|
1
|
|
|
# Author: Simon Blanke |
|
2
|
|
|
# Email: [email protected] |
|
3
|
|
|
# License: MIT License |
|
4
|
|
|
|
|
5
|
|
|
import random |
|
6
|
|
|
import numpy as np |
|
7
|
|
|
from scipy.spatial.distance import cdist |
|
8
|
|
|
|
|
9
|
|
|
from ..smb_opt.smbo import SMBO |
|
10
|
|
|
from ..local_opt import HillClimbingOptimizer |
|
11
|
|
|
|
|
12
|
|
|
|
|
13
|
|
|
class SubSpace: |
|
14
|
|
|
def __init__(self, search_space): |
|
15
|
|
|
self.search_space = search_space |
|
16
|
|
|
|
|
17
|
|
|
self.score = None |
|
18
|
|
|
self.center_pos = self.center_pos_() |
|
19
|
|
|
self.biggest_dim = self.biggest_dim_() |
|
20
|
|
|
|
|
21
|
|
|
def center_pos_(self): |
|
22
|
|
|
center_pos = [] |
|
23
|
|
|
|
|
24
|
|
|
for dim in list(self.search_space.keys()): |
|
25
|
|
|
dim_array = self.search_space[dim] |
|
26
|
|
|
array_size = dim_array.shape[0] |
|
27
|
|
|
center_idx = int(array_size / 2) |
|
28
|
|
|
|
|
29
|
|
|
center_pos.append(dim_array[center_idx]) |
|
30
|
|
|
|
|
31
|
|
|
return np.array(center_pos).astype(int) |
|
32
|
|
|
|
|
33
|
|
|
def biggest_dim_(self): |
|
34
|
|
|
largest_dim = None |
|
35
|
|
|
largest_size = 0 |
|
36
|
|
|
|
|
37
|
|
|
for dim in list(self.search_space.keys()): |
|
38
|
|
|
dim_array = self.search_space[dim] |
|
39
|
|
|
array_size = dim_array.shape[0] |
|
40
|
|
|
|
|
41
|
|
|
if array_size == largest_size: |
|
42
|
|
|
if random.randint(0, 1): |
|
43
|
|
|
largest_size = array_size |
|
44
|
|
|
largest_dim = dim |
|
45
|
|
|
|
|
46
|
|
|
elif array_size > largest_size: |
|
47
|
|
|
largest_size = array_size |
|
48
|
|
|
largest_dim = dim |
|
49
|
|
|
|
|
50
|
|
|
return largest_dim |
|
51
|
|
|
|
|
52
|
|
|
def lipschitz_bound_(self, score, K=1): |
|
53
|
|
|
self.score = score |
|
54
|
|
|
|
|
55
|
|
|
furthest_pos_ = [] |
|
56
|
|
|
|
|
57
|
|
|
for dim in list(self.search_space.keys()): |
|
58
|
|
|
dim_array = self.search_space[dim] |
|
59
|
|
|
furthest_pos_.append(dim_array[0]) |
|
60
|
|
|
furthest_pos = np.array(furthest_pos_) |
|
61
|
|
|
|
|
62
|
|
|
dist = cdist( |
|
63
|
|
|
furthest_pos.reshape(1, -1), self.center_pos.reshape(1, -1) |
|
64
|
|
|
) |
|
65
|
|
|
|
|
66
|
|
|
self.lipschitz_bound = score + K * dist |
|
67
|
|
|
|
|
68
|
|
|
|
|
69
|
|
|
class DirectAlgorithm(SMBO): |
|
70
|
|
|
name = "Direct Algorithm" |
|
71
|
|
|
_name_ = "direct_algorithm" |
|
72
|
|
|
__name__ = "DirectAlgorithm" |
|
73
|
|
|
|
|
74
|
|
|
optimizer_type = "sequential" |
|
75
|
|
|
computationally_expensive = True |
|
76
|
|
|
|
|
77
|
|
View Code Duplication |
def __init__( |
|
|
|
|
|
|
78
|
|
|
self, |
|
79
|
|
|
search_space, |
|
80
|
|
|
initialize={"grid": 4, "random": 2, "vertices": 4}, |
|
81
|
|
|
constraints=[], |
|
82
|
|
|
random_state=None, |
|
83
|
|
|
rand_rest_p=0, |
|
84
|
|
|
nth_process=None, |
|
85
|
|
|
warm_start_smbo=None, |
|
86
|
|
|
max_sample_size=10000000, |
|
87
|
|
|
sampling={"random": 1000000}, |
|
88
|
|
|
replacement=True, |
|
89
|
|
|
): |
|
90
|
|
|
super().__init__( |
|
91
|
|
|
search_space=search_space, |
|
92
|
|
|
initialize=initialize, |
|
93
|
|
|
constraints=constraints, |
|
94
|
|
|
random_state=random_state, |
|
95
|
|
|
rand_rest_p=rand_rest_p, |
|
96
|
|
|
nth_process=nth_process, |
|
97
|
|
|
warm_start_smbo=warm_start_smbo, |
|
98
|
|
|
max_sample_size=max_sample_size, |
|
99
|
|
|
sampling=sampling, |
|
100
|
|
|
replacement=replacement, |
|
101
|
|
|
) |
|
102
|
|
|
|
|
103
|
|
|
self.subspace_l = [] |
|
104
|
|
|
|
|
105
|
|
|
def select_next_subspace(self): |
|
106
|
|
|
for subspace in self.subspace_l: |
|
107
|
|
|
if subspace.score is None: |
|
108
|
|
|
return subspace |
|
109
|
|
|
|
|
110
|
|
|
def split_dim_into_n(self, subspace, n_splits=3): |
|
111
|
|
|
search_space = subspace.search_space |
|
112
|
|
|
dim_array = search_space[subspace.biggest_dim] |
|
113
|
|
|
|
|
114
|
|
|
sub_arrays = np.array_split(dim_array, n_splits) |
|
115
|
|
|
|
|
116
|
|
|
sub_search_space_l = [] |
|
117
|
|
|
for sub_array in sub_arrays: |
|
118
|
|
|
sub_search_space_ = dict(search_space) |
|
119
|
|
|
sub_search_space_[subspace.biggest_dim] = sub_array |
|
120
|
|
|
|
|
121
|
|
|
sub_search_space_l.append(sub_search_space_) |
|
122
|
|
|
|
|
123
|
|
|
for search_space_ in sub_search_space_l: |
|
124
|
|
|
try: |
|
125
|
|
|
self.subspace_l.append(SubSpace(search_space_)) |
|
126
|
|
|
except IndexError: |
|
127
|
|
|
pass |
|
128
|
|
|
|
|
129
|
|
|
self.subspace_l.remove(subspace) |
|
130
|
|
|
|
|
131
|
|
|
def select_subspace(self): |
|
132
|
|
|
lipschitz_bound_max = -np.inf |
|
133
|
|
|
next_subspace = None |
|
134
|
|
|
|
|
135
|
|
|
for subspace in self.subspace_l: |
|
136
|
|
|
if subspace.lipschitz_bound > lipschitz_bound_max: |
|
137
|
|
|
lipschitz_bound_max = subspace.lipschitz_bound |
|
138
|
|
|
next_subspace = subspace |
|
139
|
|
|
|
|
140
|
|
|
# if lipschitz_bound is nan or -inf |
|
141
|
|
|
if next_subspace is None: |
|
142
|
|
|
next_subspace = self.subspace_l[0] |
|
143
|
|
|
|
|
144
|
|
|
return next_subspace |
|
145
|
|
|
|
|
146
|
|
|
def finish_initialization(self): |
|
147
|
|
|
subspace = SubSpace(self.conv.pos_space) |
|
148
|
|
|
self.subspace_l.append(subspace) |
|
149
|
|
|
self.search_state = "iter" |
|
150
|
|
|
|
|
151
|
|
|
@SMBO.track_new_pos |
|
152
|
|
|
@SMBO.track_X_sample |
|
153
|
|
|
def iterate(self): |
|
154
|
|
|
while True: |
|
155
|
|
|
self.current_subspace = self.select_next_subspace() |
|
156
|
|
|
if self.current_subspace: |
|
157
|
|
|
pos = self.current_subspace.center_pos |
|
158
|
|
|
if self.conv.not_in_constraint(pos): |
|
159
|
|
|
return pos |
|
160
|
|
|
|
|
161
|
|
|
else: |
|
162
|
|
|
self.current_subspace = self.select_subspace() |
|
163
|
|
|
self.split_dim_into_n(self.current_subspace) |
|
164
|
|
|
|
|
165
|
|
|
pos = self.subspace_l[-1].center_pos |
|
166
|
|
|
if self.conv.not_in_constraint(pos): |
|
167
|
|
|
return pos |
|
168
|
|
|
|
|
169
|
|
|
return self.move_climb(pos, epsilon_mod=0.3) |
|
170
|
|
|
|
|
171
|
|
|
@SMBO.track_new_score |
|
172
|
|
|
def evaluate(self, score_new): |
|
173
|
|
|
if self.pos_best is None: |
|
174
|
|
|
self.pos_best = self.pos_new |
|
175
|
|
|
self.pos_current = self.pos_new |
|
176
|
|
|
|
|
177
|
|
|
self.score_best = score_new |
|
178
|
|
|
self.score_current = score_new |
|
179
|
|
|
|
|
180
|
|
|
if self.search_state == "iter": |
|
181
|
|
|
self.current_subspace.lipschitz_bound_(score_new) |
|
182
|
|
|
|
SimonBlanke /
Gradient-Free-Optimizers
Loading history...