datefmt='%H:%M:%S')
logging.getLogger().addHandler(logging.StreamHandler(sys.stdout))
logging.info(str(args))
net = VNetRec(n_channels=1,
n_classes=num_classes,
normalization='batchnorm',
has_dropout=True)
net = net.cuda()
db_train = LAHeart(base_dir=train_data_path,
split='train',
num=16,
transform=transforms.Compose([
RandomRotFlip(),
RandomCrop(patch_size),
ToTensor(),
]))
def worker_init_fn(worker_id):
random.seed(args.seed + worker_id)
trainloader = DataLoader(db_train,
batch_size=batch_size,
shuffle=True,
num_workers=4,
pin_memory=True,
worker_init_fn=worker_init_fn)
net.train()
optimizer = optim.SGD(net.parameters(),
def main():
###################
# init parameters #
###################
args = get_args()
# training path
train_data_path = args.root_path
# writer
idx = args.save.rfind('/')
log_dir = args.writer_dir + args.save[idx:]
writer = SummaryWriter(log_dir)
batch_size = args.batch_size * args.ngpu
max_iterations = args.max_iterations
base_lr = args.base_lr
patch_size = (112, 112, 80)
num_classes = 2
# random
if args.deterministic:
cudnn.benchmark = False
cudnn.deterministic = True
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
## make logger file
if os.path.exists(args.save):
shutil.rmtree(args.save)
os.makedirs(args.save, exist_ok=True)
snapshot_path = args.save
logging.basicConfig(filename=snapshot_path+"/log.txt", level=logging.INFO,
format='[%(asctime)s.%(msecs)03d] %(message)s', datefmt='%H:%M:%S')
logging.getLogger().addHandler(logging.StreamHandler(sys.stdout))
logging.info(str(args))
net = VNet(n_channels=1, n_classes=num_classes, normalization='batchnorm', has_dropout=True)
net = net.cuda()
db_train = LAHeart(base_dir=train_data_path,
split='train',
num=16,
transform = transforms.Compose([
RandomRotFlip(),
RandomCrop(patch_size),
ToTensor(),
]))
db_test = LAHeart(base_dir=train_data_path,
split='test',
transform = transforms.Compose([
CenterCrop(patch_size),
ToTensor()
]))
def worker_init_fn(worker_id):
random.seed(args.seed+worker_id)
trainloader = DataLoader(db_train, batch_size=batch_size, shuffle=True, num_workers=4, pin_memory=True, worker_init_fn=worker_init_fn)
net.train()
optimizer = optim.SGD(net.parameters(), lr=base_lr, momentum=0.9, weight_decay=0.0001)
logging.info("{} itertations per epoch".format(len(trainloader)))
iter_num = 0
max_epoch = max_iterations//len(trainloader)+1
lr_ = base_lr
net.train()
for epoch_num in tqdm(range(max_epoch), ncols=70):
time1 = time.time()
for i_batch, sampled_batch in enumerate(trainloader):
time2 = time.time()
# print('fetch data cost {}'.format(time2-time1))
volume_batch, label_batch = sampled_batch['image'], sampled_batch['label']
volume_batch, label_batch = volume_batch.cuda(), label_batch.cuda()
outputs = net(volume_batch)
loss_seg = F.cross_entropy(outputs, label_batch)
outputs_soft = F.softmax(outputs, dim=1)
loss_seg_dice = dice_loss(outputs_soft[:, 1, :, :, :], label_batch == 1)
loss = 0.5*(loss_seg+loss_seg_dice)
optimizer.zero_grad()
loss.backward()
optimizer.step()
iter_num = iter_num + 1
writer.add_scalar('lr', lr_, iter_num)
writer.add_scalar('loss/loss_seg', loss_seg, iter_num)
writer.add_scalar('loss/loss_seg_dice', loss_seg_dice, iter_num)
writer.add_scalar('loss/loss', loss, iter_num)
logging.info('iteration %d : loss : %f' % (iter_num, loss.item()))
if iter_num % 50 == 0:
image = volume_batch[0, 0:1, :, :, 20:61:10].permute(3,0,1,2).repeat(1,3,1,1)
grid_image = make_grid(image, 5, normalize=True)
writer.add_image('train/Image', grid_image, iter_num)
outputs_soft = F.softmax(outputs, 1)
image = outputs_soft[0, 1:2, :, :, 20:61:10].permute(3, 0, 1, 2).repeat(1, 3, 1, 1)
grid_image = make_grid(image, 5, normalize=False)
writer.add_image('train/Predicted_label', grid_image, iter_num)
image = label_batch[0, :, :, 20:61:10].unsqueeze(0).permute(3, 0, 1, 2).repeat(1, 3, 1, 1)
grid_image = make_grid(image, 5, normalize=False)
writer.add_image('train/Groundtruth_label', grid_image, iter_num)
## change lr
if iter_num % 2500 == 0:
lr_ = base_lr * 0.1 ** (iter_num // 2500)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_
if iter_num % 1000 == 0:
save_mode_path = os.path.join(snapshot_path, 'iter_' + str(iter_num) + '.pth')
torch.save(net.state_dict(), save_mode_path)
logging.info("save model to {}".format(save_mode_path))
if iter_num > max_iterations:
break
time1 = time.time()
if iter_num > max_iterations:
break
save_mode_path = os.path.join(snapshot_path, 'iter_'+str(max_iterations+1)+'.pth')
torch.save(net.state_dict(), save_mode_path)
logging.info("save model to {}".format(save_mode_path))
writer.close()
def debugger():
patch_size = (112, 112, 80)
training_data = data_loader(split='train')
testing_data = data_loader(split='test')
x_criterion = soft_cross_entropy #supervised loss is 0.5*(x_criterion + dice_loss)
u_criterion = nn.MSELoss() #unsupervised loss
labelled_index = np.random.permutation(LABELLED_INDEX)
unlabelled_index = np.random.permutation(
UNLABELLED_INDEX)[:len(labelled_index)]
labelled_data = [training_data[i] for i in labelled_index]
unlabelled_data = [training_data[i] for i in unlabelled_index] #size = 16
##data transformation: rotation, flip, random_crop
labelled_data = [
shape_transform(RandomRotFlip()(RandomCrop(patch_size)(sample)))
for sample in labelled_data
]
unlabelled_data = [
shape_transform(RandomRotFlip()(RandomCrop(patch_size)(sample)))
for sample in unlabelled_data
]
net = VNet(n_channels=1,
n_classes=2,
normalization='batchnorm',
has_dropout=True).cuda()
model_path = "../saved/0_supervised.pth"
net.load_state_dict(torch.load(model_path))
optimizer = optim.SGD(net.parameters(),
lr=0.01,
momentum=0.9,
weight_decay=0.0001)
training_loss = train_epoch(net=net,
labelled_data=labelled_data,
unlabelled_data=unlabelled_data,
batch_size=2,
supervised_only=True,
optimizer=optimizer,
x_criterion=x_criterion,
u_criterion=u_criterion,
K=1,
T=1,
alpha=1,
mixup_mode="__",
Lambda=0,
aug_factor=0)
net = VNet(n_channels=1,
n_classes=2,
normalization='batchnorm',
has_dropout=True).cuda()
model_path = "../saved/8_expected_supervised.pth"
net.load_state_dict(torch.load(model_path))
optimizer = optim.SGD(net.parameters(),
lr=0.01,
momentum=0.9,
weight_decay=0.0001)
training_loss = train_epoch(net=net,
labelled_data=labelled_data,
unlabelled_data=unlabelled_data,
batch_size=2,
supervised_only=False,
optimizer=optimizer,
x_criterion=x_criterion,
u_criterion=u_criterion,
K=1,
T=1,
alpha=1,
mixup_mode="__",
Lambda=0,
aug_factor=0)
def main():
###################
# init parameters #
###################
args = get_args()
# training path
train_data_path = args.root_path
# writer
idx = args.save.rfind('/')
log_dir = args.writer_dir + args.save[idx:]
writer = SummaryWriter(log_dir)
batch_size = args.batch_size * args.ngpu
max_iterations = args.max_iterations
base_lr = args.base_lr
patch_size = (112, 112, 80)
num_classes = 2
# random
if args.deterministic:
cudnn.benchmark = False
cudnn.deterministic = True
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
## make logger file
if os.path.exists(args.save):
shutil.rmtree(args.save)
os.makedirs(args.save, exist_ok=True)
snapshot_path = args.save
logging.basicConfig(filename=snapshot_path + "/log.txt",
level=logging.INFO,
format='[%(asctime)s.%(msecs)03d] %(message)s',
datefmt='%H:%M:%S')
logging.getLogger().addHandler(logging.StreamHandler(sys.stdout))
logging.info(str(args))
#training set
db_train = LAHeart(base_dir=train_data_path,
split='train',
num=16,
transform=transforms.Compose([
RandomRotFlip(),
RandomCrop(patch_size),
ToTensor()
]))
def worker_init_fn(worker_id):
random.seed(args.seed + worker_id)
trainloader = DataLoader(db_train,
batch_size=batch_size,
shuffle=True,
num_workers=4,
pin_memory=True,
worker_init_fn=worker_init_fn)
net = VNet(n_channels=1,
n_classes=num_classes,
normalization='batchnorm',
has_dropout=True)
net = net.cuda()
net.train()
optimizer = optim.SGD(net.parameters(),
lr=base_lr,
momentum=0.9,
weight_decay=0.0001)
logging.info("{} itertations per epoch".format(len(trainloader)))
iter_num = 0
alpha = 1.0
max_epoch = max_iterations // len(trainloader) + 1
lr_ = base_lr
net.train()
for epoch_num in tqdm(range(max_epoch), ncols=70):
time1 = time.time()
for i_batch, sampled_batch in enumerate(trainloader):
time2 = time.time()
# print('fetch data cost {}'.format(time2-time1))
# volume_batch.shape=(b,1,x,y,z) label_patch.shape=(b,x,y,z)
volume_batch, label_batch = sampled_batch['image'], sampled_batch[
'label']
volume_batch, label_batch = volume_batch.cuda(), label_batch.cuda()
outputs = net(volume_batch)
loss_seg = F.cross_entropy(outputs, label_batch)
outputs_soft = F.softmax(outputs, dim=1)
loss_seg_dice = dice_loss(outputs_soft[:, 1, :, :, :],
label_batch == 1)
# compute gt_signed distance function and boundary loss
with torch.no_grad():
# defalut using compute_sdf; however, compute_sdf1_1 is also worth to try;
gt_sdf_npy = compute_sdf(label_batch.cpu().numpy(),
outputs_soft.shape)
gt_sdf = torch.from_numpy(gt_sdf_npy).float().cuda(
outputs_soft.device.index)
# show signed distance map for debug
# import matplotlib.pyplot as plt
# plt.figure()
# plt.subplot(121), plt.imshow(gt_sdf_npy[0,1,:,:,40]), plt.colorbar()
# plt.subplot(122), plt.imshow(np.uint8(label_batch.cpu().numpy()[0,:,:,40]>0)), plt.colorbar()
# plt.show()
loss_boundary = boundary_loss(outputs_soft, gt_sdf)
loss = alpha * (loss_seg + loss_seg_dice) + (1 -
alpha) * loss_boundary
optimizer.zero_grad()
loss.backward()
optimizer.step()
iter_num = iter_num + 1
writer.add_scalar('lr', lr_, iter_num)
writer.add_scalar('loss/loss_seg', loss_seg, iter_num)
writer.add_scalar('loss/loss_seg_dice', loss_seg_dice, iter_num)
writer.add_scalar('loss/loss_boundary', loss_boundary, iter_num)
writer.add_scalar('loss/loss', loss, iter_num)
writer.add_scalar('loss/alpha', alpha, iter_num)
logging.info('iteration %d : alpha : %f' % (iter_num, alpha))
logging.info('iteration %d : loss_seg_dice : %f' %
(iter_num, loss_seg_dice.item()))
logging.info('iteration %d : loss_boundary : %f' %
(iter_num, loss_boundary.item()))
logging.info('iteration %d : loss : %f' % (iter_num, loss.item()))
if iter_num % 2 == 0:
image = volume_batch[0, 0:1, :, :,
20:61:10].permute(3, 0, 1,
2).repeat(1, 3, 1, 1)
grid_image = make_grid(image, 5, normalize=True)
writer.add_image('train/Image', grid_image, iter_num)
image = outputs_soft[0, 1:2, :, :,
20:61:10].permute(3, 0, 1,
2).repeat(1, 3, 1, 1)
grid_image = make_grid(image, 5, normalize=False)
writer.add_image('train/Predicted_label', grid_image, iter_num)
image = label_batch[0, :, :, 20:61:10].unsqueeze(0).permute(
3, 0, 1, 2).repeat(1, 3, 1, 1)
grid_image = make_grid(image, 5, normalize=False)
writer.add_image('train/Groundtruth_label', grid_image,
iter_num)
image = gt_sdf[0, 1:2, :, :,
20:61:10].permute(3, 0, 1,
2).repeat(1, 3, 1, 1)
grid_image = make_grid(image, 5, normalize=True)
writer.add_image('train/gt_sdf', grid_image, iter_num)
## change lr
if iter_num % 2500 == 0:
lr_ = base_lr * 0.1**(iter_num // 2500)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_
if iter_num % 1000 == 0:
save_mode_path = os.path.join(snapshot_path,
'iter_' + str(iter_num) + '.pth')
torch.save(net.state_dict(), save_mode_path)
logging.info("save model to {}".format(save_mode_path))
if iter_num > max_iterations:
break
time1 = time.time()
alpha -= 0.01
if alpha <= 0.01:
alpha = 0.01
if iter_num > max_iterations:
break
save_mode_path = os.path.join(snapshot_path,
'iter_' + str(max_iterations + 1) + '.pth')
torch.save(net.state_dict(), save_mode_path)
logging.info("save model to {}".format(save_mode_path))
writer.close()
def experiment(exp_identifier,
max_epoch,
training_data,
testing_data,
batch_size=2,
supervised_only=False,
K=2,
T=0.5,
alpha=1,
mixup_mode='all',
Lambda=1,
Lambda_ramp=None,
base_lr=0.01,
change_lr=None,
aug_factor=1,
from_saved=None,
always_do_validation=True,
decay=0):
'''
max_epoch: epochs to run. Going through labeled data once is one epoch.
batch_size: batch size of labeled data. Unlabeled data is of the same size.
training_data: data for train_epoch, list of dicts of numpy array.
training_data: data for validation, list of dicts of numpy array.
supervised_only: if True, only do supervised training on LABELLED_INDEX; otherwise, use both LABELLED_INDEX and UNLABELLED_INDEX
Hyperparameters
---------------
K: repeats of each unlabelled data
T: temperature of sharpening
alpha: mixup hyperparameter of beta distribution
mixup_mode: how mixup is performed --
'__': no mix up
'ww': x and u both mixed up with w(x+u)
'xx': both with x
'xu': x with x, u with u
'uu': both with u
... _ means no, x means with x, u means with u, w means with w(x+u)
Lambda: loss = loss_x + Lambda * loss_u, relative weight for unsupervised loss
base_lr: initial learning rate
Lambda_ramp: callable or None. Lambda is ignored if this is not None. In this case, Lambda = Lambda_ramp(epoch).
change_lr: dict, {epoch: change_multiplier}
'''
print(
f"Experiment {exp_identifier}: max_epoch = {max_epoch}, batch_size = {batch_size}, supervised_only = {supervised_only},"
f"K = {K}, T = {T}, alpha = {alpha}, mixup_mode = {mixup_mode}, Lambda = {Lambda}, Lambda_ramp = {Lambda_ramp}, base_lr = {base_lr}, aug_factor = {aug_factor}."
)
net = VNet(n_channels=1,
n_classes=2,
normalization='batchnorm',
has_dropout=True)
eval_net = VNet(n_channels=1,
n_classes=2,
normalization='batchnorm',
has_dropout=True)
if from_saved is not None:
net.load_state_dict(torch.load(from_saved))
if GPU:
net = net.cuda()
eval_net.cuda()
## eval_net is not updating
for param in eval_net.parameters():
param.detach_()
net.train()
eval_net.train()
optimizer = optim.SGD(net.parameters(),
lr=base_lr,
momentum=0.9,
weight_decay=0.0001)
x_criterion = soft_cross_entropy #supervised loss is 0.5*(x_criterion + dice_loss)
u_criterion = nn.MSELoss() #unsupervised loss
training_losses = []
testing_losses = []
testing_accuracy = [] #dice accuracy
patch_size = (112, 112, 80)
testing_data = [
shape_transform(CenterCrop(patch_size)(sample))
for sample in testing_data
]
t0 = time.time()
lr = base_lr
for epoch in range(max_epoch):
labelled_index = np.random.permutation(LABELLED_INDEX)
unlabelled_index = np.random.permutation(
UNLABELLED_INDEX)[:len(labelled_index)]
labelled_data = [training_data[i] for i in labelled_index]
unlabelled_data = [training_data[i]
for i in unlabelled_index] #size = 16
##data transformation: rotation, flip, random_crop
labelled_data = [
shape_transform(RandomRotFlip()(RandomCrop(patch_size)(sample)))
for sample in labelled_data
]
unlabelled_data = [
shape_transform(RandomRotFlip()(RandomCrop(patch_size)(sample)))
for sample in unlabelled_data
]
if Lambda_ramp is not None:
Lambda = Lambda_ramp(epoch)
print(f"Lambda ramp: Lambda = {Lambda}")
if change_lr is not None:
if epoch in change_lr:
lr_ = lr * change_lr[epoch]
print(
f"Learning rate decay at epoch {epoch}, from {lr} to {lr_}"
)
lr = lr_
#change learning rate.
for param_group in optimizer.param_groups:
param_group['lr'] = lr_
training_loss = train_epoch(net=net,
eval_net=eval_net,
labelled_data=labelled_data,
unlabelled_data=unlabelled_data,
batch_size=batch_size,
supervised_only=supervised_only,
optimizer=optimizer,
x_criterion=x_criterion,
u_criterion=u_criterion,
K=K,
T=T,
alpha=alpha,
mixup_mode=mixup_mode,
Lambda=Lambda,
aug_factor=aug_factor,
decay=decay)
training_losses.append(training_loss)
if always_do_validation or epoch % 50 == 0:
testing_dice_loss, accuracy = validation(net=net,
testing_data=testing_data,
x_criterion=x_criterion)
testing_losses.append(testing_dice_loss)
testing_accuracy.append(accuracy)
print(
f"Epoch {epoch+1}/{max_epoch}, time used: {time.time()-t0:.2f}, training loss: {training_loss:.6f}, testing dice_loss: {testing_dice_loss:.6f}, testing accuracy: {100.0*accuracy:.2f}% "
)
save_path = f"../saved/{exp_identifier}.pth"
torch.save(net.state_dict(), save_path)
print(f"Experiment {exp_identifier} finished. Model saved as {save_path}")
return training_losses, testing_losses, testing_accuracy
