-
Notifications
You must be signed in to change notification settings - Fork 3
/
train.py
176 lines (135 loc) · 7.24 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
import torch.utils.data as data
import torchvision.transforms as transforms
from model.utils import Reconstruction3DDataLoader, Reconstruction3DDataLoaderJump
from model.autoencoder import *
from utils import *
import argparse
parser = argparse.ArgumentParser(description="STEAL Net")
parser.add_argument('--batch_size', type=int, default=4, help='batch size for training')
parser.add_argument('--epochs', type=int, default=60, help='number of epochs for training')
parser.add_argument('--h', type=int, default=256, help='height of input images')
parser.add_argument('--w', type=int, default=256, help='width of input images')
parser.add_argument('--lr', type=float, default=1e-4, help='initial learning rate phase 1')
parser.add_argument('--num_workers', type=int, default=2, help='number of workers for the train loader')
parser.add_argument('--dataset_type', type=str, default='ped2', choices=['ped2','avenue', 'shanghai'], help='type of dataset: ped2, avenue, shanghai')
parser.add_argument('--dataset_path', type=str, default='dataset', help='directory of data')
parser.add_argument('--exp_dir', type=str, default='log', help='basename of folder to save weights')
parser.add_argument('--optimizer', type=str, default='adam', choices=['adam','sgd'], help='adam or sgd with momentum and cosine annealing lr')
parser.add_argument('--model_dir', type=str, default=None, help='path of model for resume')
parser.add_argument('--start_epoch', type=int, default=0, help='start epoch. usually number in filename + 1')
# related to skipping frame pseudo anomaly
parser.add_argument('--pseudo_anomaly_jump', type=float, default=0, help='pseudo anomaly jump frame (skip frame) probability. 0 no pseudo anomaly')
parser.add_argument('--jump', nargs='+', type=int, default=[3], help='Jump for pseudo anomaly (hyperparameter s)') # --jump 2 3
parser.add_argument('--print_all', action='store_true', help='print all reconstruction loss')
##################
args = parser.parse_args()
# assert 1 not in args.jump
exp_dir = args.exp_dir
exp_dir += 'lr' + str(args.lr) if args.lr != 1e-4 else ''
exp_dir += 'weight'
exp_dir += '_recon'
exp_dir += '_pajump' + str(args.pseudo_anomaly_jump) if args.pseudo_anomaly_jump != 0 else ''
exp_dir += '_jump[' + ','.join([str(args.jump[i]) for i in range(0,len(args.jump))]) + ']' if args.pseudo_anomaly_jump != 0 else ''
print('exp_dir: ', exp_dir)
torch.backends.cudnn.enabled = True # make sure to use cudnn for computational performance
train_folder = os.path.join(args.dataset_path, args.dataset_type, 'training', 'frames')
# Loading dataset
img_extension = '.tif' if args.dataset_type == 'ped1' else '.jpg'
train_dataset = Reconstruction3DDataLoader(train_folder, transforms.Compose([transforms.ToTensor()]),
resize_height=args.h, resize_width=args.w, dataset=args.dataset_type, img_extension=img_extension)
train_dataset_jump = Reconstruction3DDataLoaderJump(train_folder, transforms.Compose([transforms.ToTensor()]),
resize_height=args.h, resize_width=args.w, dataset=args.dataset_type, jump=args.jump, img_extension=img_extension)
train_size = len(train_dataset)
train_batch = data.DataLoader(train_dataset, batch_size=args.batch_size,
shuffle=True, num_workers=args.num_workers, drop_last=True)
train_batch_jump = data.DataLoader(train_dataset_jump, batch_size=args.batch_size,
shuffle=True, num_workers=args.num_workers, drop_last=True)
# Report the training process
log_dir = os.path.join('./exp', args.dataset_type, exp_dir)
if not os.path.exists(log_dir):
os.makedirs(log_dir)
orig_stdout = sys.stdout
f = open(os.path.join(log_dir, 'log.txt'), 'a')
sys.stdout = f
torch.set_printoptions(profile="full")
loss_func_mse = nn.MSELoss(reduction='none')
if args.start_epoch < args.epochs:
model = convAE()
model = nn.DataParallel(model)
model.cuda()
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)
# resume
if args.model_dir is not None:
assert args.start_epoch > 0
# Loading the trained model
model_dict = torch.load(args.model_dir)
model_weight = model_dict['model']
model.load_state_dict(model_weight.state_dict())
optimizer.load_state_dict(model_dict['optimizer'])
model.cuda()
# model.eval()
for epoch in range(args.start_epoch, args.epochs):
pseudolossepoch = 0
lossepoch = 0
pseudolosscounter = 0
losscounter = 0
for j, (imgs, imgsjump) in enumerate(zip(train_batch, train_batch_jump)):
net_in = copy.deepcopy(imgs)
net_in = net_in.cuda()
jump_pseudo_stat = []
cls_labels = []
for b in range(args.batch_size):
total_pseudo_prob = 0
rand_number = np.random.rand()
pseudo_bool = False
# skip frame pseudo anomaly
pseudo_anomaly_jump = total_pseudo_prob <= rand_number < total_pseudo_prob + args.pseudo_anomaly_jump
total_pseudo_prob += args.pseudo_anomaly_jump
if pseudo_anomaly_jump:
net_in[b] = imgsjump[0][b]
jump_pseudo_stat.append(True)
pseudo_bool = True
else:
jump_pseudo_stat.append(False)
if pseudo_bool:
cls_labels.append(0)
else:
cls_labels.append(1)
########## TRAIN
outputs = model(net_in)
cls_labels = torch.Tensor(cls_labels).unsqueeze(1).cuda()
loss_mse = loss_func_mse(outputs, net_in)
modified_loss_mse = []
for b in range(args.batch_size):
if jump_pseudo_stat[b]:
modified_loss_mse.append(torch.mean(-loss_mse[b]))
pseudolossepoch += modified_loss_mse[-1].cpu().detach().item()
pseudolosscounter += 1
else: # no pseudo anomaly
modified_loss_mse.append(torch.mean(loss_mse[b]))
lossepoch += modified_loss_mse[-1].cpu().detach().item()
losscounter += 1
assert len(modified_loss_mse) == loss_mse.size(0)
stacked_loss_mse = torch.stack(modified_loss_mse)
loss = torch.mean(stacked_loss_mse)
optimizer.zero_grad()
loss.backward()
optimizer.step()
if j % 10 == 0 or args.print_all:
print("epoch {:d} iter {:d}/{:d}".format(epoch, j, len(train_batch)))
print('Loss: {:.6f}'.format(loss.item()))
print('----------------------------------------')
print('Epoch:', epoch)
if pseudolosscounter != 0:
print('PseudoMeanLoss: Reconstruction {:.9f}'.format(pseudolossepoch/pseudolosscounter))
if losscounter != 0:
print('MeanLoss: Reconstruction {:.9f}'.format(lossepoch/losscounter))
# Save the model and the memory items
model_dict = {
'model': model,
'optimizer': optimizer.state_dict(),
}
torch.save(model_dict, os.path.join(log_dir, 'model_{:02d}.pth'.format(epoch)))
print('Training is finished')
sys.stdout = orig_stdout
f.close()