-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
175 lines (148 loc) · 7.6 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
163
164
165
166
167
168
169
170
171
172
173
174
175
import argparse
import os
import time
import traceback
from torch import optim
from torch.utils.data import DataLoader
from tqdm import tqdm
from unet.networks import define_G
from unet.unet_model import *
from utils.dataset import BrainDataset
from utils.init_logging import init_logging
dir_checkpoint = 'checkpoints/'
train_list = 'data/train_brats_only_tumor.txt'
test_list = 'data/test_brats.txt'
def train_net(reconstucter, segmenter, device, epochs=5, batch_size=1, lr=0.001, percent=0.1, save_cp=True, ):
train_set = BrainDataset(test_list, percent=percent)
train_loader = DataLoader(train_set, batch_size=batch_size, shuffle=True, num_workers=8, pin_memory=True, drop_last=True)
n_train = len(train_loader)
logger.info(f'''Starting training:
File names main.py
Starting time {starttime_r}
Epochs: {epochs}
Batch size: {batch_size}
Learning rate: {lr}
λ: {lambd}
Training size: {n_train}
Checkpoints: {save_cp}
Device: {device.type}
Training percent: {percent}
''')
optimizer_R = optim.Adam(reconstucter.parameters(), lr=lr, betas=(0.9, 0.99))
optimizer_S = optim.Adam(segmenter.parameters(), lr=lr, betas=(0.9, 0.99))
scheduler_R = optim.lr_scheduler.StepLR(optimizer_R, step_size=int(0.8 * epochs))
scheduler_S = optim.lr_scheduler.StepLR(optimizer_S, step_size=int(0.8 * epochs))
ce_loss = nn.CrossEntropyLoss()
ce_loss1 = nn.CrossEntropyLoss(reduce=False)
mse_loss = nn.MSELoss(reduce=False)
for epoch in range(epochs):
reconstucter.train()
segmenter.train()
loss1 = 0
loss2 = 0
loss3 = 0
loss4 = 0
logger.info("Epoch {:2d} learning(S) rate: {:.2e}, learning(R) rate: {:.2e}, ".format(epoch, optimizer_S.param_groups[0]['lr'],
optimizer_R.param_groups[0]['lr']))
bar = tqdm(enumerate(train_loader), total=n_train, desc=f'Epoch {epoch + 1}/{epochs}', unit='batch', ncols=120, ascii=True)
for i, data in bar:
[img, seg, seg_tumor, brain_mask, _, _] = data[0:6]
img = img.unsqueeze(dim=1).to(device=device, dtype=torch.float32)
seg_tumor = seg_tumor.to(device=device, dtype=torch.long)
brain_mask = brain_mask.to(device=device, dtype=torch.float32)
seg_no_tumor = torch.zeros_like(seg).to(device=device, dtype=torch.long)
seg_tumor_neg = 1 - seg_tumor
mask = (seg_tumor_neg * brain_mask).unsqueeze(dim=1)
brain_avg = torch.ones_like(img) * img[mask == 1].mean()
# train step 1:
optimizer_R.zero_grad()
img_pred = reconstucter(img)
seg_pred = segmenter(img_pred)
loss_seg2 = lambd[1] * ce_loss(seg_pred, seg_no_tumor)
loss_mse = lambd[2] * (mse_loss(img_pred, img) * seg_tumor_neg).mean()
loss_var = lambd[3] * (mse_loss(img_pred, brain_avg.to(dtype=torch.float)).squeeze(1) * seg_tumor).mean()
loss_recer = loss_seg2 + loss_mse + loss_var
loss_recer.backward()
optimizer_R.step()
# train step 2:
optimizer_S.zero_grad()
seg_pred = segmenter(img_pred.detach())
# Calculate weight
diff = torch.abs(img_pred.detach() - img).squeeze()
diff_norm = diff / diff.view(batch_size, -1).max(dim=-1)[0].view(batch_size, 1, 1)
weightt = 1 - diff_norm
weightt = torch.where(weightt < 0.1, 0.1 * torch.ones_like(weightt), weightt)
loss_seg1 = lambd[0] * (ce_loss1(seg_pred, seg_tumor) * weightt).mean()
loss_seg1.backward()
optimizer_S.step()
bar.set_postfix_str(
'loss(batch):{:>.2e},{:>.2e},{:>.2e},{:.2e}'.format(loss_seg1.item(), loss_seg2.item(), loss_mse.item(), loss_var.item()))
loss1 += loss_seg1
loss2 += loss_seg2
loss3 += loss_mse
loss4 += loss_var
logger.info('Epoch loss: {:.2e}, {:.2e}, {:.2e}, {:.2e}'.format(loss1 / i, loss2 / i, loss3 / i, loss4 / i))
scheduler_S.step()
scheduler_R.step()
if save_cp and (epoch + 1) % 5 == 0:
torch.save({'segmenter': segmenter.state_dict(),
'optimizer_S': optimizer_S.state_dict(),
'scheduler_S': scheduler_S.state_dict(),
'reconstucter': reconstucter.state_dict(),
'optimizer_R': optimizer_R.state_dict(),
'scheduler_R': scheduler_R.state_dict(),
'epoch': epoch},
dir_checkpoint + '%s_%s.pth' % (starttime, epoch + 1))
logger.info(f'Checkpoint {epoch + 1} saved !')
def get_args():
parser = argparse.ArgumentParser(description='Train the UNet on images and target masks',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('-e', '--epochs', metavar='E', type=int, default=10,
help='Number of epochs', dest='epochs')
parser.add_argument('-b', '--batch_size', metavar='B', type=int, nargs='?', default=2,
help='Batch size', dest='batch_size')
parser.add_argument('-l', '--learning-rate', metavar='LR', type=float, nargs='?', default=1e-3,
help='Learning rate', dest='lr')
parser.add_argument('-g', '--gpu', dest='gpu', default=0, type=int, help='witch gpu to use')
parser.add_argument('--percent', '-p', metavar='PERCENT', dest='percent', type=float, default=10.0,
help='Percent of the data that is used as training (0-100)')
parser.add_argument('--lambd', '-d', type=str, default='[1,1,10.0,5.0]', dest='lambd',
metavar='FILE', help="param of 3 losses")
return parser.parse_args()
if __name__ == '__main__':
t = time.localtime()
starttime = time.strftime("%Y-%m-%d_%H-%M-%S", t)
starttime_r = time.strftime("%Y/%m/%d %H:%M:%S", t) # readable time
logger = init_logging(starttime)
args = get_args()
lambd = eval(args.lambd)
os.environ["CUDA_VISIBLE_DEVICES"] = str(args.gpu)
torch.backends.cudnn.benchmark = True
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
logger.info(f'Using device {device}')
reconstucter = define_G(input_nc=1, output_nc=1, ngf=64, netG='resnet_9blocks', norm='instance')
segmenter = Segmenter(n_channels=1, n_classes=2, bilinear=True)
# if args.load:
# segmenter.load_state_dict(
# torch.load(args.load, map_location=device)
# )
# logger.info(f'Model loaded from {args.load}')
reconstucter.to(device=device)
segmenter.to(device=device)
try:
train_net(reconstucter=reconstucter,
segmenter=segmenter,
epochs=args.epochs,
batch_size=args.batch_size,
lr=args.lr,
device=device,
percent=args.percent / 100)
logger.info('Program done. Start on %s' % (starttime_r))
except KeyboardInterrupt:
torch.save({'reconstucter': reconstucter.state_dict(),
'segmenter': segmenter.state_dict()}, './checkpoints/INTERRUPTED.pth')
logger.error('User canceled, start on %s, Saved interrupt.' % (starttime_r))
os.remove(f'./log/{starttime[:10]}/{starttime}.txt')
except:
logger.error('\n' + (traceback.format_exc()))
logger.info('Error! Start on %s' % (starttime_r))