|
1
|
|
|
# Author: Simon Blanke |
|
2
|
|
|
# Email: [email protected] |
|
3
|
|
|
# License: MIT License |
|
4
|
|
|
|
|
5
|
|
|
import random |
|
6
|
|
|
import numpy as np |
|
7
|
|
|
|
|
8
|
|
|
from .conv import values2positions |
|
9
|
|
|
|
|
10
|
|
|
|
|
11
|
|
|
class Initializer: |
|
12
|
|
|
def __init__(self, search_space): |
|
13
|
|
|
self.search_space = search_space |
|
14
|
|
|
self.dim_sizes = np.array([array.size - 1 for array in search_space]) |
|
15
|
|
|
|
|
16
|
|
|
def set_pos(self, initialize): |
|
17
|
|
|
init_positions_list = [] |
|
18
|
|
|
|
|
19
|
|
|
if "random" in initialize: |
|
20
|
|
|
positions = self._init_random_search(initialize["random"]) |
|
21
|
|
|
init_positions_list.append(positions) |
|
22
|
|
|
if "grid" in initialize: |
|
23
|
|
|
positions = self._init_grid_search(initialize["grid"]) |
|
24
|
|
|
init_positions_list.append(positions) |
|
25
|
|
|
if "vertices" in initialize: |
|
26
|
|
|
positions = self._init_vertices(initialize["vertices"]) |
|
27
|
|
|
init_positions_list.append(positions) |
|
28
|
|
|
if "warm_start" in initialize: |
|
29
|
|
|
positions = values2positions(self.search_space, initialize["warm_start"]) |
|
30
|
|
|
init_positions_list.append(positions) |
|
31
|
|
|
|
|
32
|
|
|
return [item for sublist in init_positions_list for item in sublist] |
|
33
|
|
|
|
|
34
|
|
|
def _init_random_search(self, n_pos): |
|
35
|
|
|
positions = [] |
|
36
|
|
|
|
|
37
|
|
|
if n_pos == 0: |
|
38
|
|
|
return positions |
|
39
|
|
|
|
|
40
|
|
|
for nth_pos in range(n_pos): |
|
41
|
|
|
pos = np.random.randint(self.dim_sizes, size=self.dim_sizes.shape) |
|
42
|
|
|
positions.append(pos) |
|
43
|
|
|
|
|
44
|
|
|
return positions |
|
45
|
|
|
|
|
46
|
|
|
def _fill_rest_random(self, n_pos, positions): |
|
47
|
|
|
diff_pos = n_pos - len(positions) |
|
48
|
|
|
if diff_pos > 0: |
|
49
|
|
|
pos_rnd = self._init_random_search(n_pos=diff_pos) |
|
50
|
|
|
|
|
51
|
|
|
return positions + pos_rnd |
|
52
|
|
|
else: |
|
53
|
|
|
return positions |
|
54
|
|
|
|
|
55
|
|
|
def _init_grid_search(self, n_pos): |
|
56
|
|
|
positions = [] |
|
57
|
|
|
|
|
58
|
|
|
if n_pos == 0: |
|
59
|
|
|
return positions |
|
60
|
|
|
|
|
61
|
|
|
n_dim = len(self.dim_sizes) |
|
62
|
|
|
p_per_dim = int(np.power(n_pos, 1 / n_dim)) |
|
63
|
|
|
|
|
64
|
|
|
for dim in self.dim_sizes: |
|
65
|
|
|
dim_dist = int(dim / (p_per_dim + 1)) |
|
66
|
|
|
n_points = [n * dim_dist for n in range(1, p_per_dim + 1)] |
|
67
|
|
|
|
|
68
|
|
|
positions.append(n_points) |
|
69
|
|
|
|
|
70
|
|
|
pos_mesh = np.array(np.meshgrid(*positions)) |
|
71
|
|
|
positions = list(pos_mesh.T.reshape(-1, n_dim)) |
|
72
|
|
|
|
|
73
|
|
|
positions = self._fill_rest_random(n_pos, positions) |
|
74
|
|
|
|
|
75
|
|
|
return positions |
|
76
|
|
|
|
|
77
|
|
|
def _init_vertices(self, n_pos): |
|
78
|
|
|
positions = [] |
|
79
|
|
|
|
|
80
|
|
|
if n_pos == 0: |
|
81
|
|
|
return positions |
|
82
|
|
|
|
|
83
|
|
|
zero_array = np.zeros(self.dim_sizes.shape) |
|
84
|
|
|
sub_arrays = [] |
|
85
|
|
|
|
|
86
|
|
|
for dim in self.dim_sizes: |
|
87
|
|
|
sub_array = np.array([0, dim]) |
|
88
|
|
|
sub_arrays.append(sub_array) |
|
89
|
|
|
|
|
90
|
|
|
n_dims = len(self.dim_sizes) |
|
91
|
|
|
pos_comb_np = list(np.array(np.meshgrid(*sub_arrays)).T.reshape(-1, n_dims)) |
|
92
|
|
|
k = min(len(pos_comb_np), n_pos) |
|
93
|
|
|
|
|
94
|
|
|
positions = random.sample(pos_comb_np, k) |
|
95
|
|
|
|
|
96
|
|
|
positions = self._fill_rest_random(n_pos, positions) |
|
97
|
|
|
|
|
98
|
|
|
return positions |
|
99
|
|
|
|
|
100
|
|
|
|
SimonBlanke /
Gradient-Free-Optimizers
