def run_inference_test(root_dir, device=torch.device("cuda:0")):
images = sorted(glob(os.path.join(root_dir, "im*.nii.gz")))
segs = sorted(glob(os.path.join(root_dir, "seg*.nii.gz")))
val_files = [{"img": img, "seg": seg} for img, seg in zip(images, segs)]
# define transforms for image and segmentation
val_transforms = Compose([
LoadNiftid(keys=["img", "seg"]),
AsChannelFirstd(keys=["img", "seg"], channel_dim=-1),
# resampling with align_corners=True or dtype=float64 will generate
# slight different results between PyTorch 1.5 an 1.6
Spacingd(keys=["img", "seg"],
pixdim=[1.2, 0.8, 0.7],
mode=["bilinear", "nearest"],
dtype=np.float32),
ScaleIntensityd(keys=["img", "seg"]),
ToTensord(keys=["img", "seg"]),
])
val_ds = monai.data.Dataset(data=val_files, transform=val_transforms)
# sliding window inferene need to input 1 image in every iteration
val_loader = monai.data.DataLoader(val_ds, batch_size=1, num_workers=4)
dice_metric = DiceMetric(include_background=True,
to_onehot_y=False,
sigmoid=True,
reduction="mean")
model = UNet(
dimensions=3,
in_channels=1,
out_channels=1,
channels=(16, 32, 64, 128, 256),
strides=(2, 2, 2, 2),
num_res_units=2,
).to(device)
model_filename = os.path.join(root_dir, "best_metric_model.pth")
model.load_state_dict(torch.load(model_filename))
model.eval()
with torch.no_grad():
metric_sum = 0.0
metric_count = 0
# resampling with align_corners=True or dtype=float64 will generate
# slight different results between PyTorch 1.5 an 1.6
saver = NiftiSaver(output_dir=os.path.join(root_dir, "output"),
dtype=np.float32)
for val_data in val_loader:
val_images, val_labels = val_data["img"].to(
device), val_data["seg"].to(device)
# define sliding window size and batch size for windows inference
sw_batch_size, roi_size = 4, (96, 96, 96)
val_outputs = sliding_window_inference(val_images, roi_size,
sw_batch_size, model)
value = dice_metric(y_pred=val_outputs, y=val_labels)
not_nans = dice_metric.not_nans.item()
metric_count += not_nans
metric_sum += value.item() * not_nans
val_outputs = (val_outputs.sigmoid() >= 0.5).float()
saver.save_batch(val_outputs, val_data["img_meta_dict"])
metric = metric_sum / metric_count
return metric
def test_shape(self, spatial_dim):
spatial_dim = int(spatial_dim)
model = UNet(
spatial_dims=2, in_channels=1, out_channels=1, channels=(4, 8, 16), strides=(2, 2), num_res_units=2
)
device = "cuda:0" if torch.cuda.is_available() else "cpu"
model.to(device)
model.eval()
# Initialize a dummy 3D tensor volume with shape (N,C,D,H,W)
input_volume = torch.ones(1, 1, 64, 256, 256, device=device)
# Remove spatial dim to slide across from the roi_size
roi_size = list(input_volume.shape[2:])
roi_size.pop(spatial_dim)
# Initialize and run inferer
inferer = SliceInferer(roi_size=roi_size, spatial_dim=spatial_dim, sw_batch_size=1, cval=-1)
result = inferer(input_volume, model)
self.assertTupleEqual(result.shape, input_volume.shape)
# test that the inferer can be run multiple times
result = inferer(input_volume, model)
def run_inference_test(root_dir, device=torch.device("cuda:0")):
images = sorted(glob(os.path.join(root_dir, "im*.nii.gz")))
segs = sorted(glob(os.path.join(root_dir, "seg*.nii.gz")))
val_files = [{"img": img, "seg": seg} for img, seg in zip(images, segs)]
# define transforms for image and segmentation
val_transforms = Compose([
LoadNiftid(keys=["img", "seg"]),
AsChannelFirstd(keys=["img", "seg"], channel_dim=-1),
ScaleIntensityd(keys=["img", "seg"]),
ToTensord(keys=["img", "seg"]),
])
val_ds = monai.data.Dataset(data=val_files, transform=val_transforms)
# sliding window inferene need to input 1 image in every iteration
val_loader = DataLoader(val_ds,
batch_size=1,
num_workers=4,
collate_fn=list_data_collate,
pin_memory=torch.cuda.is_available())
model = UNet(
dimensions=3,
in_channels=1,
out_channels=1,
channels=(16, 32, 64, 128, 256),
strides=(2, 2, 2, 2),
num_res_units=2,
).to(device)
model_filename = os.path.join(root_dir, "best_metric_model.pth")
model.load_state_dict(torch.load(model_filename))
model.eval()
with torch.no_grad():
metric_sum = 0.0
metric_count = 0
saver = NiftiSaver(output_dir=os.path.join(root_dir, "output"),
dtype=int)
for val_data in val_loader:
val_images, val_labels = val_data["img"].to(
device), val_data["seg"].to(device)
# define sliding window size and batch size for windows inference
sw_batch_size, roi_size = 4, (96, 96, 96)
val_outputs = sliding_window_inference(val_images, roi_size,
sw_batch_size, model)
value = compute_meandice(y_pred=val_outputs,
y=val_labels,
include_background=True,
to_onehot_y=False,
add_sigmoid=True)
metric_count += len(value)
metric_sum += value.sum().item()
val_outputs = (val_outputs.sigmoid() >= 0.5).float()
saver.save_batch(
val_outputs, {
"filename_or_obj": val_data["img.filename_or_obj"],
"affine": val_data["img.affine"]
})
metric = metric_sum / metric_count
return metric
def main(tempdir):
monai.config.print_config()
logging.basicConfig(stream=sys.stdout, level=logging.INFO)
print(f"generating synthetic data to {tempdir} (this may take a while)")
for i in range(5):
im, seg = create_test_image_2d(128, 128, num_seg_classes=1)
Image.fromarray(im.astype("uint8")).save(os.path.join(tempdir, f"img{i:d}.png"))
Image.fromarray(seg.astype("uint8")).save(os.path.join(tempdir, f"seg{i:d}.png"))
images = sorted(glob(os.path.join(tempdir, "img*.png")))
segs = sorted(glob(os.path.join(tempdir, "seg*.png")))
val_files = [{"img": img, "seg": seg} for img, seg in zip(images, segs)]
# define transforms for image and segmentation
val_transforms = Compose(
[
LoadImaged(keys=["img", "seg"]),
AddChanneld(keys=["img", "seg"]),
ScaleIntensityd(keys="img"),
ToTensord(keys=["img", "seg"]),
]
)
val_ds = monai.data.Dataset(data=val_files, transform=val_transforms)
# sliding window inference need to input 1 image in every iteration
val_loader = DataLoader(val_ds, batch_size=1, num_workers=4, collate_fn=list_data_collate)
dice_metric = DiceMetric(include_background=True, to_onehot_y=False, sigmoid=True, reduction="mean")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = UNet(
dimensions=2,
in_channels=1,
out_channels=1,
channels=(16, 32, 64, 128, 256),
strides=(2, 2, 2, 2),
num_res_units=2,
).to(device)
model.load_state_dict(torch.load("best_metric_model_segmentation2d_dict.pth"))
model.eval()
with torch.no_grad():
metric_sum = 0.0
metric_count = 0
saver = PNGSaver(output_dir="./output")
for val_data in val_loader:
val_images, val_labels = val_data["img"].to(device), val_data["seg"].to(device)
# define sliding window size and batch size for windows inference
roi_size = (96, 96)
sw_batch_size = 4
val_outputs = sliding_window_inference(val_images, roi_size, sw_batch_size, model)
value = dice_metric(y_pred=val_outputs, y=val_labels)
metric_count += len(value)
metric_sum += value.item() * len(value)
val_outputs = val_outputs.sigmoid() >= 0.5
saver.save_batch(val_outputs)
metric = metric_sum / metric_count
print("evaluation metric:", metric)
def main():
config.print_config()
logging.basicConfig(stream=sys.stdout, level=logging.INFO)
tempdir = tempfile.mkdtemp()
print('generating synthetic data to {} (this may take a while)'.format(tempdir))
for i in range(5):
im, seg = create_test_image_3d(128, 128, 128, num_seg_classes=1)
n = nib.Nifti1Image(im, np.eye(4))
nib.save(n, os.path.join(tempdir, 'im%i.nii.gz' % i))
n = nib.Nifti1Image(seg, np.eye(4))
nib.save(n, os.path.join(tempdir, 'seg%i.nii.gz' % i))
images = sorted(glob(os.path.join(tempdir, 'im*.nii.gz')))
segs = sorted(glob(os.path.join(tempdir, 'seg*.nii.gz')))
# define transforms for image and segmentation
imtrans = Compose([ScaleIntensity(), AddChannel(), ToTensor()])
segtrans = Compose([AddChannel(), ToTensor()])
val_ds = NiftiDataset(images, segs, transform=imtrans, seg_transform=segtrans, image_only=False)
# sliding window inference for one image at every iteration
val_loader = DataLoader(val_ds, batch_size=1, num_workers=1, pin_memory=torch.cuda.is_available())
device = torch.device('cuda:0')
model = UNet(
dimensions=3,
in_channels=1,
out_channels=1,
channels=(16, 32, 64, 128, 256),
strides=(2, 2, 2, 2),
num_res_units=2,
).to(device)
model.load_state_dict(torch.load('best_metric_model.pth'))
model.eval()
with torch.no_grad():
metric_sum = 0.
metric_count = 0
saver = NiftiSaver(output_dir='./output')
for val_data in val_loader:
val_images, val_labels = val_data[0].to(device), val_data[1].to(device)
# define sliding window size and batch size for windows inference
roi_size = (96, 96, 96)
sw_batch_size = 4
val_outputs = sliding_window_inference(val_images, roi_size, sw_batch_size, model)
value = compute_meandice(y_pred=val_outputs, y=val_labels, include_background=True,
to_onehot_y=False, add_sigmoid=True)
metric_count += len(value)
metric_sum += value.sum().item()
val_outputs = (val_outputs.sigmoid() >= 0.5).float()
saver.save_batch(val_outputs, val_data[2])
metric = metric_sum / metric_count
print('evaluation metric:', metric)
shutil.rmtree(tempdir)
def main(tempdir):
config.print_config()
logging.basicConfig(stream=sys.stdout, level=logging.INFO)
print(f"generating synthetic data to {tempdir} (this may take a while)")
for i in range(5):
im, seg = create_test_image_3d(128, 128, 128, num_seg_classes=1)
n = nib.Nifti1Image(im, np.eye(4))
nib.save(n, os.path.join(tempdir, f"im{i:d}.nii.gz"))
n = nib.Nifti1Image(seg, np.eye(4))
nib.save(n, os.path.join(tempdir, f"seg{i:d}.nii.gz"))
images = sorted(glob(os.path.join(tempdir, "im*.nii.gz")))
segs = sorted(glob(os.path.join(tempdir, "seg*.nii.gz")))
# define transforms for image and segmentation
imtrans = Compose([ScaleIntensity(), AddChannel(), ToTensor()])
segtrans = Compose([AddChannel(), ToTensor()])
val_ds = ImageDataset(images, segs, transform=imtrans, seg_transform=segtrans, image_only=False)
# sliding window inference for one image at every iteration
val_loader = DataLoader(val_ds, batch_size=1, num_workers=1, pin_memory=torch.cuda.is_available())
dice_metric = DiceMetric(include_background=True, reduction="mean")
post_trans = Compose([Activations(sigmoid=True), AsDiscrete(threshold_values=True)])
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = UNet(
dimensions=3,
in_channels=1,
out_channels=1,
channels=(16, 32, 64, 128, 256),
strides=(2, 2, 2, 2),
num_res_units=2,
).to(device)
model.load_state_dict(torch.load("best_metric_model_segmentation3d_array.pth"))
model.eval()
with torch.no_grad():
metric_sum = 0.0
metric_count = 0
saver = NiftiSaver(output_dir="./output")
for val_data in val_loader:
val_images, val_labels = val_data[0].to(device), val_data[1].to(device)
# define sliding window size and batch size for windows inference
roi_size = (96, 96, 96)
sw_batch_size = 4
val_outputs = sliding_window_inference(val_images, roi_size, sw_batch_size, model)
val_outputs = post_trans(val_outputs)
value, _ = dice_metric(y_pred=val_outputs, y=val_labels)
metric_count += len(value)
metric_sum += value.item() * len(value)
saver.save_batch(val_outputs, val_data[2])
metric = metric_sum / metric_count
print("evaluation metric:", metric)
def main(tempdir):
config.print_config()
logging.basicConfig(stream=sys.stdout, level=logging.INFO)
print(f"generating synthetic data to {tempdir} (this may take a while)")
for i in range(5):
im, seg = create_test_image_2d(128, 128, num_seg_classes=1)
Image.fromarray((im * 255).astype("uint8")).save(os.path.join(tempdir, f"img{i:d}.png"))
Image.fromarray((seg * 255).astype("uint8")).save(os.path.join(tempdir, f"seg{i:d}.png"))
images = sorted(glob(os.path.join(tempdir, "img*.png")))
segs = sorted(glob(os.path.join(tempdir, "seg*.png")))
# define transforms for image and segmentation
imtrans = Compose([LoadImage(image_only=True), AddChannel(), ScaleIntensity(), EnsureType()])
segtrans = Compose([LoadImage(image_only=True), AddChannel(), ScaleIntensity(), EnsureType()])
val_ds = ArrayDataset(images, imtrans, segs, segtrans)
# sliding window inference for one image at every iteration
val_loader = DataLoader(val_ds, batch_size=1, num_workers=1, pin_memory=torch.cuda.is_available())
dice_metric = DiceMetric(include_background=True, reduction="mean", get_not_nans=False)
post_trans = Compose([EnsureType(), Activations(sigmoid=True), AsDiscrete(threshold=0.5)])
saver = SaveImage(output_dir="./output", output_ext=".png", output_postfix="seg")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = UNet(
spatial_dims=2,
in_channels=1,
out_channels=1,
channels=(16, 32, 64, 128, 256),
strides=(2, 2, 2, 2),
num_res_units=2,
).to(device)
model.load_state_dict(torch.load("best_metric_model_segmentation2d_array.pth"))
model.eval()
with torch.no_grad():
for val_data in val_loader:
val_images, val_labels = val_data[0].to(device), val_data[1].to(device)
# define sliding window size and batch size for windows inference
roi_size = (96, 96)
sw_batch_size = 4
val_outputs = sliding_window_inference(val_images, roi_size, sw_batch_size, model)
val_outputs = [post_trans(i) for i in decollate_batch(val_outputs)]
val_labels = decollate_batch(val_labels)
# compute metric for current iteration
dice_metric(y_pred=val_outputs, y=val_labels)
for val_output in val_outputs:
saver(val_output)
# aggregate the final mean dice result
print("evaluation metric:", dice_metric.aggregate().item())
# reset the status
dice_metric.reset()
def test_shape(self, input_param, input_shape, expected_shape):
net = UNet(**input_param).to(device)
net.eval()
with torch.no_grad():
result = net.forward(torch.randn(input_shape).to(device))
self.assertEqual(result.shape, expected_shape)
loss = loss_function(outputs, labels)
loss.backward()
optimizer.step()
epoch_loss += loss.item()
print(
f"{step}/{len(train_ds) // train_loader.batch_size}, train_loss: {loss.item():.4f}"
)
epoch_len = len(train_ds) // train_loader.batch_size ##
# writer.add_scalar("train_loss", loss.item(), epoch_len * epoch + step) ##
writer.add_scalar("train_loss", epoch_loss, epoch + 1)
epoch_loss /= step
epoch_loss_values.append(epoch_loss)
print(f"epoch {epoch + 1} average loss: {epoch_loss:.4f}")
if (epoch + 1) % val_interval == 0:
model.eval()
with torch.no_grad():
metric_sum = 0.0
metric_count = 0
for val_data in val_loader:
val_inputs, val_labels = (
val_data["image"].to(device),
val_data["label"].to(device),
)
roi_size = (96, 96, 96)
sw_batch_size = 4
val_outputs = sliding_window_inference(val_inputs, roi_size,
sw_batch_size, model)
#val_outputs = post_pred(val_outputs)
#val_labels = post_label(val_labels)
val_outputs = [
def test_shape(self, input_param, input_data, expected_shape):
net = UNet(**input_param)
net.eval()
with torch.no_grad():
result = net.forward(input_data)
self.assertEqual(result.shape, expected_shape)
def main(tempdir):
monai.config.print_config()
logging.basicConfig(stream=sys.stdout, level=logging.INFO)
print(f"generating synthetic data to {tempdir} (this may take a while)")
for i in range(5):
im, seg = create_test_image_3d(128, 128, 128, num_seg_classes=1, channel_dim=-1)
n = nib.Nifti1Image(im, np.eye(4))
nib.save(n, os.path.join(tempdir, f"im{i:d}.nii.gz"))
n = nib.Nifti1Image(seg, np.eye(4))
nib.save(n, os.path.join(tempdir, f"seg{i:d}.nii.gz"))
images = sorted(glob(os.path.join(tempdir, "im*.nii.gz")))
segs = sorted(glob(os.path.join(tempdir, "seg*.nii.gz")))
val_files = [{"img": img, "seg": seg} for img, seg in zip(images, segs)]
# define transforms for image and segmentation
val_transforms = Compose(
[
LoadNiftid(keys=["img", "seg"]),
AsChannelFirstd(keys=["img", "seg"], channel_dim=-1),
ScaleIntensityd(keys="img"),
ToTensord(keys=["img", "seg"]),
]
)
val_ds = monai.data.Dataset(data=val_files, transform=val_transforms)
# sliding window inference need to input 1 image in every iteration
val_loader = DataLoader(val_ds, batch_size=1, num_workers=4, collate_fn=list_data_collate)
dice_metric = DiceMetric(include_background=True, reduction="mean")
post_trans = Compose([Activations(sigmoid=True), AsDiscrete(threshold_values=True)])
# try to use all the available GPUs
devices = get_devices_spec(None)
model = UNet(
dimensions=3,
in_channels=1,
out_channels=1,
channels=(16, 32, 64, 128, 256),
strides=(2, 2, 2, 2),
num_res_units=2,
).to(devices[0])
model.load_state_dict(torch.load("best_metric_model_segmentation3d_dict.pth"))
# if we have multiple GPUs, set data parallel to execute sliding window inference
if len(devices) > 1:
model = torch.nn.DataParallel(model, device_ids=devices)
model.eval()
with torch.no_grad():
metric_sum = 0.0
metric_count = 0
saver = NiftiSaver(output_dir="./output")
for val_data in val_loader:
val_images, val_labels = val_data["img"].to(devices[0]), val_data["seg"].to(devices[0])
# define sliding window size and batch size for windows inference
roi_size = (96, 96, 96)
sw_batch_size = 4
val_outputs = sliding_window_inference(val_images, roi_size, sw_batch_size, model)
val_outputs = post_trans(val_outputs)
value, _ = dice_metric(y_pred=val_outputs, y=val_labels)
metric_count += len(value)
metric_sum += value.item() * len(value)
saver.save_batch(val_outputs, val_data["img_meta_dict"])
metric = metric_sum / metric_count
print("evaluation metric:", metric)
def main():
monai.config.print_config()
logging.basicConfig(stream=sys.stdout, level=logging.INFO)
tempdir = tempfile.mkdtemp()
print(f"generating synthetic data to {tempdir} (this may take a while)")
for i in range(5):
im, seg = create_test_image_3d(128,
128,
128,
num_seg_classes=1,
channel_dim=-1)
n = nib.Nifti1Image(im, np.eye(4))
nib.save(n, os.path.join(tempdir, f"im{i:d}.nii.gz"))
n = nib.Nifti1Image(seg, np.eye(4))
nib.save(n, os.path.join(tempdir, f"seg{i:d}.nii.gz"))
images = sorted(glob(os.path.join(tempdir, "im*.nii.gz")))
segs = sorted(glob(os.path.join(tempdir, "seg*.nii.gz")))
val_files = [{"img": img, "seg": seg} for img, seg in zip(images, segs)]
# define transforms for image and segmentation
val_transforms = Compose([
LoadNiftid(keys=["img", "seg"]),
AsChannelFirstd(keys=["img", "seg"], channel_dim=-1),
ScaleIntensityd(keys=["img", "seg"]),
ToTensord(keys=["img", "seg"]),
])
val_ds = monai.data.Dataset(data=val_files, transform=val_transforms)
# sliding window inference need to input 1 image in every iteration
val_loader = DataLoader(val_ds,
batch_size=1,
num_workers=4,
collate_fn=list_data_collate,
pin_memory=torch.cuda.is_available())
device = torch.device("cuda:0")
model = UNet(
dimensions=3,
in_channels=1,
out_channels=1,
channels=(16, 32, 64, 128, 256),
strides=(2, 2, 2, 2),
num_res_units=2,
).to(device)
model.load_state_dict(torch.load("best_metric_model.pth"))
model.eval()
with torch.no_grad():
metric_sum = 0.0
metric_count = 0
saver = NiftiSaver(output_dir="./output")
for val_data in val_loader:
val_images, val_labels = val_data["img"].to(
device), val_data["seg"].to(device)
# define sliding window size and batch size for windows inference
roi_size = (96, 96, 96)
sw_batch_size = 4
val_outputs = sliding_window_inference(val_images, roi_size,
sw_batch_size, model)
value = compute_meandice(y_pred=val_outputs,
y=val_labels,
include_background=True,
to_onehot_y=False,
add_sigmoid=True)
metric_count += len(value)
metric_sum += value.sum().item()
val_outputs = (val_outputs.sigmoid() >= 0.5).float()
saver.save_batch(
val_outputs, {
"filename_or_obj": val_data["img.filename_or_obj"],
"affine": val_data["img.affine"]
})
metric = metric_sum / metric_count
print("evaluation metric:", metric)
shutil.rmtree(tempdir)
def test_train_timing(self):
images = sorted(glob(os.path.join(self.data_dir, "img*.nii.gz")))
segs = sorted(glob(os.path.join(self.data_dir, "seg*.nii.gz")))
train_files = [{
"image": img,
"label": seg
} for img, seg in zip(images[:32], segs[:32])]
val_files = [{
"image": img,
"label": seg
} for img, seg in zip(images[-9:], segs[-9:])]
device = torch.device("cuda:0")
# define transforms for train and validation
train_transforms = Compose([
LoadImaged(keys=["image", "label"]),
EnsureChannelFirstd(keys=["image", "label"]),
Spacingd(keys=["image", "label"],
pixdim=(1.0, 1.0, 1.0),
mode=("bilinear", "nearest")),
ScaleIntensityd(keys="image"),
CropForegroundd(keys=["image", "label"], source_key="image"),
# pre-compute foreground and background indexes
# and cache them to accelerate training
FgBgToIndicesd(keys="label", fg_postfix="_fg", bg_postfix="_bg"),
# change to execute transforms with Tensor data
EnsureTyped(keys=["image", "label"]),
# move the data to GPU and cache to avoid CPU -> GPU sync in every epoch
ToDeviced(keys=["image", "label"], device=device),
# randomly crop out patch samples from big
# image based on pos / neg ratio
# the image centers of negative samples
# must be in valid image area
RandCropByPosNegLabeld(
keys=["image", "label"],
label_key="label",
spatial_size=(64, 64, 64),
pos=1,
neg=1,
num_samples=4,
fg_indices_key="label_fg",
bg_indices_key="label_bg",
),
RandFlipd(keys=["image", "label"], prob=0.5, spatial_axis=[1, 2]),
RandAxisFlipd(keys=["image", "label"], prob=0.5),
RandRotate90d(keys=["image", "label"],
prob=0.5,
spatial_axes=(1, 2)),
RandZoomd(keys=["image", "label"],
prob=0.5,
min_zoom=0.8,
max_zoom=1.2,
keep_size=True),
RandRotated(
keys=["image", "label"],
prob=0.5,
range_x=np.pi / 4,
mode=("bilinear", "nearest"),
align_corners=True,
dtype=np.float64,
),
RandAffined(keys=["image", "label"],
prob=0.5,
rotate_range=np.pi / 2,
mode=("bilinear", "nearest")),
RandGaussianNoised(keys="image", prob=0.5),
RandStdShiftIntensityd(keys="image",
prob=0.5,
factors=0.05,
nonzero=True),
])
val_transforms = Compose([
LoadImaged(keys=["image", "label"]),
EnsureChannelFirstd(keys=["image", "label"]),
Spacingd(keys=["image", "label"],
pixdim=(1.0, 1.0, 1.0),
mode=("bilinear", "nearest")),
ScaleIntensityd(keys="image"),
CropForegroundd(keys=["image", "label"], source_key="image"),
EnsureTyped(keys=["image", "label"]),
# move the data to GPU and cache to avoid CPU -> GPU sync in every epoch
ToDeviced(keys=["image", "label"], device=device),
])
max_epochs = 5
learning_rate = 2e-4
val_interval = 1 # do validation for every epoch
# set CacheDataset, ThreadDataLoader and DiceCE loss for MONAI fast training
train_ds = CacheDataset(data=train_files,
transform=train_transforms,
cache_rate=1.0,
num_workers=8)
val_ds = CacheDataset(data=val_files,
transform=val_transforms,
cache_rate=1.0,
num_workers=5)
# disable multi-workers because `ThreadDataLoader` works with multi-threads
train_loader = ThreadDataLoader(train_ds,
num_workers=0,
batch_size=4,
shuffle=True)
val_loader = ThreadDataLoader(val_ds, num_workers=0, batch_size=1)
loss_function = DiceCELoss(to_onehot_y=True,
softmax=True,
squared_pred=True,
batch=True)
model = UNet(
spatial_dims=3,
in_channels=1,
out_channels=2,
channels=(16, 32, 64, 128, 256),
strides=(2, 2, 2, 2),
num_res_units=2,
norm=Norm.BATCH,
).to(device)
# Novograd paper suggests to use a bigger LR than Adam,
# because Adam does normalization by element-wise second moments
optimizer = Novograd(model.parameters(), learning_rate * 10)
scaler = torch.cuda.amp.GradScaler()
post_pred = Compose(
[EnsureType(), AsDiscrete(argmax=True, to_onehot=2)])
post_label = Compose([EnsureType(), AsDiscrete(to_onehot=2)])
dice_metric = DiceMetric(include_background=True,
reduction="mean",
get_not_nans=False)
best_metric = -1
total_start = time.time()
for epoch in range(max_epochs):
epoch_start = time.time()
print("-" * 10)
print(f"epoch {epoch + 1}/{max_epochs}")
model.train()
epoch_loss = 0
step = 0
for batch_data in train_loader:
step_start = time.time()
step += 1
optimizer.zero_grad()
# set AMP for training
with torch.cuda.amp.autocast():
outputs = model(batch_data["image"])
loss = loss_function(outputs, batch_data["label"])
scaler.scale(loss).backward()
scaler.step(optimizer)
scaler.update()
epoch_loss += loss.item()
epoch_len = math.ceil(len(train_ds) / train_loader.batch_size)
print(f"{step}/{epoch_len}, train_loss: {loss.item():.4f}"
f" step time: {(time.time() - step_start):.4f}")
epoch_loss /= step
print(f"epoch {epoch + 1} average loss: {epoch_loss:.4f}")
if (epoch + 1) % val_interval == 0:
model.eval()
with torch.no_grad():
for val_data in val_loader:
roi_size = (96, 96, 96)
sw_batch_size = 4
# set AMP for validation
with torch.cuda.amp.autocast():
val_outputs = sliding_window_inference(
val_data["image"], roi_size, sw_batch_size,
model)
val_outputs = [
post_pred(i) for i in decollate_batch(val_outputs)
]
val_labels = [
post_label(i)
for i in decollate_batch(val_data["label"])
]
dice_metric(y_pred=val_outputs, y=val_labels)
metric = dice_metric.aggregate().item()
dice_metric.reset()
if metric > best_metric:
best_metric = metric
print(
f"epoch: {epoch + 1} current mean dice: {metric:.4f}, best mean dice: {best_metric:.4f}"
)
print(
f"time consuming of epoch {epoch + 1} is: {(time.time() - epoch_start):.4f}"
)
total_time = time.time() - total_start
print(
f"train completed, best_metric: {best_metric:.4f} total time: {total_time:.4f}"
)
# test expected metrics
self.assertGreater(best_metric, 0.95)
def train_process(fast=False):
epoch_num = 10
val_interval = 1
train_trans, val_trans = transformations()
train_ds = Dataset(data=train_files, transform=train_trans)
val_ds = Dataset(data=val_files, transform=val_trans)
train_loader = DataLoader(train_ds, batch_size=2, shuffle=True)
val_loader = DataLoader(val_ds, batch_size=1)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
n1 = 16
model = UNet(dimensions=3,
in_channels=1,
out_channels=2,
channels=(n1 * 1, n1 * 2, n1 * 4, n1 * 8, n1 * 16),
strides=(2, 2, 2, 2)).to(device)
loss_function = DiceLoss(to_onehot_y=True, softmax=True)
post_pred = AsDiscrete(argmax=True, to_onehot=True, n_classes=2)
post_label = AsDiscrete(to_onehot=True, n_classes=2)
optimizer = torch.optim.Adam(model.parameters(), 1e-4, weight_decay=1e-5)
best_metric = -1
best_metric_epoch = -1
best_metrics_epochs_and_time = [[], [], []]
epoch_loss_values = list()
metric_values = list()
for epoch in range(epoch_num):
print(f"epoch {epoch + 1}/{epoch_num}")
model.train()
epoch_loss = 0
step = 0
for batch_data in train_loader:
step += 1
inputs, labels = batch_data['image'].to(
device), batch_data['label'].to(device)
optimizer.zero_grad()
outputs = model(inputs)
loss = loss_function(outputs, labels)
loss.backward()
optimizer.step()
epoch_loss += loss.item()
epoch_len = math.ceil(len(train_ds) / train_loader.batch_size)
print(f"{step}/{epoch_len}, train_loss: {loss.item():.4f}")
epoch_loss /= step
epoch_loss_values.append(epoch_loss)
print(f"epoch {epoch + 1} average loss: {epoch_loss:.4f}")
if (epoch + 1) % val_interval == 0:
model.eval()
with torch.no_grad():
metric_sum = 0.
metric_count = 0
for val_data in val_loader:
val_inputs, val_labels = val_data['image'].to(
device), val_data['label'].to(device)
val_outputs = model(val_inputs)
val_outputs = post_pred(val_outputs)
val_labels = post_label(val_labels)
value = compute_meandice(y_pred=val_outputs,
y=val_labels,
include_background=False)
metric_count += len(value)
metric_sum += value.sum().item()
metric = metric_sum / metric_count
metric_values.append(metric)
if metric > best_metric:
best_metric = metric
epochs_no_improve = 0
best_metric_epoch = epoch + 1
best_metrics_epochs_and_time[0].append(best_metric)
best_metrics_epochs_and_time[1].append(best_metric_epoch)
torch.save(model.state_dict(), 'sLUMRTL644.pth')
else:
epochs_no_improve += 1
print(
f"current epoch: {epoch + 1} current mean dice: {metric:.4f}"
f" best mean dice: {best_metric:.4f} at epoch: {best_metric_epoch}"
)
print(
f"train completed, best_metric: {best_metric:.4f} at epoch: {best_metric_epoch}"
)
return epoch_num, epoch_loss_values, metric_values, best_metrics_epochs_and_time
def main(tempdir):
print_config()
logging.basicConfig(stream=sys.stdout, level=logging.INFO)
print(f"generating synthetic data to {tempdir} (this may take a while)")
for i in range(5):
im, _ = create_test_image_3d(128,
128,
128,
num_seg_classes=1,
channel_dim=-1)
n = nib.Nifti1Image(im, np.eye(4))
nib.save(n, os.path.join(tempdir, f"im{i:d}.nii.gz"))
images = sorted(glob(os.path.join(tempdir, "im*.nii.gz")))
files = [{"img": img} for img in images]
# define pre transforms
pre_transforms = Compose([
LoadImaged(keys="img"),
EnsureChannelFirstd(keys="img"),
Orientationd(keys="img", axcodes="RAS"),
Resized(keys="img",
spatial_size=(96, 96, 96),
mode="trilinear",
align_corners=True),
ScaleIntensityd(keys="img"),
EnsureTyped(keys="img"),
])
# define dataset and dataloader
dataset = Dataset(data=files, transform=pre_transforms)
dataloader = DataLoader(dataset, batch_size=2, num_workers=4)
# define post transforms
post_transforms = Compose([
EnsureTyped(keys="pred"),
Activationsd(keys="pred", sigmoid=True),
Invertd(
keys=
"pred", # invert the `pred` data field, also support multiple fields
transform=pre_transforms,
orig_keys=
"img", # get the previously applied pre_transforms information on the `img` data field,
# then invert `pred` based on this information. we can use same info
# for multiple fields, also support different orig_keys for different fields
meta_keys=
"pred_meta_dict", # key field to save inverted meta data, every item maps to `keys`
orig_meta_keys=
"img_meta_dict", # get the meta data from `img_meta_dict` field when inverting,
# for example, may need the `affine` to invert `Spacingd` transform,
# multiple fields can use the same meta data to invert
meta_key_postfix=
"meta_dict", # if `meta_keys=None`, use "{keys}_{meta_key_postfix}" as the meta key,
# if `orig_meta_keys=None`, use "{orig_keys}_{meta_key_postfix}",
# otherwise, no need this arg during inverting
nearest_interp=
False, # don't change the interpolation mode to "nearest" when inverting transforms
# to ensure a smooth output, then execute `AsDiscreted` transform
to_tensor=True, # convert to PyTorch Tensor after inverting
),
AsDiscreted(keys="pred", threshold=0.5),
SaveImaged(keys="pred",
meta_keys="pred_meta_dict",
output_dir="./out",
output_postfix="seg",
resample=False),
])
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
net = UNet(
spatial_dims=3,
in_channels=1,
out_channels=1,
channels=(16, 32, 64, 128, 256),
strides=(2, 2, 2, 2),
num_res_units=2,
).to(device)
net.load_state_dict(
torch.load("best_metric_model_segmentation3d_dict.pth"))
net.eval()
with torch.no_grad():
for d in dataloader:
images = d["img"].to(device)
# define sliding window size and batch size for windows inference
d["pred"] = sliding_window_inference(inputs=images,
roi_size=(96, 96, 96),
sw_batch_size=4,
predictor=net)
# decollate the batch data into a list of dictionaries, then execute postprocessing transforms
d = [post_transforms(i) for i in decollate_batch(d)]
def main(tempdir):
print_config()
logging.basicConfig(stream=sys.stdout, level=logging.INFO)
print(f"generating synthetic data to {tempdir} (this may take a while)")
for i in range(5):
im, _ = create_test_image_3d(128,
128,
128,
num_seg_classes=1,
channel_dim=-1)
n = nib.Nifti1Image(im, np.eye(4))
nib.save(n, os.path.join(tempdir, f"im{i:d}.nii.gz"))
images = sorted(glob(os.path.join(tempdir, "im*.nii.gz")))
files = [{"img": img} for img in images]
# define pre transforms
pre_transforms = Compose([
LoadImaged(keys="img"),
EnsureChannelFirstd(keys="img"),
Orientationd(keys="img", axcodes="RAS"),
Resized(keys="img",
spatial_size=(96, 96, 96),
mode="trilinear",
align_corners=True),
ScaleIntensityd(keys="img"),
ToTensord(keys="img"),
])
# define dataset and dataloader
dataset = Dataset(data=files, transform=pre_transforms)
dataloader = DataLoader(dataset, batch_size=2, num_workers=4)
# define post transforms
post_transforms = Compose([
Activationsd(keys="pred", sigmoid=True),
AsDiscreted(keys="pred", threshold_values=True),
Invertd(keys="pred",
transform=pre_transforms,
loader=dataloader,
orig_keys="img",
nearest_interp=True),
SaveImaged(keys="pred_inverted",
output_dir="./output",
output_postfix="seg",
resample=False),
])
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
net = UNet(
dimensions=3,
in_channels=1,
out_channels=1,
channels=(16, 32, 64, 128, 256),
strides=(2, 2, 2, 2),
num_res_units=2,
).to(device)
net.load_state_dict(
torch.load("best_metric_model_segmentation3d_dict.pth"))
net.eval()
with torch.no_grad():
for d in dataloader:
images = d["img"].to(device)
# define sliding window size and batch size for windows inference
d["pred"] = sliding_window_inference(inputs=images,
roi_size=(96, 96, 96),
sw_batch_size=4,
predictor=net)
# execute post transforms to invert spatial transforms and save to NIfTI files
post_transforms(d)
def main(tempdir):
monai.config.print_config()
logging.basicConfig(stream=sys.stdout, level=logging.INFO)
print(f"generating synthetic data to {tempdir} (this may take a while)")
for i in range(5):
im, seg = create_test_image_3d(128, 128, 128, num_seg_classes=1, channel_dim=-1)
n = nib.Nifti1Image(im, np.eye(4))
nib.save(n, os.path.join(tempdir, f"im{i:d}.nii.gz"))
n = nib.Nifti1Image(seg, np.eye(4))
nib.save(n, os.path.join(tempdir, f"seg{i:d}.nii.gz"))
images = sorted(glob(os.path.join(tempdir, "im*.nii.gz")))
segs = sorted(glob(os.path.join(tempdir, "seg*.nii.gz")))
val_files = [{"img": img, "seg": seg} for img, seg in zip(images, segs)]
# define transforms for image and segmentation
val_transforms = Compose(
[
LoadImaged(keys=["img", "seg"]),
AsChannelFirstd(keys=["img", "seg"], channel_dim=-1),
ScaleIntensityd(keys="img"),
EnsureTyped(keys=["img", "seg"]),
]
)
val_ds = monai.data.Dataset(data=val_files, transform=val_transforms)
# sliding window inference need to input 1 image in every iteration
val_loader = DataLoader(val_ds, batch_size=1, num_workers=4, collate_fn=list_data_collate)
dice_metric = DiceMetric(include_background=True, reduction="mean", get_not_nans=False)
post_trans = Compose([EnsureType(), Activations(sigmoid=True), AsDiscrete(threshold=0.5)])
saver = SaveImage(output_dir="./output", output_ext=".nii.gz", output_postfix="seg")
# try to use all the available GPUs
devices = [torch.device("cuda" if torch.cuda.is_available() else "cpu")]
#devices = get_devices_spec(None)
model = UNet(
spatial_dims=3,
in_channels=1,
out_channels=1,
channels=(16, 32, 64, 128, 256),
strides=(2, 2, 2, 2),
num_res_units=2,
).to(devices[0])
model.load_state_dict(torch.load("best_metric_model_segmentation3d_dict.pth"))
# if we have multiple GPUs, set data parallel to execute sliding window inference
if len(devices) > 1:
model = torch.nn.DataParallel(model, device_ids=devices)
model.eval()
with torch.no_grad():
for val_data in val_loader:
val_images, val_labels = val_data["img"].to(devices[0]), val_data["seg"].to(devices[0])
# define sliding window size and batch size for windows inference
roi_size = (96, 96, 96)
sw_batch_size = 4
val_outputs = sliding_window_inference(val_images, roi_size, sw_batch_size, model)
val_outputs = [post_trans(i) for i in decollate_batch(val_outputs)]
val_labels = decollate_batch(val_labels)
meta_data = decollate_batch(val_data["img_meta_dict"])
# compute metric for current iteration
dice_metric(y_pred=val_outputs, y=val_labels)
for val_output, data in zip(val_outputs, meta_data):
saver(val_output, data)
# aggregate the final mean dice result
print("evaluation metric:", dice_metric.aggregate().item())
# reset the status
dice_metric.reset()