-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
176 lines (162 loc) · 6.99 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
import shutil
import torch
import time
import torch.nn as nn
from models.deeplab import Deeplab
from torch.autograd import Variable
from torch.utils import data
from loader.image_label_loader import imageLabelLoader
from util.confusion_matrix import ConfusionMatrix
import torchvision.models as models
def weights_init(m):
classname = m.__class__.__name__
if classname.find('Conv') != -1:
m.weight.data.normal_(0.0, 0.02)
elif classname.find('BatchNorm2d') != -1:
m.weight.data.normal_(1.0, 0.02)
m.bias.data.fill_(0)
def save_checkpoint(state, is_best, filename):
torch.save(state, filename)
if is_best:
shutil.copyfile(filename, './checkpoint/model_best.pth.tar')
def update_confusion_matrix(matrix, output, target):
values, indices = output.max(1)
output = indices
target = target.cpu().numpy()
output = output.cpu().numpy()
matrix.update(target, output)
return matrix
def train(train_loader, model, criterion, optimizer, epoch):
# switch to train mode
model.train()
for i, (images, labels) in enumerate(train_loader):
run_time = time.time()
labels = labels.cuda(async=True)
input_var = torch.autograd.Variable(images)
target_var = torch.autograd.Variable(labels)
# compute output
output = model.forward(input_var)
loss = criterion(output, target_var)/args['batch_size']
# compute gradient and do SGD step
optimizer.zero_grad()
loss.backward()
optimizer.step()
if i % args['print_freq'] == 0:
matrix = ConfusionMatrix()
update_confusion_matrix(matrix, output.data, labels)
run_time = time.time() - run_time
print('Epoch/Iter: [{epoch}/{iter}]\t'
'loss: {loss:.4f}\t'
'acc: {accuracy:.4f}\t'
'fg_acc: {fg_accuracy:.4f}\t'
'avg_prec: {avg_precision:.4f}\t'
'avg_rec: {avg_recall:.4f}\t'
'avg_f1: {avg_f1:.4f}\t'
'run_time:{run_time:.2f}\t'.format(
epoch=epoch, iter=i+epoch*len(train_loader), loss=loss.data[0], accuracy=matrix.accuracy(),
fg_accuracy=matrix.fg_accuracy(), avg_precision=matrix.avg_precision(),
avg_recall=matrix.avg_recall(), avg_f1core=matrix.avg_f1score(), run_time=run_time))
def validate(val_loader, model, criterion):
# switch to evaluate mode
model.eval()
run_time = time.time()
matrix = ConfusionMatrix(args['label_nums'])
loss = 0
for i, (images, labels) in enumerate(val_loader):
labels = labels.cuda(async=True)
input_var = torch.autograd.Variable(images, volatile=True)
target_var = torch.autograd.Variable(labels, volatile=True)
# compute output
output = model(input_var)
loss += criterion(output, target_var)/args['batch_size']
matrix = update_confusion_matrix(matrix, output.data, labels)
loss /= (i+1)
run_time = time.time() - run_time
print('=================================================')
print('val:'
'loss: {0:.4f}\t'
'accuracy: {1:.4f}\t'
'fg_accuracy: {2:.4f}\t'
'avg_precision: {3:.4f}\t'
'avg_recall: {4:.4f}\t'
'avg_f1score: {5:.4f}\t'
'run_time:{run_time:.2f}\t'
.format(loss.data[0], matrix.accuracy(),
matrix.fg_accuracy(), matrix.avg_precision(), matrix.avg_recall(), matrix.avg_f1score(),run_time=run_time))
print('=================================================')
return matrix.avg_f1score()
def get_parameters(model, parameter_name):
for name, param in model.named_parameters():
if name in [parameter_name]:
return param
def main():
train_loader = data.DataLoader(imageLabelLoader(args['data_path'],dataName=args['domainB'], phase='train'), batch_size=args['batch_size'],
num_workers=args['num_workers'], shuffle=True)
val_loader = data.DataLoader(imageLabelLoader(args['data_path'], dataName=args['domainB'], phase='val'), batch_size=args['batch_size'],
num_workers=args['num_workers'], shuffle=False)
model = Deeplab()
print(model)
if args['pretrain_model'] != '':
pretrained_dict = torch.load(args['weigths_pool'] + '/' + args['pretrain_model'])
model.weights_init(pretrained_dict=pretrained_dict)
else:
model.apply(weights_init())
ignored_params = list(map(id, model.fc8_1.parameters()))
ignored_params.extend(list(map(id, model.fc8_2.parameters())))
ignored_params.extend(list(map(id, model.fc8_3.parameters())))
ignored_params.extend(list(map(id, model.fc8_4.parameters())))
base_params = filter(lambda p: id(p) not in ignored_params,
model.parameters())
optimizer = torch.optim.SGD([
{'params': base_params},
{'params': get_parameters(model.fc8_1, 'weight'), 'lr': args['l_rate'] * 10},
{'params': get_parameters(model.fc8_2, 'weight'), 'lr': args['l_rate'] * 10},
{'params': get_parameters(model.fc8_3, 'weight'), 'lr': args['l_rate'] * 10},
{'params': get_parameters(model.fc8_4, 'weight'), 'lr': args['l_rate'] * 10},
{'params': get_parameters(model.fc8_1, 'bias'), 'lr': args['l_rate'] * 20},
{'params': get_parameters(model.fc8_2, 'bias'), 'lr': args['l_rate'] * 20},
{'params': get_parameters(model.fc8_3, 'bias'), 'lr': args['l_rate'] * 20},
{'params': get_parameters(model.fc8_4, 'bias'), 'lr': args['l_rate'] * 20},
], lr=args['l_rate'], momentum=0.9, weight_decay=5e-4)
criterion = nn.CrossEntropyLoss(size_average=False).cuda()
# multi GPUS
model = torch.nn.DataParallel(model,device_ids=args['device_ids']).cuda()
best_prec = 0
for epoch in range(args['n_epoch']):
train(train_loader, model, criterion, optimizer, epoch)
if epoch > 0 and epoch % 9 == 0:
prec = validate(val_loader, model, criterion)
is_best = prec > best_prec
best_prec = max(prec, best_prec)
save_checkpoint({
'epoch': epoch + 1,
'arch': 'deeplab(indoor)',
'state_dict': model.state_dict(),
'best_prec1': best_prec,
'optimizer': optimizer.state_dict(),
}, is_best,filename='./checkpoint/indoor_epoch_'+str(epoch)+'.pth.tar')
#break
if __name__ == '__main__':
global args
args = {
'test_init': False,
'label_nums': 12,
'l_rate': 1e-8,
'data_path': 'datasets',
'n_epoch': 1000,
'batch_size': 10,
'num_workers': 10,
'print_freq': 10,
'device_ids': [0],
'domainA': 'Lip',
'domainB': 'Indoor',
'weigths_pool': 'pretrain_models',
'pretrain_model': 'deeplab.pth',
'loadSizeH': 241,
'loadSizeW': 121,
'fineSizeH': 241,
'fineSizeW': 121,
'name':'v3',
'checkpoints_dir':'checkpoints',
}
main()