-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
executable file
·231 lines (193 loc) · 5.45 KB
/
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
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
import argparse
import os
from copy import deepcopy
from active.app import active_learn
from active.utils import make_dirs, set_logger
SIZE = {
1000: "small",
2500: "medium",
5000: "large",
}
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument(
"--seed", type=int, default=3061, help="random seed for initial labels"
)
parser.add_argument(
"--gpu_id",
type=int,
default=0,
choices=[0, 1, 2, 3],
help="index of the gpu id",
)
parser.add_argument(
"--export_id",
type=str,
default="cifar10",
choices=[
"cifar10",
"cifar100",
"svhn",
],
help="the data to be explored",
)
parser.add_argument(
"--algorithm",
type=str,
default="Badge",
choices=[
"Random",
"KCenter",
"Entropy",
"Badge",
],
help="the active learning algorithm to select subset",
)
parser.add_argument(
"--n_worker", type=int, default=4, help="number of workers to load data"
)
parser.add_argument(
"--n_ev_batch", type=int, default=256, help="evaluate batch size"
)
parser.add_argument(
"--t_epoch",
type=int,
default=50,
help="time interval between epochs to evaluate model",
)
parser.add_argument(
"--r_round",
type=int,
default=0,
help="the resume round of active learning",
)
parser.add_argument(
"--n_round", type=int, default=5, help="number of rounds to proceed"
)
parser.add_argument(
"--n_init", type=int, default=1000, help="number of initial labels"
)
parser.add_argument(
"--n_query", type=int, default=1000, help="number of queries per round"
)
parser.add_argument(
"--percentile",
type=float,
default=0.1,
help="percentile of data variance to be filtered",
)
parser.add_argument(
"--n_filter",
type=int,
default=5,
help="number of filterng step to update hard-level mask",
)
parser.add_argument(
"--n_MCdrop",
type=int,
default=1,
help="number of MC dropout samples",
)
parser.add_argument(
"--dp_rate",
default=0.0,
type=float,
help="dropout rate for the classifier",
)
parser.add_argument("--momentum", default=0.9, type=float, help="momentum")
parser.add_argument(
"--proxy_arc",
type=str,
choices=["res-18", "wrn-28-2", "vgg-16"],
help="the proxy architecture for AL training",
)
parser.add_argument(
"--target_arc",
type=str,
choices=["res-18", "vgg-16"],
help="the target architecture for AL evaluation",
)
parser.add_argument(
"--semi",
action="store_true",
help="whether to use semi-supervised learning",
)
args = parser.parse_args()
net_dir = f"{args.proxy_arc}-{'semi' if args.semi else 'sup'}-{args.target_arc}-sup"
algorithm = f"{args.algorithm}_{args.n_filter}_{args.percentile}_{args.seed}"
args.size = f"{SIZE[args.n_init]}_{SIZE[args.n_query]}"
args.path = f"{args.export_id}/{net_dir}/{algorithm}/{args.size}/"
if args.export_id in ["cifar10", "svhn"]:
args.n_classes = 10
elif args.export_id == "cifar100":
args.n_classes = 100
else:
raise NotImplementedError
return args
def get_semisup_args(args):
ss_args = deepcopy(args)
assert (
args.proxy_arc == "wrn-28-2"
), "Architecture of the proxy is constrained to be 'wrn-28-2'."
ss_args.arc = "wrn-28-2"
ss_args.mu = 7
ss_args.n_tr_batch = 64
ss_args.l2_reg = 5e-4
if args.export_id in ["cifar100", "svhn"]:
ss_args.w_lr = 0.1
ss_args.w_epoch = 50
else:
ss_args.w_epoch = 0
ss_args.n_epoch = 200
ss_args.lr = 0.03
ss_args.cycle = 7 / 16
ss_args.semi = True
ss_args.ema_m = 0.999
ss_args.eta = 0.01
ss_args.lamb = 1.0
return ss_args
def get_sup_args(args, arc):
s_args = deepcopy(args)
s_args.arc = arc
s_args.mu = 1
s_args.n_tr_batch = 128
s_args.l2_reg = 5e-4
s_args.semi = False
s_args.w_epoch = 0
if arc == "wrn-28-2":
s_args.n_epoch = 300
s_args.lr = 0.1
s_args.milestone = [160, 240, 280]
s_args.gamma = 0.5
elif arc == "res-18":
s_args.n_epoch = 200
if args.export_id == "svhn":
s_args.lr = 0.01
s_args.milestone = []
s_args.gamma = 1.0
else:
s_args.lr = 0.1
s_args.milestone = [160]
s_args.gamma = 0.1
elif arc == "vgg-16":
s_args.n_epoch = 200
s_args.lr = 0.01
s_args.milestone = []
s_args.gamma = 1.0
else:
raise NotImplementedError
return s_args
if __name__ == "__main__":
args = parse_args()
if args.semi:
p_args = get_semisup_args(args)
else:
p_args = get_sup_args(args, args.proxy_arc)
t_args = get_sup_args(args, args.target_arc)
p_args.path = os.path.join(args.path, "train")
t_args.path = os.path.join(args.path, f"eval-{t_args.arc}")
make_dirs(args, ["log", "idxs"])
make_dirs(p_args, ["ckpt", "metric"])
make_dirs(t_args, ["ckpt", "metric"])
set_logger(args, record_args=True)
active_learn(args, p_args, t_args)