-
Notifications
You must be signed in to change notification settings - Fork 2
/
train_model.py
155 lines (133 loc) · 6.37 KB
/
train_model.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
import argparse
from loader import get_loader
from networks import UNet, ImageDiscriminator
from loss import SNDisLoss, SNGenLoss, NCMSE
import torch
import time
parser = argparse.ArgumentParser()
parser.add_argument('--mode', type=str, default='train')
parser.add_argument('--load_mode', type=int, default=1)
parser.add_argument('--data_path', type=str, default='/home/sutanu/CT_d/CT')
parser.add_argument('--saved_path', type=str, default='/home/sutanu/CT_d/data/npy_img/')
parser.add_argument('--save_path', type=str, default='/home/sutanu/CT_d/Checkpoint/save/')
parser.add_argument('--test_patient', type=str, default='L506')
parser.add_argument('--norm_range_min', type=float, default=-1024.0)
parser.add_argument('--norm_range_max', type=float, default=3072.0)
parser.add_argument('--trunc_min', type=float, default=-160.0)
parser.add_argument('--trunc_max', type=float, default=240.0)
parser.add_argument('--transform', type=bool, default=False)
# if patch training, batch size is (--patch_n * --batch_size)
parser.add_argument('--patch_n', type=int, default=4)
parser.add_argument('--patch_size', type=int, default=100)
parser.add_argument('--batch_size', type=int, default=5)
parser.add_argument('--num_epochs', type=int, default=1000)
parser.add_argument('--print_iters', type=int, default=20)
parser.add_argument('--decay_iters', type=int, default=6000)
parser.add_argument('--save_iters', type=int, default=1000)
parser.add_argument('--test_iters', type=int, default=1000)
parser.add_argument('--lr', type=float, default=1e-3)
parser.add_argument('--device', type=str)
parser.add_argument('--num_workers', type=int, default=7)
parser.add_argument('--multi_gpu', type=bool, default=False)
parser.add_argument('--load_chkpt', type=bool, default=False)
args = parser.parse_args()
data_loader = get_loader(mode=args.mode,
load_mode=args.load_mode,
saved_path=args.saved_path,
test_patient=args.test_patient,
patch_n=(args.patch_n if args.mode=='train' else None),
patch_size=(args.patch_size if args.mode=='train' else None),
transform=args.transform,
batch_size=(args.batch_size if args.mode=='train' else 1),
num_workers=args.num_workers)
if args.load_chkpt:
print('Loading Chekpoint')
whole_model = torch.load(args.save_path+ 'latest_ckpt.pth.tar')
netG_state_dict,optG_state_dict = whole_model['netG_state_dict'], whole_model['optG_state_dict']
netD_state_dict,optD_state_dict = whole_model['netD_state_dict'], whole_model['optD_state_dict']
netG = UNet()
netG = netG.cuda()
netD= ImageDiscriminator()
netD=netD.cuda()
optG = torch.optim.Adam(netG.parameters())
optD= torch.optim.Adam(netD.parameters())
netD.load_state_dict(netD_state_dict)
netG.load_state_dict(netG_state_dict)
optG.load_state_dict(optG_state_dict)
optD.load_state_dict(optD_state_dict)
cur_epoch = whole_model['epoch']
total_iters = whole_model['total_iters']
lr = whole_model['lr']
# netG = torch.nn.DataParallel(netG, device_ids=[0, 1])
# netD = torch.nn.DataParallel(netD, device_ids=[0, 1])
print(cur_epoch)
else:
print('Training model from scrath')
netG = UNet()
netG = netG.cuda()
netD= ImageDiscriminator()
netD=netD.cuda()
optG = torch.optim.Adam(netG.parameters(), lr=args.lr)
optD = torch.optim.Adam(netD.parameters(), lr=4*args.lr)
cur_epoch = 0
total_iters = 0
lr=args.lr
train_losses = []
criterion= NCMSE()
GANLoss=SNGenLoss()
DLoss=SNDisLoss()
criterion = criterion.cuda()
start_time = time.time()
for epoch in range(cur_epoch, args.num_epochs):
netG.train()
for iter_, (x, y) in enumerate(data_loader):
total_iters += 1
# add 1 channel
x = x.unsqueeze(0).float()
y = y.unsqueeze(0).float()
if args.patch_size: #patch training
x = x.view(-1, 1, args.patch_size, args.patch_size)
y = y.view(-1, 1, args.patch_size, args.patch_size)
x = x.cuda()
y = y.cuda()
pred = netG(x)
optD.zero_grad(),
netD.zero_grad(),
pos_neg_imgs = torch.cat([y, pred], dim=0)
pred_pos_neg = netD(pos_neg_imgs)
pred_pos, pred_neg = torch.chunk(pred_pos_neg, 2, dim=0)
d_loss = DLoss(pred_pos, pred_neg)
d_loss.backward(retain_graph=True)
optD.step()
optG.zero_grad()
netG.zero_grad()
g_loss = GANLoss(pred_neg)
rloss= criterion(pred, y, x)
loss = g_loss + 0.1*rloss
loss.backward()
optG.step()
# print
if total_iters % args.print_iters == 0:
print("STEP [{}], EPOCH [{}/{}], ITER [{}/{}] \nR_LOSS: {:.8f}, G_LOSS: {:.8f}, D_LOSS: {:.14f}, TIME: {:.1f}s".format(total_iters, epoch,
args.num_epochs, iter_+1,
len(data_loader), rloss.item(), g_loss.item(), d_loss.item()
,time.time() - start_time))
if total_iters % args.decay_iters == 0:
lr = lr * 0.5
for param_group in optG.param_groups:
param_group['lr'] = lr
for param_group in optD.param_groups:
param_group['lr'] = 4*lr
# save model
if total_iters % args.save_iters == 0:
saved_model = {
'epoch': epoch ,
'netG_state_dict': netG.state_dict(),
'optG_state_dict': optG.state_dict(),
'netD_state_dict': netD.state_dict(),
'optD_state_dict': optD.state_dict(),
'lr': lr,
'total_iters': total_iters
}
torch.save(saved_model, '{}iter_{}_ckpt.pth.tar'.format(args.save_path, total_iters))
torch.save(saved_model, '{}latest_ckpt.pth.tar'.format(args.save_path))