|
1
|
|
|
# Author: Simon Blanke |
|
2
|
|
|
# Email: [email protected] |
|
3
|
|
|
# License: MIT License |
|
4
|
|
|
|
|
5
|
|
|
import warnings |
|
6
|
|
|
import numpy as np |
|
7
|
|
|
from itertools import compress |
|
8
|
|
|
|
|
9
|
|
|
np.seterr(divide="ignore", invalid="ignore") |
|
10
|
|
|
|
|
11
|
|
|
from ..base_optimizer import BaseOptimizer |
|
12
|
|
|
from ...search import Search |
|
13
|
|
|
|
|
14
|
|
|
|
|
15
|
|
|
def memory_warning_1(search_space_size): |
|
16
|
|
|
if search_space_size > 1000000: |
|
17
|
|
|
warning_message0 = "\n Warning:" |
|
18
|
|
|
warning_message1 = "\n search space too large for smb-optimization." |
|
19
|
|
|
warning_message3 = "\n Please reduce search space size for better performance." |
|
20
|
|
|
print(warning_message0 + warning_message1 + warning_message3) |
|
21
|
|
|
|
|
22
|
|
|
|
|
23
|
|
|
def memory_warning_2(all_pos_comb): |
|
24
|
|
|
all_pos_comb_gbyte = all_pos_comb.nbytes / 1000000000 |
|
25
|
|
|
if all_pos_comb_gbyte > 1: |
|
26
|
|
|
warning_message0 = "\n Warning:" |
|
27
|
|
|
warning_message2 = "\n Memory-load exceeding recommended limit." |
|
28
|
|
|
print(warning_message0 + warning_message2) |
|
29
|
|
|
|
|
30
|
|
|
|
|
31
|
|
|
class SMBO(BaseOptimizer, Search): |
|
32
|
|
|
def __init__( |
|
33
|
|
|
self, |
|
34
|
|
|
search_space, |
|
35
|
|
|
initialize={"grid": 4, "random": 2, "vertices": 4}, |
|
36
|
|
|
warm_start_smbo=None, |
|
37
|
|
|
): |
|
38
|
|
|
super().__init__(search_space, initialize) |
|
39
|
|
|
self.warm_start_smbo = warm_start_smbo |
|
40
|
|
|
|
|
41
|
|
|
search_space_size = 1 |
|
42
|
|
|
for value_ in search_space.values(): |
|
43
|
|
|
search_space_size *= len(value_) |
|
44
|
|
|
|
|
45
|
|
|
self.X_sample = [] |
|
46
|
|
|
self.Y_sample = [] |
|
47
|
|
|
|
|
48
|
|
|
memory_warning_1(search_space_size) |
|
49
|
|
|
self.all_pos_comb = self._all_possible_pos() |
|
50
|
|
|
memory_warning_2(self.all_pos_comb) |
|
51
|
|
|
|
|
52
|
|
|
def init_warm_start_smbo(self): |
|
53
|
|
|
if self.warm_start_smbo is not None: |
|
54
|
|
|
X_sample_values = self.warm_start_smbo[self.conv.para_names].values |
|
55
|
|
|
Y_sample = self.warm_start_smbo["score"].values |
|
56
|
|
|
|
|
57
|
|
|
self.X_sample = self.conv.values2positions(X_sample_values) |
|
58
|
|
|
self.Y_sample = list(Y_sample) |
|
59
|
|
|
|
|
60
|
|
|
# filter out nan |
|
61
|
|
|
mask = ~np.isnan(Y_sample) |
|
62
|
|
|
self.X_sample = list(compress(self.X_sample, mask)) |
|
63
|
|
|
self.Y_sample = list(compress(self.Y_sample, mask)) |
|
64
|
|
|
|
|
65
|
|
|
def track_X_sample(func): |
|
66
|
|
|
def wrapper(self, *args, **kwargs): |
|
67
|
|
|
pos = func(self, *args, **kwargs) |
|
68
|
|
|
self.X_sample.append(pos) |
|
69
|
|
|
return pos |
|
70
|
|
|
|
|
71
|
|
|
return wrapper |
|
72
|
|
|
|
|
73
|
|
|
def _all_possible_pos(self): |
|
74
|
|
|
pos_space = [] |
|
75
|
|
|
for dim_ in self.conv.max_positions: |
|
76
|
|
|
pos_space.append(np.arange(dim_)) |
|
77
|
|
|
|
|
78
|
|
|
n_dim = len(pos_space) |
|
79
|
|
|
return np.array(np.meshgrid(*pos_space)).T.reshape(-1, n_dim) |
|
80
|
|
|
|
|
81
|
|
|
@track_X_sample |
|
82
|
|
|
def init_pos(self, pos): |
|
83
|
|
|
super().init_pos(pos) |
|
84
|
|
|
return pos |
|
85
|
|
|
|
SimonBlanke /
Gradient-Free-Optimizers
