|
1
|
|
|
# Author: Simon Blanke |
|
2
|
|
|
# Email: [email protected] |
|
3
|
|
|
# License: MIT License |
|
4
|
|
|
|
|
5
|
|
|
import numpy as np |
|
6
|
|
|
import pandas as pd |
|
7
|
|
|
|
|
8
|
|
|
|
|
9
|
|
|
class HyperGradientConv: |
|
10
|
|
|
def __init__(self, s_space): |
|
11
|
|
|
self.s_space = s_space |
|
12
|
|
|
|
|
13
|
|
|
def value2position(self, value: list) -> list: |
|
14
|
|
|
position = [] |
|
15
|
|
|
for n, space_dim in enumerate(self.s_space.values_l): |
|
16
|
|
|
pos = np.abs(value[n] - np.array(space_dim)).argmin() |
|
17
|
|
|
position.append(int(pos)) |
|
18
|
|
|
|
|
19
|
|
|
return position |
|
20
|
|
|
|
|
21
|
|
|
def value2para(self, value: list) -> dict: |
|
22
|
|
|
para = {} |
|
23
|
|
|
for key, p_ in zip(self.s_space.dim_keys, value): |
|
24
|
|
|
para[key] = p_ |
|
25
|
|
|
|
|
26
|
|
|
return para |
|
27
|
|
|
|
|
28
|
|
|
def para2value(self, para: dict) -> list: |
|
29
|
|
|
value = [] |
|
30
|
|
|
for para_name in self.s_space.dim_keys: |
|
31
|
|
|
value.append(para[para_name]) |
|
32
|
|
|
|
|
33
|
|
|
return value |
|
34
|
|
|
|
|
35
|
|
|
def position2value(self, position): |
|
36
|
|
|
value = [] |
|
37
|
|
|
|
|
38
|
|
|
for n, space_dim in enumerate(self.s_space.values_l): |
|
39
|
|
|
value.append(space_dim[position[n]]) |
|
40
|
|
|
|
|
41
|
|
|
return value |
|
42
|
|
|
|
|
43
|
|
|
def para_func2str(self, para): |
|
44
|
|
|
para_conv = {} |
|
45
|
|
|
for dim_key in self.s_space.dim_keys: |
|
46
|
|
|
if self.s_space.data_types[dim_key] == "number": |
|
47
|
|
|
continue |
|
48
|
|
|
|
|
49
|
|
|
try: |
|
50
|
|
|
value_conv = para[dim_key].__name__ |
|
51
|
|
|
except: |
|
52
|
|
|
value_conv = para[dim_key] |
|
53
|
|
|
|
|
54
|
|
|
para_conv[dim_key] = value_conv |
|
55
|
|
|
|
|
56
|
|
|
def value_func2str(self, value): |
|
57
|
|
|
try: |
|
58
|
|
|
return value.__name__ |
|
59
|
|
|
except: |
|
60
|
|
|
return value |
|
61
|
|
|
|
|
62
|
|
|
def conv_para(self, para_hyper): |
|
63
|
|
|
para_gfo = {} |
|
64
|
|
|
for para in self.s_space.dim_keys: |
|
65
|
|
|
value_hyper = para_hyper[para] |
|
66
|
|
|
space_dim = list(self.s_space.func2str[para]) |
|
67
|
|
|
|
|
68
|
|
|
if self.s_space.data_types[para] == "number": |
|
69
|
|
|
value_gfo = np.abs(value_hyper - np.array(space_dim)).argmin() |
|
70
|
|
|
else: |
|
71
|
|
|
value_hyper = self.value_func2str(value_hyper) |
|
72
|
|
|
|
|
73
|
|
|
if value_hyper in space_dim: |
|
74
|
|
|
value_gfo = space_dim.index(value_hyper) |
|
75
|
|
|
else: |
|
76
|
|
|
raise ValueError( |
|
77
|
|
|
"'{}' was not found in '{}'".format(value_hyper, para) |
|
78
|
|
|
) |
|
79
|
|
|
|
|
80
|
|
|
para_gfo[para] = value_gfo |
|
81
|
|
|
return para_gfo |
|
82
|
|
|
|
|
83
|
|
|
def conv_initialize(self, initialize): |
|
84
|
|
|
if "warm_start" in list(initialize.keys()): |
|
85
|
|
|
warm_start_l = initialize["warm_start"] |
|
86
|
|
|
warm_start_gfo = [] |
|
87
|
|
|
for warm_start in warm_start_l: |
|
88
|
|
|
para_gfo = self.conv_para(warm_start) |
|
89
|
|
|
warm_start_gfo.append(para_gfo) |
|
90
|
|
|
|
|
91
|
|
|
initialize["warm_start"] = warm_start_gfo |
|
92
|
|
|
|
|
93
|
|
|
return initialize |
|
94
|
|
|
|
|
95
|
|
|
def get_list_positions(self, list1_values, search_dim): |
|
96
|
|
|
list_positions = [] |
|
97
|
|
|
|
|
98
|
|
|
for value2 in list1_values: |
|
99
|
|
|
list_positions.append(search_dim.index(value2)) |
|
100
|
|
|
|
|
101
|
|
|
return list_positions |
|
102
|
|
|
|
|
103
|
|
|
def values2positions(self, values, search_dim): |
|
104
|
|
|
return np.array(search_dim).searchsorted(values) |
|
105
|
|
|
|
|
106
|
|
|
def positions2results(self, positions): |
|
107
|
|
|
results_dict = {} |
|
108
|
|
|
|
|
109
|
|
|
for para_name in self.s_space.dim_keys: |
|
110
|
|
|
values_list = self.s_space[para_name] |
|
111
|
|
|
pos_ = positions[para_name].values |
|
112
|
|
|
values_ = [values_list[idx] for idx in pos_] |
|
113
|
|
|
results_dict[para_name] = values_ |
|
114
|
|
|
|
|
115
|
|
|
results = pd.DataFrame.from_dict(results_dict) |
|
116
|
|
|
|
|
117
|
|
|
diff_list = np.setdiff1d(positions.columns, results.columns) |
|
118
|
|
|
results[diff_list] = positions[diff_list] |
|
119
|
|
|
|
|
120
|
|
|
return results |
|
121
|
|
|
|
|
122
|
|
|
def conv_memory_warm_start(self, results): |
|
123
|
|
|
if results is None: |
|
124
|
|
|
return results |
|
125
|
|
|
|
|
126
|
|
|
results.reset_index(inplace=True, drop=True) |
|
127
|
|
|
|
|
128
|
|
|
df_positions_dict = {} |
|
129
|
|
|
for dim_key in self.s_space.dim_keys: |
|
130
|
|
|
result_dim_values = list(results[dim_key].values) |
|
131
|
|
|
search_dim = self.s_space.func2str[dim_key] |
|
132
|
|
|
|
|
133
|
|
|
if self.s_space.data_types[dim_key] == "object": |
|
134
|
|
|
result_dim_values_tmp = [] |
|
135
|
|
|
for value in result_dim_values: |
|
136
|
|
|
try: |
|
137
|
|
|
value = value.__name__ |
|
138
|
|
|
except: |
|
139
|
|
|
pass |
|
140
|
|
|
|
|
141
|
|
|
result_dim_values_tmp.append(value) |
|
142
|
|
|
|
|
143
|
|
|
result_dim_values = result_dim_values_tmp |
|
144
|
|
|
|
|
145
|
|
|
list1_positions = self.get_list_positions(result_dim_values, search_dim) |
|
146
|
|
|
else: |
|
147
|
|
|
list1_positions = self.values2positions(result_dim_values, search_dim) |
|
148
|
|
|
|
|
149
|
|
|
# remove None |
|
150
|
|
|
# list1_positions_ = [x for x in list1_positions if x is not None] |
|
151
|
|
|
df_positions_dict[dim_key] = list1_positions |
|
152
|
|
|
|
|
153
|
|
|
results_new = pd.DataFrame(df_positions_dict) |
|
154
|
|
|
|
|
155
|
|
|
results_new["score"] = results["score"] |
|
156
|
|
|
results_new.dropna(how="any", inplace=True) |
|
157
|
|
|
|
|
158
|
|
|
return results_new |
|
159
|
|
|
|
SimonBlanke /
Hyperactive
