|
1
|
|
|
# Author: Simon Blanke |
|
2
|
|
|
# Email: [email protected] |
|
3
|
|
|
# License: MIT License |
|
4
|
|
|
|
|
5
|
|
|
import time |
|
6
|
|
|
import numpy as np |
|
7
|
|
|
import multiprocessing |
|
8
|
|
|
|
|
9
|
|
|
from functools import partial |
|
10
|
|
|
|
|
11
|
|
|
from .base_positioner import BasePositioner |
|
12
|
|
|
from .util import initialize_search, finish_search_, sort_for_best |
|
13
|
|
|
from meta_learn import HyperactiveWrapper |
|
14
|
|
|
|
|
15
|
|
|
|
|
16
|
|
|
class BaseOptimizer: |
|
17
|
|
|
def __init__(self, _core_, _arg_): |
|
18
|
|
|
|
|
19
|
|
|
""" |
|
20
|
|
|
|
|
21
|
|
|
Parameters |
|
22
|
|
|
---------- |
|
23
|
|
|
|
|
24
|
|
|
search_config: dict |
|
25
|
|
|
A dictionary providing the model and hyperparameter search space for the |
|
26
|
|
|
optimization process. |
|
27
|
|
|
n_iter: int |
|
28
|
|
|
The number of iterations the optimizer performs. |
|
29
|
|
|
metric: string, optional (default: "accuracy") |
|
30
|
|
|
The metric the model is evaluated by. |
|
31
|
|
|
n_jobs: int, optional (default: 1) |
|
32
|
|
|
The number of searches to run in parallel. |
|
33
|
|
|
cv: int, optional (default: 3) |
|
34
|
|
|
The number of folds for the cross validation. |
|
35
|
|
|
verbosity: int, optional (default: 1) |
|
36
|
|
|
Verbosity level. 1 prints out warm_start points and their scores. |
|
37
|
|
|
random_state: int, optional (default: None) |
|
38
|
|
|
Sets the random seed. |
|
39
|
|
|
warm_start: dict, optional (default: False) |
|
40
|
|
|
Dictionary that definies a start point for the optimizer. |
|
41
|
|
|
memory: bool, optional (default: True) |
|
42
|
|
|
A memory, that saves the evaluation during the optimization to save time when |
|
43
|
|
|
optimizer returns to position. |
|
44
|
|
|
scatter_init: int, optional (default: False) |
|
45
|
|
|
Defines the number n of random positions that should be evaluated with 1/n the |
|
46
|
|
|
training data, to find a better initial position. |
|
47
|
|
|
|
|
48
|
|
|
Returns |
|
49
|
|
|
------- |
|
50
|
|
|
None |
|
51
|
|
|
|
|
52
|
|
|
""" |
|
53
|
|
|
|
|
54
|
|
|
self._core_ = _core_ |
|
55
|
|
|
self._arg_ = _arg_ |
|
56
|
|
|
|
|
57
|
|
|
self.search_config = self._core_.search_config |
|
58
|
|
|
self.n_iter = self._core_.n_iter |
|
59
|
|
|
|
|
60
|
|
|
if self._core_.meta_learn: |
|
61
|
|
|
self._meta_ = HyperactiveWrapper(self._core_.search_config) |
|
62
|
|
|
|
|
63
|
|
|
self.pos_list = [] |
|
64
|
|
|
self.score_list = [] |
|
65
|
|
|
|
|
66
|
|
|
def _init_base_positioner(self, _cand_, positioner=None): |
|
67
|
|
|
if positioner: |
|
68
|
|
|
_p_ = positioner(**self._arg_.kwargs_opt) |
|
69
|
|
|
else: |
|
70
|
|
|
_p_ = BasePositioner(**self._arg_.kwargs_opt) |
|
71
|
|
|
|
|
72
|
|
|
_p_.pos_current = _cand_.pos_best |
|
73
|
|
|
_p_.score_current = _cand_.score_best |
|
74
|
|
|
|
|
75
|
|
|
return _p_ |
|
76
|
|
|
|
|
77
|
|
|
def _update_pos(self, _cand_, _p_): |
|
78
|
|
|
_cand_.pos_best = _p_.pos_new |
|
79
|
|
|
_cand_.score_best = _p_.score_new |
|
80
|
|
|
|
|
81
|
|
|
_p_.pos_current = _p_.pos_new |
|
82
|
|
|
_p_.score_current = _p_.score_new |
|
83
|
|
|
|
|
84
|
|
|
return _cand_, _p_ |
|
85
|
|
|
|
|
86
|
|
|
def search(self, nth_process, X, y): |
|
87
|
|
|
self._core_, _cand_ = initialize_search(self._core_, nth_process, X, y) |
|
88
|
|
|
_p_ = self._init_opt_positioner(_cand_, X, y) |
|
89
|
|
|
|
|
90
|
|
|
for i in range(self._core_.n_iter): |
|
91
|
|
|
_cand_ = self._iterate(i, _cand_, _p_, X, y) |
|
92
|
|
|
self._core_.update_p_bar(1, _cand_) |
|
93
|
|
|
|
|
94
|
|
|
run_time = time.time() - self.start_time |
|
95
|
|
|
if self._core_.max_time and run_time > self._core_.max_time: |
|
96
|
|
|
break |
|
97
|
|
|
|
|
98
|
|
|
# get_search_path |
|
99
|
|
|
if self._core_.get_search_path: |
|
100
|
|
|
pos_list = [] |
|
101
|
|
|
score_list = [] |
|
102
|
|
|
if isinstance(_p_, list): |
|
103
|
|
|
for p in _p_: |
|
104
|
|
|
pos_list.append(p.pos_new) |
|
105
|
|
|
score_list.append(p.score_new) |
|
106
|
|
|
|
|
107
|
|
|
pos_list_ = np.array(pos_list) |
|
108
|
|
|
score_list_ = np.array(score_list) |
|
109
|
|
|
|
|
110
|
|
|
self.pos_list.append(pos_list_) |
|
|
|
|
|
|
111
|
|
|
self.score_list.append(score_list_) |
|
|
|
|
|
|
112
|
|
|
else: |
|
113
|
|
|
pos_list.append(_p_.pos_new) |
|
114
|
|
|
score_list.append(_p_.score_new) |
|
115
|
|
|
|
|
116
|
|
|
pos_list_ = np.array(pos_list) |
|
117
|
|
|
score_list_ = np.array(score_list) |
|
118
|
|
|
|
|
119
|
|
|
self.pos_list.append(pos_list_) |
|
120
|
|
|
self.score_list.append(score_list_) |
|
121
|
|
|
|
|
122
|
|
|
_cand_ = finish_search_(self._core_, _cand_, X, y) |
|
123
|
|
|
|
|
124
|
|
|
return _cand_ |
|
125
|
|
|
|
|
126
|
|
|
def _search_multiprocessing(self, X, y): |
|
127
|
|
|
"""Wrapper for the parallel search. Passes integer that corresponds to process number""" |
|
128
|
|
|
pool = multiprocessing.Pool(self._core_.n_jobs) |
|
129
|
|
|
search = partial(self.search, X=X, y=y) |
|
130
|
|
|
|
|
131
|
|
|
_cand_list = pool.map(search, self._core_._n_process_range) |
|
132
|
|
|
|
|
133
|
|
|
return _cand_list |
|
134
|
|
|
|
|
135
|
|
|
def _run_one_job(self, X, y): |
|
136
|
|
|
_cand_ = self.search(0, X, y) |
|
137
|
|
|
|
|
138
|
|
|
self.model_best = _cand_.model_best |
|
139
|
|
|
self.score_best = _cand_.score_best |
|
140
|
|
|
start_point = _cand_._get_warm_start() |
|
141
|
|
|
|
|
142
|
|
|
self.results[self.score_best] = start_point |
|
143
|
|
|
|
|
144
|
|
|
if self._core_.verbosity: |
|
145
|
|
|
print("\nbest para =", start_point) |
|
146
|
|
|
print("score =", self.score_best) |
|
147
|
|
|
|
|
148
|
|
|
if self._core_.meta_learn: |
|
149
|
|
|
self._meta_.collect(X, y, _cand_list=[_cand_]) |
|
150
|
|
|
|
|
151
|
|
|
def _run_multiple_jobs(self, X, y): |
|
152
|
|
|
_cand_list = self._search_multiprocessing(X, y) |
|
153
|
|
|
|
|
154
|
|
|
start_point_list = [] |
|
155
|
|
|
score_best_list = [] |
|
156
|
|
|
model_best_list = [] |
|
157
|
|
|
for _cand_ in _cand_list: |
|
158
|
|
|
model_best = _cand_.model_best |
|
159
|
|
|
score_best = _cand_.score_best |
|
160
|
|
|
start_point = _cand_._get_warm_start() |
|
161
|
|
|
|
|
162
|
|
|
self.results[score_best] = start_point |
|
163
|
|
|
|
|
164
|
|
|
start_point_list.append(start_point) |
|
165
|
|
|
score_best_list.append(score_best) |
|
166
|
|
|
model_best_list.append(model_best) |
|
167
|
|
|
|
|
168
|
|
|
start_point_sorted, score_best_sorted = sort_for_best( |
|
169
|
|
|
start_point_list, score_best_list |
|
170
|
|
|
) |
|
171
|
|
|
|
|
172
|
|
|
model_best_sorted, score_best_sorted = sort_for_best( |
|
173
|
|
|
model_best_list, score_best_list |
|
174
|
|
|
) |
|
175
|
|
|
|
|
176
|
|
|
if self._core_.verbosity: |
|
177
|
|
|
for i in range(int(self._core_.n_jobs / 2)): |
|
178
|
|
|
print("\n") |
|
179
|
|
|
print("\nList of start points (best first):\n") |
|
180
|
|
|
for start_point, score_best in zip(start_point_sorted, score_best_sorted): |
|
181
|
|
|
|
|
182
|
|
|
print("best para =", start_point) |
|
183
|
|
|
print("score =", score_best, "\n") |
|
184
|
|
|
|
|
185
|
|
|
self.score_best = score_best_sorted[0] |
|
186
|
|
|
self.model_best = model_best_sorted[0] |
|
187
|
|
|
|
|
188
|
|
|
def _fit(self, X, y): |
|
189
|
|
|
"""Public method for starting the search with the training data (X, y) |
|
190
|
|
|
|
|
191
|
|
|
Parameters |
|
192
|
|
|
---------- |
|
193
|
|
|
X : array-like or sparse matrix of shape = [n_samples, n_features] |
|
194
|
|
|
|
|
195
|
|
|
y : array-like, shape = [n_samples] or [n_samples, n_outputs] |
|
196
|
|
|
|
|
197
|
|
|
Returns |
|
198
|
|
|
------- |
|
199
|
|
|
None |
|
200
|
|
|
""" |
|
201
|
|
|
self.start_time = time.time() |
|
202
|
|
|
self.results = {} |
|
203
|
|
|
|
|
204
|
|
|
if self._core_.n_jobs == 1: |
|
205
|
|
|
self._run_one_job(X, y) |
|
206
|
|
|
else: |
|
207
|
|
|
self._run_multiple_jobs(X, y) |
|
208
|
|
|
|
SimonBlanke /
Hyperactive
Loading history...