-
Notifications
You must be signed in to change notification settings - Fork 0
/
main2_anytime.py
152 lines (123 loc) · 4.87 KB
/
main2_anytime.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
import torch
import torch.nn as nn, torch.optim as optim
from torch.autograd import Variable
from torch.utils.data import DataLoader
import models
from options import parse_args
from utils import Logger, AverageMeter, ClassErrorMeter
from datasets import get_dataset
from utils.measure_v2 import measure
torch.backends.cudnn.benchmark = True
###########################
# DEFINE GLOBAL VARIABLES #
###########################
# parse arguments
args, model_args = parse_args()
# define logger
logdir = args.logdir
logger = Logger(logdir, read_only=args.test_only)
logger.log('args: %s'%str(args))
logger.log('model args: %s'%str(model_args))
# define model
model = models.get_model(args.model, model_args).cuda()
# logger.log('full-model FLOPs: %d' % measure(model, torch.zeros(1, 3, 32, 32).cuda(), k=-1)[0])
# define datasets - 0: train, 1: val, 2: test
datasets = get_dataset(args.dataset, val_size=args.valsize)
dataloaders = []
for d in datasets:
dataloaders.append(DataLoader(d,
batch_size=args.batch_size,
shuffle=True,
num_workers=4))
# define loss
criterion = nn.CrossEntropyLoss().cuda()
# define optimizer
optimizer = optim.SGD(model.parameters(),
lr=args.lr,
momentum=args.momentum,
weight_decay=args.wd,
nesterov=args.nesterov)
# define lr scheduler
lr_scheduler = optim.lr_scheduler.MultiStepLR(
optimizer,
milestones=[args.num_epochs/2, args.num_epochs*3/4],
gamma=0.1)
####################
# DEFINE FUNCTIONS #
####################
def log_iter(epoch, mode, i, losses, errors):
prompt = '[epoch %3d] [%s] [iter %3d] [loss %.6f %.6f] [error %.4f]'
logger.log(prompt%(epoch, mode, i, sum(losses), losses[-1], errors[-1]))
def log_epoch(epoch, mode, losses, errors):
prompt = '[epoch %3d] [%s] [anytime %d] [loss %.6f] [error %.4f]'
for k, (l, e) in enumerate(zip(losses, errors)):
logger.log(prompt%(epoch, mode, k, l(), e()))
logger.scalar_summary('loss/%s'%mode, losses[-1](), epoch)
logger.scalar_summary('error/%s'%mode, errors[-1](), epoch)
def train(epoch, dataloader):
model.train()
loss_average = [AverageMeter() for _ in range(model.anytime)]
error_average = [ClassErrorMeter() for _ in range(model.anytime)]
for i, (images, labels) in enumerate(dataloader):
n = images.size()[0]
optimizer.zero_grad()
losses = []
errors = []
for k in range(model.anytime):
y = model(Variable(images.cuda()), k)
loss = criterion(y, Variable(labels.cuda()))
losses.append(loss_average[k].add(loss.item(), n))
errors.append(error_average[k].add(y.cpu().data, labels))
loss.backward()
optimizer.step()
log_iter(epoch, 'train', i, losses, errors)
log_epoch(epoch, 'train', loss_average, error_average)
return error_average[-1]()
def test(epoch, dataloader, val=False):
if len(dataloader) == 0:
return 1.0
mode = 'val' if val else 'test'
model.eval()
loss_average = [AverageMeter() for _ in range(model.anytime)]
error_average = [ClassErrorMeter() for _ in range(model.anytime)]
for i, (images, labels) in enumerate(dataloader):
with torch.no_grad():
n = images.size()[0]
losses = []
errors = []
for k in range(model.anytime):
y = model(Variable(images.cuda()), k)
loss = criterion(y, Variable(labels.cuda()))
losses.append(loss_average[k].add(loss.item(), n))
errors.append(error_average[k].add(y.cpu().data, labels))
log_iter(epoch, mode, i, losses, errors)
log_epoch(epoch, mode, loss_average, error_average)
return error_average[-1]()
###############
# MAIN SCRIPT #
###############
if args.test_only:
model.load_state_dict(logger.load('best.model'))
val_error = test(0, dataloaders[1], val=True)
test_error = test(0, dataloaders[2], val=False)
exit()
last_epoch = -1
if args.resume:
model_state, optim_state, last_epoch = logger.load_checkpoint()
model.load_state_dict(model_state)
optimizer.load_state_dict(optim_state)
val_best = 1.0
test_best = 1.0
for epoch in range(last_epoch+1, args.num_epochs):
lr_scheduler.step(epoch=epoch)
train(epoch, dataloaders[0])
val_error = test(epoch, dataloaders[1], val=True)
test_error = test(epoch, dataloaders[2], val=False)
state_dict = model.state_dict()
is_best = False
if (args.valsize > 0 and val_best > val_error) or (args.valsize == 0 and test_best > test_error):
val_best = val_error
test_best = test_error
is_best = True
logger.save_checkpoint(epoch, model, optimizer, is_best)
logger.log('[epoch %3d] [best] [val %.4f] [test %.4f]'%(epoch, val_best, test_best))