def train(args, snapshot_path):
base_lr = args.base_lr
train_data_path = args.root_path
batch_size = args.batch_size
max_iterations = args.max_iterations
def create_model(ema=False):
# Network definition
net = unet_3D(n_classes=2, in_channels=1)
model = net.cuda()
if ema:
for param in model.parameters():
param.detach_()
return model
model = create_model()
ema_model = create_model(ema=True)
db_train = BraTS2019(base_dir=train_data_path,
split='train',
num=None,
transform=transforms.Compose([
RandomRotFlip(),
RandomCrop(args.patch_size),
ToTensor(),
]))
def worker_init_fn(worker_id):
random.seed(args.seed + worker_id)
labeled_idxs = list(range(0, args.labeled_num))
unlabeled_idxs = list(range(args.labeled_num, 250))
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()
ema_model.train()
optimizer = optim.SGD(model.parameters(),
lr=base_lr,
momentum=0.9,
weight_decay=0.0001)
ce_loss = CrossEntropyLoss()
dice_loss = losses.DiceLoss(2)
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)
loss_ce = ce_loss(outputs, label_batch[:])
loss_dice = dice_loss(outputs_soft, label_batch.unsqueeze(1))
supervised_loss = 0.5 * (loss_dice + loss_ce)
consistency_weight = get_current_consistency_weight(iter_num //
150)
consistency_loss = F.mse_loss(outputs[args.labeled_bs:],
ema_output)
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()))
writer.add_scalar('loss/loss', loss, iter_num)
if iter_num % 20 == 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)
if iter_num > 0 and iter_num % 200 == 0:
model.eval()
avg_metric = test_all_case(model,
args.root_path,
test_list="val.txt",
num_classes=2,
patch_size=args.patch_size,
stride_xy=64,
stride_z=64)
if avg_metric[:, 0].mean() > best_performance:
best_performance = avg_metric[:, 0].mean()
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)
writer.add_scalar('info/val_dice_score', avg_metric[0, 0],
iter_num)
writer.add_scalar('info/val_hd95', avg_metric[0, 1], iter_num)
logging.info('iteration %d : dice_score : %f hd95 : %f' %
(iter_num, avg_metric[0, 0].mean(),
avg_metric[0, 1].mean()))
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!"
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_dis[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_DistMap',
grid_image, iter_num)
if iter_num > 0 and iter_num % 200 == 0:
model.eval()
avg_metric = test_all_case(
model, args.root_path, test_list="test.list", num_classes=num_classes, patch_size=patch_size, stride_xy=18, stride_z=4)
print(avg_metric)
if avg_metric[:, 0].mean() > best_performance:
best_performance = avg_metric[:, 0].mean()
save_mode_path = os.path.join(snapshot_path,
'best_model.pth')
torch.save(model.state_dict(), save_mode_path)
logging.info("save model to {}".format(save_mode_path))
writer.add_scalar('val/la_heart_dice',
avg_metric[0, 0], iter_num)
writer.add_scalar('val/la_heart_hd95',
avg_metric[0, 1], iter_num)
model.train()
# change lr
if iter_num % 2500 == 0:
