|
1
|
|
|
# Author: Simon Blanke |
|
2
|
|
|
# Email: [email protected] |
|
3
|
|
|
# License: MIT License |
|
4
|
|
|
|
|
5
|
|
|
import os |
|
6
|
|
|
import glob |
|
7
|
|
|
import json |
|
8
|
|
|
import dill |
|
9
|
|
|
import pickle |
|
10
|
|
|
import datetime |
|
11
|
|
|
import hashlib |
|
12
|
|
|
import inspect |
|
13
|
|
|
|
|
14
|
|
|
import numpy as np |
|
15
|
|
|
import pandas as pd |
|
16
|
|
|
|
|
17
|
|
|
|
|
18
|
|
|
class Memory: |
|
19
|
|
|
def __init__(self, _space_, _main_args_, _cand_): |
|
20
|
|
|
self._space_ = _space_ |
|
21
|
|
|
self._main_args_ = _main_args_ |
|
22
|
|
|
|
|
23
|
|
|
self.pos_best = None |
|
24
|
|
|
self.score_best = -np.inf |
|
25
|
|
|
|
|
26
|
|
|
self.memory_type = _main_args_.memory |
|
27
|
|
|
self.memory_dict = {} |
|
28
|
|
|
|
|
29
|
|
|
self.meta_data_found = False |
|
30
|
|
|
|
|
31
|
|
|
|
|
32
|
|
|
class ShortTermMemory(Memory): |
|
33
|
|
|
def __init__(self, _space_, _main_args_, _cand_): |
|
34
|
|
|
super().__init__(_space_, _main_args_, _cand_) |
|
35
|
|
|
|
|
36
|
|
|
|
|
37
|
|
|
class LongTermMemory(Memory): |
|
38
|
|
|
def __init__(self, _space_, _main_args_, _cand_): |
|
39
|
|
|
super().__init__(_space_, _main_args_, _cand_) |
|
40
|
|
|
|
|
41
|
|
|
self.score_col_name = "mean_test_score" |
|
42
|
|
|
|
|
43
|
|
|
current_path = os.path.realpath(__file__) |
|
44
|
|
|
meta_learn_path, _ = current_path.rsplit("/", 1) |
|
45
|
|
|
|
|
46
|
|
|
self.datetime = datetime.datetime.now().strftime("%d.%m.%Y - %H:%M:%S") + "/" |
|
47
|
|
|
func_str = self._get_func_str(_cand_.func_) |
|
48
|
|
|
self.func_path_ = self._get_hash(func_str.encode("utf-8")) + "/" |
|
49
|
|
|
|
|
50
|
|
|
self.meta_path = meta_learn_path + "/meta_data/" |
|
51
|
|
|
self.func_path = self.meta_path + self.func_path_ |
|
52
|
|
|
self.date_path = self.meta_path + self.func_path_ + self.datetime |
|
53
|
|
|
|
|
54
|
|
|
if not os.path.exists(self.date_path): |
|
55
|
|
|
os.makedirs(self.date_path, exist_ok=True) |
|
56
|
|
|
|
|
57
|
|
|
def load_memory(self, model_func): |
|
58
|
|
|
para, score = self._read_func_metadata(model_func) |
|
59
|
|
|
if para is None or score is None: |
|
60
|
|
|
return |
|
61
|
|
|
|
|
62
|
|
|
self._load_data_into_memory(para, score) |
|
63
|
|
|
|
|
64
|
|
|
def save_memory(self, _main_args_, _opt_args_, _cand_): |
|
65
|
|
|
|
|
66
|
|
|
path = self._get_file_path(_cand_.func_) |
|
67
|
|
|
meta_data = self._collect(_cand_) |
|
68
|
|
|
|
|
69
|
|
|
self._save_toCSV(meta_data, path) |
|
70
|
|
|
|
|
71
|
|
|
obj_func_path = self.func_path + "objective_function.py" |
|
72
|
|
|
if not os.path.exists(obj_func_path): |
|
73
|
|
|
file = open(obj_func_path, "w") |
|
74
|
|
|
file.write(self._get_func_str(_cand_.func_)) |
|
75
|
|
|
file.close() |
|
76
|
|
|
|
|
77
|
|
|
search_config_path = self.date_path + "search_config.py" |
|
78
|
|
|
search_config_temp = dict(self._main_args_.search_config) |
|
79
|
|
|
|
|
80
|
|
|
for key in search_config_temp.keys(): |
|
81
|
|
|
if isinstance(key, str): |
|
82
|
|
|
continue |
|
83
|
|
|
search_config_temp[key.__name__] = search_config_temp[key] |
|
84
|
|
|
del search_config_temp[key] |
|
85
|
|
|
|
|
86
|
|
|
search_config_str = "search_config = " + str(search_config_temp) |
|
87
|
|
|
|
|
88
|
|
|
if not os.path.exists(search_config_path): |
|
89
|
|
|
file = open(search_config_path, "w") |
|
90
|
|
|
file.write(search_config_str) |
|
91
|
|
|
file.close() |
|
92
|
|
|
|
|
93
|
|
|
os.chdir(self.date_path) |
|
94
|
|
|
os.system("black search_config.py") |
|
95
|
|
|
os.getcwd() |
|
96
|
|
|
|
|
97
|
|
|
run_data = { |
|
98
|
|
|
"random_state": self._main_args_.random_state, |
|
99
|
|
|
"max_time": self._main_args_.random_state, |
|
100
|
|
|
"n_iter": self._main_args_.n_iter, |
|
101
|
|
|
"optimizer": self._main_args_.optimizer, |
|
102
|
|
|
"n_jobs": self._main_args_.n_jobs, |
|
103
|
|
|
"eval_time": np.array(_cand_.eval_time).sum(), |
|
104
|
|
|
"total_time": _cand_.total_time, |
|
105
|
|
|
} |
|
106
|
|
|
|
|
107
|
|
|
with open("run_data.json", "w") as f: |
|
108
|
|
|
json.dump(run_data, f, indent=4) |
|
109
|
|
|
|
|
110
|
|
|
""" |
|
111
|
|
|
|
|
112
|
|
|
print("_opt_args_.kwargs_opt", _opt_args_.kwargs_opt) |
|
113
|
|
|
|
|
114
|
|
|
opt_para = pd.DataFrame.from_dict(_opt_args_.kwargs_opt, dtype=object) |
|
115
|
|
|
print("opt_para", opt_para) |
|
116
|
|
|
opt_para.to_csv( |
|
117
|
|
|
self.meta_data_path + self.func_path + self.datetime + "opt_para", |
|
118
|
|
|
index=False, |
|
119
|
|
|
) |
|
120
|
|
|
""" |
|
121
|
|
|
|
|
122
|
|
|
def _save_toCSV(self, meta_data_new, path): |
|
123
|
|
|
if os.path.exists(path): |
|
124
|
|
|
meta_data_old = pd.read_csv(path) |
|
125
|
|
|
meta_data = meta_data_old.append(meta_data_new) |
|
126
|
|
|
|
|
127
|
|
|
columns = list(meta_data.columns) |
|
128
|
|
|
noScore = ["mean_test_score", "cv_default_score"] |
|
129
|
|
|
columns_noScore = [c for c in columns if c not in noScore] |
|
130
|
|
|
|
|
131
|
|
|
meta_data = meta_data.drop_duplicates(subset=columns_noScore) |
|
132
|
|
|
else: |
|
133
|
|
|
meta_data = meta_data_new |
|
134
|
|
|
|
|
135
|
|
|
meta_data.to_csv(path, index=False) |
|
136
|
|
|
|
|
137
|
|
|
def _read_func_metadata(self, model_func): |
|
138
|
|
|
paths = self._get_func_data_names() |
|
139
|
|
|
|
|
140
|
|
|
meta_data_list = [] |
|
141
|
|
|
for path in paths: |
|
142
|
|
|
meta_data = pd.read_csv(path) |
|
143
|
|
|
meta_data_list.append(meta_data) |
|
144
|
|
|
self.meta_data_found = True |
|
145
|
|
|
|
|
146
|
|
|
if len(meta_data_list) > 0: |
|
147
|
|
|
meta_data = pd.concat(meta_data_list, ignore_index=True) |
|
148
|
|
|
|
|
149
|
|
|
column_names = meta_data.columns |
|
150
|
|
|
score_name = [name for name in column_names if self.score_col_name in name] |
|
151
|
|
|
|
|
152
|
|
|
para = meta_data.drop(score_name, axis=1) |
|
153
|
|
|
score = meta_data[score_name] |
|
154
|
|
|
|
|
155
|
|
|
print("Loading meta data successful") |
|
156
|
|
|
return para, score |
|
157
|
|
|
|
|
158
|
|
|
else: |
|
159
|
|
|
print("Warning: No meta data found for following function:", model_func) |
|
160
|
|
|
return None, None |
|
161
|
|
|
|
|
162
|
|
|
def _get_opt_meta_data(self): |
|
163
|
|
|
results_dict = {} |
|
164
|
|
|
para_list = [] |
|
165
|
|
|
score_list = [] |
|
166
|
|
|
|
|
167
|
|
|
for key in self.memory_dict.keys(): |
|
168
|
|
|
pos = np.fromstring(key, dtype=int) |
|
169
|
|
|
para = self._space_.pos2para(pos) |
|
170
|
|
|
score = self.memory_dict[key] |
|
171
|
|
|
|
|
172
|
|
|
for key in para.keys(): |
|
173
|
|
|
if ( |
|
174
|
|
|
not isinstance(para[key], int) |
|
175
|
|
|
and not isinstance(para[key], float) |
|
176
|
|
|
and not isinstance(para[key], str) |
|
177
|
|
|
): |
|
178
|
|
|
|
|
179
|
|
|
para_dill = dill.dumps(para[key]) |
|
180
|
|
|
para_hash = self._get_hash(para_dill) |
|
181
|
|
|
|
|
182
|
|
|
with open( |
|
183
|
|
|
self.date_path + str(para_hash) + ".pkl", "wb" |
|
184
|
|
|
) as pickle_file: |
|
185
|
|
|
dill.dump(para_dill, pickle_file) |
|
186
|
|
|
|
|
187
|
|
|
para[key] = para_hash |
|
188
|
|
|
|
|
189
|
|
|
if score != 0: |
|
190
|
|
|
para_list.append(para) |
|
191
|
|
|
score_list.append(score) |
|
192
|
|
|
|
|
193
|
|
|
results_dict["params"] = para_list |
|
194
|
|
|
results_dict["mean_test_score"] = score_list |
|
195
|
|
|
|
|
196
|
|
|
return results_dict |
|
197
|
|
|
|
|
198
|
|
|
def _load_data_into_memory(self, paras, scores): |
|
199
|
|
|
|
|
200
|
|
|
for idx in range(paras.shape[0]): |
|
201
|
|
|
para = paras.iloc[[idx]] |
|
202
|
|
|
|
|
203
|
|
|
pos = self._space_.para2pos(paras.iloc[[idx]], self._get_pkl_hash) |
|
204
|
|
|
pos_str = pos.tostring() |
|
205
|
|
|
|
|
206
|
|
|
score = float(scores.values[idx]) |
|
207
|
|
|
self.memory_dict[pos_str] = score |
|
208
|
|
|
|
|
209
|
|
|
if score > self.score_best: |
|
210
|
|
|
self.score_best = score |
|
211
|
|
|
self.pos_best = pos |
|
212
|
|
|
|
|
213
|
|
|
def _get_para(self): |
|
214
|
|
|
results_dict = self._get_opt_meta_data() |
|
215
|
|
|
|
|
216
|
|
|
return pd.DataFrame(results_dict["params"]) |
|
217
|
|
|
|
|
218
|
|
|
def _get_score(self): |
|
219
|
|
|
results_dict = self._get_opt_meta_data() |
|
220
|
|
|
return pd.DataFrame( |
|
221
|
|
|
results_dict["mean_test_score"], columns=["mean_test_score"] |
|
222
|
|
|
) |
|
223
|
|
|
|
|
224
|
|
|
def _collect(self, _cand_): |
|
225
|
|
|
para_pd = self._get_para() |
|
226
|
|
|
metric_pd = self._get_score() |
|
227
|
|
|
|
|
228
|
|
|
eval_time = pd.DataFrame(_cand_.eval_time, columns=["eval_time"]) |
|
229
|
|
|
md_model = pd.concat( |
|
230
|
|
|
[para_pd, metric_pd, eval_time], axis=1, ignore_index=False |
|
231
|
|
|
) |
|
232
|
|
|
|
|
233
|
|
|
return md_model |
|
234
|
|
|
|
|
235
|
|
|
def _get_hash(self, object): |
|
236
|
|
|
return hashlib.sha1(object).hexdigest() |
|
237
|
|
|
|
|
238
|
|
|
def _get_func_str(self, func): |
|
239
|
|
|
return inspect.getsource(func) |
|
240
|
|
|
|
|
241
|
|
|
def _get_subdirs(self): |
|
242
|
|
|
subdirs = glob.glob(self.func_path + "*/") |
|
243
|
|
|
|
|
244
|
|
|
return subdirs |
|
245
|
|
|
|
|
246
|
|
|
def _get_func_data_names(self): |
|
247
|
|
|
subdirs = self._get_subdirs() |
|
248
|
|
|
|
|
249
|
|
|
path_list = [] |
|
250
|
|
|
for subdir in subdirs: |
|
251
|
|
|
paths = glob.glob(subdir + "*.csv") |
|
252
|
|
|
path_list = path_list + paths |
|
253
|
|
|
|
|
254
|
|
|
return path_list |
|
255
|
|
|
|
|
256
|
|
|
def _get_pkl_hash(self, hash): |
|
257
|
|
|
subdirs = self._get_subdirs() |
|
258
|
|
|
|
|
259
|
|
|
path_list = [] |
|
260
|
|
|
for subdir in subdirs: |
|
261
|
|
|
paths = glob.glob(subdir + hash + "*.pkl") |
|
262
|
|
|
path_list = path_list + paths |
|
263
|
|
|
|
|
264
|
|
|
return path_list |
|
265
|
|
|
|
|
266
|
|
|
def _get_file_path(self, model_func): |
|
267
|
|
|
feature_hash = self._get_hash(self._main_args_.X) |
|
268
|
|
|
label_hash = self._get_hash(self._main_args_.y) |
|
269
|
|
|
|
|
270
|
|
|
if not os.path.exists(self.date_path): |
|
271
|
|
|
os.makedirs(self.date_path) |
|
272
|
|
|
|
|
273
|
|
|
return self.date_path + (feature_hash + "_" + label_hash + ".csv") |
|
274
|
|
|
|
SimonBlanke /
Hyperactive
