-
Notifications
You must be signed in to change notification settings - Fork 16
/
train.py
executable file
·228 lines (177 loc) · 8 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
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
import os
import time
import argparse
import numpy as np
from tqdm import tqdm
from glob import glob
import torch
from torch.utils import data
from tensorboardX import SummaryWriter
from dataset import PreTrain_DS, DAVIS_Train_DS, YouTube_Train_DS
from model import AFB_URR, FeatureBank
import myutils
def get_args():
parser = argparse.ArgumentParser(description='Train AFB-URR')
parser.add_argument('--gpu', type=int, default=0,
help='GPU card id.')
parser.add_argument('--dataset', type=str, default=None, required=True,
help='Dataset folder.')
parser.add_argument('--seed', type=int, default=-1,
help='Random seed.')
parser.add_argument('--log', action='store_true',
help='Save the training results.')
parser.add_argument('--level', type=int, default=0,
help='0: pretrain. 1: DAVIS. 2: Youtube-VOS.')
parser.add_argument('--lr', type=float, default=1e-5,
help='Learning rate, default 1e-5.')
parser.add_argument('--lu', type=float, default=0.5,
help='Regularization factor, default 0.5.')
parser.add_argument('--resume', type=str,
help='Path to the checkpoint (default: none)')
parser.add_argument('--new', action='store_true',
help='Train the model from the begining.')
parser.add_argument('--scheduler-step', type=int, default=25,
help='Scheduler step size. Default 25.')
parser.add_argument('--total-epochs', type=int, default=100,
help='Total running epochs. Default 100.')
parser.add_argument('--budget', type=int, default=300000,
help='Max number of features that feature bank can store. Default: 300000')
parser.add_argument('--obj-n', type=int, default=3,
help='Max number of objects that will be trained at the same time.')
parser.add_argument('--clip-n', type=int, default=6,
help='Max frames that will be sampled as a batch.')
return parser.parse_args()
def train_model(model, dataloader, criterion, optimizer, desc):
stats = myutils.AvgMeter()
uncertainty_stats = myutils.AvgMeter()
progress_bar = tqdm(dataloader, desc=desc)
for iter_idx, sample in enumerate(progress_bar):
frames, masks, obj_n, info = sample
obj_n = obj_n.item()
if obj_n == 1:
continue
frames, masks = frames[0].to(device), masks[0].to(device)
fb_global = FeatureBank(obj_n, args.budget, device)
k4_list, v4_list = model.memorize(frames[0:1], masks[0:1])
fb_global.init_bank(k4_list, v4_list)
scores, uncertainty = model.segment(frames[1:], fb_global)
label = torch.argmax(masks[1:], dim=1).long()
optimizer.zero_grad()
loss = criterion(scores, label)
loss = loss + args.lu * uncertainty
loss.backward()
optimizer.step()
uncertainty_stats.update(uncertainty.item())
stats.update(loss.item())
progress_bar.set_postfix(loss=f'{loss.item():.5f} ({stats.avg:.5f} {uncertainty_stats.avg:.5f})')
# For debug
# print(info)
# myutils.vis_result(frames, masks, scores)
progress_bar.close()
return stats.avg
def main():
# torch.autograd.set_detect_anomaly(True)
if args.level == 0:
dataset = PreTrain_DS(args.dataset, output_size=400, clip_n=args.clip_n, max_obj_n=args.obj_n)
desc = 'Pre Train'
elif args.level == 1:
dataset = DAVIS_Train_DS(args.dataset, output_size=400, clip_n=args.clip_n, max_obj_n=args.obj_n)
desc = 'Train DAVIS17'
elif args.level == 2:
dataset = YouTube_Train_DS(args.dataset, output_size=400, clip_n=args.clip_n, max_obj_n=args.obj_n)
desc = 'Train YV18'
else:
raise ValueError(f'{args.level} is unknown.')
dataloader = data.DataLoader(dataset, batch_size=1, shuffle=True, num_workers=2, pin_memory=True)
print(myutils.gct(), f'Load level {args.level} dataset: {len(dataset)} training cases.')
model = AFB_URR(device, update_bank=False, load_imagenet_params=True)
model = model.to(device)
model.train()
model.apply(myutils.set_bn_eval) # turn-off BN
params = model.parameters()
optimizer = torch.optim.AdamW(filter(lambda x: x.requires_grad, params), args.lr)
start_epoch = 0
best_loss = 100000000
if args.resume:
if os.path.isfile(args.resume):
checkpoint = torch.load(args.resume)
model.load_state_dict(checkpoint['model'], strict=False)
seed = checkpoint['seed']
if not args.new:
start_epoch = checkpoint['epoch'] + 1
optimizer.load_state_dict(checkpoint['optimizer'])
best_loss = checkpoint['loss']
print(myutils.gct(),
f'Loaded checkpoint {args.resume} (epoch: {start_epoch-1}, best loss: {best_loss})')
else:
if args.seed < 0:
seed = int(time.time())
else:
seed = args.seed
print(myutils.gct(), f'Loaded checkpoint {args.resume}. Train from the beginning.')
else:
print(myutils.gct(), f'No checkpoint found at {args.resume}')
raise IOError
else:
if args.seed < 0:
seed = int(time.time())
else:
seed = args.seed
print(myutils.gct(), 'Random seed:', seed)
torch.manual_seed(seed)
np.random.seed(seed)
criterion = torch.nn.CrossEntropyLoss().to(device)
scheduler = torch.optim.lr_scheduler.StepLR(
optimizer, step_size=args.scheduler_step, gamma=0.5, last_epoch=start_epoch - 1)
for epoch in range(start_epoch, args.total_epochs):
lr = scheduler.get_last_lr()[0]
print('')
print(myutils.gct(), f'Epoch: {epoch} lr: {lr}')
loss = train_model(model, dataloader, criterion, optimizer, desc)
if args.log:
checkpoint = {
'epoch': epoch,
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'loss': loss,
'seed': seed,
}
checkpoint_path = f'{model_path}/final.pth'
torch.save(checkpoint, checkpoint_path)
if best_loss > loss:
best_loss = loss
checkpoint_path = f'{model_path}/epoch_{epoch:03d}_loss_{loss:.03f}.pth'
torch.save(checkpoint, checkpoint_path)
checkpoint_path = f'{model_path}/best.pth'
torch.save(checkpoint, checkpoint_path)
print('Best model updated.')
scheduler.step()
if __name__ == '__main__':
args = get_args()
print(myutils.gct(), f'Args = {args}')
if args.gpu >= 0 and torch.cuda.is_available():
device = torch.device('cuda', args.gpu)
else:
raise ValueError('CUDA is required. --gpu must be >= 0.')
if args.log:
if not os.path.exists('logs'):
os.makedirs('logs')
prefix = f'level{args.level}'
log_dir = 'logs/{}'.format(time.strftime(prefix + '_%Y%m%d-%H%M%S'))
log_path = os.path.join(log_dir, 'log')
model_path = os.path.join(log_dir, 'model')
if not os.path.exists(log_path):
os.makedirs(log_path)
if not os.path.exists(model_path):
os.makedirs(model_path)
myutils.save_scripts(log_dir, scripts_to_save=glob('*.*'))
myutils.save_scripts(log_dir, scripts_to_save=glob('dataset/*.py', recursive=True))
myutils.save_scripts(log_dir, scripts_to_save=glob('model/*.py', recursive=True))
myutils.save_scripts(log_dir, scripts_to_save=glob('myutils/*.py', recursive=True))
vis_writer = SummaryWriter(log_path)
vis_writer_step = 0
print(myutils.gct(), f'Create log dir: {log_dir}')
main()
if args.log:
vis_writer.close()
print(myutils.gct(), 'Training done.')