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
net = unet_3D(n_classes=2, in_channels=1)
model = net.cuda()
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()
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:]
outputs = model(volume_batch)
outputs_soft = torch.softmax(outputs, dim=1)
loss_ce = ce_loss(outputs[:args.labeled_bs],
label_batch[:args.labeled_bs][:])
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=2)
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[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!"
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!"
image_a_adapt = FDA_source_to_target(image_a, image_b, 0.02).cuda()
image_b_adapt = FDA_source_to_target(image_b, image_a, 0.02).cuda()
optimiser.zero_grad()
a1, a2, a3, a4, a5 = net.downsample(image_a)
pred_seg_a = net.upsample(a1, a2, a3, a4, a5)
# pred_seg_a = net(image_a)
loss_seg_a = criterion(pred_seg_a, target_a)
res2 = net.downsample(image_b)
pred_seg_b = net.upsample(*res2)
ent_a = entropy_loss(pred_seg_a)
ent_b = entropy_loss(pred_seg_b)
loss = loss_seg_a + LAMBDA_ENT * (ent_a + ent_b)
loss.backward()
optimiser.step()
# dice_score = dice_coeff(torch.round(pred), l).item()
# epoch_train_loss_rec.append(loss_recon.item())
epoch_train_loss_seg.append(loss_seg_a.item())
# mean_loss_rec = np.mean(epoch_train_loss_rec)
mean_loss_seg = np.mean(epoch_train_loss_seg)
# print('Train A - avg seg:{}'.format(np.mean(epoch_train_loss_seg)))