def test_shape(self, input_param, input_shape, expected_shape):
net = SegResNet(**input_param)
net.eval()
with torch.no_grad():
result = net(torch.randn(input_shape))
self.assertEqual(result.shape, expected_shape)
def test_ill_arg(self):
with self.assertRaises(AssertionError):
SegResNet(spatial_dims=4)
def main_worker(args):
# disable logging for processes except 0 on every node
if args.local_rank != 0:
f = open(os.devnull, "w")
sys.stdout = sys.stderr = f
if not os.path.exists(args.dir):
raise FileNotFoundError(f"Missing directory {args.dir}")
# initialize the distributed training process, every GPU runs in a process
dist.init_process_group(backend="nccl", init_method="env://")
total_start = time.time()
train_transforms = Compose([
# load 4 Nifti images and stack them together
LoadNiftid(keys=["image", "label"]),
AsChannelFirstd(keys="image"),
ConvertToMultiChannelBasedOnBratsClassesd(keys="label"),
Spacingd(keys=["image", "label"],
pixdim=(1.5, 1.5, 2.0),
mode=("bilinear", "nearest")),
Orientationd(keys=["image", "label"], axcodes="RAS"),
RandSpatialCropd(keys=["image", "label"],
roi_size=[128, 128, 64],
random_size=False),
NormalizeIntensityd(keys="image", nonzero=True, channel_wise=True),
RandFlipd(keys=["image", "label"], prob=0.5, spatial_axis=0),
RandScaleIntensityd(keys="image", factors=0.1, prob=0.5),
RandShiftIntensityd(keys="image", offsets=0.1, prob=0.5),
ToTensord(keys=["image", "label"]),
])
# create a training data loader
train_ds = BratsCacheDataset(
root_dir=args.dir,
transform=train_transforms,
section="training",
num_workers=4,
cache_rate=args.cache_rate,
shuffle=True,
)
train_loader = DataLoader(train_ds,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
# validation transforms and dataset
val_transforms = Compose([
LoadNiftid(keys=["image", "label"]),
AsChannelFirstd(keys="image"),
ConvertToMultiChannelBasedOnBratsClassesd(keys="label"),
Spacingd(keys=["image", "label"],
pixdim=(1.5, 1.5, 2.0),
mode=("bilinear", "nearest")),
Orientationd(keys=["image", "label"], axcodes="RAS"),
CenterSpatialCropd(keys=["image", "label"], roi_size=[128, 128, 64]),
NormalizeIntensityd(keys="image", nonzero=True, channel_wise=True),
ToTensord(keys=["image", "label"]),
])
val_ds = BratsCacheDataset(
root_dir=args.dir,
transform=val_transforms,
section="validation",
num_workers=4,
cache_rate=args.cache_rate,
shuffle=False,
)
val_loader = DataLoader(val_ds,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
if dist.get_rank() == 0:
# Logging for TensorBoard
writer = SummaryWriter(log_dir=args.log_dir)
# create UNet, DiceLoss and Adam optimizer
device = torch.device(f"cuda:{args.local_rank}")
if args.network == "UNet":
model = UNet(
dimensions=3,
in_channels=4,
out_channels=3,
channels=(16, 32, 64, 128, 256),
strides=(2, 2, 2, 2),
num_res_units=2,
).to(device)
else:
model = SegResNet(in_channels=4,
out_channels=3,
init_filters=16,
dropout_prob=0.2).to(device)
loss_function = DiceLoss(to_onehot_y=False,
sigmoid=True,
squared_pred=True)
optimizer = torch.optim.Adam(model.parameters(),
lr=args.lr,
weight_decay=1e-5,
amsgrad=True)
# wrap the model with DistributedDataParallel module
model = DistributedDataParallel(model, device_ids=[args.local_rank])
# start a typical PyTorch training
total_epoch = args.epochs
best_metric = -1000000
best_metric_epoch = -1
epoch_time = AverageMeter("Time", ":6.3f")
progress = ProgressMeter(total_epoch, [epoch_time], prefix="Epoch: ")
end = time.time()
print(f"Time elapsed before training: {end-total_start}")
for epoch in range(total_epoch):
train_loss = train(train_loader, model, loss_function, optimizer,
epoch, args, device)
epoch_time.update(time.time() - end)
if epoch % args.print_freq == 0:
progress.display(epoch)
if dist.get_rank() == 0:
writer.add_scalar("Loss/train", train_loss, epoch)
if (epoch + 1) % args.val_interval == 0:
metric, metric_tc, metric_wt, metric_et = evaluate(
model, val_loader, device)
if dist.get_rank() == 0:
writer.add_scalar("Mean Dice/val", metric, epoch)
writer.add_scalar("Mean Dice TC/val", metric_tc, epoch)
writer.add_scalar("Mean Dice WT/val", metric_wt, epoch)
writer.add_scalar("Mean Dice ET/val", metric_et, epoch)
if metric > best_metric:
best_metric = metric
best_metric_epoch = epoch + 1
print(
f"current epoch: {epoch + 1} current mean dice: {metric:.4f}"
f" tc: {metric_tc:.4f} wt: {metric_wt:.4f} et: {metric_et:.4f}"
f"\nbest mean dice: {best_metric:.4f} at epoch: {best_metric_epoch}"
)
end = time.time()
print(f"Time elapsed after epoch {epoch + 1} is {end - total_start}")
if dist.get_rank() == 0:
print(
f"train completed, best_metric: {best_metric:.4f} at epoch: {best_metric_epoch}"
)
# all processes should see same parameters as they all start from same
# random parameters and gradients are synchronized in backward passes,
# therefore, saving it in one process is sufficient
torch.save(model.state_dict(), "final_model.pth")
writer.flush()
dist.destroy_process_group()
def test_script(self):
input_param, input_shape, expected_shape = TEST_CASE_SEGRESNET[0]
net = SegResNet(**input_param)
test_data = torch.randn(input_shape)
out_orig, out_reloaded = test_script_save(net, test_data)
assert torch.allclose(out_orig, out_reloaded)
def main_worker(args):
# disable logging for processes except 0 on every node
if args.local_rank != 0:
f = open(os.devnull, "w")
sys.stdout = sys.stderr = f
if not os.path.exists(args.dir):
raise FileNotFoundError(f"missing directory {args.dir}")
# initialize the distributed training process, every GPU runs in a process
dist.init_process_group(backend="nccl", init_method="env://")
device = torch.device(f"cuda:{args.local_rank}")
torch.cuda.set_device(device)
# use amp to accelerate training
scaler = torch.cuda.amp.GradScaler()
torch.backends.cudnn.benchmark = True
total_start = time.time()
train_transforms = Compose([
# load 4 Nifti images and stack them together
LoadImaged(keys=["image", "label"]),
EnsureChannelFirstd(keys="image"),
ConvertToMultiChannelBasedOnBratsClassesd(keys="label"),
Orientationd(keys=["image", "label"], axcodes="RAS"),
Spacingd(
keys=["image", "label"],
pixdim=(1.0, 1.0, 1.0),
mode=("bilinear", "nearest"),
),
EnsureTyped(keys=["image", "label"]),
ToDeviced(keys=["image", "label"], device=device),
RandSpatialCropd(keys=["image", "label"],
roi_size=[224, 224, 144],
random_size=False),
RandFlipd(keys=["image", "label"], prob=0.5, spatial_axis=0),
RandFlipd(keys=["image", "label"], prob=0.5, spatial_axis=1),
RandFlipd(keys=["image", "label"], prob=0.5, spatial_axis=2),
NormalizeIntensityd(keys="image", nonzero=True, channel_wise=True),
RandScaleIntensityd(keys="image", factors=0.1, prob=0.5),
RandShiftIntensityd(keys="image", offsets=0.1, prob=0.5),
])
# create a training data loader
train_ds = BratsCacheDataset(
root_dir=args.dir,
transform=train_transforms,
section="training",
num_workers=4,
cache_rate=args.cache_rate,
shuffle=True,
)
# ThreadDataLoader can be faster if no IO operations when caching all the data in memory
train_loader = ThreadDataLoader(train_ds,
num_workers=0,
batch_size=args.batch_size,
shuffle=True)
# validation transforms and dataset
val_transforms = Compose([
LoadImaged(keys=["image", "label"]),
EnsureChannelFirstd(keys="image"),
ConvertToMultiChannelBasedOnBratsClassesd(keys="label"),
Orientationd(keys=["image", "label"], axcodes="RAS"),
Spacingd(
keys=["image", "label"],
pixdim=(1.0, 1.0, 1.0),
mode=("bilinear", "nearest"),
),
NormalizeIntensityd(keys="image", nonzero=True, channel_wise=True),
EnsureTyped(keys=["image", "label"]),
ToDeviced(keys=["image", "label"], device=device),
])
val_ds = BratsCacheDataset(
root_dir=args.dir,
transform=val_transforms,
section="validation",
num_workers=4,
cache_rate=args.cache_rate,
shuffle=False,
)
# ThreadDataLoader can be faster if no IO operations when caching all the data in memory
val_loader = ThreadDataLoader(val_ds,
num_workers=0,
batch_size=args.batch_size,
shuffle=False)
# create network, loss function and optimizer
if args.network == "SegResNet":
model = SegResNet(
blocks_down=[1, 2, 2, 4],
blocks_up=[1, 1, 1],
init_filters=16,
in_channels=4,
out_channels=3,
dropout_prob=0.0,
).to(device)
else:
model = UNet(
spatial_dims=3,
in_channels=4,
out_channels=3,
channels=(16, 32, 64, 128, 256),
strides=(2, 2, 2, 2),
num_res_units=2,
).to(device)
loss_function = DiceFocalLoss(
smooth_nr=1e-5,
smooth_dr=1e-5,
squared_pred=True,
to_onehot_y=False,
sigmoid=True,
batch=True,
)
optimizer = Novograd(model.parameters(), lr=args.lr)
lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer, T_max=args.epochs)
# wrap the model with DistributedDataParallel module
model = DistributedDataParallel(model, device_ids=[device])
dice_metric = DiceMetric(include_background=True, reduction="mean")
dice_metric_batch = DiceMetric(include_background=True,
reduction="mean_batch")
post_trans = Compose(
[EnsureType(),
Activations(sigmoid=True),
AsDiscrete(threshold=0.5)])
# start a typical PyTorch training
best_metric = -1
best_metric_epoch = -1
print(f"time elapsed before training: {time.time() - total_start}")
train_start = time.time()
for epoch in range(args.epochs):
epoch_start = time.time()
print("-" * 10)
print(f"epoch {epoch + 1}/{args.epochs}")
epoch_loss = train(train_loader, model, loss_function, optimizer,
lr_scheduler, scaler)
print(f"epoch {epoch + 1} average loss: {epoch_loss:.4f}")
if (epoch + 1) % args.val_interval == 0:
metric, metric_tc, metric_wt, metric_et = evaluate(
model, val_loader, dice_metric, dice_metric_batch, post_trans)
if metric > best_metric:
best_metric = metric
best_metric_epoch = epoch + 1
if dist.get_rank() == 0:
torch.save(model.state_dict(), "best_metric_model.pth")
print(
f"current epoch: {epoch + 1} current mean dice: {metric:.4f}"
f" tc: {metric_tc:.4f} wt: {metric_wt:.4f} et: {metric_et:.4f}"
f"\nbest mean dice: {best_metric:.4f} at epoch: {best_metric_epoch}"
)
print(
f"time consuming of epoch {epoch + 1} is: {(time.time() - epoch_start):.4f}"
)
print(
f"train completed, best_metric: {best_metric:.4f} at epoch: {best_metric_epoch},"
f" total train time: {(time.time() - train_start):.4f}")
dist.destroy_process_group()
def test_shape(self, input_param, input_shape, expected_shape):
net = SegResNet(**input_param).to(device)
with eval_mode(net):
result = net(torch.randn(input_shape).to(device))
self.assertEqual(result.shape, expected_shape)
def test_script(self):
input_param, input_shape, expected_shape = TEST_CASE_SEGRESNET[0]
net = SegResNet(**input_param)
test_data = torch.randn(input_shape)
test_script_save(net, test_data)