-
Notifications
You must be signed in to change notification settings - Fork 4
/
train.py
180 lines (134 loc) · 6.2 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
import datetime
import os
import torch
from torch import nn
from torch import optim
from torch.autograd import Variable
from torch.backends import cudnn
from torch.utils.data import DataLoader
from torchvision import transforms
import triple_transforms
from nets import depth_predciton, basic, basic_NL, DGNLNet
from config import train_raincityscapes_path, test_raincityscapes_path
from dataset3 import ImageFolder
from misc import AvgMeter, check_mkdir
# torch.cuda.set_device(0)
cudnn.benchmark = True
ckpt_path = './ckpt'
exp_name = 'DGNLNet'
args = {
'iter_num': 40000,
'train_batch_size': 4,
'last_iter': 0,
'lr': 5e-4,
'lr_decay': 0.9,
'weight_decay': 0,
'momentum': 0.9,
'resume_snapshot': '',
'val_freq': 50000000,
'img_size_h': 512,
'img_size_w': 1024,
'crop_size': 512,
'snapshot_epochs': 10000
}
transform = transforms.Compose([
transforms.ToTensor()
#transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
to_pil = transforms.ToPILImage()
triple_transform = triple_transforms.Compose([
triple_transforms.Resize((args['img_size_h'], args['img_size_w'])),
#triple_transforms.RandomCrop(args['crop_size']),
triple_transforms.RandomHorizontallyFlip()
])
train_set = ImageFolder(train_raincityscapes_path, transform=transform, target_transform=transform, triple_transform=triple_transform, is_train=True)
train_loader = DataLoader(train_set, batch_size=args['train_batch_size'], num_workers=8, shuffle=True)
test1_set = ImageFolder(test_raincityscapes_path, transform=transform, target_transform=transform, is_train=False)
test1_loader = DataLoader(test1_set, batch_size=2)
criterion = nn.L1Loss()
criterion_depth = nn.L1Loss()
log_path = os.path.join(ckpt_path, exp_name, str(datetime.datetime.now()) + '.txt')
def main():
net = DGNLNet().cuda().train()
optimizer = optim.Adam([
{'params': [param for name, param in net.named_parameters() if name[-4:] == 'bias' and param.requires_grad],
'lr': 2 * args['lr']},
{'params': [param for name, param in net.named_parameters() if name[-4:] != 'bias' and param.requires_grad],
'lr': args['lr'], 'weight_decay': args['weight_decay']}
])
if len(args['resume_snapshot']) > 0:
print('training resumes from \'%s\'' % args['resume_snapshot'])
net.load_state_dict(torch.load(os.path.join(ckpt_path, exp_name, args['resume_snapshot'] + '.pth')))
optimizer.load_state_dict(torch.load(os.path.join(ckpt_path, exp_name, args['resume_snapshot'] + '_optim.pth')))
optimizer.param_groups[0]['lr'] = 2 * args['lr']
optimizer.param_groups[1]['lr'] = args['lr']
check_mkdir(ckpt_path)
check_mkdir(os.path.join(ckpt_path, exp_name))
open(log_path, 'w').write(str(args) + '\n\n')
train(net, optimizer)
def train(net, optimizer):
curr_iter = args['last_iter']
while True:
train_loss_record = AvgMeter()
train_net_loss_record = AvgMeter()
train_depth_loss_record = AvgMeter()
for i, data in enumerate(train_loader):
optimizer.param_groups[0]['lr'] = 2 * args['lr'] * (1 - float(curr_iter) / args['iter_num']
) ** args['lr_decay']
optimizer.param_groups[1]['lr'] = args['lr'] * (1 - float(curr_iter) / args['iter_num']
) ** args['lr_decay']
inputs, gts, dps = data
batch_size = inputs.size(0)
inputs = Variable(inputs).cuda()
gts = Variable(gts).cuda()
dps = Variable(dps).cuda()
optimizer.zero_grad()
result, depth_pred = net(inputs)
loss_net = criterion(result, gts)
loss_depth = criterion_depth(depth_pred, dps)
loss = loss_net + loss_depth
loss.backward()
optimizer.step()
# for n, p in net.named_parameters():
# if n[-5:] == 'alpha':
# print(p.grad.data)
# print(p.data)
train_loss_record.update(loss.data, batch_size)
train_net_loss_record.update(loss_net.data, batch_size)
train_depth_loss_record.update(loss_depth.data, batch_size)
curr_iter += 1
log = '[iter %d], [train loss %.5f], [lr %.13f], [loss_net %.5f], [loss_depth %.5f]' % \
(curr_iter, train_loss_record.avg, optimizer.param_groups[1]['lr'],
train_net_loss_record.avg, train_depth_loss_record.avg)
print(log)
open(log_path, 'a').write(log + '\n')
if (curr_iter + 1) % args['val_freq'] == 0:
validate(net, curr_iter, optimizer)
if (curr_iter + 1) % args['snapshot_epochs'] == 0:
torch.save(net.state_dict(), os.path.join(ckpt_path, exp_name, ('%d.pth' % (curr_iter + 1) )))
torch.save(optimizer.state_dict(), os.path.join(ckpt_path, exp_name, ('%d_optim.pth' % (curr_iter + 1) )))
if curr_iter > args['iter_num']:
return
def validate(net, curr_iter, optimizer):
print('validating...')
net.eval()
loss_record1, loss_record2 = AvgMeter(), AvgMeter()
iter_num1 = len(test1_loader)
with torch.no_grad():
for i, data in enumerate(test1_loader):
inputs, gts, dps = data
inputs = Variable(inputs).cuda()
gts = Variable(gts).cuda()
dps = Variable(dps).cuda()
res = net(inputs)
loss = criterion(res, gts)
loss_record1.update(loss.data, inputs.size(0))
print('processed test1 %d / %d' % (i + 1, iter_num1))
snapshot_name = 'iter_%d_loss1_%.5f_loss2_%.5f_lr_%.6f' % (curr_iter + 1, loss_record1.avg, loss_record2.avg,
optimizer.param_groups[1]['lr'])
print('[validate]: [iter %d], [loss1 %.5f], [loss2 %.5f]' % (curr_iter + 1, loss_record1.avg, loss_record2.avg))
torch.save(net.state_dict(), os.path.join(ckpt_path, exp_name, snapshot_name + '.pth'))
torch.save(optimizer.state_dict(), os.path.join(ckpt_path, exp_name, snapshot_name + '_optim.pth'))
net.train()
if __name__ == '__main__':
main()