|
1
|
|
|
# Author: Simon Blanke |
|
2
|
|
|
# Email: [email protected] |
|
3
|
|
|
# License: MIT License |
|
4
|
|
|
|
|
5
|
|
|
|
|
6
|
|
|
from ..local_opt import HillClimbingOptimizer |
|
7
|
|
|
from .sampling import InitialSampler |
|
8
|
|
|
|
|
9
|
|
|
import warnings |
|
10
|
|
|
import numpy as np |
|
11
|
|
|
|
|
12
|
|
|
np.seterr(divide="ignore", invalid="ignore") |
|
13
|
|
|
|
|
14
|
|
|
|
|
15
|
|
|
class SMBO(HillClimbingOptimizer): |
|
16
|
|
|
def __init__( |
|
17
|
|
|
self, |
|
18
|
|
|
*args, |
|
19
|
|
|
warm_start_smbo=None, |
|
20
|
|
|
max_sample_size=10000000, |
|
21
|
|
|
sampling={"random": 1000000}, |
|
22
|
|
|
replacement=True, |
|
23
|
|
|
**kwargs |
|
24
|
|
|
): |
|
25
|
|
|
super().__init__(*args, **kwargs) |
|
26
|
|
|
|
|
27
|
|
|
self.warm_start_smbo = warm_start_smbo |
|
28
|
|
|
self.max_sample_size = max_sample_size |
|
29
|
|
|
self.sampling = sampling |
|
30
|
|
|
self.replacement = replacement |
|
31
|
|
|
|
|
32
|
|
|
self.sampler = InitialSampler(self.conv, max_sample_size) |
|
33
|
|
|
|
|
34
|
|
|
self.init_warm_start_smbo(warm_start_smbo) |
|
35
|
|
|
|
|
36
|
|
|
def init_warm_start_smbo(self, search_data): |
|
37
|
|
|
if search_data is not None: |
|
38
|
|
|
# filter out nan and inf |
|
39
|
|
|
warm_start_smbo = search_data[ |
|
40
|
|
|
~search_data.isin([np.nan, np.inf, -np.inf]).any(axis=1) |
|
41
|
|
|
] |
|
42
|
|
|
|
|
43
|
|
|
# filter out elements that are not in search space |
|
44
|
|
|
int_idx_list = [] |
|
45
|
|
|
for para_name in self.conv.para_names: |
|
46
|
|
|
search_data_dim = warm_start_smbo[para_name].values |
|
47
|
|
|
search_space_dim = self.conv.search_space[para_name] |
|
48
|
|
|
|
|
49
|
|
|
int_idx = np.nonzero(np.in1d(search_data_dim, search_space_dim))[0] |
|
50
|
|
|
int_idx_list.append(int_idx) |
|
51
|
|
|
|
|
52
|
|
|
intersec = int_idx_list[0] |
|
53
|
|
|
for int_idx in int_idx_list[1:]: |
|
54
|
|
|
intersec = np.intersect1d(intersec, int_idx) |
|
55
|
|
|
warm_start_smbo_f = warm_start_smbo.iloc[intersec] |
|
56
|
|
|
|
|
57
|
|
|
X_sample_values = warm_start_smbo_f[self.conv.para_names].values |
|
58
|
|
|
Y_sample = warm_start_smbo_f["score"].values |
|
59
|
|
|
|
|
60
|
|
|
self.X_sample = self.conv.values2positions(X_sample_values) |
|
61
|
|
|
self.Y_sample = list(Y_sample) |
|
62
|
|
|
|
|
63
|
|
|
else: |
|
64
|
|
|
self.X_sample = [] |
|
65
|
|
|
self.Y_sample = [] |
|
66
|
|
|
|
|
67
|
|
|
def track_X_sample(iterate): |
|
68
|
|
|
def wrapper(self, *args, **kwargs): |
|
69
|
|
|
pos = iterate(self, *args, **kwargs) |
|
70
|
|
|
self.X_sample.append(pos) |
|
71
|
|
|
return pos |
|
72
|
|
|
|
|
73
|
|
|
return wrapper |
|
74
|
|
|
|
|
75
|
|
|
def track_y_sample(evaluate): |
|
76
|
|
|
def wrapper(self, score): |
|
77
|
|
|
evaluate(self, score) |
|
78
|
|
|
|
|
79
|
|
|
if np.isnan(score) or np.isinf(score): |
|
80
|
|
|
del self.X_sample[-1] |
|
81
|
|
|
else: |
|
82
|
|
|
self.Y_sample.append(score) |
|
83
|
|
|
|
|
84
|
|
|
return wrapper |
|
85
|
|
|
|
|
86
|
|
|
def _sampling(self, all_pos_comb): |
|
87
|
|
|
if self.sampling is False: |
|
88
|
|
|
return all_pos_comb |
|
89
|
|
|
elif "random" in self.sampling: |
|
90
|
|
|
return self.random_sampling(all_pos_comb) |
|
91
|
|
|
|
|
92
|
|
|
def random_sampling(self, pos_comb): |
|
93
|
|
|
n_samples = self.sampling["random"] |
|
94
|
|
|
n_pos_comb = pos_comb.shape[0] |
|
95
|
|
|
|
|
96
|
|
|
if n_pos_comb <= n_samples: |
|
97
|
|
|
return pos_comb |
|
98
|
|
|
else: |
|
99
|
|
|
_idx_sample = np.random.choice(n_pos_comb, n_samples, replace=False) |
|
100
|
|
|
pos_comb_sampled = pos_comb[_idx_sample, :] |
|
101
|
|
|
return pos_comb_sampled |
|
102
|
|
|
|
|
103
|
|
|
def _all_possible_pos(self): |
|
104
|
|
|
pos_space = self.sampler.get_pos_space() |
|
105
|
|
|
n_dim = len(pos_space) |
|
106
|
|
|
all_pos_comb = np.array(np.meshgrid(*pos_space)).T.reshape(-1, n_dim) |
|
107
|
|
|
|
|
108
|
|
|
all_pos_comb_constr = [] |
|
109
|
|
|
for pos in all_pos_comb: |
|
110
|
|
|
if self.conv.not_in_constraint(pos): |
|
111
|
|
|
all_pos_comb_constr.append(pos) |
|
112
|
|
|
|
|
113
|
|
|
all_pos_comb_constr = np.array(all_pos_comb_constr) |
|
114
|
|
|
return all_pos_comb_constr |
|
115
|
|
|
|
|
116
|
|
|
def memory_warning(self, max_sample_size): |
|
117
|
|
|
if ( |
|
118
|
|
|
self.conv.search_space_size > self.warnings |
|
119
|
|
|
and max_sample_size > self.warnings |
|
120
|
|
|
): |
|
121
|
|
|
warning_message0 = "\n Warning:" |
|
122
|
|
|
warning_message1 = ( |
|
123
|
|
|
"\n search space size of " |
|
124
|
|
|
+ str(self.conv.search_space_size) |
|
125
|
|
|
+ " exceeding recommended limit." |
|
126
|
|
|
) |
|
127
|
|
|
warning_message3 = "\n Reduce search space size for better performance." |
|
128
|
|
|
warnings.warn(warning_message0 + warning_message1 + warning_message3) |
|
129
|
|
|
|
|
130
|
|
|
@track_X_sample |
|
131
|
|
|
def init_pos(self): |
|
132
|
|
|
return super().init_pos() |
|
133
|
|
|
|
|
134
|
|
|
@HillClimbingOptimizer.track_new_pos |
|
135
|
|
|
@track_X_sample |
|
136
|
|
|
def iterate(self): |
|
137
|
|
|
pos_new = self._propose_location() |
|
138
|
|
|
if not self.replacement: |
|
139
|
|
|
self._remove_position(pos_new) |
|
140
|
|
|
return pos_new |
|
141
|
|
|
|
|
142
|
|
|
def _remove_position(self, position): |
|
143
|
|
|
mask = np.all(self.all_pos_comb == position, axis=1) |
|
144
|
|
|
self.all_pos_comb = self.all_pos_comb[np.invert(mask)] |
|
145
|
|
|
|
|
146
|
|
|
@HillClimbingOptimizer.track_new_score |
|
147
|
|
|
@track_y_sample |
|
148
|
|
|
def evaluate(self, score_new): |
|
149
|
|
|
self._evaluate_new2current(score_new) |
|
150
|
|
|
self._evaluate_current2best() |
|
151
|
|
|
|
|
152
|
|
|
@HillClimbingOptimizer.track_new_score |
|
153
|
|
|
@track_y_sample |
|
154
|
|
|
def evaluate_init(self, score_new): |
|
155
|
|
|
self._evaluate_new2current(score_new) |
|
156
|
|
|
self._evaluate_current2best() |
|
157
|
|
|
|
|
158
|
|
|
def _propose_location(self): |
|
159
|
|
|
try: |
|
160
|
|
|
self._training() |
|
161
|
|
|
except ValueError: |
|
162
|
|
|
print( |
|
163
|
|
|
"Warning: training sequential model failed. Performing random iteration instead." |
|
164
|
|
|
) |
|
165
|
|
|
return self.move_random() |
|
166
|
|
|
|
|
167
|
|
|
exp_imp = self._expected_improvement() |
|
168
|
|
|
|
|
169
|
|
|
index_best = list(exp_imp.argsort()[::-1]) |
|
170
|
|
|
all_pos_comb_sorted = self.pos_comb[index_best] |
|
171
|
|
|
pos_best = all_pos_comb_sorted[0] |
|
172
|
|
|
|
|
173
|
|
|
return pos_best |
|
174
|
|
|
|
SimonBlanke /
Gradient-Free-Optimizers
