|
1
|
|
|
"""Quasi-Monte Carlo optimizer.""" |
|
2
|
|
|
# copyright: hyperactive developers, MIT License (see LICENSE file) |
|
3
|
|
|
|
|
4
|
|
|
from .._adapters._base_optuna_adapter import _BaseOptunaAdapter |
|
5
|
|
|
|
|
6
|
|
|
|
|
7
|
|
|
class QMCOptimizer(_BaseOptunaAdapter): |
|
8
|
|
|
"""Quasi-Monte Carlo optimizer. |
|
9
|
|
|
|
|
10
|
|
|
Parameters |
|
11
|
|
|
---------- |
|
12
|
|
|
param_space : dict[str, tuple or list or optuna distributions] |
|
13
|
|
|
The search space to explore. Dictionary with parameter names |
|
14
|
|
|
as keys and either tuples/lists of (low, high) or |
|
15
|
|
|
optuna distribution objects as values. |
|
16
|
|
|
n_trials : int, default=100 |
|
17
|
|
|
Number of optimization trials. |
|
18
|
|
|
initialize : dict[str, int], default=None |
|
19
|
|
|
The method to generate initial positions. A dictionary with |
|
20
|
|
|
the following key literals and the corresponding value type: |
|
21
|
|
|
{"grid": int, "vertices": int, "random": int, "warm_start": list[dict]} |
|
22
|
|
|
random_state : None, int, default=None |
|
23
|
|
|
If None, create a new random state. If int, create a new random state |
|
24
|
|
|
seeded with the value. |
|
25
|
|
|
early_stopping : int, default=None |
|
26
|
|
|
Number of trials after which to stop if no improvement. |
|
27
|
|
|
max_score : float, default=None |
|
28
|
|
|
Maximum score threshold. Stop optimization when reached. |
|
29
|
|
|
qmc_type : str, default="sobol" |
|
30
|
|
|
Type of QMC sequence. Options: "sobol", "halton". |
|
31
|
|
|
scramble : bool, default=True |
|
32
|
|
|
Whether to scramble the QMC sequence. |
|
33
|
|
|
experiment : BaseExperiment, optional |
|
34
|
|
|
The experiment to optimize parameters for. |
|
35
|
|
|
Optional, can be passed later via ``set_params``. |
|
36
|
|
|
|
|
37
|
|
|
Examples |
|
38
|
|
|
-------- |
|
39
|
|
|
Basic usage of QMCOptimizer with a scikit-learn experiment: |
|
40
|
|
|
|
|
41
|
|
|
>>> from hyperactive.experiment.integrations import SklearnCvExperiment |
|
42
|
|
|
>>> from hyperactive.opt.optuna import QMCOptimizer |
|
43
|
|
|
>>> from sklearn.datasets import load_iris |
|
44
|
|
|
>>> from sklearn.svm import SVC |
|
45
|
|
|
>>> X, y = load_iris(return_X_y=True) |
|
46
|
|
|
>>> sklearn_exp = SklearnCvExperiment(estimator=SVC(), X=X, y=y) |
|
47
|
|
|
>>> param_space = { |
|
48
|
|
|
... "C": (0.01, 10), |
|
49
|
|
|
... "gamma": (0.0001, 10), |
|
50
|
|
|
... } |
|
51
|
|
|
>>> optimizer = QMCOptimizer( |
|
52
|
|
|
... param_space=param_space, n_trials=50, experiment=sklearn_exp |
|
53
|
|
|
... ) |
|
54
|
|
|
>>> best_params = optimizer.run() |
|
55
|
|
|
""" |
|
56
|
|
|
|
|
57
|
|
|
_tags = { |
|
58
|
|
|
"info:name": "Quasi-Monte Carlo Optimizer", |
|
59
|
|
|
"info:local_vs_global": "global", |
|
60
|
|
|
"info:explore_vs_exploit": "explore", |
|
61
|
|
|
"info:compute": "low", |
|
62
|
|
|
"python_dependencies": ["optuna"], |
|
63
|
|
|
} |
|
64
|
|
|
|
|
65
|
|
|
def __init__( |
|
66
|
|
|
self, |
|
67
|
|
|
param_space=None, |
|
68
|
|
|
n_trials=100, |
|
69
|
|
|
initialize=None, |
|
70
|
|
|
random_state=None, |
|
71
|
|
|
early_stopping=None, |
|
72
|
|
|
max_score=None, |
|
73
|
|
|
qmc_type="sobol", |
|
74
|
|
|
scramble=True, |
|
75
|
|
|
experiment=None, |
|
76
|
|
|
): |
|
77
|
|
|
self.qmc_type = qmc_type |
|
78
|
|
|
self.scramble = scramble |
|
79
|
|
|
|
|
80
|
|
|
super().__init__( |
|
81
|
|
|
param_space=param_space, |
|
82
|
|
|
n_trials=n_trials, |
|
83
|
|
|
initialize=initialize, |
|
84
|
|
|
random_state=random_state, |
|
85
|
|
|
early_stopping=early_stopping, |
|
86
|
|
|
max_score=max_score, |
|
87
|
|
|
experiment=experiment, |
|
88
|
|
|
) |
|
89
|
|
|
|
|
90
|
|
|
def _get_optimizer(self): |
|
91
|
|
|
"""Get the QMC optimizer. |
|
92
|
|
|
|
|
93
|
|
|
Returns |
|
94
|
|
|
------- |
|
95
|
|
|
optimizer |
|
96
|
|
|
The Optuna QMCOptimizer instance |
|
97
|
|
|
""" |
|
98
|
|
|
import optuna |
|
99
|
|
|
|
|
100
|
|
|
optimizer_kwargs = { |
|
101
|
|
|
"qmc_type": self.qmc_type, |
|
102
|
|
|
"scramble": self.scramble, |
|
103
|
|
|
} |
|
104
|
|
|
|
|
105
|
|
|
if self.random_state is not None: |
|
106
|
|
|
optimizer_kwargs["seed"] = self.random_state |
|
107
|
|
|
|
|
108
|
|
|
return optuna.samplers.QMCSampler(**optimizer_kwargs) |
|
109
|
|
|
|
|
110
|
|
View Code Duplication |
@classmethod |
|
|
|
|
|
|
111
|
|
|
def get_test_params(cls, parameter_set="default"): |
|
112
|
|
|
"""Return testing parameter settings for the optimizer.""" |
|
113
|
|
|
from sklearn.datasets import load_iris |
|
114
|
|
|
from sklearn.linear_model import LogisticRegression |
|
115
|
|
|
|
|
116
|
|
|
from hyperactive.experiment.integrations import SklearnCvExperiment |
|
117
|
|
|
|
|
118
|
|
|
# Test case 1: Halton sequence without scrambling |
|
119
|
|
|
params = super().get_test_params(parameter_set) |
|
120
|
|
|
params[0].update( |
|
121
|
|
|
{ |
|
122
|
|
|
"qmc_type": "halton", |
|
123
|
|
|
"scramble": False, |
|
124
|
|
|
} |
|
125
|
|
|
) |
|
126
|
|
|
|
|
127
|
|
|
# Test case 2: Sobol sequence with scrambling |
|
128
|
|
|
X, y = load_iris(return_X_y=True) |
|
129
|
|
|
lr_exp = SklearnCvExperiment( |
|
130
|
|
|
estimator=LogisticRegression(random_state=42, max_iter=1000), X=X, y=y |
|
131
|
|
|
) |
|
132
|
|
|
|
|
133
|
|
|
mixed_param_space = { |
|
134
|
|
|
"C": (0.01, 100), # Continuous |
|
135
|
|
|
"penalty": [ |
|
136
|
|
|
"l1", |
|
137
|
|
|
"l2", |
|
138
|
|
|
], # Categorical - removed elasticnet to avoid solver conflicts |
|
139
|
|
|
"solver": ["liblinear", "saga"], # Categorical |
|
140
|
|
|
} |
|
141
|
|
|
|
|
142
|
|
|
params.append( |
|
143
|
|
|
{ |
|
144
|
|
|
"param_space": mixed_param_space, |
|
145
|
|
|
"n_trials": 16, # Power of 2 for better QMC properties |
|
146
|
|
|
"experiment": lr_exp, |
|
147
|
|
|
"qmc_type": "sobol", # Different sequence type |
|
148
|
|
|
"scramble": True, # With scrambling for randomization |
|
149
|
|
|
} |
|
150
|
|
|
) |
|
151
|
|
|
|
|
152
|
|
|
# Test case 3: Different sampler configuration with same experiment |
|
153
|
|
|
params.append( |
|
154
|
|
|
{ |
|
155
|
|
|
"param_space": mixed_param_space, |
|
156
|
|
|
"n_trials": 8, # Power of 2, good for QMC |
|
157
|
|
|
"experiment": lr_exp, |
|
158
|
|
|
"qmc_type": "halton", # Different QMC type |
|
159
|
|
|
"scramble": False, |
|
160
|
|
|
} |
|
161
|
|
|
) |
|
162
|
|
|
|
|
163
|
|
|
return params |
|
164
|
|
|
|
SimonBlanke /
Hyperactive
Loading history...