forked from jcdubron/scnn_pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain.py
162 lines (153 loc) · 8.5 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
import math
from torch import optim
from torch.utils.data import DataLoader
from torchvision import transforms
from tensorboardX import SummaryWriter
# import sys
# sys.path.append(os.getcwd())
from python.model import *
from python.transforms import *
from python.dataset import *
from python.train import *
from python.test import *
from python.prob2lines import *
# parser
parser = argparse.ArgumentParser(description='PyTorch SCNN Model')
parser.add_argument('--train_data_dir', metavar='DIR', default='/home/dwt/scnn_pytorch',
help='path to train dataset (default: /home/dwt/scnn_pytorch)')
parser.add_argument('--eval_data_dir', metavar='DIR', default='/home/dwt/scnn_pytorch',
help='path to eval dataset (default: /home/dwt/scnn_pytorch)')
parser.add_argument('--test_data_dir', metavar='DIR', default=None,
help='path to test dataset')
parser.add_argument('--train_list_file', metavar='DIR', default='train.txt',
help='train list file (default: train.txt)')
parser.add_argument('--eval_list_file', metavar='DIR', default='eval.txt',
help='eval list file (default: eval.txt)')
parser.add_argument('--test_list_file', metavar='DIR', nargs='+', default=None,
help='test list file')
parser.add_argument('--batch_size', type=int, default=8, metavar='N',
help='batch size for training (default: 8)')
parser.add_argument('--epoches', type=int, default=10, metavar='N',
help='number of epoches to train (default: 10)')
parser.add_argument('--batches', type=int, default=60000, metavar='N',
help='number of batches to train (default: 60000)')
parser.add_argument('--lr', type=float, default=0.01, metavar='LR',
help='learning rate (default: 0.01)')
parser.add_argument('--momentum', type=float, default=0.9, metavar='M',
help='SGD momentum (default: 0.9)')
parser.add_argument('--weight_decay', default=1e-4, type=float, metavar='W',
help='weight decay (default: 1e-4)')
parser.add_argument('--bce_weight', type=float, default=0.1, metavar='M',
help='binary cross entropy loss weight (default: 0.1)')
parser.add_argument('--bg_weight', type=float, default=0.4, metavar='M',
help='background loss weight (default: 0.4)')
parser.add_argument('--gpu', metavar='N', type=int, nargs='+', default=-1,
help='GPU ids')
parser.add_argument('--seed', type=int, default=1, metavar='S',
help='random seed (default: 1)')
parser.add_argument('--checkpoint', metavar='DIR', default=None,
help='use pre-trained model')
parser.add_argument('--weights', metavar='DIR', default=None,
help='use finetuned model')
parser.add_argument('--log_interval', type=int, default=100, metavar='N',
help='how many batches to wait before logging training status (default: 100)')
parser.add_argument('--snapshot_interval', type=int, default=1, metavar='N',
help='how many epoches to wait before saving snapshot (default: 2)')
parser.add_argument('--snapshot_prefix', type=str, default='./snapshot/model', metavar='PATH',
help='snapshot prefix (default: ./snapshot/model)')
parser.add_argument('--tensorboard', type=str, default='log', metavar='PATH',
help='tensorboard log path (default: log)')
parser.add_argument('--test_log', type=str, default='test_result.txt', metavar='PATH',
help='tensorboard log path (default: test_result.txt)')
args = parser.parse_args()
# tensorboardX
writer = SummaryWriter(args.tensorboard)
# random seed
np.random.seed(args.seed)
torch.manual_seed(args.seed)
# cuda and seed
use_cuda = args.gpu[0]>=0 and torch.cuda.is_available()
device = torch.device('cuda:{0}'.format(args.gpu[0]) if use_cuda else 'cpu')
torch.manual_seed(args.seed)
if use_cuda:
print('Use Device: GPU', args.gpu)
else:
print('Use Device: CPU')
# model
model = SCNN().to(device)
if len(args.gpu) > 1:
model = torch.nn.DataParallel(model, device_ids=args.gpu)
# scheduler
# optimizer = optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay)
fc9_params = [id(p) for p in model.fc9.parameters()]
fc10_params = [id(p) for p in model.fc10.parameters()]
base_params = filter(lambda p: id(p) not in fc9_params + fc10_params, model.parameters())
optimizer = optim.SGD([{'params': base_params},
{'params': model.fc9.parameters(), 'lr': args.lr * 10},
{'params': model.fc10.parameters(), 'lr': args.lr * 10}],
lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay)
scheduler = optim.lr_scheduler.LambdaLR(optimizer, lambda epoch: math.pow(1-epoch/args.batches, 0.9))
epoch_start = 1
# continue training from checkpoint
if args.checkpoint is not None:
assert os.path.isfile(args.checkpoint)
print('Start loading weights from [{0}].'.format(args.checkpoint))
checkpoint = torch.load(args.checkpoint)
model.load_state_dict(checkpoint['model_state_dict'])
# scheduler.load_state_dict(checkpoint['scheduler_state_dict'])
for k,v in checkpoint['model_state_dict'].items():
print('load weights', k, v.shape)
if 'epoch' in checkpoint:
epoch_start = checkpoint['epoch'] + 1
print('Loading checkpoint done.')
# finetune training with weights
elif args.weights is not None:
assert os.path.isfile(args.weights)
print('Start loading weights from [{0}].'.format(args.weights))
model_dict = model.state_dict()
# for k,v in model_dict.items():
# print('original model', k, v.shape)
weights = torch.load(args.weights)
weights = {k: v for k, v in weights.items() if k in model.state_dict()}
for k,v in weights.items():
print('load weights', k, v.shape)
model_dict.update(weights)
# for k,v in model_dict.items():
# print('After load', k, v.shape)
model.load_state_dict(model_dict)
print('Loading weights done.')
# train or test
# mean=[0.37042467, 0.36758537, 0.3584016]
mean = [0.3598, 0.3653, 0.3662]
std = [0.2573, 0.2663, 0.2756]
if args.test_list_file is not None:
for idx in range(len(args.test_list_file)):
print('Start test dataset initialization [{0}].'.format(args.test_list_file[idx]))
test_dataset = TestLaneDataset(img_dir=args.test_data_dir, list_file=args.test_list_file[idx],
transform=transforms.Compose([TestSampleResize((800, 288)),
TestSampleToTensor(),
TestSampleNormalize(mean=mean, std=std)]))
test_loader = DataLoader(test_dataset, batch_size=args.batch_size,
shuffle=False, num_workers=4, drop_last=False)
print('Test Dataset initialization done.')
test(model, device, test_loader, args, idx)
else:
print('Start train dataset initialization [{0}].'.format(args.train_list_file))
train_dataset = LaneDataset(img_dir=args.train_data_dir, prob_dir=args.train_data_dir+'_labelmap',
list_file=args.train_list_file, tag=False,
transform=transforms.Compose([SampleResize((800, 288)),
SampleToTensor(),
SampleNormalize(mean=mean, std=std)]))
train_loader = DataLoader(train_dataset, batch_size=args.batch_size,
shuffle=True, num_workers=4, drop_last=False)
print('Train Dataset initialization done.')
print('Start train dataset initialization ({0}).'.format(args.eval_list_file))
eval_dataset = LaneDataset(img_dir=args.eval_data_dir, prob_dir=args.eval_data_dir+'_labelmap',
list_file=args.eval_list_file, tag=True,
transform=transforms.Compose([SampleResize((800, 288)),
SampleToTensor(),
SampleNormalize(mean=mean, std=std)]))
eval_loader = DataLoader(eval_dataset, batch_size=args.batch_size,
shuffle=False, num_workers=4, drop_last=False)
print('Eval Dataset initialization done.')
train(model, writer, args, device, train_loader, eval_loader, scheduler, epoch_start, loss_weight=(1, args.bce_weight))