def main(mode='train', args=None, writer=None):
# create dataset and dataloader
dataset_path = os.path.join(args.data, args.dataset)
dataset_train = CNV_2d5(dataset_path, scale=(
args.crop_height, args.crop_width), mode='train')
dataloader_train = DataLoader(dataset_train, batch_size=args.batch_size,
shuffle=True, num_workers=args.num_workers, pin_memory=True, drop_last=True)
dataset_val = CNV_2d5(dataset_path, scale=(
args.crop_height, args.crop_width), mode='val')
dataloader_val = DataLoader(dataset_val, batch_size=1, shuffle=True,
num_workers=args.num_workers, pin_memory=True, drop_last=True)
dataset_test = CNV_2d5(dataset_path, scale=(
args.crop_height, args.crop_width), mode='test')
dataloader_test = DataLoader(dataset_test, batch_size=1, shuffle=True,
num_workers=args.num_workers, pin_memory=True, drop_last=True)
# build model
os.environ['CUDA_VISIBLE_DEVICES'] = args.cuda
# load model
model_all = {'unet_2.5d': UNet(in_channels=3, n_classes=args.num_classes),
'resunetplusplus': ResUnetPlusPlus(channel=3),
'cpfnet_2.5d':CPFNet(),
}
model = model_all[args.net_work].cuda()
cudnn.benchmark = True
if(args.optimizer=="SGD"):
optimizer = torch.optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay)
elif(args.optimizer=="Adam"):
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr,betas=(0.9, 0.999), eps=1e-08, weight_decay=args.weight_decay, amsgrad=False)
criterion_aux = nn.BCELoss()
criterion_main = LS.DiceLoss()
# criterion_0 = LS.SD_Loss()
criterion = [criterion_aux, criterion_main]
if mode == 'train': # tv
train(args, model, optimizer, criterion, dataloader_train, dataloader_val, writer)
if mode == 'test': # 单独使用测试集
eval(args, model, dataloader_test)
if mode == 'train_test':
train(args, model, optimizer, criterion, dataloader_train, dataloader_val,writer)
eval(args, model, dataloader_test)
def test_dice(args, epoch, model, testLoader, optimizer, testF, config):
model.eval()
test_loss = 0
dice = [0.0, 0.0]
# no gradient computation
dice_loss = dloss.DiceLoss(nclass=3)
with torch.no_grad():
for data, target in testLoader:
if args.cuda:
data, target = data.cuda(), target.type(
torch.LongTensor).cuda()
data, target = Variable(data,
requires_grad=False), Variable(target)
output = model(data)
# -------------- Process multi supervision in a model -----------------------------------------
if isinstance(output, list):
loss1 = []
loss2 = []
tar_numeled = target.view(target.numel())
tar_size = target.size()
for i, out in enumerate(output):
if i < (len(output) - 1):
out = F.interpolate(out,
size=tar_size[2:5],
mode='trilinear')
loss1.append(
dice_loss(out, target).detach().data.cpu().numpy())
out = out.permute(0, 2, 3, 4, 1).contiguous()
out = out.view(out.numel() // 3, 3)
loss2.append(
F.cross_entropy(
out, tar_numeled).detach().data.cpu().numpy())
test_loss += reduce(lambda x, y: x + y, loss1)
test_loss += reduce(lambda x, y: x + y, loss2)
target = target[:, 0, :, :, :]
pred = torch.argmax(output[-1], dim=1)
assert len(target.size()) == len(pred.size())
else:
output = F.softmax(output, dim=1)
test_loss += dice_loss(output,
target).detach().data.cpu().numpy()
# output = output.permute(0, 2, 3, 4, 1).contiguous()
# output = output.view(output.numel()//3, 3)
# target = target.view(target.numel())
target = target.view(target.size(0), target.size(2),
target.size(3), target.size(4))
# test_loss += F.cross_entropy(output,
# target).detach().data.cpu().numpy()
# get the index of the max probability
pred = torch.argmax(output, dim=1)
d = evaluate_dice(pred, target, cpu=True)
dice[0] += d[0]
dice[1] += d[1]
# loss function already averages over batch size
test_loss /= len(testLoader)
dice = [c / len(testLoader) for c in dice] # average dice on every sample
print(
'\nTest set: Average loss: {:.4f}, Kidney_Dice: {:.6f}, Tumor_Dice: {:.6f}\n'
.format(test_loss, dice[0], dice[1]))
#
testF.write('{},{},{},{}\n'.format(epoch, test_loss, dice[0], dice[1]))
testF.flush()
return (dice[0] + dice[1])
def train_dice(args, epoch, model, trainLoader, optimizer, trainF, config,
scheduler):
model.train()
nProcessed = 0
nTrain = len(trainLoader.dataset)
# nIter_per_epoch = nTrain // batch_size
dice_loss = dloss.DiceLoss(nclass=3)
for batch_idx, (data, target) in enumerate(trainLoader):
if args.cuda:
data, target = data.cuda(), target.type(torch.LongTensor).cuda()
data, target = Variable(data), Variable(target)
optimizer.zero_grad()
# output is already the result of softmax function
# may add CrossEntropy here
output = model(data)
# tar_numeled = False
# -------------- Process multi supervision in a model -----------------------------------------
if isinstance(output, list):
loss1 = []
loss2 = []
tar_size = target.size()
tar_numeled = target.view(target.numel())
for i, out in enumerate(output):
if i < (len(output) - 1):
out = F.interpolate(out,
size=tar_size[2:5],
mode='trilinear')
loss1.append(dice_loss(out, target))
out = out.permute(0, 2, 3, 4, 1).contiguous()
out = out.view(out.numel() // 3, 3)
loss2.append(F.cross_entropy(out, tar_numeled))
loss = reduce(lambda x, y: x + y, loss1)
loss += reduce(lambda x, y: x + y, loss2)
target = target[:, 0, :, :, :]
pred = torch.argmax(output[-1], dim=1)
assert len(target.size()) == len(pred.size())
else:
output = F.softmax(output, dim=1)
loss = dice_loss(output, target)
# output = output.permute(0, 2, 3, 4, 1).contiguous()
# output = output.view(output.numel()//3, 3)
# target = target.view(target.numel())
target = target.view(target.size(0), target.size(2),
target.size(3), target.size(4))
# loss += F.cross_entropy(output, target)
# get the index of the max log-probability
pred = torch.argmax(output, dim=1)
# ----------- apex mixed precision ------------------------------------------------
# loss.backward() becomes:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
optimizer.step()
# update learning rate
scheduler(optimizer, i=batch_idx, epoch=epoch)
nProcessed += len(data)
# print(output.size(), pred.size(), target.size())
dice = evaluate_dice(pred, target, cpu=True)
incorrect = pred.ne(target.data).cpu().sum()
partialEpoch = (int)(epoch + batch_idx / len(trainLoader))
loss_data = loss.detach().data.cpu().numpy()
print(
'Train Epoch: {} [{}/{} ({:.0f}%)], Loss: {:.4f}, Kidney_Dice: {:.6f}, Tumor_Dice: {:.6f}'
.format(partialEpoch, nProcessed, nTrain,
100. * batch_idx / len(trainLoader), loss_data, dice[0],
dice[1]))
#
trainF.write('{},{},{},{}\n'.format(partialEpoch, loss_data, dice[0],
dice[1]))
trainF.flush()
def main(mode='train', args=None, writer=None, k_fold=1):
# create dataset and dataloader
dataset_path = os.path.join(args.data, args.dataset)
dataset_train = LinearLesion(dataset_path,
scale=(args.crop_height, args.crop_width),
k_fold_test=k_fold,
mode='train')
dataloader_train = DataLoader(dataset_train,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
pin_memory=True,
drop_last=True)
dataset_val = LinearLesion(dataset_path,
scale=(args.crop_height, args.crop_width),
k_fold_test=k_fold,
mode='val')
dataloader_val = DataLoader(
dataset_val,
batch_size=len(
args.cuda.split(',')
), # the default is 1(the number of gpu), you can set it to what you want
shuffle=True,
num_workers=args.num_workers,
pin_memory=True,
drop_last=True)
dataset_test = LinearLesion(dataset_path,
scale=(args.crop_height, args.crop_width),
k_fold_test=k_fold,
mode='test')
dataloader_test = DataLoader(
dataset_test,
batch_size=len(
args.cuda.split(',')
), # the default is 1(the number of gpu), you can set it to what you want
shuffle=True,
num_workers=args.num_workers,
pin_memory=True,
drop_last=True)
# build model
os.environ['CUDA_VISIBLE_DEVICES'] = args.cuda
# load model
model_all = {
'UNet': UNet(in_channels=args.input_channel,
n_classes=args.num_classes)
}
model = model_all[args.net_work]
cudnn.benchmark = True
# model._initialize_weights()
if torch.cuda.is_available() and args.use_gpu:
model = torch.nn.DataParallel(model).cuda()
# load pretrained model if exists
if args.pretrained_model_path and mode == 'test':
print("=> loading pretrained model '{}'".format(
args.pretrained_model_path))
checkpoint = torch.load(args.pretrained_model_path)
model.load_state_dict(checkpoint['state_dict'])
print('Done!')
optimizer = torch.optim.SGD(model.parameters(),
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
# criterion_aux=nn.NLLLoss(weight=None)
criterion_aux = nn.BCEWithLogitsLoss(weight=None)
criterion_main = LS.DiceLoss()
criterion = [criterion_aux, criterion_main]
if mode == 'train':
train(args, model, optimizer, criterion, dataloader_train,
dataloader_val, writer, k_fold)
if mode == 'test':
eval(model, dataloader_test, args)
if mode == 'train_test':
train(args, model, optimizer, criterion, dataloader_train,
dataloader_val)
eval(model, dataloader_test, args)
def main(mode='train', args=None, writer=None, k_fold=1):
# create dataset and dataloader
dataset_path = os.path.join(args.data, args.dataset)
dataset_train = CNV_AND_SRF(dataset_path,
scale=(args.crop_height, args.crop_width),
k_fold_test=k_fold,
mode='train')
dataloader_train = DataLoader(dataset_train,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
pin_memory=True,
drop_last=True)
dataset_val = CNV_AND_SRF(dataset_path,
scale=(args.crop_height, args.crop_width),
k_fold_test=k_fold,
mode='val')
dataloader_val = DataLoader(dataset_val,
batch_size=1,
shuffle=True,
num_workers=args.num_workers,
pin_memory=True,
drop_last=True)
dataset_test = CNV_AND_SRF(dataset_path,
scale=(args.crop_height, args.crop_width),
k_fold_test=k_fold,
mode='test')
dataloader_test = DataLoader(dataset_test,
batch_size=1,
shuffle=True,
num_workers=args.num_workers,
pin_memory=True,
drop_last=True)
# build model
os.environ['CUDA_VISIBLE_DEVICES'] = args.cuda
# load model
model_all = {
'UNet': UNet(in_channels=args.input_channel,
n_classes=args.num_classes),
'ResUnetPlusPlus': ResUnetPlusPlus(channel=args.input_channel)
}
model = model_all[args.net_work].cuda()
cudnn.benchmark = True
# if torch.cuda.is_available() and args.use_gpu:
# model = torch.nn.DataParallel(model).cuda()
if args.pretrained_model_path and model == 'test':
print("=> loading pretrained model '{}'".format(
args.pretrained_model_path))
checkpoint = torch.load(args.pretrained_model_path)
model.load_state_dict(checkpoint['state_dict'])
print('Done!')
optimizer = torch.optim.SGD(model.parameters(),
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
criterion_aux = nn.BCEWithLogitsLoss(weight=None)
criterion_main = LS.DiceLoss()
criterion = [criterion_aux, criterion_main]
if mode == 'train': # 交叉验证
train(args, model, optimizer, criterion, dataloader_train,
dataloader_val, writer, k_fold)
if mode == 'test': # 单独使用测试集
eval(model, dataloader_test, args)
shuffle=True)
val_loader = DataLoader(dataset=val_set,
batch_size=1,
num_workers=args.n_threads,
shuffle=False)
# model info
model = UNet3D(in_channels=1,
filter_num_list=[16, 32, 48, 64, 96],
class_num=args.n_labels).to(device)
optimizer = optim.Adam(model.parameters(), lr=args.lr)
# common.print_network(model)
# model = nn.DataParallel(model, device_ids=[0,1]) # multi-GPU
# loss=loss.DiceLoss(weight=np.array([0.2,0.3,0.5]))
loss = loss.DiceLoss(weight=np.array([0.3, 0.7]))
log = logger.Train_Logger(save_path, "train_log")
best = [0, np.inf] # 初始化最优模型的epoch和performance
trigger = 0 # early stop 计数器
for epoch in range(1, args.epochs + 1):
common.adjust_learning_rate(optimizer, epoch, args)
train_log = train(model, train_loader, optimizer, loss, args.n_labels)
val_log = val(model, val_loader, loss, args.n_labels)
log.update(epoch, train_log, val_log)
# Save checkpoint.
state = {
'net': model.state_dict(),
'optimizer': optimizer.state_dict(),