|
1
|
|
|
# Author: Simon Blanke |
|
2
|
|
|
# Email: [email protected] |
|
3
|
|
|
# License: MIT License |
|
4
|
|
|
|
|
5
|
|
|
|
|
6
|
|
|
import numpy as np |
|
7
|
|
|
|
|
8
|
|
|
from sklearn.neighbors import KernelDensity |
|
9
|
|
|
from .smbo import SMBO |
|
10
|
|
|
|
|
11
|
|
|
|
|
12
|
|
|
class TreeStructuredParzenEstimators(SMBO): |
|
13
|
|
|
name = "Tree Structured Parzen Estimators" |
|
14
|
|
|
_name_ = "tree_structured_parzen_estimators" |
|
15
|
|
|
__name__ = "TreeStructuredParzenEstimators" |
|
16
|
|
|
|
|
17
|
|
|
optimizer_type = "sequential" |
|
18
|
|
|
computationally_expensive = True |
|
19
|
|
|
|
|
20
|
|
|
def __init__( |
|
21
|
|
|
self, |
|
22
|
|
|
search_space, |
|
23
|
|
|
initialize={"grid": 4, "random": 2, "vertices": 4}, |
|
24
|
|
|
constraints=[], |
|
25
|
|
|
random_state=None, |
|
26
|
|
|
rand_rest_p=0, |
|
27
|
|
|
nth_process=None, |
|
28
|
|
|
warm_start_smbo=None, |
|
29
|
|
|
max_sample_size=10000000, |
|
30
|
|
|
sampling={"random": 1000000}, |
|
31
|
|
|
replacement=True, |
|
32
|
|
|
gamma_tpe=0.2, |
|
33
|
|
|
): |
|
34
|
|
|
super().__init__( |
|
35
|
|
|
search_space=search_space, |
|
36
|
|
|
initialize=initialize, |
|
37
|
|
|
constraints=constraints, |
|
38
|
|
|
random_state=random_state, |
|
39
|
|
|
rand_rest_p=rand_rest_p, |
|
40
|
|
|
nth_process=nth_process, |
|
41
|
|
|
warm_start_smbo=warm_start_smbo, |
|
42
|
|
|
max_sample_size=max_sample_size, |
|
43
|
|
|
sampling=sampling, |
|
44
|
|
|
replacement=replacement, |
|
45
|
|
|
) |
|
46
|
|
|
|
|
47
|
|
|
self.gamma_tpe = gamma_tpe |
|
48
|
|
|
|
|
49
|
|
|
kde_para = { |
|
50
|
|
|
"kernel": "gaussian", |
|
51
|
|
|
"bandwidth": 1, |
|
52
|
|
|
"rtol": 0.001, |
|
53
|
|
|
} |
|
54
|
|
|
|
|
55
|
|
|
self.kd_best = KernelDensity(**kde_para) |
|
56
|
|
|
self.kd_worst = KernelDensity(**kde_para) |
|
57
|
|
|
|
|
58
|
|
|
def finish_initialization(self): |
|
59
|
|
|
self.all_pos_comb = self._all_possible_pos() |
|
60
|
|
|
return super().finish_initialization() |
|
61
|
|
|
|
|
62
|
|
|
def _get_samples(self): |
|
63
|
|
|
n_samples = len(self.X_sample) |
|
64
|
|
|
n_best = max(round(n_samples * self.gamma_tpe), 1) |
|
65
|
|
|
|
|
66
|
|
|
Y_sample = np.array(self.Y_sample) |
|
67
|
|
|
index_best = Y_sample.argsort()[-n_best:] |
|
68
|
|
|
n_worst = int(n_samples - n_best) |
|
69
|
|
|
index_worst = Y_sample.argsort()[:n_worst] |
|
70
|
|
|
|
|
71
|
|
|
best_samples = [self.X_sample[i] for i in index_best] |
|
72
|
|
|
worst_samples = [self.X_sample[i] for i in index_worst] |
|
73
|
|
|
|
|
74
|
|
|
return best_samples, worst_samples |
|
75
|
|
|
|
|
76
|
|
|
def _expected_improvement(self): |
|
77
|
|
|
self.pos_comb = self._sampling(self.all_pos_comb) |
|
78
|
|
|
|
|
79
|
|
|
logprob_best = self.kd_best.score_samples(self.pos_comb) |
|
80
|
|
|
logprob_worst = self.kd_worst.score_samples(self.pos_comb) |
|
81
|
|
|
|
|
82
|
|
|
prob_best = np.exp(logprob_best) |
|
83
|
|
|
prob_worst = np.exp(logprob_worst) |
|
84
|
|
|
|
|
85
|
|
|
WorstOverbest = np.divide( |
|
86
|
|
|
prob_worst, |
|
87
|
|
|
prob_best, |
|
88
|
|
|
out=np.zeros_like(prob_worst), |
|
89
|
|
|
where=prob_worst != 0, |
|
90
|
|
|
) |
|
91
|
|
|
|
|
92
|
|
|
exp_imp_inv = self.gamma_tpe + WorstOverbest * (1 - self.gamma_tpe) |
|
93
|
|
|
exp_imp = 1 / exp_imp_inv |
|
94
|
|
|
|
|
95
|
|
|
return exp_imp |
|
96
|
|
|
|
|
97
|
|
|
def _training(self): |
|
98
|
|
|
best_samples, worst_samples = self._get_samples() |
|
99
|
|
|
|
|
100
|
|
|
self.kd_best.fit(best_samples) |
|
101
|
|
|
self.kd_worst.fit(worst_samples) |
|
102
|
|
|
|
SimonBlanke /
Gradient-Free-Optimizers
