-
Notifications
You must be signed in to change notification settings - Fork 2
/
de_main.py
211 lines (193 loc) · 10.5 KB
/
de_main.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
import argparse
import logging
import os
import numpy as np
from ipec.cnn.evaluator import initialise_cnn_evaluator
from ipec.cnn.layers import initialise_cnn_layers_3_bytes, initialise_cnn_layers_with_xavier_weights
from ipec.de.agent import save_agent, load_agent
from ipec.de.population import initialise_cnn_population
def main(args):
_filter_args(args)
if args.optimise == 1:
_optimise_learned_agent(args)
else:
_de_search(args)
def _optimise_learned_agent(args):
"""
optimise the learned agent
:param args: arguments
"""
if args.log_file is None:
log_file_path = 'log/ipde_cnn_optimise.log'
else:
log_file_path = args.log_file
tensorboard_path = None
if args.use_tensorboard == 1:
tensorboard_path = os.path.join(os.path.splitext(log_file_path)[0], 'tensorboard')
if not os.path.exists(tensorboard_path):
os.makedirs(tensorboard_path)
logging.basicConfig(filename=log_file_path, level=logging.DEBUG)
logging.info('===Load data - dataset:%s, mode:%s===', args.dataset, args.mode)
loaded_data = _load_data(args.dataset, args.mode)
logging.info('===Data loaded===')
logging.info('===Started===')
if loaded_data is not None:
evaluator = initialise_cnn_evaluator(training_epoch=args.training_epoch,
training_data=loaded_data.train['images'],
training_label=loaded_data.train['labels'],
validation_data=loaded_data.validation['images'],
validation_label=loaded_data.validation['labels'],
max_gpu=args.max_gpu,
first_gpu_id=args.first_gpu_id,
class_num=args.class_num,
regularise=args.regularise,
dropout=args.dropout,
mean_centre=7,
mean_divisor=80,
stddev_divisor=16,
test_data=loaded_data.test['images'],
test_label=loaded_data.test['labels'],
optimise=True,
tensorboard_path=tensorboard_path)
else:
evaluator = initialise_cnn_evaluator(training_epoch=args.training_epoch,
max_gpu=args.max_gpu,
first_gpu_id=args.first_gpu_id,
class_num=args.class_num,
regularise=args.regularise,
dropout=args.dropout,
mean_centre=7,
mean_divisor=80,
stddev_divisor=16,
optimise=True,
tensorboard_path=tensorboard_path)
loaded_agent = load_agent(args.gbest_file)
evaluator.eval(loaded_agent)
logging.info('===Finished===')
def _de_search(args):
"""
use IPDE to search the best agent
:param args: arguments
"""
if args.log_file is None:
logging.basicConfig(filename='log/ipde_cnn.log', level=logging.DEBUG)
else:
logging.basicConfig(filename=args.log_file, level=logging.DEBUG)
logging.info('===Load data - dataset:%s, mode:%s===', args.dataset, args.mode)
loaded_data = _load_data(args.dataset, args.mode, args.partial_dataset)
logging.info('===Data loaded===')
logging.info('===Started===')
if loaded_data is not None:
evaluator = initialise_cnn_evaluator(training_epoch=args.training_epoch,
training_data=loaded_data.train['images'],
training_label=loaded_data.train['labels'],
validation_data=loaded_data.validation['images'],
validation_label=loaded_data.validation['labels'],
max_gpu=args.max_gpu,
first_gpu_id=args.first_gpu_id,
class_num=args.class_num,
regularise=args.regularise,
dropout=args.dropout,
mean_centre=7,
mean_divisor=80,
stddev_divisor=16,
test_data=loaded_data.test['images'],
test_label=loaded_data.test['labels'])
else:
evaluator = initialise_cnn_evaluator(training_epoch=args.training_epoch,
max_gpu=args.max_gpu,
first_gpu_id=args.first_gpu_id,
class_num=args.class_num,
regularise=args.regularise,
dropout=args.dropout,
mean_centre=7,
mean_divisor=80,
stddev_divisor=16
)
if args.ip_structure == 1:
layers = initialise_cnn_layers_3_bytes()
elif args.ip_structure == 2:
layers = initialise_cnn_layers_with_xavier_weights()
else:
layers = None
de_pop = initialise_cnn_population(pop_size=args.pop_size, agent_length=args.agent_length,
evaluator=evaluator, f=args.f_weight, cr=args.cr,
max_fully_connected_length=args.max_full,
layers=layers, max_generation=args.max_generation)
best_agent = de_pop.evolve()
save_agent(best_agent, args.gbest_file)
logging.info('===Finished===')
def _load_data(dataset_name, mode, partial_dataset=None):
"""
load the dataset
:param dataset_name: dataset name
:param mode: mode
:return: loaded data
"""
loaded_data = None
from ipec.data.core import DataLoader
DataLoader.mode = mode
DataLoader.partial_dataset = partial_dataset
if dataset_name == 'mb':
from ipec.data.mb import loaded_data
elif dataset_name == 'mbi':
from ipec.data.mbi import loaded_data
elif dataset_name == 'mdrbi':
from ipec.data.mdrbi import loaded_data
elif dataset_name == 'mrb':
from ipec.data.mrb import loaded_data
elif dataset_name == 'mrd':
from ipec.data.mrd import loaded_data
elif dataset_name == 'convex':
from ipec.data.convex import loaded_data
return loaded_data
def _filter_args(args):
"""
filter the arguments
:param args: arguments
"""
args.class_num = int(args.class_num) if args.class_num is not None else None
args.pop_size = int(args.pop_size) if args.pop_size is not None else None
args.agent_length = int(args.agent_length) if args.agent_length is not None else None
args.max_full = int(args.max_full) if args.max_full is not None else None
args.max_generation = int(args.max_generation) if args.max_generation is not None else None
args.training_epoch = int(args.training_epoch) if args.training_epoch is not None else None
args.first_gpu_id = int(args.first_gpu_id) if args.first_gpu_id is not None else None
args.max_gpu = int(args.max_gpu) if args.max_gpu is not None else None
args.optimise = int(args.optimise) if args.optimise is not None else 0
args.f_weight = float(args.f_weight) if args.f_weight is not None else None
args.cr = float(args.cr) if args.cr is not None else None
args.regularise = float(args.regularise) if args.regularise is not None else 0
args.dropout = float(args.dropout) if args.dropout is not None else 0
args.ip_structure = int(args.ip_structure) if args.ip_structure is not None else 0
args.partial_dataset = float(args.partial_dataset) if args.partial_dataset is not None else None
args.use_tensorboard = int(args.use_tensorboard) if args.use_tensorboard is not None else 0
# main entrance
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('-d', '--dataset', help='choose a dataset among mb, mbi, mdrbi, mrb, mrd or convex')
parser.add_argument('-c', '--class_num', help='# of classes for the classification problem')
parser.add_argument('-m', '--mode', help='default:None, 1: production (load full data)')
parser.add_argument('-s', '--pop_size', help='population size')
parser.add_argument('-l', '--agent_length', help='agent max length')
parser.add_argument('--max_full', help='max fully connected layers')
parser.add_argument('--max_generation', help='max fly steps')
parser.add_argument('-e', '--training_epoch', help='training epoch for the evaluation')
parser.add_argument('-f', '--first_gpu_id', help='first gpu id')
parser.add_argument('-g', '--max_gpu', help='max number of gpu')
parser.add_argument('-o', '--optimise',
help='optimise the learned CNN architecture. Default: None. 1: optimise; otherwise IPDE search')
parser.add_argument('--log_file', help='the path of log file')
parser.add_argument('--gbest_file', help='the path of gbest file')
parser.add_argument('--f_weight', help='differential weight')
parser.add_argument('--cr', help='crossover rate')
parser.add_argument('-r', '--regularise', help='weight regularisation hyper-parameter. ')
parser.add_argument('--dropout', help='enable dropout and set dropout rate')
parser.add_argument('--ip_structure',
help='IP structure. default: 5 bytes, 1: 3 bytes, 2: 2 bytes with xavier weight initialisation')
parser.add_argument('--partial_dataset',
help='Use partial dataset for learning CNN architecture to speed up the learning process.')
parser.add_argument('--use_tensorboard',
help='indicate whether to use tensorboard. default: not use, 1: use')
args = parser.parse_args()
main(args)