|
1
|
|
|
# Author: Simon Blanke |
|
2
|
|
|
# Email: [email protected] |
|
3
|
|
|
# License: MIT License |
|
4
|
|
|
|
|
5
|
|
|
from ._optimizer_attributes import OptimizerAttributes |
|
6
|
|
|
from ._constraint import Constraint |
|
7
|
|
|
|
|
8
|
|
|
from ..optimization_backend.gradient_free_optimizers._objective_function import ( |
|
9
|
|
|
ObjectiveFunction, |
|
10
|
|
|
) |
|
11
|
|
|
from ..optimization_backend.gradient_free_optimizers._hyper_gradient_conv import ( |
|
12
|
|
|
HyperGradientConv, |
|
13
|
|
|
) |
|
14
|
|
|
|
|
15
|
|
|
|
|
16
|
|
|
class Search(OptimizerAttributes): |
|
17
|
|
|
max_time: float |
|
18
|
|
|
nth_process: int |
|
19
|
|
|
|
|
20
|
|
|
def __init__(self, optimizer_class, opt_params): |
|
21
|
|
|
super().__init__() |
|
22
|
|
|
self.optimizer_class = optimizer_class |
|
23
|
|
|
self.opt_params = opt_params |
|
24
|
|
|
|
|
25
|
|
|
def setup( |
|
26
|
|
|
self, |
|
27
|
|
|
experiment, |
|
28
|
|
|
s_space, |
|
29
|
|
|
n_iter, |
|
30
|
|
|
initialize, |
|
31
|
|
|
constraints, |
|
32
|
|
|
pass_through, |
|
33
|
|
|
max_score, |
|
34
|
|
|
early_stopping, |
|
35
|
|
|
random_state, |
|
36
|
|
|
memory, |
|
37
|
|
|
memory_warm_start, |
|
38
|
|
|
): |
|
39
|
|
|
self.experiment = experiment |
|
40
|
|
|
self.s_space = s_space |
|
41
|
|
|
self.n_iter = n_iter |
|
42
|
|
|
|
|
43
|
|
|
self.initialize = initialize |
|
44
|
|
|
self.constraints = constraints |
|
45
|
|
|
self.pass_through = pass_through |
|
46
|
|
|
self.max_score = max_score |
|
47
|
|
|
self.early_stopping = early_stopping |
|
48
|
|
|
self.random_state = random_state |
|
49
|
|
|
self.memory = memory |
|
50
|
|
|
self.memory_warm_start = memory_warm_start |
|
51
|
|
|
|
|
52
|
|
|
def pass_args(self, max_time, nth_process, verbosity): |
|
53
|
|
|
self.max_time = max_time |
|
54
|
|
|
self.nth_process = nth_process |
|
55
|
|
|
|
|
56
|
|
|
if "progress_bar" in verbosity: |
|
57
|
|
|
self.verbosity = ["progress_bar"] |
|
58
|
|
|
else: |
|
59
|
|
|
self.verbosity = [] |
|
60
|
|
|
|
|
61
|
|
|
def convert_results2hyper(self): |
|
62
|
|
|
self.eval_times = sum(self.gfo_optimizer.eval_times) |
|
63
|
|
|
self.iter_times = sum(self.gfo_optimizer.iter_times) |
|
64
|
|
|
|
|
65
|
|
|
if self.gfo_optimizer.best_para is not None: |
|
66
|
|
|
value = self.hg_conv.para2value(self.gfo_optimizer.best_para) |
|
67
|
|
|
position = self.hg_conv.position2value(value) |
|
68
|
|
|
best_para = self.hg_conv.value2para(position) |
|
69
|
|
|
self.best_para = best_para |
|
70
|
|
|
else: |
|
71
|
|
|
self.best_para = None |
|
72
|
|
|
|
|
73
|
|
|
self.best_score = self.gfo_optimizer.best_score |
|
74
|
|
|
self.positions = self.gfo_optimizer.search_data |
|
75
|
|
|
self.search_data = self.hg_conv.positions2results(self.positions) |
|
76
|
|
|
|
|
77
|
|
|
results_dd = self.gfo_optimizer.search_data.drop_duplicates( |
|
78
|
|
|
subset=self.s_space.dim_keys, keep="first" |
|
79
|
|
|
) |
|
80
|
|
|
self.memory_values_df = results_dd[ |
|
81
|
|
|
self.s_space.dim_keys + ["score"] |
|
82
|
|
|
].reset_index(drop=True) |
|
83
|
|
|
|
|
84
|
|
|
def _setup_process(self): |
|
85
|
|
|
self.hg_conv = HyperGradientConv(self.s_space) |
|
86
|
|
|
|
|
87
|
|
|
initialize = self.hg_conv.conv_initialize(self.initialize) |
|
88
|
|
|
search_space_positions = self.s_space.positions |
|
89
|
|
|
|
|
90
|
|
|
# conv warm start for smbo from values into positions |
|
91
|
|
|
if "warm_start_smbo" in self.opt_params: |
|
92
|
|
|
self.opt_params["warm_start_smbo"] = self.hg_conv.conv_memory_warm_start( |
|
93
|
|
|
self.opt_params["warm_start_smbo"] |
|
94
|
|
|
) |
|
95
|
|
|
|
|
96
|
|
|
gfo_constraints = [ |
|
97
|
|
|
Constraint(constraint, self.s_space) for constraint in self.constraints |
|
98
|
|
|
] |
|
99
|
|
|
|
|
100
|
|
|
self.gfo_optimizer = self.optimizer_class( |
|
101
|
|
|
search_space=search_space_positions, |
|
102
|
|
|
initialize=initialize, |
|
103
|
|
|
constraints=gfo_constraints, |
|
104
|
|
|
random_state=self.random_state, |
|
105
|
|
|
nth_process=self.nth_process, |
|
106
|
|
|
**self.opt_params, |
|
107
|
|
|
) |
|
108
|
|
|
|
|
109
|
|
|
self.conv = self.gfo_optimizer.conv |
|
110
|
|
|
|
|
111
|
|
|
def _search(self, p_bar): |
|
112
|
|
|
self._setup_process() |
|
113
|
|
|
|
|
114
|
|
|
memory_warm_start = self.hg_conv.conv_memory_warm_start(self.memory_warm_start) |
|
115
|
|
|
|
|
116
|
|
|
self.experiment.backend_adapter(ObjectiveFunction, self.s_space) |
|
117
|
|
|
|
|
118
|
|
|
self.gfo_optimizer.init_search( |
|
119
|
|
|
self.experiment.gfo_objective_function, |
|
120
|
|
|
self.n_iter, |
|
121
|
|
|
self.max_time, |
|
122
|
|
|
self.max_score, |
|
123
|
|
|
self.early_stopping, |
|
124
|
|
|
self.memory, |
|
125
|
|
|
memory_warm_start, |
|
126
|
|
|
False, |
|
127
|
|
|
) |
|
128
|
|
|
for nth_iter in range(self.n_iter): |
|
129
|
|
|
if p_bar: |
|
130
|
|
|
p_bar.set_description( |
|
131
|
|
|
"[" |
|
132
|
|
|
+ str(self.nth_process) |
|
133
|
|
|
+ "] " |
|
134
|
|
|
+ str(self.experiment.__class__.__name__) |
|
135
|
|
|
+ " (" |
|
136
|
|
|
+ self.optimizer_class.name |
|
137
|
|
|
+ ")", |
|
138
|
|
|
) |
|
139
|
|
|
|
|
140
|
|
|
self.gfo_optimizer.search_step(nth_iter) |
|
141
|
|
|
if self.gfo_optimizer.stop.check(): |
|
142
|
|
|
break |
|
143
|
|
|
|
|
144
|
|
|
if p_bar: |
|
145
|
|
|
p_bar.set_postfix( |
|
146
|
|
|
best_score=str(self.gfo_optimizer.score_best), |
|
147
|
|
|
best_pos=str(self.gfo_optimizer.pos_best), |
|
148
|
|
|
best_iter=str(self.gfo_optimizer.p_bar._best_since_iter), |
|
149
|
|
|
) |
|
150
|
|
|
|
|
151
|
|
|
p_bar.update(1) |
|
152
|
|
|
p_bar.refresh() |
|
153
|
|
|
|
|
154
|
|
|
self.gfo_optimizer.finish_search() |
|
155
|
|
|
|
|
156
|
|
|
self.convert_results2hyper() |
|
157
|
|
|
|
|
158
|
|
|
self._add_result_attributes( |
|
159
|
|
|
self.best_para, |
|
160
|
|
|
self.best_score, |
|
161
|
|
|
self.gfo_optimizer.p_bar._best_since_iter, |
|
162
|
|
|
self.eval_times, |
|
163
|
|
|
self.iter_times, |
|
164
|
|
|
self.search_data, |
|
165
|
|
|
self.gfo_optimizer.random_seed, |
|
166
|
|
|
) |
|
167
|
|
|
|
SimonBlanke /
Hyperactive
