-
Notifications
You must be signed in to change notification settings - Fork 10
/
main.py
163 lines (125 loc) · 5.38 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
import os
from pprint import pprint
from tqdm import tqdm
import torch.nn as nn
import torch.nn.parallel
import torch.backends.cudnn as cudnn
import torch.optim
import torch.utils.data
from torch.utils.tensorboard import SummaryWriter
import utils
from model import resnet32
from config import get_arguments
parser = get_arguments()
args = parser.parse_args()
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
args.device = device
exp_loc, model_loc = utils.log_folders(args)
writer = SummaryWriter(log_dir=exp_loc)
def main():
"""Main script"""
assert not (args.logit_adj_post and args.logit_adj_train)
train_loader, val_loader = utils.get_loaders(args)
num_class = len(args.class_names)
model = torch.nn.DataParallel(resnet32(num_classes=num_class))
model = model.to(device)
cudnn.benchmark = True
criterion = nn.CrossEntropyLoss().to(device)
if args.logit_adj_post:
if os.path.isfile(os.path.join(model_loc, "model.th")):
print("=> loading pretrained model ")
checkpoint = torch.load(os.path.join(model_loc, "model.th"))
model.load_state_dict(checkpoint['state_dict'])
for tro in args.tro_post_range:
args.tro = tro
args.logit_adjustments = utils.compute_adjustment(train_loader, tro, args)
val_loss, val_acc = validate(val_loader, model, criterion)
results = utils.class_accuracy(val_loader, model, args)
results["OA"] = val_acc
pprint(results)
hyper_param = utils.log_hyperparameter(args, tro)
writer.add_hparams(hparam_dict=hyper_param, metric_dict=results)
writer.close()
else:
print("=> No pre trained model found")
return
args.logit_adjustments = utils.compute_adjustment(train_loader, args.tro_train, args)
optimizer = torch.optim.SGD(model.parameters(),
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay,
nesterov=True)
lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer,
milestones=args.scheduler_steps)
loop = tqdm(range(0, args.epochs), total=args.epochs, leave=False)
val_loss, val_acc = 0, 0
for epoch in loop:
# train for one epoch
train_loss, train_acc = train(train_loader, model, criterion, optimizer)
writer.add_scalar("train/acc", train_acc, epoch)
writer.add_scalar("train/loss", train_loss, epoch)
lr_scheduler.step()
# evaluate on validation set
if (epoch % args.log_val) == 0 or (epoch == (args.epochs - 1)):
val_loss, val_acc = validate(val_loader, model, criterion)
writer.add_scalar("val/acc", val_acc, epoch)
writer.add_scalar("val/loss", val_loss, epoch)
loop.set_description(f"Epoch [{epoch}/{args.epochs}")
loop.set_postfix(train_loss=f"{train_loss:.2f}", val_loss=f"{val_loss:.2f}",
train_acc=f"{train_acc:.2f}",
val_acc=f"{val_acc:.2f}")
file_name = 'model.th'
mdel_data = {"state_dict": model.state_dict()}
torch.save(mdel_data, os.path.join(model_loc, file_name))
results = utils.class_accuracy(val_loader, model, args)
results["OA"] = val_acc
hyper_param = utils.log_hyperparameter(args, args.tro_train)
pprint(results)
writer.add_hparams(hparam_dict=hyper_param, metric_dict=results)
writer.close()
def train(train_loader, model, criterion, optimizer):
""" Run one train epoch """
losses = utils.AverageMeter()
accuracies = utils.AverageMeter()
model.train()
for _, (inputs, target) in enumerate(train_loader):
target = target.to(device)
input_var = inputs.to(device)
target_var = target
output = model(input_var)
acc = utils.accuracy(output.data, target)
if args.logit_adj_train:
output = output + args.logit_adjustments
loss = criterion(output, target_var)
loss_r = 0
for parameter in model.parameters():
loss_r += torch.sum(parameter ** 2)
loss = loss + args.weight_decay * loss_r
optimizer.zero_grad()
loss.backward()
optimizer.step()
losses.update(loss.item(), inputs.size(0))
accuracies.update(acc, inputs.size(0))
return losses.avg, accuracies.avg
def validate(val_loader, model, criterion):
""" Run evaluation """
losses = utils.AverageMeter()
accuracies = utils.AverageMeter()
model.eval()
with torch.no_grad():
for _, (inputs, target) in enumerate(val_loader):
target = target.to(device)
input_var = inputs.to(device)
target_var = target.to(device)
output = model(input_var)
loss = criterion(output, target_var)
if args.logit_adj_post:
output = output - args.logit_adjustments
elif args.logit_adj_train:
loss = criterion(output + args.logit_adjustments, target_var)
acc = utils.accuracy(output.data, target)
losses.update(loss.item(), inputs.size(0))
accuracies.update(acc, inputs.size(0))
return losses.avg, accuracies.avg
if __name__ == '__main__':
main()