/
main.py
254 lines (233 loc) · 12.3 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
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
# Copyright 2022 Garena Online Private Limited.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from _preamble_ import *
from models import get_networks
from models.LookAhead import Lookahead
from dataset import get_dataset
from utils import timer, get_epoch_logger, get_logger
parser = argparse.ArgumentParser(description='Configuration')
parser.add_argument('--SEED', default=0, type=int)
parser.add_argument('--is_Train', action='store_true')
parser.add_argument('--dataset', default='CIFAR10')
parser.add_argument('--data_path', default='./Data/CIFAR')
parser.add_argument('--network', default='ResNet-18', choices=['ResNet-50', 'ResNet-34', 'ResNet-18', 'ResNet-12', 'ResNet-10', \
'WRN-28-10', 'WRN-16-10', 'WRN-16-8', 'VGG-11', 'VGG-19'])
parser.add_argument('--root', default='./experiments')
parser.add_argument('--exp_name', default='cifar10_resnet18_sgd_zp')
parser.add_argument('--tr_epochs', default=100, type=int)
parser.add_argument('--momentum', type=float, default=0.9, metavar='M', help='SGD momentum')
parser.add_argument('--weight_decay', '--wd', default=1e-3, type=float, metavar='W')
parser.add_argument('--lr', type=float, default=0.1, metavar='LR', help='learning rate')
parser.add_argument('--milestones', nargs='+', default=[0.3, 0.6, 0.8], type=float)
parser.add_argument('--gamma', type=float, default=0.2)
parser.add_argument('--train_batch_size', default=128, type=int)
parser.add_argument('--input_norm', action='store_true')
parser.add_argument('--optim', default='SGD', type=str)
parser.add_argument('--lookahead', default=None)
parser.add_argument('--slr', default=None)
parser.add_argument('--resume', action='store_true')
parser.add_argument('--checkpoint', type=str, default='./experiments/')
parser.add_argument('--net_only', default=True, type=lambda x: bool(int(x)))
args, _ = parser.parse_known_args()
np.random.seed(args.SEED)
torch.manual_seed(args.SEED)
if args.lookahead is None:
Exp_Path = os.path.join(args.root, '_'+args.dataset+'_'+args.network, args.exp_name +
'_wd' + '{:.1e}'.format(args.weight_decay) + '_lr'+'{:.1e}'.format(args.lr))
elif args.slr is None:
Exp_Path = os.path.join( args.root, '_'+args.dataset+'_'+args.network, args.exp_name +
'_wd' + '{:.1e}'.format(args.weight_decay) + '_lr'+'{:.1e}'.format(args.lr) + '_LA' + args.lookahead )
else:
Exp_Path = os.path.join( args.root, '_'+args.dataset+'_'+args.network, args.exp_name +
'_wd' + '{:.1e}'.format(args.weight_decay) + '_lr'+'{:.1e}'.format(args.lr) + '_LA' + args.lookahead +'_SLR' + args.slr )
pathlib.Path(os.path.join(Exp_Path, 'nets')).mkdir(parents=True, exist_ok=True)
def topk_accuracy(predict, device='cuda', test_loader=None, num_batch=None, topk=5, criterion=None, target_transform=None):
clncorrect = 0
idx_batch = 0
lst_label = []
lst_pred = []
lst_loss = []
assert topk>=5
for clndata, target in tqdm.tqdm(test_loader):
if target_transform is not None:
target = target_transform(target)
clndata, target = clndata.to(device), target.to(device)
num = target.shape[0]
with torch.no_grad():
output = predict(clndata)
if criterion is not None:
lst_loss.append(criterion(output, target))
else:
lst_loss.append(torch.ones(num,1)*1e6)
pred = output.topk(topk, dim=1)[1]
lst_label.append(target)
lst_pred.append(pred)
idx_batch += 1
if idx_batch == num_batch:
break
label = torch.cat(lst_label).view(-1, 1).cpu().numpy()
pred = torch.cat(lst_pred).view(-1, topk).cpu().numpy()
loss = torch.cat(lst_loss).view(-1,1).cpu().numpy()
num = label.shape[0]
top1_acc = (label == pred[:, :1]).sum() / num
top5_acc = (label == pred[:, :5]).sum() / num
avg_loss = np.mean(loss)
message = f'***** Val set Top-1 acc: {top1_acc :.2f}; Top-5 acc: {top5_acc :.2f}'
return top1_acc, avg_loss, message
class regularized_LA_loss():
def __init__(self, beta=0.1, step=5) -> None:
self.beta = beta
self.local_penalty = nn.MSELoss(size_average=False)
self._steps = 0
self.K = step
self.params_cache = defaultdict(dict)
def __call__(self, data, labels, net, is_step=False):
loss = F.cross_entropy(net(data), labels)
reg_loss = 0
if self._steps % self.K == 0:
for name, param in net.named_parameters():
current_param = param.clone().detach()
self.params_cache[name] = current_param
reg_loss += self.local_penalty(param, current_param)
else:
for name, param in net.named_parameters():
reg_loss += self.local_penalty(param, self.params_cache[name])
loss += reg_loss * self.beta
if is_step:
self.step()
return loss
def step(self):
self._steps += 1
class Classifier():
def __init__(self, network_type, in_channels, num_classes, is_train, tr_epochs, optimizer, lookahead, slr,
lr, momentum, milestones_ratio, gamma, weight_decay, device='cuda') -> None:
super().__init__()
self.net = get_networks(network_type)(in_channels, num_classes).to(device)
self.device = device
self.num_GPUs = torch.cuda.device_count()
self.lookahead = lookahead
self.slr = slr
if self.num_GPUs > 1:
self.net = nn.DataParallel(module=self.net, device_ids=list(range(self.num_GPUs)))
if is_train:
if optimizer == 'SGD':
self.optimizer = optim.SGD(self.net.parameters(), lr=lr, momentum=momentum, weight_decay=weight_decay)
elif optimizer == 'Adam':
self.optimizer = optim.Adam(self.net.parameters(), lr=lr, betas=(0.9, 0.999), eps=1e-08, weight_decay=weight_decay, amsgrad=False)
elif optimizer == 'AdamW':
self.optimizer = optim.AdamW(self.net.parameters(), lr=lr, betas=(0.9, 0.999), eps=1e-08, weight_decay=weight_decay, amsgrad=False)
else:
assert False, 'Optimizer not found!'
if lookahead is not None:
la_steps, la_alpha = lookahead.split('_')
self.optimizer = Lookahead(self.optimizer, la_steps=int(la_steps), la_alpha=float(la_alpha))
if slr is not None:
slr_steps, slr_beta = args.slr.split('_')
self.criterion = regularized_LA_loss(float(slr_beta), int(slr_steps))
else:
self.criterion = nn.CrossEntropyLoss()
milestones_epoch = [int(s * tr_epochs) for s in milestones_ratio]
self.scheduler = optim.lr_scheduler.MultiStepLR(self.optimizer, milestones=milestones_epoch, gamma=gamma)
self.tr_epochs = tr_epochs
self.start_epoch = 0
self.log_interval = 50
else:
self.eval_mode()
self.set_requires_grad([self.net], False)
def eval_mode(self):
self.net.eval()
def train_mode(self):
self.net.train()
def load_networks(self, path):
self.checkpoint = torch.load(path)
if self.num_GPUs == 1:
self.net.load_state_dict(self.checkpoint['state_dict'])
else:
self.net.module.load_state_dict(self.checkpoint['state_dict'])
def resume_training(self, path, net_only=True):
self.load_networks(path)
if not net_only:
self.start_epoch = self.checkpoint['stop_epoch']
self.optimizer.load_state_dict(self.checkpoint['optimizer'])
self.scheduler.last_epoch = self.start_epoch
def set_requires_grad(self, nets, requires_grad=False):
if not isinstance(nets, list):
nets = [nets]
for net in nets:
if net is not None:
for param in net.parameters():
param.requires_grad = requires_grad
def fit(self, train_loader, test_loader, logger, save_path='.', logging_tr=True, target_transform=None):
tic_toc = timer()
epoch_logger = get_epoch_logger()
for epoch in range(self.start_epoch, self.tr_epochs):
logger.info('Training Epoch: {}; Learning rate: {:0.8f} .....'.format(epoch, self.optimizer.param_groups[0]['lr']))
self.train_mode()
self._fit_one_epoch(train_loader, epoch, logger=logger)
if logging_tr:
_, _ = self._evaluate(epoch, train_loader, tag='train', logger=logger, target_transform=target_transform)
result, _ = self._evaluate(epoch, test_loader, logger=logger, target_transform=target_transform)
epoch_logger.append_results([epoch, result])
best_epoch, message = epoch_logger.update_best_epoch()
model_state_dict = self.net.state_dict() if self.num_GPUs<=1 else self.net.module.state_dict()
checkpoint = {'state_dict':model_state_dict, 'stop_epoch':epoch, 'optimizer': self.optimizer.state_dict()}
torch.save(checkpoint, os.path.join(save_path, "ckp_latest.pt"))
if best_epoch == epoch:
torch.save(checkpoint, os.path.join(save_path, "ckp_best.pt"))
logger.info(message + '\t' + 'Time for an epoch: {:.2f}s.\n'.format(tic_toc.toc()))
def _fit_one_epoch(self, train_loader, epoch, logger):
tic_toc = timer()
for batch_idx, (data, target) in enumerate(train_loader):
data, target = data.to(self.device), target.to(self.device)
self.optimizer.zero_grad()
if self.slr is None:
loss = self.criterion(self.net(data), target)
else:
loss = self.criterion(data, target, self.net, is_step=True)
loss.backward()
self.optimizer.step()
if (batch_idx+1) % self.log_interval == 0:
logger.info('[{}/{} ({:.0f}%)], Loss: {:.6f}, Time for Batches: {:03f}'.format(
batch_idx *len(data), len(train_loader.dataset), 100. * batch_idx / len(train_loader),
loss.item(),
tic_toc.toc()))
if self.scheduler is not None:
self.scheduler.step()
def _evaluate(self, epoch, loader, tag='test', logger=None, target_transform=None):
if self.lookahead is not None:
self.optimizer._backup_and_load_cache()
self.eval_mode()
accuracy, _, message = topk_accuracy(
self.net, device=self.device , test_loader=loader, criterion=nn.CrossEntropyLoss(reduction='none'))
if self.lookahead is not None:
self.optimizer._clear_and_load_backup()
if logger is not None:
logger.info(message)
return accuracy, message
def main():
logger = get_logger(os.path.join(Exp_Path, 'logging.txt'))
logger.info(args)
loader, train_loader, test_loader = get_dataset(name=args.dataset, batch_size=args.train_batch_size, num_workers=6, input_norm=args.input_norm, data_path=args.data_path)
model = Classifier(network_type=args.network, in_channels=loader.img_channels, num_classes=loader.num_classes,
is_train=True, tr_epochs=args.tr_epochs,
optimizer=args.optim, lookahead=args.lookahead, slr=args.slr,
lr=args.lr, momentum=args.momentum, milestones_ratio=args.milestones, gamma=args.gamma,
weight_decay=args.weight_decay)
logger.info(model.net)
if args.resume:
model.resume_training(path=args.checkpoint, net_only=args.net_only)
model.fit(train_loader=train_loader, test_loader=test_loader,
logger=logger, save_path=os.path.join(Exp_Path, 'nets'), logging_tr=False)
if __name__ == '__main__':
main()