|
1
|
|
|
# Author: Simon Blanke |
|
2
|
|
|
# Email: [email protected] |
|
3
|
|
|
# License: MIT License |
|
4
|
|
|
|
|
5
|
|
|
from ..util import sort_for_best |
|
6
|
|
|
import numpy as np |
|
7
|
|
|
|
|
8
|
|
|
|
|
9
|
|
|
class InitSearchPosition: |
|
10
|
|
|
def __init__(self, space, model, warm_start, scatter_init): |
|
11
|
|
|
self._space_ = space |
|
12
|
|
|
self._model_ = model |
|
13
|
|
|
self.warm_start = warm_start |
|
14
|
|
|
self.scatter_init = scatter_init |
|
15
|
|
|
|
|
16
|
|
|
if self.warm_start: |
|
17
|
|
|
self.n_warm_start_keys = len(list(self.warm_start.keys())) |
|
18
|
|
|
else: |
|
19
|
|
|
self.n_warm_start_keys = 0 |
|
20
|
|
|
|
|
21
|
|
|
def _create_warm_start(self, nth_process): |
|
22
|
|
|
pos = [] |
|
23
|
|
|
|
|
24
|
|
|
for layer_key in self._space_.para_space.keys(): |
|
25
|
|
|
layer_str, para_str = layer_key.rsplit(".", 1) |
|
26
|
|
|
|
|
27
|
|
|
search_position = self._space_.para_space[layer_key].index( |
|
28
|
|
|
*self.warm_start[layer_str][para_str] |
|
29
|
|
|
) |
|
30
|
|
|
|
|
31
|
|
|
pos.append(search_position) |
|
32
|
|
|
|
|
33
|
|
|
return np.array(pos) |
|
34
|
|
|
|
|
35
|
|
|
def _set_start_pos(self, nth_process, X, y): |
|
36
|
|
|
if self.warm_start and self.scatter_init: |
|
37
|
|
|
pos = self._warm_start_scatter_init(nth_process, X, y) |
|
38
|
|
|
elif self.warm_start: |
|
39
|
|
|
pos = self._warm_start(nth_process) |
|
40
|
|
|
elif self.scatter_init: |
|
41
|
|
|
pos = self._scatter_init(nth_process, X, y) |
|
42
|
|
|
else: |
|
43
|
|
|
pos = self._space_.get_random_pos() |
|
44
|
|
|
|
|
45
|
|
|
return pos |
|
46
|
|
|
|
|
47
|
|
|
def _warm_start_scatter_init(self, nth_process, X, y): |
|
48
|
|
|
if self.n_warm_start_keys > nth_process: |
|
49
|
|
|
pos = self._create_warm_start(nth_process) |
|
50
|
|
|
else: |
|
51
|
|
|
pos = self._scatter_init(nth_process, X, y) |
|
52
|
|
|
|
|
53
|
|
|
return pos |
|
54
|
|
|
|
|
55
|
|
|
def _warm_start(self, nth_process): |
|
56
|
|
|
if self.n_warm_start_keys > nth_process: |
|
57
|
|
|
pos = self._create_warm_start(nth_process) |
|
58
|
|
|
else: |
|
59
|
|
|
pos = self._space_.get_random_pos() |
|
60
|
|
|
|
|
61
|
|
|
return pos |
|
62
|
|
|
|
|
63
|
|
|
def _scatter_init(self, nth_process, X, y): |
|
64
|
|
|
pos_list = [] |
|
65
|
|
|
for _ in range(self.scatter_init): |
|
66
|
|
|
pos = self._space_.get_random_pos() |
|
67
|
|
|
pos_list.append(pos) |
|
68
|
|
|
|
|
69
|
|
|
pos_best_list, score_best_list = self._scatter_train(X, y, pos_list) |
|
70
|
|
|
|
|
71
|
|
|
pos_best_sorted, _ = sort_for_best(pos_best_list, score_best_list) |
|
72
|
|
|
|
|
73
|
|
|
nth_best_pos = nth_process - self.n_warm_start_keys |
|
74
|
|
|
|
|
75
|
|
|
return pos_best_sorted[nth_best_pos] |
|
76
|
|
|
|
|
77
|
|
|
def _scatter_train(self, X, y, pos_list): |
|
78
|
|
|
pos_best_list = [] |
|
79
|
|
|
score_best_list = [] |
|
80
|
|
|
|
|
81
|
|
|
X, y = self._get_random_sample(X, y) |
|
82
|
|
|
|
|
83
|
|
|
for pos in pos_list: |
|
84
|
|
|
para = self._space_.pos2para(pos) |
|
85
|
|
|
score, _ = self._model_.train_model(para, X, y) |
|
86
|
|
|
|
|
87
|
|
|
pos_best_list.append(pos) |
|
88
|
|
|
score_best_list.append(score) |
|
89
|
|
|
|
|
90
|
|
|
return pos_best_list, score_best_list |
|
91
|
|
|
|
|
92
|
|
|
def _get_random_sample(self, X, y): |
|
93
|
|
|
if isinstance(X, np.ndarray) and isinstance(y, np.ndarray): |
|
94
|
|
|
n_samples = int(X.shape[0] / self.scatter_init) |
|
95
|
|
|
|
|
96
|
|
|
idx = np.random.choice(np.arange(len(X)), n_samples, replace=False) |
|
97
|
|
|
|
|
98
|
|
|
X_sample = X[idx] |
|
99
|
|
|
y_sample = y[idx] |
|
100
|
|
|
|
|
101
|
|
|
return X_sample, y_sample |
|
102
|
|
|
|
SimonBlanke /
Hyperactive
