-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
207 lines (156 loc) · 7.65 KB
/
train.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
import argparse
import glob
import os
import torch
import torch.nn as nn
import torch.optim as optim
import numpy as np
#from PIL import Image
import transforms
#from torchvision import transforms
# from tensorboardX import SummaryWriter
from conf import settings
from utils import *
# from lr_scheduler import WarmUpLR
import timm
from timm.scheduler.cosine_lr import CosineLRScheduler
from timm.scheduler.step_lr import StepLRScheduler
from criterion import LSR
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--net', type=str, required=True, help='net type')
parser.add_argument('--num_workers', type=int, default=2, help='number of workers for dataloader')
parser.add_argument('--batch_size', type=int, default=256, help='batch size for dataloader')
parser.add_argument('--loss', type=str, default='label_smooth', choices=['label_smooth'], help='loss function')
parser.add_argument('--weight_decay', action='store_true', help='1-D. No bias decay (regularization)')
parser.add_argument('--optimizer', type=str, default='SGD', choices=['SGD', 'AdamW'], help='Optimizer')
parser.add_argument('--lr', type=float, default=0.04, help='learning rate')
parser.add_argument('--init_lr', type=float, default=0.001, help='initial learning rate when using learning rate scheduler')
parser.add_argument('--decay_rate', type=float, default=0.9, help='learning rate decay rate when using multi-step LR scheduler')
parser.add_argument('--lr_scheduler', type=str, default='cosinelr', choices=['cosinelr', 'steplr'], help='learning rate scheduler')
parser.add_argument('--epochs', type=int, default=450, help='training epoches')
parser.add_argument('--warm_t', type=int, default=5, help='warm up phase')
parser.add_argument('--decay_t', type=int, default=10, help='Decay LR for every decay_t epochs in StepLR')
parser.add_argument('--gpus', type=str, default=0, help='gpu device')
parser.add_argument('--log_step', type=int, default=1, help='printing loss step')
parser.add_argument('--val_step', type=int, default=1, help='validation step')
parser.add_argument('--save_step', type=int, default=1, help='save checkpoint step')
parser.add_argument('--wandb', action='store_true', help='tracking with wandb')
parser.add_argument('--run_name', type=str, default='scy_exp3', help='wandb run name')
args = parser.parse_args()
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpus
#checkpoint directory
checkpoint_path = os.path.join(settings.CHECKPOINT_PATH, args.net, settings.TIME_NOW)
if not os.path.exists(checkpoint_path):
os.makedirs(checkpoint_path)
checkpoint_path = os.path.join(checkpoint_path, '{net}-{epoch}-{type}.pth')
#tensorboard log directory
log_path = os.path.join(settings.LOG_DIR, args.net, settings.TIME_NOW)
if not os.path.exists(log_path):
os.makedirs(log_path)
if args.wandb:
import wandb
wandb.init(project='scy_test', entity="dnn_22_2", name=args.run_name, settings=wandb.Settings(code_dir="."))
wandb.run.log_code(".")
#get dataloader
train_transforms = transforms.Compose([
#transforms.ToPILImage(),
transforms.ToCVImage(),
transforms.RandomResizedCrop(settings.IMAGE_SIZE),
transforms.RandomHorizontalFlip(),
transforms.ColorJitter(brightness=0.4, saturation=0.4, hue=0.4),
#transforms.RandomErasing(),
#transforms.CutOut(56),
transforms.ToTensor(),
transforms.Normalize(settings.TRAIN_MEAN, settings.TRAIN_STD)
])
test_transforms = transforms.Compose([
transforms.ToCVImage(),
transforms.CenterCrop(settings.IMAGE_SIZE),
transforms.ToTensor(),
transforms.Normalize(settings.TRAIN_MEAN, settings.TRAIN_STD)
])
train_dataloader = get_train_dataloader(
settings.DATA_PATH,
train_transforms,
args.batch_size,
args.num_workers
)
test_dataloader = get_test_dataloader(
settings.DATA_PATH,
test_transforms,
args.batch_size,
args.num_workers
)
#device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
net = get_network(args)
net = init_weights(net)
# if isinstance(args.gpus, int):
# args.gpus = [args.gpus]
# net = nn.DataParallel(net, device_ids=args.gpus)
net = net.cuda()
#cross_entropy = nn.CrossEntropyLoss()
if args.loss == 'label_smooth':
lsr_loss = LSR() # Label smoothing
#apply no weight decay on bias
if args.weight_decay:
params = split_weights(net)
else:
params = net.parameters()
if args.optimizer == 'SGD':
optimizer = optim.SGD(params, lr=args.lr, momentum=0.9, weight_decay=1e-4, nesterov=True)
elif args.optimizer == 'AdamW':
optimizer = optim.AdamW(params, lr=args.lr, betas=(0.9, 0.999), eps=1e-08, weight_decay=0.01)
# warmup_scheduler = WarmUpLR(optimizer, iter_per_epoch * args.warm)
if args.lr_scheduler == 'cosinelr':
warmup_scheduler = CosineLRScheduler(optimizer, t_initial=args.epochs, warmup_t=args.warm_t, warmup_lr_init=args.init_lr)
elif args.lr_scheduler == 'steplr':
warmup_scheduler = StepLRScheduler(optimizer, decay_t=args.decay_t, warmup_t=args.warm_t, warmup_lr_init=args.init_lr, decay_rate=args.decay_rate)
#set up training phase learning rate scheduler
# train_scheduler = optim.lr_scheduler.MultiStepLR(optimizer, milestones=settings.MILESTONES)
#train_scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, args.epochs - args.warm)
num_iters = len(train_dataloader)
best_acc = 0.0
for epoch in range(1, args.epochs + 1):
#training procedure
net.train()
for batch_index, (images, labels) in enumerate(train_dataloader):
images = images.cuda()
labels = labels.cuda()
optimizer.zero_grad()
predicts = net(images)
loss = lsr_loss(predicts, labels)
loss.backward()
optimizer.step()
n_iter = (epoch - 1) * len(train_dataloader) + batch_index + 1
if batch_index % args.log_step == 0:
print(f'Epoch : [{epoch} / {args.epochs}], \tIter : [{batch_index} / {num_iters}], \tLoss : {loss.item()}')
if args.wandb:
wandb.log({'Epoch' : epoch, 'Iter' : batch_index, 'Train Loss': loss.item()})
warmup_scheduler.step(epoch)
if args.wandb:
wandb.log({'LR' : optimizer.param_groups[0]['lr']})
if epoch % args.val_step == 0:
net.eval()
total_loss = 0
correct = 0
for images, labels in test_dataloader:
images = images.cuda()
labels = labels.cuda()
predicts = net(images)
_, preds = predicts.max(1)
correct += preds.eq(labels).sum().float()
loss = lsr_loss(predicts, labels)
total_loss += loss.item()
test_loss = total_loss / len(test_dataloader)
acc = correct / len(test_dataloader.dataset)
print('Test set: loss: {:.4f}, Accuracy: {:.4f}'.format(test_loss, acc))
if args.wandb:
wandb.log({'Validation Loss': test_loss, 'Validation acc': acc})
#save weights file
if epoch % args.save_step == 0:
if best_acc < acc:
print(f'Saving checkpoint ... accuracy = {acc}')
torch.save(net.state_dict(), checkpoint_path.format(net=args.net, epoch=epoch, type='best'))
best_acc = acc
continue