|
1
|
|
|
"""Grid search optimizer.""" |
|
2
|
|
|
|
|
3
|
|
|
from collections.abc import Sequence |
|
4
|
|
|
|
|
5
|
|
|
import numpy as np |
|
6
|
|
|
|
|
7
|
|
|
from sklearn.model_selection import ParameterGrid, ParameterSampler, check_cv |
|
8
|
|
|
|
|
9
|
|
|
from hyperactive.base import BaseOptimizer |
|
10
|
|
|
|
|
11
|
|
|
|
|
12
|
|
|
class GridSearch(BaseOptimizer): |
|
13
|
|
|
"""Grid search optimizer. |
|
14
|
|
|
|
|
15
|
|
|
Parameters |
|
16
|
|
|
---------- |
|
17
|
|
|
experiment : BaseExperiment, optional |
|
18
|
|
|
The experiment to optimize parameters for. |
|
19
|
|
|
Optional, can be passed later in ``add_search``. |
|
20
|
|
|
param_grid : dict[str, list] |
|
21
|
|
|
The search space to explore. A dictionary with parameter |
|
22
|
|
|
names as keys and a numpy array as values. |
|
23
|
|
|
error_score : float, default=np.nan |
|
24
|
|
|
The score to assign if an error occurs during the evaluation of a parameter set. |
|
25
|
|
|
|
|
26
|
|
|
Example |
|
27
|
|
|
------- |
|
28
|
|
|
Grid search applied to scikit-learn parameter tuning: |
|
29
|
|
|
|
|
30
|
|
|
1. defining the experiment to optimize: |
|
31
|
|
|
>>> from hyperactive.experiment.integrations import SklearnCvExperiment |
|
32
|
|
|
>>> from sklearn.datasets import load_iris |
|
33
|
|
|
>>> from sklearn.svm import SVC |
|
34
|
|
|
>>> |
|
35
|
|
|
>>> X, y = load_iris(return_X_y=True) |
|
36
|
|
|
>>> |
|
37
|
|
|
>>> sklearn_exp = SklearnCvExperiment( |
|
38
|
|
|
... estimator=SVC(), |
|
39
|
|
|
... X=X, |
|
40
|
|
|
... y=y, |
|
41
|
|
|
... ) |
|
42
|
|
|
|
|
43
|
|
|
2. setting up the grid search optimizer: |
|
44
|
|
|
>>> from hyperactive.opt import GridSearch |
|
45
|
|
|
>>> param_grid = { |
|
46
|
|
|
... "C": [0.01, 0.1, 1, 10], |
|
47
|
|
|
... "gamma": [0.0001, 0.01, 0.1, 1, 10], |
|
48
|
|
|
... } |
|
49
|
|
|
>>> grid_search = GridSearch(sklearn_exp, param_grid=param_grid) |
|
50
|
|
|
|
|
51
|
|
|
3. running the grid search: |
|
52
|
|
|
>>> best_params = grid_search.run() |
|
53
|
|
|
|
|
54
|
|
|
Best parameters can also be accessed via the attributes: |
|
55
|
|
|
>>> best_params = grid_search.best_params_ |
|
56
|
|
|
""" |
|
57
|
|
|
|
|
58
|
|
|
def __init__( |
|
59
|
|
|
self, |
|
60
|
|
|
experiment=None, |
|
61
|
|
|
param_grid=None, |
|
62
|
|
|
error_score=np.nan, |
|
63
|
|
|
): |
|
64
|
|
|
self.experiment = experiment |
|
65
|
|
|
self.param_grid = param_grid |
|
66
|
|
|
self.error_score = error_score |
|
67
|
|
|
|
|
68
|
|
|
super().__init__() |
|
69
|
|
|
|
|
70
|
|
|
def _check_param_grid(self, param_grid): |
|
71
|
|
|
"""_check_param_grid from sklearn 1.0.2, before it was removed.""" |
|
72
|
|
|
if hasattr(param_grid, "items"): |
|
73
|
|
|
param_grid = [param_grid] |
|
74
|
|
|
|
|
75
|
|
|
for p in param_grid: |
|
76
|
|
|
for name, v in p.items(): |
|
77
|
|
|
if isinstance(v, np.ndarray) and v.ndim > 1: |
|
78
|
|
|
raise ValueError("Parameter array should be one-dimensional.") |
|
79
|
|
|
|
|
80
|
|
|
if isinstance(v, str) or not isinstance(v, (np.ndarray, Sequence)): |
|
81
|
|
|
raise ValueError( |
|
82
|
|
|
f"Parameter grid for parameter ({name}) needs to" |
|
83
|
|
|
f" be a list or numpy array, but got ({type(v)})." |
|
84
|
|
|
" Single values need to be wrapped in a list" |
|
85
|
|
|
" with one element." |
|
86
|
|
|
) |
|
87
|
|
|
|
|
88
|
|
|
if len(v) == 0: |
|
89
|
|
|
raise ValueError( |
|
90
|
|
|
f"Parameter values for parameter ({name}) need " |
|
91
|
|
|
"to be a non-empty sequence." |
|
92
|
|
|
) |
|
93
|
|
|
|
|
94
|
|
|
def _run(self, experiment, param_grid, error_score): |
|
95
|
|
|
"""Run the optimization search process.""" |
|
96
|
|
|
self._check_param_grid(param_grid) |
|
97
|
|
|
candidate_params = list(ParameterGrid(param_grid)) |
|
98
|
|
|
|
|
99
|
|
|
scores = [] |
|
100
|
|
|
for candidate_param in candidate_params: |
|
101
|
|
|
try: |
|
102
|
|
|
score = experiment(**candidate_param) |
|
103
|
|
|
except Exception: # noqa: B904 |
|
104
|
|
|
# Catch all exceptions and assign error_score |
|
105
|
|
|
score = error_score |
|
106
|
|
|
scores.append(score) |
|
107
|
|
|
|
|
108
|
|
|
best_index = np.argmin(scores) |
|
109
|
|
|
best_params = candidate_params[best_index] |
|
110
|
|
|
|
|
111
|
|
|
self.best_index_ = best_index |
|
112
|
|
|
self.best_params_ = best_params |
|
113
|
|
|
self.best_score_ = scores[best_index] |
|
114
|
|
|
|
|
115
|
|
|
return best_params |
|
116
|
|
|
|
|
117
|
|
|
@classmethod |
|
118
|
|
|
def get_test_params(cls, parameter_set="default"): |
|
119
|
|
|
"""Return testing parameter settings for the skbase object. |
|
120
|
|
|
|
|
121
|
|
|
``get_test_params`` is a unified interface point to store |
|
122
|
|
|
parameter settings for testing purposes. This function is also |
|
123
|
|
|
used in ``create_test_instance`` and ``create_test_instances_and_names`` |
|
124
|
|
|
to construct test instances. |
|
125
|
|
|
|
|
126
|
|
|
``get_test_params`` should return a single ``dict``, or a ``list`` of ``dict``. |
|
127
|
|
|
|
|
128
|
|
|
Each ``dict`` is a parameter configuration for testing, |
|
129
|
|
|
and can be used to construct an "interesting" test instance. |
|
130
|
|
|
A call to ``cls(**params)`` should |
|
131
|
|
|
be valid for all dictionaries ``params`` in the return of ``get_test_params``. |
|
132
|
|
|
|
|
133
|
|
|
The ``get_test_params`` need not return fixed lists of dictionaries, |
|
134
|
|
|
it can also return dynamic or stochastic parameter settings. |
|
135
|
|
|
|
|
136
|
|
|
Parameters |
|
137
|
|
|
---------- |
|
138
|
|
|
parameter_set : str, default="default" |
|
139
|
|
|
Name of the set of test parameters to return, for use in tests. If no |
|
140
|
|
|
special parameters are defined for a value, will return `"default"` set. |
|
141
|
|
|
|
|
142
|
|
|
Returns |
|
143
|
|
|
------- |
|
144
|
|
|
params : dict or list of dict, default = {} |
|
145
|
|
|
Parameters to create testing instances of the class |
|
146
|
|
|
Each dict are parameters to construct an "interesting" test instance, i.e., |
|
147
|
|
|
`MyClass(**params)` or `MyClass(**params[i])` creates a valid test instance. |
|
148
|
|
|
`create_test_instance` uses the first (or only) dictionary in `params` |
|
149
|
|
|
""" |
|
150
|
|
|
from hyperactive.experiment.integrations import SklearnCvExperiment |
|
151
|
|
|
|
|
152
|
|
|
sklearn_exp = SklearnCvExperiment.create_test_instance() |
|
153
|
|
|
param_grid = { |
|
154
|
|
|
"C": [0.01, 0.1, 1, 10], |
|
155
|
|
|
"gamma": [0.0001, 0.01, 0.1, 1, 10], |
|
156
|
|
|
} |
|
157
|
|
|
params_sklearn = { |
|
158
|
|
|
"experiment": sklearn_exp, |
|
159
|
|
|
"param_grid": param_grid, |
|
160
|
|
|
} |
|
161
|
|
|
|
|
162
|
|
|
from hyperactive.experiment.toy import Ackley |
|
163
|
|
|
|
|
164
|
|
|
ackley_exp = Ackley.create_test_instance() |
|
165
|
|
|
param_grid = { |
|
166
|
|
|
"x0": np.linspace(-5, 5, 10), |
|
167
|
|
|
"x1": np.linspace(-5, 5, 10), |
|
168
|
|
|
} |
|
169
|
|
|
params_ackley = { |
|
170
|
|
|
"experiment": ackley_exp, |
|
171
|
|
|
"param_grid": param_grid, |
|
172
|
|
|
} |
|
173
|
|
|
|
|
174
|
|
|
return [params_sklearn, params_ackley] |
|
175
|
|
|
|
SimonBlanke /
Hyperactive
