|
1
|
|
|
# Author: Simon Blanke |
|
2
|
|
|
# Email: [email protected] |
|
3
|
|
|
# License: MIT License |
|
4
|
|
|
|
|
5
|
|
|
|
|
6
|
|
|
from ..base_optimizer import BaseOptimizer |
|
7
|
|
|
from ...search import Search |
|
8
|
|
|
|
|
9
|
|
|
import numpy as np |
|
10
|
|
|
from itertools import compress |
|
11
|
|
|
|
|
12
|
|
|
np.seterr(divide="ignore", invalid="ignore") |
|
13
|
|
|
|
|
14
|
|
|
|
|
15
|
|
|
class SMBO(BaseOptimizer, Search): |
|
16
|
|
|
def __init__( |
|
17
|
|
|
self, |
|
18
|
|
|
search_space, |
|
19
|
|
|
initialize={"grid": 4, "random": 2, "vertices": 4}, |
|
20
|
|
|
warm_start_smbo=None, |
|
21
|
|
|
sampling={"random": 100000}, |
|
22
|
|
|
warnings=100000000, |
|
23
|
|
|
): |
|
24
|
|
|
super().__init__(search_space, initialize) |
|
25
|
|
|
self.warm_start_smbo = warm_start_smbo |
|
26
|
|
|
self.sampling = sampling |
|
27
|
|
|
self.warnings = warnings |
|
28
|
|
|
|
|
29
|
|
|
def init_position_combinations(self): |
|
30
|
|
|
search_space_size = 1 |
|
31
|
|
|
for value_ in self.conv.search_space.values(): |
|
32
|
|
|
search_space_size *= len(value_) |
|
33
|
|
|
|
|
34
|
|
|
self.X_sample = [] |
|
35
|
|
|
self.Y_sample = [] |
|
36
|
|
|
|
|
37
|
|
|
if self.warnings: |
|
38
|
|
|
self.memory_warning(search_space_size) |
|
39
|
|
|
self.all_pos_comb = self._all_possible_pos() |
|
40
|
|
|
|
|
41
|
|
|
def init_warm_start_smbo(self): |
|
42
|
|
|
if self.warm_start_smbo is not None: |
|
43
|
|
|
X_sample_values = self.warm_start_smbo[self.conv.para_names].values |
|
44
|
|
|
Y_sample = self.warm_start_smbo["score"].values |
|
45
|
|
|
|
|
46
|
|
|
self.X_sample = self.conv.values2positions(X_sample_values) |
|
47
|
|
|
self.Y_sample = list(Y_sample) |
|
48
|
|
|
|
|
49
|
|
|
# filter out nan |
|
50
|
|
|
mask = ~np.isnan(Y_sample) |
|
51
|
|
|
self.X_sample = list(compress(self.X_sample, mask)) |
|
52
|
|
|
self.Y_sample = list(compress(self.Y_sample, mask)) |
|
53
|
|
|
|
|
54
|
|
|
def track_X_sample(func): |
|
55
|
|
|
def wrapper(self, *args, **kwargs): |
|
56
|
|
|
pos = func(self, *args, **kwargs) |
|
57
|
|
|
self.X_sample.append(pos) |
|
58
|
|
|
return pos |
|
59
|
|
|
|
|
60
|
|
|
return wrapper |
|
61
|
|
|
|
|
62
|
|
|
def random_sampling(self): |
|
63
|
|
|
n_samples = self.sampling["random"] |
|
64
|
|
|
n_pos_comb = self.all_pos_comb.shape[0] |
|
65
|
|
|
|
|
66
|
|
|
if n_pos_comb <= n_samples: |
|
67
|
|
|
return self.all_pos_comb |
|
68
|
|
|
else: |
|
69
|
|
|
_idx_sample = np.random.choice(n_pos_comb, n_samples, replace=False) |
|
70
|
|
|
pos_comb_sampled = self.all_pos_comb[_idx_sample, :] |
|
71
|
|
|
return pos_comb_sampled |
|
72
|
|
|
|
|
73
|
|
|
def _all_possible_pos(self): |
|
74
|
|
|
if self.conv.max_dim < 255: |
|
75
|
|
|
_dtype = np.uint8 |
|
76
|
|
|
elif self.conv.max_dim < 65535: |
|
77
|
|
|
_dtype = np.uint16 |
|
78
|
|
|
elif self.conv.max_dim < 4294967295: |
|
79
|
|
|
_dtype = np.uint32 |
|
80
|
|
|
else: |
|
81
|
|
|
_dtype = np.uint64 |
|
82
|
|
|
|
|
83
|
|
|
pos_space = [] |
|
84
|
|
|
for dim_ in self.conv.dim_sizes: |
|
85
|
|
|
pos_space.append(np.arange(dim_, dtype=_dtype)) |
|
86
|
|
|
|
|
87
|
|
|
n_dim = len(pos_space) |
|
88
|
|
|
return np.array(np.meshgrid(*pos_space)).T.reshape(-1, n_dim) |
|
89
|
|
|
|
|
90
|
|
|
def memory_warning(self, search_space_size): |
|
91
|
|
|
if search_space_size > self.warnings: |
|
92
|
|
|
warning_message0 = "\n Warning:" |
|
93
|
|
|
warning_message1 = ( |
|
94
|
|
|
"\n search space size of " |
|
95
|
|
|
+ str(search_space_size) |
|
96
|
|
|
+ " exceeding recommended limit." |
|
97
|
|
|
) |
|
98
|
|
|
warning_message3 = "\n Reduce search space size for better performance." |
|
99
|
|
|
print(warning_message0 + warning_message1 + warning_message3) |
|
100
|
|
|
|
|
101
|
|
|
@track_X_sample |
|
102
|
|
|
def init_pos(self, pos): |
|
103
|
|
|
super().init_pos(pos) |
|
104
|
|
|
return pos |
|
105
|
|
|
|
SimonBlanke /
Gradient-Free-Optimizers
