def train(args, snapshot_path):
base_lr = args.base_lr
num_classes = args.num_classes
batch_size = args.batch_size
max_iterations = args.max_iterations
def worker_init_fn(worker_id):
random.seed(args.seed + worker_id)
model = net_factory(net_type=args.model, in_chns=1, class_num=num_classes)
db_train = BaseDataSets(base_dir=args.root_path,
split="train",
num=None,
transform=transforms.Compose(
[RandomGenerator(args.patch_size)]))
total_slices = len(db_train)
labeled_slice = patients_to_slices(args.root_path, args.labeled_num)
print("Total silices is: {}, labeled slices is: {}".format(
total_slices, labeled_slice))
labeled_idxs = list(range(0, labeled_slice))
unlabeled_idxs = list(range(labeled_slice, total_slices))
batch_sampler = TwoStreamBatchSampler(labeled_idxs, unlabeled_idxs,
batch_size,
batch_size - args.labeled_bs)
trainloader = DataLoader(db_train,
batch_sampler=batch_sampler,
num_workers=16,
pin_memory=True,
worker_init_fn=worker_init_fn)
db_val = BaseDataSets(base_dir=args.root_path, split="val")
valloader = DataLoader(db_val, batch_size=1, shuffle=False, num_workers=1)
model.train()
optimizer = optim.SGD(model.parameters(),
lr=base_lr,
momentum=0.9,
weight_decay=0.0001)
ce_loss = CrossEntropyLoss()
dice_loss = losses.DiceLoss(num_classes)
writer = SummaryWriter(snapshot_path + '/log')
logging.info("{} iterations per epoch".format(len(trainloader)))
iter_num = 0
max_epoch = max_iterations // len(trainloader) + 1
best_performance = 0.0
iterator = tqdm(range(max_epoch), ncols=70)
for epoch_num in iterator:
for i_batch, sampled_batch in enumerate(trainloader):
volume_batch, label_batch = sampled_batch['image'], sampled_batch[
'label']
volume_batch, label_batch = volume_batch.cuda(), label_batch.cuda()
unlabeled_volume_batch = volume_batch[args.labeled_bs:]
outputs = model(volume_batch)
outputs_soft = torch.softmax(outputs, dim=1)
loss_ce = ce_loss(outputs[:args.labeled_bs],
label_batch[:][:args.labeled_bs].long())
loss_dice = dice_loss(outputs_soft[:args.labeled_bs],
label_batch[:args.labeled_bs].unsqueeze(1))
supervised_loss = 0.5 * (loss_dice + loss_ce)
consistency_weight = get_current_consistency_weight(iter_num //
150)
consistency_loss = losses.entropy_loss(outputs_soft, C=4)
loss = supervised_loss + consistency_weight * consistency_loss
optimizer.zero_grad()
loss.backward()
optimizer.step()
lr_ = base_lr * (1.0 - iter_num / max_iterations)**0.9
for param_group in optimizer.param_groups:
param_group['lr'] = lr_
iter_num = iter_num + 1
writer.add_scalar('info/lr', lr_, iter_num)
writer.add_scalar('info/total_loss', loss, iter_num)
writer.add_scalar('info/loss_ce', loss_ce, iter_num)
writer.add_scalar('info/loss_dice', loss_dice, iter_num)
writer.add_scalar('info/consistency_loss', consistency_loss,
iter_num)
writer.add_scalar('info/consistency_weight', consistency_weight,
iter_num)
logging.info(
'iteration %d : loss : %f, loss_ce: %f, loss_dice: %f' %
(iter_num, loss.item(), loss_ce.item(), loss_dice.item()))
if iter_num % 20 == 0:
image = volume_batch[1, 0:1, :, :]
writer.add_image('train/Image', image, iter_num)
outputs = torch.argmax(torch.softmax(outputs, dim=1),
dim=1,
keepdim=True)
writer.add_image('train/Prediction', outputs[1, ...] * 50,
iter_num)
labs = label_batch[1, ...].unsqueeze(0) * 50
writer.add_image('train/GroundTruth', labs, iter_num)
if iter_num > 0 and iter_num % 200 == 0:
model.eval()
metric_list = 0.0
for i_batch, sampled_batch in enumerate(valloader):
metric_i = test_single_volume(sampled_batch["image"],
sampled_batch["label"],
model,
classes=num_classes)
metric_list += np.array(metric_i)
metric_list = metric_list / len(db_val)
for class_i in range(num_classes - 1):
writer.add_scalar('info/val_{}_dice'.format(class_i + 1),
metric_list[class_i, 0], iter_num)
writer.add_scalar('info/val_{}_hd95'.format(class_i + 1),
metric_list[class_i, 1], iter_num)
performance = np.mean(metric_list, axis=0)[0]
mean_hd95 = np.mean(metric_list, axis=0)[1]
writer.add_scalar('info/val_mean_dice', performance, iter_num)
writer.add_scalar('info/val_mean_hd95', mean_hd95, iter_num)
if performance > best_performance:
best_performance = performance
save_mode_path = os.path.join(
snapshot_path, 'iter_{}_dice_{}.pth'.format(
iter_num, round(best_performance, 4)))
save_best = os.path.join(
snapshot_path, '{}_best_model.pth'.format(args.model))
torch.save(model.state_dict(), save_mode_path)
torch.save(model.state_dict(), save_best)
logging.info('iteration %d : mean_dice : %f mean_hd95 : %f' %
(iter_num, performance, mean_hd95))
model.train()
if iter_num % 3000 == 0:
save_mode_path = os.path.join(snapshot_path,
'iter_' + str(iter_num) + '.pth')
torch.save(model.state_dict(), save_mode_path)
logging.info("save model to {}".format(save_mode_path))
if iter_num >= max_iterations:
break
if iter_num >= max_iterations:
iterator.close()
break
writer.close()
return "Training Finished!"
transform=transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
normalize,
]))
test_dataset = dataset.CheXpertDataset(root_dir=args.root_path,
csv_file=args.csv_file_test,
transform=transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
normalize,
]))
labeled_idxs = list(range(args.labeled_num))
unlabeled_idxs = list(range(args.labeled_num, 7000))
batch_sampler = TwoStreamBatchSampler(labeled_idxs, unlabeled_idxs,
batch_size, batch_size - labeled_bs)
def worker_init_fn(worker_id):
random.seed(args.seed + worker_id)
train_dataloader = DataLoader(dataset=train_dataset,
batch_sampler=batch_sampler,
num_workers=8,
pin_memory=True,
worker_init_fn=worker_init_fn)
val_dataloader = DataLoader(dataset=val_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=8,
pin_memory=True,
worker_init_fn=worker_init_fn)
def train(args, snapshot_path):
base_lr = args.base_lr
num_classes = args.num_classes
batch_size = args.batch_size
max_iterations = args.max_iterations
def create_model(ema=False):
# Network definition
model = net_factory(net_type=args.model,
in_chns=1,
class_num=num_classes)
if ema:
for param in model.parameters():
param.detach_()
return model
model = create_model()
ema_model = create_model(ema=True)
def worker_init_fn(worker_id):
random.seed(args.seed + worker_id)
db_train = BaseDataSets(base_dir=args.root_path,
split="train",
num=None,
transform=transforms.Compose(
[RandomGenerator(args.patch_size)]))
db_val = BaseDataSets(base_dir=args.root_path, split="val")
total_slices = len(db_train)
labeled_slice = patients_to_slices(args.root_path, args.labeled_num)
print("Total silices is: {}, labeled slices is: {}".format(
total_slices, labeled_slice))
labeled_idxs = list(range(0, labeled_slice))
unlabeled_idxs = list(range(labeled_slice, total_slices))
batch_sampler = TwoStreamBatchSampler(labeled_idxs, unlabeled_idxs,
batch_size,
batch_size - args.labeled_bs)
trainloader = DataLoader(db_train,
batch_sampler=batch_sampler,
num_workers=4,
pin_memory=True,
worker_init_fn=worker_init_fn)
model.train()
valloader = DataLoader(db_val, batch_size=1, shuffle=False, num_workers=1)
optimizer = optim.SGD(model.parameters(),
lr=base_lr,
momentum=0.9,
weight_decay=0.0001)
ce_loss = CrossEntropyLoss()
dice_loss = losses.DiceLoss(num_classes)
writer = SummaryWriter(snapshot_path + '/log')
logging.info("{} iterations per epoch".format(len(trainloader)))
iter_num = 0
max_epoch = max_iterations // len(trainloader) + 1
best_performance = 0.0
iterator = tqdm(range(max_epoch), ncols=70)
for epoch_num in iterator:
for i_batch, sampled_batch in enumerate(trainloader):
volume_batch, label_batch = sampled_batch['image'], sampled_batch[
'label']
volume_batch, label_batch = volume_batch.cuda(), label_batch.cuda()
unlabeled_volume_batch = volume_batch[args.labeled_bs:]
noise = torch.clamp(
torch.randn_like(unlabeled_volume_batch) * 0.1, -0.2, 0.2)
ema_inputs = unlabeled_volume_batch + noise
outputs = model(volume_batch)
outputs_soft = torch.softmax(outputs, dim=1)
with torch.no_grad():
ema_output = ema_model(ema_inputs)
T = 8
_, _, w, h = unlabeled_volume_batch.shape
volume_batch_r = unlabeled_volume_batch.repeat(2, 1, 1, 1)
stride = volume_batch_r.shape[0] // 2
preds = torch.zeros([stride * T, num_classes, w, h]).cuda()
for i in range(T // 2):
ema_inputs = volume_batch_r + \
torch.clamp(torch.randn_like(
volume_batch_r) * 0.1, -0.2, 0.2)
with torch.no_grad():
preds[2 * stride * i:2 * stride *
(i + 1)] = ema_model(ema_inputs)
preds = F.softmax(preds, dim=1)
preds = preds.reshape(T, stride, num_classes, w, h)
preds = torch.mean(preds, dim=0)
uncertainty = -1.0 * \
torch.sum(preds*torch.log(preds + 1e-6), dim=1, keepdim=True)
loss_ce = ce_loss(outputs[:args.labeled_bs],
label_batch[:args.labeled_bs][:].long())
loss_dice = dice_loss(outputs_soft[:args.labeled_bs],
label_batch[:args.labeled_bs].unsqueeze(1))
supervised_loss = 0.5 * (loss_dice + loss_ce)
consistency_weight = get_current_consistency_weight(iter_num //
150)
consistency_dist = losses.softmax_mse_loss(
outputs[args.labeled_bs:],
ema_output) # (batch, 2, 112,112,80)
threshold = (0.75 + 0.25 * ramps.sigmoid_rampup(
iter_num, max_iterations)) * np.log(2)
mask = (uncertainty < threshold).float()
consistency_loss = torch.sum(
mask * consistency_dist) / (2 * torch.sum(mask) + 1e-16)
loss = supervised_loss + consistency_weight * consistency_loss
optimizer.zero_grad()
loss.backward()
optimizer.step()
update_ema_variables(model, ema_model, args.ema_decay, iter_num)
lr_ = base_lr * (1.0 - iter_num / max_iterations)**0.9
for param_group in optimizer.param_groups:
param_group['lr'] = lr_
iter_num = iter_num + 1
writer.add_scalar('info/lr', lr_, iter_num)
writer.add_scalar('info/total_loss', loss, iter_num)
writer.add_scalar('info/loss_ce', loss_ce, iter_num)
writer.add_scalar('info/loss_dice', loss_dice, iter_num)
writer.add_scalar('info/consistency_loss', consistency_loss,
iter_num)
writer.add_scalar('info/consistency_weight', consistency_weight,
iter_num)
logging.info(
'iteration %d : loss : %f, loss_ce: %f, loss_dice: %f' %
(iter_num, loss.item(), loss_ce.item(), loss_dice.item()))
if iter_num % 20 == 0:
image = volume_batch[1, 0:1, :, :]
writer.add_image('train/Image', image, iter_num)
outputs = torch.argmax(torch.softmax(outputs, dim=1),
dim=1,
keepdim=True)
writer.add_image('train/Prediction', outputs[1, ...] * 50,
iter_num)
labs = label_batch[1, ...].unsqueeze(0) * 50
writer.add_image('train/GroundTruth', labs, iter_num)
if iter_num > 0 and iter_num % 200 == 0:
model.eval()
metric_list = 0.0
for i_batch, sampled_batch in enumerate(valloader):
metric_i = test_single_volume(sampled_batch["image"],
sampled_batch["label"],
model,
classes=num_classes)
metric_list += np.array(metric_i)
metric_list = metric_list / len(db_val)
for class_i in range(num_classes - 1):
writer.add_scalar('info/val_{}_dice'.format(class_i + 1),
metric_list[class_i, 0], iter_num)
writer.add_scalar('info/val_{}_hd95'.format(class_i + 1),
metric_list[class_i, 1], iter_num)
performance = np.mean(metric_list, axis=0)[0]
mean_hd95 = np.mean(metric_list, axis=0)[1]
writer.add_scalar('info/val_mean_dice', performance, iter_num)
writer.add_scalar('info/val_mean_hd95', mean_hd95, iter_num)
if performance > best_performance:
best_performance = performance
save_mode_path = os.path.join(
snapshot_path, 'iter_{}_dice_{}.pth'.format(
iter_num, round(best_performance, 4)))
save_best = os.path.join(
snapshot_path, '{}_best_model.pth'.format(args.model))
torch.save(model.state_dict(), save_mode_path)
torch.save(model.state_dict(), save_best)
logging.info('iteration %d : mean_dice : %f mean_hd95 : %f' %
(iter_num, performance, mean_hd95))
model.train()
if iter_num % 3000 == 0:
save_mode_path = os.path.join(snapshot_path,
'iter_' + str(iter_num) + '.pth')
torch.save(model.state_dict(), save_mode_path)
logging.info("save model to {}".format(save_mode_path))
if iter_num >= max_iterations:
break
if iter_num >= max_iterations:
iterator.close()
break
writer.close()
return "Training Finished!"
def train(args, snapshot_path):
base_lr = args.base_lr
num_classes = args.num_classes
batch_size = args.batch_size
max_iterations = args.max_iterations
def create_model(ema=False):
# Network definition
model = net_factory(net_type=args.model,
in_chns=1,
class_num=num_classes)
if ema:
for param in model.parameters():
param.detach_()
return model
model1 = kaiming_normal_init_weight(create_model())
model2 = xavier_normal_init_weight(create_model())
def worker_init_fn(worker_id):
random.seed(args.seed + worker_id)
db_train = BaseDataSets(base_dir=args.root_path,
split="train",
num=None,
transform=transforms.Compose(
[RandomGenerator(args.patch_size)]))
db_val = BaseDataSets(base_dir=args.root_path, split="val")
total_slices = len(db_train)
labeled_slice = patients_to_slices(args.root_path, args.labeled_num)
print("Total silices is: {}, labeled slices is: {}".format(
total_slices, labeled_slice))
labeled_idxs = list(range(0, labeled_slice))
unlabeled_idxs = list(range(labeled_slice, total_slices))
batch_sampler = TwoStreamBatchSampler(labeled_idxs, unlabeled_idxs,
batch_size,
batch_size - args.labeled_bs)
trainloader = DataLoader(db_train,
batch_sampler=batch_sampler,
num_workers=4,
pin_memory=True,
worker_init_fn=worker_init_fn)
model1.train()
model2.train()
valloader = DataLoader(db_val, batch_size=1, shuffle=False, num_workers=1)
optimizer1 = optim.SGD(model1.parameters(),
lr=base_lr,
momentum=0.9,
weight_decay=0.0001)
optimizer2 = optim.SGD(model2.parameters(),
lr=base_lr,
momentum=0.9,
weight_decay=0.0001)
ce_loss = CrossEntropyLoss()
dice_loss = losses.DiceLoss(num_classes)
writer = SummaryWriter(snapshot_path + '/log')
logging.info("{} iterations per epoch".format(len(trainloader)))
iter_num = 0
max_epoch = max_iterations // len(trainloader) + 1
best_performance1 = 0.0
best_performance2 = 0.0
iterator = tqdm(range(max_epoch), ncols=70)
for epoch_num in iterator:
for i_batch, sampled_batch in enumerate(trainloader):
volume_batch, label_batch = sampled_batch['image'], sampled_batch[
'label']
volume_batch, label_batch = volume_batch.cuda(), label_batch.cuda()
outputs1 = model1(volume_batch)
outputs_soft1 = torch.softmax(outputs1, dim=1)
outputs2 = model2(volume_batch)
outputs_soft2 = torch.softmax(outputs2, dim=1)
consistency_weight = get_current_consistency_weight(iter_num //
150)
loss1 = 0.5 * (
ce_loss(outputs1[:args.labeled_bs],
label_batch[:][:args.labeled_bs].long()) +
dice_loss(outputs_soft1[:args.labeled_bs],
label_batch[:args.labeled_bs].unsqueeze(1)))
loss2 = 0.5 * (
ce_loss(outputs2[:args.labeled_bs],
label_batch[:][:args.labeled_bs].long()) +
dice_loss(outputs_soft2[:args.labeled_bs],
label_batch[:args.labeled_bs].unsqueeze(1)))
pseudo_outputs1 = torch.argmax(
outputs_soft1[args.labeled_bs:].detach(), dim=1, keepdim=False)
pseudo_outputs2 = torch.argmax(
outputs_soft2[args.labeled_bs:].detach(), dim=1, keepdim=False)
pseudo_supervision1 = ce_loss(outputs1[args.labeled_bs:],
pseudo_outputs2)
pseudo_supervision2 = ce_loss(outputs2[args.labeled_bs:],
pseudo_outputs1)
model1_loss = loss1 + consistency_weight * pseudo_supervision1
model2_loss = loss2 + consistency_weight * pseudo_supervision2
loss = model1_loss + model2_loss
optimizer1.zero_grad()
optimizer2.zero_grad()
loss.backward()
optimizer1.step()
optimizer2.step()
iter_num = iter_num + 1
lr_ = base_lr * (1.0 - iter_num / max_iterations)**0.9
for param_group in optimizer1.param_groups:
param_group['lr'] = lr_
for param_group in optimizer2.param_groups:
param_group['lr'] = lr_
writer.add_scalar('lr', lr_, iter_num)
writer.add_scalar('consistency_weight/consistency_weight',
consistency_weight, iter_num)
writer.add_scalar('loss/model1_loss', model1_loss, iter_num)
writer.add_scalar('loss/model2_loss', model2_loss, iter_num)
logging.info('iteration %d : model1 loss : %f model2 loss : %f' %
(iter_num, model1_loss.item(), model2_loss.item()))
if iter_num % 50 == 0:
image = volume_batch[1, 0:1, :, :]
writer.add_image('train/Image', image, iter_num)
outputs = torch.argmax(torch.softmax(outputs1, dim=1),
dim=1,
keepdim=True)
writer.add_image('train/model1_Prediction',
outputs[1, ...] * 50, iter_num)
outputs = torch.argmax(torch.softmax(outputs2, dim=1),
dim=1,
keepdim=True)
writer.add_image('train/model2_Prediction',
outputs[1, ...] * 50, iter_num)
labs = label_batch[1, ...].unsqueeze(0) * 50
writer.add_image('train/GroundTruth', labs, iter_num)
if iter_num > 0 and iter_num % 200 == 0:
model1.eval()
metric_list = 0.0
for i_batch, sampled_batch in enumerate(valloader):
metric_i = test_single_volume(sampled_batch["image"],
sampled_batch["label"],
model1,
classes=num_classes)
metric_list += np.array(metric_i)
metric_list = metric_list / len(db_val)
for class_i in range(num_classes - 1):
writer.add_scalar(
'info/model1_val_{}_dice'.format(class_i + 1),
metric_list[class_i, 0], iter_num)
writer.add_scalar(
'info/model1_val_{}_hd95'.format(class_i + 1),
metric_list[class_i, 1], iter_num)
performance1 = np.mean(metric_list, axis=0)[0]
mean_hd951 = np.mean(metric_list, axis=0)[1]
writer.add_scalar('info/model1_val_mean_dice', performance1,
iter_num)
writer.add_scalar('info/model1_val_mean_hd95', mean_hd951,
iter_num)
if performance1 > best_performance1:
best_performance1 = performance1
save_mode_path = os.path.join(
snapshot_path, 'model1_iter_{}_dice_{}.pth'.format(
iter_num, round(best_performance1, 4)))
save_best = os.path.join(
snapshot_path, '{}_best_model1.pth'.format(args.model))
torch.save(model1.state_dict(), save_mode_path)
torch.save(model1.state_dict(), save_best)
logging.info(
'iteration %d : model1_mean_dice : %f model1_mean_hd95 : %f'
% (iter_num, performance1, mean_hd951))
model1.train()
model2.eval()
metric_list = 0.0
for i_batch, sampled_batch in enumerate(valloader):
metric_i = test_single_volume(sampled_batch["image"],
sampled_batch["label"],
model2,
classes=num_classes)
metric_list += np.array(metric_i)
metric_list = metric_list / len(db_val)
for class_i in range(num_classes - 1):
writer.add_scalar(
'info/model2_val_{}_dice'.format(class_i + 1),
metric_list[class_i, 0], iter_num)
writer.add_scalar(
'info/model2_val_{}_hd95'.format(class_i + 1),
metric_list[class_i, 1], iter_num)
performance2 = np.mean(metric_list, axis=0)[0]
mean_hd952 = np.mean(metric_list, axis=0)[1]
writer.add_scalar('info/model2_val_mean_dice', performance2,
iter_num)
writer.add_scalar('info/model2_val_mean_hd95', mean_hd952,
iter_num)
if performance2 > best_performance2:
best_performance2 = performance2
save_mode_path = os.path.join(
snapshot_path, 'model2_iter_{}_dice_{}.pth'.format(
iter_num, round(best_performance2)))
save_best = os.path.join(
snapshot_path, '{}_best_model2.pth'.format(args.model))
torch.save(model2.state_dict(), save_mode_path)
torch.save(model2.state_dict(), save_best)
logging.info(
'iteration %d : model2_mean_dice : %f model2_mean_hd95 : %f'
% (iter_num, performance2, mean_hd952))
model2.train()
# change lr
if iter_num % 2500 == 0:
lr_ = base_lr * 0.1**(iter_num // 2500)
for param_group in optimizer1.param_groups:
param_group['lr'] = lr_
for param_group in optimizer2.param_groups:
param_group['lr'] = lr_
if iter_num % 3000 == 0:
save_mode_path = os.path.join(
snapshot_path, 'model1_iter_' + str(iter_num) + '.pth')
torch.save(model1.state_dict(), save_mode_path)
logging.info("save model1 to {}".format(save_mode_path))
save_mode_path = os.path.join(
snapshot_path, 'model2_iter_' + str(iter_num) + '.pth')
torch.save(model2.state_dict(), save_mode_path)
logging.info("save model2 to {}".format(save_mode_path))
if iter_num >= max_iterations:
break
time1 = time.time()
if iter_num >= max_iterations:
iterator.close()
break
writer.close()