|
1
|
|
|
# Author: Simon Blanke |
|
2
|
|
|
# Email: [email protected] |
|
3
|
|
|
# License: MIT License |
|
4
|
|
|
|
|
5
|
|
|
import time |
|
6
|
|
|
import random |
|
7
|
|
|
import numpy as np |
|
8
|
|
|
import pandas as pd |
|
9
|
|
|
|
|
10
|
|
|
from importlib import import_module |
|
11
|
|
|
|
|
12
|
|
|
|
|
13
|
|
|
optimizer_dict = { |
|
14
|
|
|
"HillClimbing": "HillClimbingOptimizer", |
|
15
|
|
|
"StochasticHillClimbing": "StochasticHillClimbingOptimizer", |
|
16
|
|
|
"TabuSearch": "TabuOptimizer", |
|
17
|
|
|
"RandomSearch": "RandomSearchOptimizer", |
|
18
|
|
|
"RandomRestartHillClimbing": "RandomRestartHillClimbingOptimizer", |
|
19
|
|
|
"RandomAnnealing": "RandomAnnealingOptimizer", |
|
20
|
|
|
"SimulatedAnnealing": "SimulatedAnnealingOptimizer", |
|
21
|
|
|
"StochasticTunneling": "StochasticTunnelingOptimizer", |
|
22
|
|
|
"ParallelTempering": "ParallelTemperingOptimizer", |
|
23
|
|
|
"ParticleSwarm": "ParticleSwarmOptimizer", |
|
24
|
|
|
"EvolutionStrategy": "EvolutionStrategyOptimizer", |
|
25
|
|
|
"Bayesian": "BayesianOptimizer", |
|
26
|
|
|
"TPE": "TreeStructuredParzenEstimators", |
|
27
|
|
|
"DecisionTree": "DecisionTreeOptimizer", |
|
28
|
|
|
} |
|
29
|
|
|
|
|
30
|
|
|
|
|
31
|
|
|
class SearchProcess: |
|
32
|
|
|
def __init__( |
|
33
|
|
|
self, |
|
34
|
|
|
nth_process, |
|
35
|
|
|
verb, |
|
36
|
|
|
objective_function, |
|
37
|
|
|
search_space, |
|
38
|
|
|
n_iter, |
|
39
|
|
|
function_parameter, |
|
40
|
|
|
optimizer, |
|
41
|
|
|
n_jobs, |
|
42
|
|
|
init_para, |
|
43
|
|
|
memory, |
|
44
|
|
|
hyperactive, |
|
45
|
|
|
random_state, |
|
46
|
|
|
): |
|
47
|
|
|
self.nth_process = nth_process |
|
48
|
|
|
self.verb = verb |
|
49
|
|
|
self.objective_function = objective_function |
|
50
|
|
|
self.search_space = search_space |
|
51
|
|
|
self.n_iter = n_iter |
|
52
|
|
|
self.function_parameter = function_parameter |
|
53
|
|
|
self.optimizer = optimizer |
|
54
|
|
|
self.n_jobs = n_jobs |
|
55
|
|
|
self.init_para = init_para |
|
56
|
|
|
self.memory = memory |
|
57
|
|
|
self.hyperactive = hyperactive |
|
58
|
|
|
self.random_state = random_state |
|
59
|
|
|
|
|
60
|
|
|
self._process_arguments() |
|
61
|
|
|
|
|
62
|
|
|
self.iter_times = [] |
|
63
|
|
|
self.eval_times = [] |
|
64
|
|
|
|
|
65
|
|
|
module = import_module("gradient_free_optimizers") |
|
66
|
|
|
self.opt_class = getattr(module, optimizer_dict[optimizer]) |
|
67
|
|
|
|
|
68
|
|
|
self.res = ResultsManager(objective_function, search_space, function_parameter) |
|
69
|
|
|
|
|
70
|
|
|
def _results_dict(self): |
|
71
|
|
|
results_dict = { |
|
72
|
|
|
"eval_times": self.eval_times, |
|
73
|
|
|
"iter_times": self.iter_times, |
|
74
|
|
|
"memory": self.cand.memory_dict_new, |
|
75
|
|
|
"para_best": self.cand.para_best, |
|
76
|
|
|
"score_best": self.cand.score_best, |
|
77
|
|
|
} |
|
78
|
|
|
|
|
79
|
|
|
return results_dict |
|
80
|
|
|
|
|
81
|
|
|
def _time_exceeded(self, start_time, max_time): |
|
82
|
|
|
run_time = time.time() - start_time |
|
83
|
|
|
return max_time and run_time > max_time |
|
84
|
|
|
|
|
85
|
|
|
def _initialize_search(self, nth_process): |
|
86
|
|
|
init_positions = self.cand.init.set_start_pos(self.n_positions) |
|
87
|
|
|
self.opt = self.opt_class(init_positions, self.cand.space.dim, opt_para={}) |
|
88
|
|
|
|
|
89
|
|
|
self.verb.p_bar.init_p_bar(nth_process, self.n_iter, self.objective_function) |
|
90
|
|
|
|
|
91
|
|
|
def _process_arguments(self): |
|
92
|
|
|
self._set_random_seed() |
|
93
|
|
|
|
|
94
|
|
|
if isinstance(self.optimizer, dict): |
|
95
|
|
|
optimizer = list(self.optimizer.keys())[0] |
|
96
|
|
|
self.opt_para = self.optimizer[optimizer] |
|
97
|
|
|
self.optimizer = optimizer |
|
98
|
|
|
|
|
99
|
|
|
self.n_positions = self._get_n_positions() |
|
100
|
|
|
else: |
|
101
|
|
|
self.opt_para = {} |
|
102
|
|
|
self.n_positions = self._get_n_positions() |
|
103
|
|
|
|
|
104
|
|
|
def _get_n_positions(self): |
|
105
|
|
|
n_positions_strings = [ |
|
106
|
|
|
"n_positions", |
|
107
|
|
|
"system_temperatures", |
|
108
|
|
|
"n_particles", |
|
109
|
|
|
"individuals", |
|
110
|
|
|
] |
|
111
|
|
|
|
|
112
|
|
|
n_positions = 1 |
|
113
|
|
|
for n_pos_name in n_positions_strings: |
|
114
|
|
|
if n_pos_name in list(self.opt_para.keys()): |
|
115
|
|
|
n_positions = self.opt_para[n_pos_name] |
|
116
|
|
|
if n_positions == "system_temperatures": |
|
117
|
|
|
n_positions = len(n_positions) |
|
118
|
|
|
|
|
119
|
|
|
return n_positions |
|
120
|
|
|
|
|
121
|
|
|
def _set_random_seed(self): |
|
122
|
|
|
"""Sets the random seed separately for each thread (to avoid getting the same results in each thread)""" |
|
123
|
|
|
if self.random_state is None: |
|
124
|
|
|
self.random_state = np.random.randint(0, high=2 ** 32 - 2) |
|
125
|
|
|
|
|
126
|
|
|
random.seed(self.random_state + self.nth_process) |
|
127
|
|
|
np.random.seed(self.random_state + self.nth_process) |
|
128
|
|
|
|
|
129
|
|
|
def store_memory(self, memory): |
|
130
|
|
|
pass |
|
131
|
|
|
|
|
132
|
|
|
def print_best_para(self): |
|
133
|
|
|
self.verb.info.print_start_point() |
|
134
|
|
|
|
|
135
|
|
|
def search(self, start_time, max_time, nth_process): |
|
136
|
|
|
start_time_search = time.time() |
|
137
|
|
|
self._initialize_search(nth_process) |
|
138
|
|
|
|
|
139
|
|
|
# loop to initialize N positions |
|
140
|
|
|
for nth_init in range(len(self.opt.init_positions)): |
|
141
|
|
|
start_time_iter = time.time() |
|
142
|
|
|
pos_new = self.opt.init_pos(nth_init) |
|
143
|
|
|
|
|
144
|
|
|
start_time_eval = time.time() |
|
145
|
|
|
score_new = self.cand.get_score(pos_new, nth_init) |
|
146
|
|
|
self.eval_times.append(time.time() - start_time_eval) |
|
147
|
|
|
|
|
148
|
|
|
self.opt.evaluate(score_new) |
|
149
|
|
|
self.iter_times.append(time.time() - start_time_iter) |
|
150
|
|
|
|
|
151
|
|
|
# loop to do the iterations |
|
152
|
|
|
for nth_iter in range(len(self.opt.init_positions), self.n_iter): |
|
153
|
|
|
start_time_iter = time.time() |
|
154
|
|
|
pos_new = self.opt.iterate(nth_iter) |
|
155
|
|
|
|
|
156
|
|
|
start_time_eval = time.time() |
|
157
|
|
|
score_new = self.cand.get_score(pos_new, nth_iter) |
|
158
|
|
|
self.eval_times.append(time.time() - start_time_eval) |
|
159
|
|
|
|
|
160
|
|
|
self.opt.evaluate(score_new) |
|
161
|
|
|
self.iter_times.append(time.time() - start_time_search) |
|
162
|
|
|
|
|
163
|
|
|
if self._time_exceeded(start_time, max_time): |
|
164
|
|
|
break |
|
165
|
|
|
|
|
166
|
|
|
self.verb.p_bar.close_p_bar() |
|
167
|
|
|
|
|
168
|
|
|
self.res.memory_dict_new = self.cand.memory_dict_new |
|
169
|
|
|
self.res.results_dict = self._results_dict() |
|
170
|
|
|
|
|
171
|
|
|
return self.res |
|
172
|
|
|
|
|
173
|
|
|
|
|
174
|
|
|
from optimization_metadata import HyperactiveWrapper |
|
175
|
|
|
from ..meta_data.meta_data_path import meta_data_path |
|
176
|
|
|
|
|
177
|
|
|
|
|
178
|
|
|
class ResultsManager: |
|
179
|
|
|
def __init__( |
|
180
|
|
|
self, objective_function, search_space, function_parameter, |
|
181
|
|
|
): |
|
182
|
|
|
self.objective_function = objective_function |
|
183
|
|
|
self.search_space = search_space |
|
184
|
|
|
self.function_parameter = function_parameter |
|
185
|
|
|
|
|
186
|
|
|
self.memory_dict_new = {} |
|
187
|
|
|
|
|
188
|
|
|
self.hypermem = HyperactiveWrapper( |
|
189
|
|
|
main_path=meta_data_path(), |
|
190
|
|
|
X=function_parameter["features"], |
|
191
|
|
|
y=function_parameter["target"], |
|
192
|
|
|
model=self.objective_function, |
|
193
|
|
|
search_space=search_space, |
|
194
|
|
|
) |
|
195
|
|
|
|
|
196
|
|
|
def load_long_term_memory(self): |
|
197
|
|
|
return self.hypermem.load() |
|
198
|
|
|
|
|
199
|
|
|
def save_long_term_memory(self): |
|
200
|
|
|
self.hypermem.save(self.memory_dict_new) |
|
201
|
|
|
|
SimonBlanke /
Hyperactive
