-
Notifications
You must be signed in to change notification settings - Fork 7
/
Copy pathargs.py
141 lines (129 loc) · 7.81 KB
/
args.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
import os
import glob
import time
import argparse
# model_names = list(map(lambda n: os.path.basename(n)[:-3],
# glob.glob('models/[A-Za-z]*.py')))
model_names = ['msdnet', 'msdnet_ge', 'IMTA_MSDNet', 'mobilenet_imagenet']
arg_parser = argparse.ArgumentParser(
description='Image classification PK main script')
exp_group = arg_parser.add_argument_group('exp', 'experiment setting')
exp_group.add_argument('--save', default='save/default-{}'.format(time.time()),
type=str, metavar='SAVE',
help='path to the experiment logging directory'
'(default: save/debug)')
exp_group.add_argument('--resume', action='store_true',
help='path to latest checkpoint (default: none)')
exp_group.add_argument('--eval', '--evaluate', dest='evalmode', default=None,
choices=['anytime', 'dynamic'],
help='way to evaluate')
exp_group.add_argument('--evaluate-from', default=None, type=str, metavar='PATH',
help='path to saved checkpoint (default: none)')
exp_group.add_argument('--print-freq', '-p', default=10, type=int,
metavar='N', help='print frequency (default: 100)')
exp_group.add_argument('--seed', default=0, type=int,
help='random seed')
exp_group.add_argument('--gpu',
help='GPU available.')
# dataset related
data_group = arg_parser.add_argument_group('data', 'dataset setting')
data_group.add_argument('--data', metavar='D', default='cifar10',
choices=['cifar10', 'cifar100', 'ImageNet'],
help='data to work on')
data_group.add_argument('--data-root', metavar='DIR', default='data',
help='path to dataset (default: data)')
data_group.add_argument('--use-valid', action='store_true',
help='use validation set or not')
data_group.add_argument('-j', '--workers', default=4, type=int, metavar='N',
help='number of data loading workers (default: 4)')
# data_group.add_argument('--normalized', action='store_true',
# help='normalize the data into zero mean and unit std')
# data_group.add_argument('--augmentation', default=0.08, type=float, metavar='M',
# help='')
# model arch related
arch_group = arg_parser.add_argument_group('arch',
'model architecture setting')
arch_group.add_argument('--arch', '-a', metavar='ARCH', default='resnet',
type=str, choices=model_names,
help='model architecture: ' +
' | '.join(model_names) +
' (default: resnet)')
arch_group.add_argument('-d', '--depth', default=56, type=int, metavar='D',
help='depth (default=56)')
arch_group.add_argument('--drop-rate', default=0.0, type=float,
metavar='DROPRATE', help='dropout rate (default: 0.2)')
arch_group.add_argument('--death-mode', default='none',
choices=['none', 'linear', 'uniform'],
help='death mode (default: none)')
arch_group.add_argument('--death-rate', default=0.5, type=float,
help='death rate rate (default: 0.5)')
# arch_group.add_argument('--growth-rate', default=12, type=int,
# metavar='GR', help='Growth rate of DenseNet'
# ' (1 means dot\'t use compression) (default: 0.5)')
arch_group.add_argument('--bn-size', default=4, type=int,
metavar='B', help='bottle neck ratio of DenseNet'
' (0 means dot\'t use bottle necks) (default: 4)')
arch_group.add_argument('--reduction', default=0.5, type=float,
metavar='C', help='compression ratio of DenseNet'
' (1 means dot\'t use compression) (default: 0.5)')
# used to set the argument when to resume automatically
arch_resume_names = ['arch', 'depth', 'death_mode', 'death_rate', 'death_rate',
'growth_rate', 'bn_size', 'compression']
# msdnet config
arch_group.add_argument('--nBlocks', type=int, default=1)
arch_group.add_argument('--nChannels', type=int, default=32)
arch_group.add_argument('--base', type=int,default=4)
arch_group.add_argument('--stepmode', type=str, choices=['even', 'lin_grow'])
arch_group.add_argument('--step', type=int, default=1)
arch_group.add_argument('--growthRate', type=int, default=6)
arch_group.add_argument('--grFactor', default='1-2-4', type=str)
arch_group.add_argument('--prune', default='max', choices=['min', 'max'])
arch_group.add_argument('--bnFactor', default='1-2-4')
arch_group.add_argument('--bottleneck', default=True, type=bool)
arch_group.add_argument('--pretrained', default=None, type=str, metavar='PATH',
help='path to load pretrained msdnet (default: none)')
arch_group.add_argument('--priornet', default=None, type=str, metavar='PATH',
help='path to load pretrained priornet (default: none)')
# training related
optim_group = arg_parser.add_argument_group('optimization',
'optimization setting')
optim_group.add_argument('--trainer', default='train', type=str,
help='trainer file name without ".py"'
' (default: train)')
optim_group.add_argument('--epochs', default=164, type=int, metavar='N',
help='number of total epochs to run (default: 164)')
optim_group.add_argument('--start-epoch', default=0, type=int, metavar='N',
help='manual epoch number (useful on restarts)')
optim_group.add_argument('--switch-mode', default=300, type=int, metavar='N',
help='number of epochs to switch mode (default: 300)')
optim_group.add_argument('--patience', default=0, type=int, metavar='N',
help='patience for early stopping'
'(0 means no early stopping)')
optim_group.add_argument('-b', '--batch-size', default=64, type=int,
metavar='N', help='mini-batch size (default: 64)')
optim_group.add_argument('--optimizer', default='sgd',
choices=['sgd', 'rmsprop', 'adam'], metavar='N',
help='optimizer (default=sgd)')
optim_group.add_argument('--lr', '--learning-rate', default=0.1, type=float,
metavar='LR',
help='initial learning rate (default: 0.1)')
optim_group.add_argument('--lr-type', default='multistep', type=str, metavar='T',
help='learning rate strategy (default: multistep)',
choices=['cosine', 'multistep'])
optim_group.add_argument('--decay-rate', default=0.1, type=float, metavar='N',
help='decay rate of learning rate (default: 0.1)')
optim_group.add_argument('--momentum', default=0.9, type=float, metavar='M',
help='momentum (default=0.9)')
optim_group.add_argument('--alpha', default=0.99, type=float, metavar='M',
help='alpha for ')
optim_group.add_argument('--beta1', default=0.9, type=float, metavar='M',
help='beta1 for Adam (default: 0.9)')
optim_group.add_argument('--beta2', default=0.999, type=float, metavar='M',
help='beta2 for Adam (default: 0.999)')
optim_group.add_argument('--weight-decay', '--wd', default=1e-4, type=float,
metavar='W', help='weight decay (default: 1e-4)')
### add kd hyperparameters
optim_group.add_argument('--gamma', default=0.9, type=float, metavar='M',
help='gamma for kld loss')
optim_group.add_argument('-T', default=3.0, type=float, metavar='M',
help='Temperature for KD')