107
|
|
|
else: |
|
108
|
|
|
from sklearn.metrics import make_scorer |
|
109
|
|
|
|
|
110
|
|
|
self._scoring = make_scorer(scoring) |
|
111
|
|
|
self.scorer_ = self._scoring |
|
112
|
|
|
|
|
113
|
|
|
def _paramnames(self): |
|
114
|
|
|
"""Return the parameter names of the search. |
|
115
|
|
|
|
|
116
|
|
|
Returns |
|
117
|
|
|
------- |
|
118
|
|
|
list of str |
|
119
|
|
|
The parameter names of the search parameters. |
|
120
|
|
|
""" |
|
121
|
|
|
return list(self.estimator.get_params().keys()) |
|
122
|
|
|
|
|
123
|
|
|
def _score(self, params): |
|
124
|
|
|
"""Score the parameters. |
|
125
|
|
|
|
|
126
|
|
|
Parameters |
|
127
|
|
|
---------- |
|
128
|
|
|
params : dict with string keys |
|
129
|
|
|
Parameters to score. |
|
130
|
|
|
|
|
131
|
|
|
Returns |
|
132
|
|
|
------- |
|
133
|
|
|
float |
|
134
|
|
|
The score of the parameters. |
|
135
|
|
|
dict |
|
136
|
|
|
Additional metadata about the search. |
|
137
|
|
|
""" |
|
138
|
|
|
estimator = clone(self.estimator) |
|
139
|
|
|
estimator.set_params(**params) |
|
140
|
|
|
|
|
141
|
|
|
cv_results = cross_validate( |
|
142
|
|
|
estimator, |
|
143
|
|
|
self.X, |
|
144
|
|
|
self.y, |
|
145
|
|
|
scoring=self._scoring, |
|
146
|
|
|
cv=self._cv, |
|
147
|
|
|
) |
|
148
|
|
|
|
|
149
|
|
|
add_info_d = { |
|
150
|
|
|
"score_time": cv_results["score_time"], |
|
151
|
|
|
"fit_time": cv_results["fit_time"], |
|
152
|
|
|
"n_test_samples": _num_samples(self.X), |
|
153
|
|
|
} |
|
154
|
|
|
|
|
155
|
|
|
return cv_results["test_score"].mean(), add_info_d |
|
156
|
|
|
|
|
157
|
|
|
@classmethod |
|
158
|
|
|
def get_test_params(cls, parameter_set="default"): |
|
159
|
|
|
"""Return testing parameter settings for the skbase object. |
|
160
|
|
|
|
|
161
|
|
|
``get_test_params`` is a unified interface point to store |
|
162
|
|
|
parameter settings for testing purposes. This function is also |
|
163
|
|
|
used in ``create_test_instance`` and ``create_test_instances_and_names`` |
|
164
|
|
|
to construct test instances. |
|
165
|
|
|
|
|
166
|
|
|
``get_test_params`` should return a single ``dict``, or a ``list`` of ``dict``. |
|
167
|
|
|
|
|
168
|
|
|
Each ``dict`` is a parameter configuration for testing, |
|
169
|
|
|
and can be used to construct an "interesting" test instance. |
|
170
|
|
|
A call to ``cls(**params)`` should |
|
171
|
|
|
be valid for all dictionaries ``params`` in the return of ``get_test_params``. |
|
172
|
|
|
|
|
173
|
|
|
The ``get_test_params`` need not return fixed lists of dictionaries, |
|
174
|
|
|
it can also return dynamic or stochastic parameter settings. |
|
175
|
|
|
|
|
176
|
|
|
Parameters |
|
177
|
|
|
---------- |
|
178
|
|
|
parameter_set : str, default="default" |
|
179
|
|
|
Name of the set of test parameters to return, for use in tests. If no |
|
180
|
|
|
special parameters are defined for a value, will return `"default"` set. |
|
181
|
|
|
|
|
182
|
|
|
Returns |
|
183
|
|
|
------- |
|
184
|
|
|
params : dict or list of dict, default = {} |
|
185
|
|
|
Parameters to create testing instances of the class |
|
186
|
|
|
Each dict are parameters to construct an "interesting" test instance, i.e., |
|
187
|
|
|
`MyClass(**params)` or `MyClass(**params[i])` creates a valid test instance. |
|
188
|
|
|
`create_test_instance` uses the first (or only) dictionary in `params` |
|
189
|
|
|
""" |
|
190
|
|
|
from sklearn.datasets import load_diabetes, load_iris |
|
191
|
|
|
from sklearn.svm import SVC, SVR |
|
192
|
|
|
from sklearn.metrics import accuracy_score, mean_absolute_error |
|
193
|
|
|
from sklearn.model_selection import KFold |
|
194
|
|
|
|
|
195
|
|
|
X, y = load_iris(return_X_y=True) |
|
196
|
|
|
params_classif = { |
|
197
|
|
|
"estimator": SVC(), |
|
198
|
|
|
"scoring": accuracy_score, |
|
199
|
|
|
"cv": KFold(n_splits=3, shuffle=True), |
|
200
|
|
|
"X": X, |
|
201
|
|
|
"y": y, |
|
202
|
|
|
} |
|
203
|
|
|
|
|
204
|
|
|
X, y = load_diabetes(return_X_y=True) |
|
205
|
|
|
params_regress = { |
|
206
|
|
|
"estimator": SVR(), |
|
207
|
|
|
"scoring": mean_absolute_error, |
|
208
|
|
|
"cv": 2, |
|
209
|
|
|
"X": X, |
|
210
|
|
|
"y": y, |
|
211
|
|
|
} |
|
212
|
|
|
|
|
213
|
|
|
X, y = load_diabetes(return_X_y=True) |
|
214
|
|
|
params_all_default = { |
|
215
|
|
|
"estimator": SVR(), |
|
216
|
|
|
"X": X, |
|
217
|
|
|
"y": y, |
|
218
|
|
|
} |
|
219
|
|
|
|
|
220
|
|
|
return [params_classif, params_regress, params_all_default] |
|
221
|
|
|
|
|
222
|
|
|
@classmethod |
|
223
|
|
|
def _get_score_params(self): |
|
224
|
|
|
"""Return settings for testing the score function. Used in tests only. |
|
225
|
|
|
|
|
226
|
|
|
Returns a list, the i-th element corresponds to self.get_test_params()[i]. |
|
227
|
|
|
It should be a valid call for self.score. |
|
228
|
|
|
|
|
229
|
|
|
Returns |
|
230
|
|
|
------- |
|
231
|
|
|
list of dict |
|
232
|
|
|
The parameters to be used for scoring. |
|
233
|
|
|
""" |
|
234
|
|
|
score_params_classif = {"C": 1.0, "kernel": "linear"} |
|
235
|
|
|
score_params_regress = {"C": 1.0, "kernel": "linear"} |
|
236
|
|
|
score_params_defaults = {"C": 1.0, "kernel": "linear"} |
|
237
|
|
|
return [score_params_classif, score_params_regress, score_params_defaults] |
|
238
|
|
|
|