-
Notifications
You must be signed in to change notification settings - Fork 2
/
main.py
246 lines (214 loc) · 12.5 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
import argparse
import os
import torchvision.models as models
from copy import deepcopy
from utils import preresnets
from utils.checkpoint import save_checkpoint, check_local
from utils.dataset import get_dataset_train
from utils.train_func import train_fdg, train_decouple, validate
from utils.util import warm_update
import torch
import torch.nn as nn
import torch.distributed as dist
import torch.multiprocessing as mp
from torch.nn.parallel import DistributedDataParallel as DDP
parser = argparse.ArgumentParser(description='PyTorch NICO Training')
parser.add_argument('-a', '--arch', metavar='ARCH', default='resnet50',help='[densenet121/resnet34/wide_resnet50_2]')
parser.add_argument('--epochs', default=90, type=int, metavar='N',help='number of total epochs to run')
parser.add_argument('--total_epoch', default=200, type=int, metavar='N',help='number of total epochs to run')
parser.add_argument('-gas', "--gradient_accumulation_steps", type=float, default=1.0, help='control accumulate batch')
parser.add_argument('--start-epoch', default=0, type=int, metavar='N',help='manual epoch number (useful on restarts)')
parser.add_argument('-b', '--batch-size', default=256, type=int,metavar='N',help='mini-batch size (default: 256)',dest='batch_size')
parser.add_argument('--num_workers', default=64, type=int,help='processings for data loader')
parser.add_argument("--update_epoch",default=1, type=int, help='per "update_epoch" epoch to use swa update')
parser.add_argument('--lr', '--learning-rate', default=0.1, type=float,metavar='LR', help='initial learning rate', dest='lr')
parser.add_argument('--momentum', default=0.9, type=float, metavar='M',help='momentum')
parser.add_argument('--wd', '--weight-decay', default=1e-4, type=float,metavar='W', help='weight decay (default: 1e-4)',dest='weight_decay')
parser.add_argument('-p', '--print-freq', default=10, type=int,metavar='N', help='print frequency (default: 10)')
parser.add_argument("--log_freq", type=int, default=500,help="log frequency to file")
parser.add_argument("--ckpt", default="./ckpts/",help="folder to output checkpoints")
parser.add_argument("--scheme", default="fdg",help="[fdg/decouple]")
parser.add_argument("--ratio", type=float, default=1.0, help='control foriour range')
parser.add_argument("--cos_lr", action='store_true',help='use cosine learning rate')
parser.add_argument("--swa", action='store_true', help='use swa')
parser.add_argument("--use_seg", action='store_true', help='use seg')
parser.add_argument("--track",default="1",help='[1/2]')
parser.add_argument("--fine_tune", default= None, type=str, metavar='PATH',help='path to load pretrained model and fintune(default: none)')
parser.add_argument("--pretrained", default= None, type=str, metavar='PATH',help='path to load pretrained model (default: none)')
parser.add_argument("--moco_pretrained", default= None, type=str, metavar='PATH',help='path to load moco pretrained model (default: none)')
parser.add_argument("--root", default='', type=str, metavar='PATH',help='root path to load datasets')
parser.add_argument("--json_path", default='./dataset_json/', type=str, metavar='PATH',help='path to load json files')
parser.add_argument('--img_size',nargs='+',help='the list for model ckpts')
parser.add_argument('--dist-url', default='tcp://127.0.0.1:4455', type=str, help='url used to set up distributed training')
parser.add_argument('--dist-backend', default='nccl', type=str, help='distributed backend')
parser.add_argument('--rank', default=-1, type=int, help='node rank for distributed training')
parser.add_argument('--world-size', default=-1, type=int, help='number of nodes for distributed training')
best_acc1 = 0
def main():
args = parser.parse_args()
global best_acc1
args.ckpt += "NICO"
args.ckpt += "-" + args.arch
args.ckpt += "-track" + args.track
print("Use MODEL: {} for training".format(args.arch))
print("Use Track: {} for training".format(args.track))
print("Use Ratio: {} for training".format(args.ratio))
print("Save Path: {}".format(args.ckpt))
if not os.path.isdir(args.ckpt):
os.makedirs(args.ckpt)
if args.scheme == 'fdg':
student_ckpt = args.ckpt+'/student/'
teacher_ckpt = args.ckpt+'/teacher/'
if not os.path.isdir(student_ckpt):
os.makedirs(student_ckpt)
if not os.path.isdir(teacher_ckpt):
os.makedirs(teacher_ckpt)
# |-DDP initilization
ngpus_per_node = torch.cuda.device_count()
args.world_size = ngpus_per_node * args.world_size
mp.spawn(main_worker, nprocs=ngpus_per_node, args=(ngpus_per_node, args))
def main_worker(gpu, ngpus_per_node, args):
global best_acc1
args.gpu = gpu
args.ngpus_per_node = ngpus_per_node
if args.gpu is not None:
print("Use GPU: {} for training".format(args.gpu))
# For multiprocessing distributed training, rank needs to be the
# global rank among all the processes
args.rank = args.rank * ngpus_per_node + gpu
dist.init_process_group(backend=args.dist_backend, init_method=args.dist_url,
world_size=args.world_size, rank=args.rank)
# |-model initilization
if 'preresnet' in args.arch:
model = preresnets.models[args.arch](pretrained=False, num_classes=60)
else:
model = models.__dict__[args.arch](pretrained=False,num_classes=60)
if args.scheme=='fdg':
if 'preresnet' in args.arch:
model_teacher = preresnets.models[args.arch](pretrained=False,num_classes=60)
else:
model_teacher = models.__dict__[args.arch](pretrained=False,num_classes=60)
# When using a single GPU per process and per
# DistributedDataParallel, we need to divide the batch size
# ourselves based on the total number of GPUs we have
args.batch_size = int(args.batch_size / ngpus_per_node)
args.num_workers = int((args.num_workers + ngpus_per_node - 1) / ngpus_per_node)
criterion = nn.CrossEntropyLoss().cuda(args.gpu)
optimizer = torch.optim.SGD(model.parameters(), args.lr,momentum=args.momentum,weight_decay=args.weight_decay)
# |-Check local weight
if args.scheme=='fdg':
model, model_teacher, optimizer, best_acc1, start_epoch = check_local(args,model,model_teacher,optimizer)
else:
model, optimizer, best_acc1, start_epoch = check_local(args,model,model,optimizer)
args.start_epoch = start_epoch
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
torch.cuda.set_device(args.gpu)
model.cuda(args.gpu)
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu])
if args.scheme=='fdg':
model_teacher.cuda(args.gpu)
# |-Log files initilization
if args.rank % ngpus_per_node == 0:
log_file = os.path.join(args.ckpt, "log.txt")
if os.path.exists(log_file):
args.log_file = open(log_file, mode="a")
else:
args.log_file = open(log_file, mode="w")
args.log_file.write("Network - " + args.arch + "\n")
args.log_file.write("Learning Rate - " + str(args.lr) + "\n")
args.log_file.write("Batch Size - " + str(args.batch_size) + "\n")
args.log_file.write("Weight Decay - " + str(args.weight_decay) + "\n")
args.log_file.write("--------------------------------------------------" + "\n")
args.log_file.close()
# |-Training initilization
if args.cos_lr:
for param_group in optimizer.param_groups:
param_group['lr']=args.lr
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=args.total_epoch, last_epoch=-1)
for i in range(start_epoch):
scheduler.step()
# |-Data initilization
if args.track == '2':
train_dataset1, valid_dataset1 = get_dataset_train(int(args.img_size[0]), 'track2', args.scheme, args.use_seg, args.root, json_path=args.json_path)
train_dataset2, valid_dataset2 = get_dataset_train(int(args.img_size[1]), 'track2', args.scheme, args.use_seg, args.root, json_path=args.json_path)
train_dataset3, valid_dataset3 = get_dataset_train(int(args.img_size[2]), 'track2', args.scheme, args.use_seg, args.root, json_path=args.json_path)
else:
train_dataset1, valid_dataset1 = get_dataset_train(int(args.img_size[0]), 'track1', args.scheme, args.use_seg, args.root, json_path=args.json_path)
train_dataset2, valid_dataset2 = get_dataset_train(int(args.img_size[1]), 'track1', args.scheme, args.use_seg, args.root, json_path=args.json_path)
train_dataset3, valid_dataset3 = get_dataset_train(int(args.img_size[2]), 'track1', args.scheme, args.use_seg, args.root, json_path=args.json_path)
train_sampler1 = torch.utils.data.distributed.DistributedSampler(train_dataset1)
train_sampler2 = torch.utils.data.distributed.DistributedSampler(train_dataset2)
train_sampler3 = torch.utils.data.distributed.DistributedSampler(train_dataset3)
train_loader1 = torch.utils.data.DataLoader(
train_dataset1, batch_size=args.batch_size, shuffle=(train_sampler1 is None),
num_workers=args.num_workers, pin_memory=True, sampler=train_sampler1)
train_loader2 = torch.utils.data.DataLoader(
train_dataset2, batch_size=args.batch_size, shuffle=(train_sampler2 is None),
num_workers=args.num_workers, pin_memory=True, sampler=train_sampler2)
train_loader3 = torch.utils.data.DataLoader(
train_dataset3, batch_size=args.batch_size, shuffle=(train_sampler3 is None),
num_workers=args.num_workers, pin_memory=True, sampler=train_sampler3)
val_loader1 = torch.utils.data.DataLoader(valid_dataset1, batch_size=args.batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True, drop_last=False)
val_loader2 = torch.utils.data.DataLoader(valid_dataset2, batch_size=args.batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True, drop_last=False)
val_loader3 = torch.utils.data.DataLoader(valid_dataset3, batch_size=args.batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True, drop_last=False)
# |-Training and online test
for epoch in range(args.start_epoch, args.epochs):
if epoch<200:
train_loader = train_loader3
val_loader = val_loader3
train_sampler =train_sampler3
if epoch<120:
train_loader = train_loader2
val_loader = val_loader2
train_sampler =train_sampler2
if epoch<60:
train_loader = train_loader1
val_loader = val_loader1
train_sampler =train_sampler1
train_sampler.set_epoch(epoch)
if args.rank % ngpus_per_node == 0:
args.log_file = open(log_file, mode="a")
if args.scheme == 'fdg':
train_fdg(train_loader, model, model_teacher, criterion, optimizer, epoch, args)
else:
if args.swa:
model_t = deepcopy(model)
train_decouple(train_loader, model, criterion, optimizer, epoch, args)
if epoch % args.update_epoch == 0:
warm_update(model, model_t, 0.0005, epoch+1)
else:
train_decouple(train_loader, model, criterion, optimizer, epoch, args)
if(args.cos_lr):
scheduler.step()
print('[%03d] %.7f'%(epoch, scheduler.get_lr()[0]))
if args.rank % ngpus_per_node == 0:
acc1 = validate(val_loader, model, criterion, args)
# |-Save checkpoint
is_best = acc1 > best_acc1
best_acc1 = max(acc1, best_acc1)
args.log_file.close()
if args.scheme == 'fdg':
save_checkpoint({
"epoch": epoch + 1,
"arch": args.arch,
"state_dict": model.state_dict(),
"best_acc": best_acc1,
"optimizer" : optimizer.state_dict(),
}, is_best, epoch, save_path=args.ckpt+'/student/')
save_checkpoint({
"epoch": epoch + 1,
"arch": args.arch,
"state_dict": model_teacher.state_dict(),
"best_acc": best_acc1,
"optimizer" : optimizer.state_dict(),
}, is_best, epoch, save_path=args.ckpt+'/teacher/')
else:
save_checkpoint({
"epoch": epoch + 1,
"arch": args.arch,
"state_dict": model.module.state_dict(),
"best_acc": best_acc1,
"optimizer" : optimizer.state_dict(),
}, is_best, epoch, save_path=args.ckpt)
if __name__ == '__main__':
main()