def run_inference_test(root_dir, test_x, test_y, device="cuda:0"):
# define transforms for image and classification
val_transforms = Compose([LoadPNG(image_only=True), AddChannel(), ScaleIntensity(), ToTensor()])
val_ds = MedNISTDataset(test_x, test_y, val_transforms)
val_loader = DataLoader(val_ds, batch_size=300, num_workers=10)
model = densenet121(spatial_dims=2, in_channels=1, out_channels=len(np.unique(test_y))).to(device)
model_filename = os.path.join(root_dir, "best_metric_model.pth")
model.load_state_dict(torch.load(model_filename))
model.eval()
y_true = list()
y_pred = list()
with torch.no_grad():
for test_data in val_loader:
test_images, test_labels = test_data[0].to(device), test_data[1].to(device)
pred = model(test_images).argmax(dim=1)
for i in range(len(pred)):
y_true.append(test_labels[i].item())
y_pred.append(pred[i].item())
tps = [np.sum((np.asarray(y_true) == idx) & (np.asarray(y_pred) == idx)) for idx in np.unique(test_y)]
return tps
train_x = [image_files_list[i] for i in train_indices]
train_y = [image_class[i] for i in train_indices]
val_x = [image_files_list[i] for i in val_indices]
val_y = [image_class[i] for i in val_indices]
test_x = [image_files_list[i] for i in test_indices]
test_y = [image_class[i] for i in test_indices]
# MONAI transforms, Dataset and Dataloader for preprocessing
train_transforms = Compose([
LoadImage(image_only=True),
AddChannel(),
ScaleIntensity(),
RandRotate(range_x=np.pi / 12, prob=0.5, keep_size=True),
RandFlip(spatial_axis=0, prob=0.5),
RandZoom(min_zoom=0.9, max_zoom=1.1, prob=0.5),
ToTensor(),
])
val_transforms = Compose([
LoadImage(image_only=True),
AddChannel(),
ScaleIntensity(),
ToTensor()
])
act = Activations(softmax=True)
to_onehot = AsDiscrete(to_onehot=True, n_classes=num_class)
class MedNISTDataset(torch.utils.data.Dataset):
def __init__(self, image_files, labels, transforms):
self.image_files = image_files
def main():
monai.config.print_config()
logging.basicConfig(stream=sys.stdout, level=logging.INFO)
# IXI dataset as a demo, downloadable from https://brain-development.org/ixi-dataset/
images = [
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI314-IOP-0889-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI249-Guys-1072-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI609-HH-2600-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI173-HH-1590-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI020-Guys-0700-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI342-Guys-0909-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI134-Guys-0780-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI577-HH-2661-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI066-Guys-0731-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI130-HH-1528-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI607-Guys-1097-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI175-HH-1570-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI385-HH-2078-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI344-Guys-0905-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI409-Guys-0960-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI584-Guys-1129-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI253-HH-1694-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI092-HH-1436-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI574-IOP-1156-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI585-Guys-1130-T1.nii.gz"
]),
]
# 2 binary labels for gender classification: man and woman
labels = np.array(
[0, 0, 0, 1, 0, 0, 0, 1, 1, 0, 0, 0, 1, 0, 1, 0, 1, 0, 1, 0],
dtype=np.int64)
# Define transforms
train_transforms = Compose([
ScaleIntensity(),
AddChannel(),
Resize((96, 96, 96)),
RandRotate90(),
ToTensor()
])
val_transforms = Compose(
[ScaleIntensity(),
AddChannel(),
Resize((96, 96, 96)),
ToTensor()])
# Define nifti dataset, data loader
check_ds = NiftiDataset(image_files=images,
labels=labels,
transform=train_transforms)
check_loader = DataLoader(check_ds,
batch_size=2,
num_workers=2,
pin_memory=torch.cuda.is_available())
im, label = monai.utils.misc.first(check_loader)
print(type(im), im.shape, label)
# create a training data loader
train_ds = NiftiDataset(image_files=images[:10],
labels=labels[:10],
transform=train_transforms)
train_loader = DataLoader(train_ds,
batch_size=2,
shuffle=True,
num_workers=2,
pin_memory=torch.cuda.is_available())
# create a validation data loader
val_ds = NiftiDataset(image_files=images[-10:],
labels=labels[-10:],
transform=val_transforms)
val_loader = DataLoader(val_ds,
batch_size=2,
num_workers=2,
pin_memory=torch.cuda.is_available())
# Create DenseNet121, CrossEntropyLoss and Adam optimizer
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = monai.networks.nets.densenet.densenet121(spatial_dims=3,
in_channels=1,
out_channels=2).to(device)
loss_function = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), 1e-5)
# start a typical PyTorch training
val_interval = 2
best_metric = -1
best_metric_epoch = -1
epoch_loss_values = list()
metric_values = list()
writer = SummaryWriter()
for epoch in range(5):
print("-" * 10)
print(f"epoch {epoch + 1}/{5}")
model.train()
epoch_loss = 0
step = 0
for batch_data in train_loader:
step += 1
inputs, labels = batch_data[0].to(device), batch_data[1].to(device)
optimizer.zero_grad()
outputs = model(inputs)
loss = loss_function(outputs, labels)
loss.backward()
optimizer.step()
epoch_loss += loss.item()
epoch_len = len(train_ds) // train_loader.batch_size
print(f"{step}/{epoch_len}, train_loss: {loss.item():.4f}")
writer.add_scalar("train_loss", loss.item(),
epoch_len * epoch + step)
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():
num_correct = 0.0
metric_count = 0
for val_data in val_loader:
val_images, val_labels = val_data[0].to(
device), val_data[1].to(device)
val_outputs = model(val_images)
value = torch.eq(val_outputs.argmax(dim=1), val_labels)
metric_count += len(value)
num_correct += value.sum().item()
metric = num_correct / metric_count
metric_values.append(metric)
if metric > best_metric:
best_metric = metric
best_metric_epoch = epoch + 1
torch.save(model.state_dict(),
"best_metric_model_classification3d_array.pth")
print("saved new best metric model")
print(
"current epoch: {} current accuracy: {:.4f} best accuracy: {:.4f} at epoch {}"
.format(epoch + 1, metric, best_metric, best_metric_epoch))
writer.add_scalar("val_accuracy", metric, epoch + 1)
print(
f"train completed, best_metric: {best_metric:.4f} at epoch: {best_metric_epoch}"
)
writer.close()
def main():
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)
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 = 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())
dice_metric = DiceMetric(include_background=True,
to_onehot_y=False,
sigmoid=True,
reduction="mean")
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[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 = 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).float()
saver.save_batch(val_outputs, val_data[2])
metric = metric_sum / metric_count
print("evaluation metric:", metric)
shutil.rmtree(tempdir)
def default_collate(batch):
data = torch.stack([item[0] for item in batch])
target = torch.stack([item[1] for item in batch]) # image labels.
return data, target
train_transforms = Compose([ScaleIntensity(),
Resize((image_size, image_size, image_size)),
RandAffine(
prob=0.5,
# rotate_range=(np.pi * 2, np.pi * 2, np.pi * 2),
scale_range=(0.15, 0.15, 0.15),
padding_mode='border'),
ToTensor()])
val_transforms = Compose([ScaleIntensity(),Resize((image_size, image_size, image_size)),ToTensor()])
dataset_show = RSNADataset3D(df_study.head(5), 'train', transform=val_transforms)
dataset_show_aug = RSNADataset3D(df_study.head(5), 'train', transform=train_transforms)
bce = nn.BCEWithLogitsLoss()
def criterion(logits, target):
loss = bce(logits.cuda(), target.cuda())
return loss
# In[21]:
def main(tempdir):
monai.config.print_config()
logging.basicConfig(stream=sys.stdout, level=logging.INFO)
# create a temporary directory and 40 random image, mask pairs
print(f"generating synthetic data to {tempdir} (this may take a while)")
for i in range(40):
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")))
# define transforms for image and segmentation
train_imtrans = Compose([
LoadImage(image_only=True),
ScaleIntensity(),
AddChannel(),
RandSpatialCrop((96, 96), random_size=False),
RandRotate90(prob=0.5, spatial_axes=(0, 1)),
ToTensor(),
])
train_segtrans = Compose([
LoadImage(image_only=True),
AddChannel(),
RandSpatialCrop((96, 96), random_size=False),
RandRotate90(prob=0.5, spatial_axes=(0, 1)),
ToTensor(),
])
val_imtrans = Compose([
LoadImage(image_only=True),
ScaleIntensity(),
AddChannel(),
ToTensor()
])
val_segtrans = Compose(
[LoadImage(image_only=True),
AddChannel(), ToTensor()])
# define array dataset, data loader
check_ds = ArrayDataset(images, train_imtrans, segs, train_segtrans)
check_loader = DataLoader(check_ds,
batch_size=10,
num_workers=2,
pin_memory=torch.cuda.is_available())
im, seg = monai.utils.misc.first(check_loader)
print(im.shape, seg.shape)
# create a training data loader
train_ds = ArrayDataset(images[:20], train_imtrans, segs[:20],
train_segtrans)
train_loader = DataLoader(train_ds,
batch_size=4,
shuffle=True,
num_workers=8,
pin_memory=torch.cuda.is_available())
# create a validation data loader
val_ds = ArrayDataset(images[-20:], val_imtrans, segs[-20:], val_segtrans)
val_loader = DataLoader(val_ds,
batch_size=1,
num_workers=4,
pin_memory=torch.cuda.is_available())
dice_metric = DiceMetric(include_background=True, reduction="mean")
post_trans = Compose(
[Activations(sigmoid=True),
AsDiscrete(threshold_values=True)])
# create UNet, DiceLoss and Adam optimizer
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = monai.networks.nets.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)
loss_function = monai.losses.DiceLoss(sigmoid=True)
optimizer = torch.optim.Adam(model.parameters(), 1e-3)
# start a typical PyTorch training
val_interval = 2
best_metric = -1
best_metric_epoch = -1
epoch_loss_values = list()
metric_values = list()
writer = SummaryWriter()
for epoch in range(10):
print("-" * 10)
print(f"epoch {epoch + 1}/{10}")
model.train()
epoch_loss = 0
step = 0
for batch_data in train_loader:
step += 1
inputs, labels = batch_data[0].to(device), batch_data[1].to(device)
optimizer.zero_grad()
outputs = model(inputs)
loss = loss_function(outputs, labels)
loss.backward()
optimizer.step()
epoch_loss += loss.item()
epoch_len = len(train_ds) // train_loader.batch_size
print(f"{step}/{epoch_len}, train_loss: {loss.item():.4f}")
writer.add_scalar("train_loss", loss.item(),
epoch_len * epoch + step)
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
val_images = None
val_labels = None
val_outputs = None
for val_data in val_loader:
val_images, val_labels = val_data[0].to(
device), val_data[1].to(device)
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(val_outputs)
value, _ = dice_metric(y_pred=val_outputs, y=val_labels)
metric_count += len(value)
metric_sum += value.item() * len(value)
metric = metric_sum / metric_count
metric_values.append(metric)
if metric > best_metric:
best_metric = metric
best_metric_epoch = epoch + 1
torch.save(model.state_dict(),
"best_metric_model_segmentation2d_array.pth")
print("saved new best metric model")
print(
"current epoch: {} current mean dice: {:.4f} best mean dice: {:.4f} at epoch {}"
.format(epoch + 1, metric, best_metric, best_metric_epoch))
writer.add_scalar("val_mean_dice", metric, epoch + 1)
# plot the last model output as GIF image in TensorBoard with the corresponding image and label
plot_2d_or_3d_image(val_images,
epoch + 1,
writer,
index=0,
tag="image")
plot_2d_or_3d_image(val_labels,
epoch + 1,
writer,
index=0,
tag="label")
plot_2d_or_3d_image(val_outputs,
epoch + 1,
writer,
index=0,
tag="output")
print(
f"train completed, best_metric: {best_metric:.4f} at epoch: {best_metric_epoch}"
)
writer.close()
def test_cupy(self):
test_data = [[1, 2], [3, 4]]
cupy_array = cp.ascontiguousarray(cp.asarray(test_data))
result = ToTensor()(cupy_array)
self.assertTrue(isinstance(result, torch.Tensor))
assert_allclose(result, test_data, type_test=False)
def _define_transforms(self):
"""Define and initialize all data transforms.
* training set images transform
* training set targets transform
* validation set images transform
* validation set targets transform
* validation set images post-transform
* test set images transform
* test set targets transform
* test set images post-transform
* prediction set images transform
* prediction set images post-transform
@return True if data transforms could be instantiated, False otherwise.
"""
# Define transforms for training
self._train_image_transforms = Compose([
LoadImage(image_only=True),
ScaleIntensityRange(0, 65535, 0.0, 1.0, clip=False),
AddChannel(),
RandSpatialCrop(self._roi_size, random_size=False),
RandRotate90(prob=0.5, spatial_axes=(0, 1)),
ToTensor()
])
self._train_target_transforms = Compose([
LoadImage(image_only=True),
ScaleIntensityRange(0, 65535, 0.0, 1.0, clip=False),
AddChannel(),
RandSpatialCrop(self._roi_size, random_size=False),
RandRotate90(prob=0.5, spatial_axes=(0, 1)),
ToTensor()
])
# Define transforms for validation
self._validation_image_transforms = Compose([
LoadImage(image_only=True),
ScaleIntensityRange(0, 65535, 0.0, 1.0, clip=False),
AddChannel(),
ToTensor()
])
self._validation_target_transforms = Compose([
LoadImage(image_only=True),
ScaleIntensityRange(0, 65535, 0.0, 1.0, clip=False),
AddChannel(),
ToTensor()
])
# Define transforms for testing
self._test_image_transforms = Compose([
LoadImage(image_only=True),
ScaleIntensityRange(0, 65535, 0.0, 1.0, clip=False),
AddChannel(),
ToTensor()
])
self._test_target_transforms = Compose([
LoadImage(image_only=True),
ScaleIntensityRange(0, 65535, 0.0, 1.0, clip=False),
AddChannel(),
ToTensor()
])
# Define transforms for prediction
self._prediction_image_transforms = Compose(
[LoadImage(image_only=True),
AddChannel(),
ToTensor()])
# Post transforms
self._validation_post_transforms = Compose([Identity()])
self._test_post_transforms = Compose(
[ToNumpy(), ScaleIntensity(0, 65535)])
self._prediction_post_transforms = Compose(
[ToNumpy(), ScaleIntensity(0, 65535)])
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
train_imtrans = Compose([
ScaleIntensity(),
AddChannel(),
RandSpatialCrop((96, 96, 96), random_size=False),
ToTensor()
])
train_segtrans = Compose([
AddChannel(),
RandSpatialCrop((96, 96, 96), random_size=False),
ToTensor()
])
val_imtrans = Compose(
[ScaleIntensity(),
AddChannel(),
Resize((96, 96, 96)),
ToTensor()])
val_segtrans = Compose([AddChannel(), Resize((96, 96, 96)), ToTensor()])
# define nifti dataset, data loader
check_ds = NiftiDataset(images,
def __init__(self,
data: str,
split: str,
extension: str,
classes: int,
column: str,
regression: bool,
debug: bool = False):
if regression and classes != 1:
print('Support for multi-class regression is not available.')
sys.exit(1)
self.datapath = wsl_data_dir / data
self.data = data
self.classes = classes
if data in known_extensions.keys():
self.extension = known_extensions[data]
else:
self.extension = extension
df = pd.read_csv(wsl_csv_dir / data / 'info.csv',
converters={
column: literal_eval,
'box': literal_eval
})
self.df = df
df = df.drop_duplicates(subset='Id', keep='first', ignore_index=True)
Ids = pd.read_csv(wsl_csv_dir / data / f'{split}.csv').Id.tolist()
df = df[df.Id.isin(Ids)]
self.names = df.Id.to_list()
self.labels = df[column].tolist()
if debug:
self.names = self.names[0:100]
self.labels = self.labels[0:100]
self.image_transforms = Compose([
Resize((224, 224)),
RepeatChannel(repeats=3),
CastToType(dtype=np.float32),
ToTensor()
])
if regression:
self.lmax = df[column].max()
self.lmin = df[column].min()
self.labels = [[round((x - self.lmin) / self.lmax, 2)]
for x in self.labels]
else:
if classes == 1:
self.labels = [[x] for x in self.labels]
else:
self.class_names = self.labels[0].keys()
self.labels = [list(x.values()) for x in self.labels]
self.pos_weight = [
round((len(col) - sum(col)) / sum(col), 2)
for col in zip(*self.labels)
]
])))
TESTS.append(
(list, pad_collate, RandSpatialCrop(roi_size=[8, 7],
random_size=True)))
TESTS.append((list, pad_collate,
RandRotate(prob=1,
range_x=np.pi,
keep_size=False,
dtype=np.float64)))
TESTS.append((list, pad_collate,
RandZoom(prob=1, min_zoom=1.1, max_zoom=2.0,
keep_size=False)))
TESTS.append((list, pad_collate,
Compose([RandRotate90(prob=1, max_k=2),
ToTensor()])))
class _Dataset(torch.utils.data.Dataset):
def __init__(self, images, labels, transforms):
self.images = images
self.labels = labels
self.transform = transforms
def __len__(self):
return len(self.images)
def __getitem__(self, index):
return self.transform(self.images[index]), self.labels[index]
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.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")))
# define transforms for image and segmentation
imtrans = Compose([
LoadImage(image_only=True),
ScaleIntensity(),
AddChannel(),
ToTensor()
])
segtrans = Compose([LoadImage(image_only=True), AddChannel(), ToTensor()])
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")
post_trans = Compose(
[Activations(sigmoid=True),
AsDiscrete(threshold_values=True)])
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_array.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[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(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.to(dtype=torch.int))
metric = metric_sum / metric_count
print("evaluation metric:", metric)
def main(cfg):
# -------------------------------------------------------------------------
# Configs
# -------------------------------------------------------------------------
# Create log/model dir
log_dir = create_log_dir(cfg)
# Set the logger
logging.basicConfig(
format="%(asctime)s %(levelname)2s %(message)s",
level=logging.INFO,
datefmt="%Y-%m-%d %H:%M:%S",
)
log_name = os.path.join(log_dir, "logs.txt")
logger = logging.getLogger()
fh = logging.FileHandler(log_name)
fh.setLevel(logging.INFO)
logger.addHandler(fh)
# Set TensorBoard summary writter
writer = SummaryWriter(log_dir)
# Save configs
logging.info(json.dumps(cfg))
with open(os.path.join(log_dir, "config.json"), "w") as fp:
json.dump(cfg, fp, indent=4)
# Set device cuda/cpu
device = set_device(cfg)
# Set cudnn benchmark/deterministic
if cfg["benchmark"]:
torch.backends.cudnn.benchmark = True
else:
set_determinism(seed=0)
# -------------------------------------------------------------------------
# Transforms and Datasets
# -------------------------------------------------------------------------
# Pre-processing
preprocess_cpu_train = None
preprocess_gpu_train = None
preprocess_cpu_valid = None
preprocess_gpu_valid = None
if cfg["backend"] == "cucim":
preprocess_cpu_train = Compose([ToTensorD(keys="label")])
preprocess_gpu_train = Compose([
ToCupy(),
RandCuCIM(name="color_jitter",
brightness=64.0 / 255.0,
contrast=0.75,
saturation=0.25,
hue=0.04),
RandCuCIM(name="image_flip",
apply_prob=cfg["prob"],
spatial_axis=-1),
RandCuCIM(name="rand_image_rotate_90",
prob=cfg["prob"],
max_k=3,
spatial_axis=(-2, -1)),
CastToType(dtype=np.float32),
RandCuCIM(name="rand_zoom", min_zoom=0.9, max_zoom=1.1),
CuCIM(name="scale_intensity_range",
a_min=0.0,
a_max=255.0,
b_min=-1.0,
b_max=1.0),
ToTensor(device=device),
])
preprocess_cpu_valid = Compose([ToTensorD(keys="label")])
preprocess_gpu_valid = Compose([
ToCupy(dtype=np.float32),
CuCIM(name="scale_intensity_range",
a_min=0.0,
a_max=255.0,
b_min=-1.0,
b_max=1.0),
ToTensor(device=device),
])
elif cfg["backend"] == "numpy":
preprocess_cpu_train = Compose([
ToTensorD(keys=("image", "label")),
TorchVisionD(keys="image",
name="ColorJitter",
brightness=64.0 / 255.0,
contrast=0.75,
saturation=0.25,
hue=0.04),
ToNumpyD(keys="image"),
RandFlipD(keys="image", prob=cfg["prob"], spatial_axis=-1),
RandRotate90D(keys="image", prob=cfg["prob"]),
CastToTypeD(keys="image", dtype=np.float32),
RandZoomD(keys="image",
prob=cfg["prob"],
min_zoom=0.9,
max_zoom=1.1),
ScaleIntensityRangeD(keys="image",
a_min=0.0,
a_max=255.0,
b_min=-1.0,
b_max=1.0),
ToTensorD(keys="image"),
])
preprocess_cpu_valid = Compose([
CastToTypeD(keys="image", dtype=np.float32),
ScaleIntensityRangeD(keys="image",
a_min=0.0,
a_max=255.0,
b_min=-1.0,
b_max=1.0),
ToTensorD(keys=("image", "label")),
])
else:
raise ValueError(
f"Backend should be either numpy or cucim! ['{cfg['backend']}' is provided.]"
)
# Post-processing
postprocess = Compose([
Activations(sigmoid=True),
AsDiscrete(threshold=0.5),
])
# Create MONAI dataset
train_json_info_list = load_decathlon_datalist(
data_list_file_path=cfg["dataset_json"],
data_list_key="training",
base_dir=cfg["data_root"],
)
valid_json_info_list = load_decathlon_datalist(
data_list_file_path=cfg["dataset_json"],
data_list_key="validation",
base_dir=cfg["data_root"],
)
train_dataset = PatchWSIDataset(
data=train_json_info_list,
region_size=cfg["region_size"],
grid_shape=cfg["grid_shape"],
patch_size=cfg["patch_size"],
transform=preprocess_cpu_train,
image_reader_name="openslide" if cfg["use_openslide"] else "cuCIM",
)
valid_dataset = PatchWSIDataset(
data=valid_json_info_list,
region_size=cfg["region_size"],
grid_shape=cfg["grid_shape"],
patch_size=cfg["patch_size"],
transform=preprocess_cpu_valid,
image_reader_name="openslide" if cfg["use_openslide"] else "cuCIM",
)
# DataLoaders
train_dataloader = DataLoader(train_dataset,
num_workers=cfg["num_workers"],
batch_size=cfg["batch_size"],
pin_memory=cfg["pin"])
valid_dataloader = DataLoader(valid_dataset,
num_workers=cfg["num_workers"],
batch_size=cfg["batch_size"],
pin_memory=cfg["pin"])
# Get sample batch and some info
first_sample = first(train_dataloader)
if first_sample is None:
raise ValueError("First sample is None!")
for d in ["image", "label"]:
logging.info(f"[{d}] \n"
f" {d} shape: {first_sample[d].shape}\n"
f" {d} type: {type(first_sample[d])}\n"
f" {d} dtype: {first_sample[d].dtype}")
logging.info(f"Batch size: {cfg['batch_size']}")
logging.info(f"[Training] number of batches: {len(train_dataloader)}")
logging.info(f"[Validation] number of batches: {len(valid_dataloader)}")
# -------------------------------------------------------------------------
# Deep Learning Model and Configurations
# -------------------------------------------------------------------------
# Initialize model
model = TorchVisionFCModel("resnet18",
n_classes=1,
use_conv=True,
pretrained=cfg["pretrain"])
model = model.to(device)
# Loss function
loss_func = torch.nn.BCEWithLogitsLoss()
loss_func = loss_func.to(device)
# Optimizer
if cfg["novograd"] is True:
optimizer = Novograd(model.parameters(), lr=cfg["lr"])
else:
optimizer = SGD(model.parameters(), lr=cfg["lr"], momentum=0.9)
# AMP scaler
cfg["amp"] = cfg["amp"] and monai.utils.get_torch_version_tuple() >= (1, 6)
if cfg["amp"] is True:
scaler = GradScaler()
else:
scaler = None
# Learning rate scheduler
if cfg["cos"] is True:
scheduler = lr_scheduler.CosineAnnealingLR(optimizer,
T_max=cfg["n_epochs"])
else:
scheduler = None
# -------------------------------------------------------------------------
# Training/Evaluating
# -------------------------------------------------------------------------
train_counter = {"n_epochs": cfg["n_epochs"], "epoch": 1, "step": 1}
total_valid_time, total_train_time = 0.0, 0.0
t_start = time.perf_counter()
metric_summary = {"loss": np.Inf, "accuracy": 0, "best_epoch": 1}
# Training/Validation Loop
for _ in range(cfg["n_epochs"]):
t_epoch = time.perf_counter()
logging.info(
f"[Training] learning rate: {optimizer.param_groups[0]['lr']}")
# Training
train_counter = training(
train_counter,
model,
loss_func,
optimizer,
scaler,
cfg["amp"],
train_dataloader,
preprocess_gpu_train,
postprocess,
device,
writer,
cfg["print_step"],
)
if scheduler is not None:
scheduler.step()
if cfg["save"]:
torch.save(
model.state_dict(),
os.path.join(log_dir,
f"model_epoch_{train_counter['epoch']}.pt"))
t_train = time.perf_counter()
train_time = t_train - t_epoch
total_train_time += train_time
# Validation
if cfg["validate"]:
valid_loss, valid_acc = validation(
model,
loss_func,
cfg["amp"],
valid_dataloader,
preprocess_gpu_valid,
postprocess,
device,
cfg["print_step"],
)
t_valid = time.perf_counter()
valid_time = t_valid - t_train
total_valid_time += valid_time
if valid_loss < metric_summary["loss"]:
metric_summary["loss"] = min(valid_loss,
metric_summary["loss"])
metric_summary["accuracy"] = max(valid_acc,
metric_summary["accuracy"])
metric_summary["best_epoch"] = train_counter["epoch"]
writer.add_scalar("valid/loss", valid_loss, train_counter["epoch"])
writer.add_scalar("valid/accuracy", valid_acc,
train_counter["epoch"])
logging.info(
f"[Epoch: {train_counter['epoch']}/{cfg['n_epochs']}] loss: {valid_loss:.3f}, accuracy: {valid_acc:.2f}, "
f"time: {t_valid - t_epoch:.1f}s (train: {train_time:.1f}s, valid: {valid_time:.1f}s)"
)
else:
logging.info(
f"[Epoch: {train_counter['epoch']}/{cfg['n_epochs']}] Train time: {train_time:.1f}s"
)
writer.flush()
t_end = time.perf_counter()
# Save final metrics
metric_summary["train_time_per_epoch"] = total_train_time / cfg["n_epochs"]
metric_summary["total_time"] = t_end - t_start
writer.add_hparams(hparam_dict=cfg,
metric_dict=metric_summary,
run_name=log_dir)
writer.close()
logging.info(f"Metric Summary: {metric_summary}")
# Save the best and final model
if cfg["validate"] is True:
copyfile(
os.path.join(log_dir,
f"model_epoch_{metric_summary['best_epoch']}.pth"),
os.path.join(log_dir, "model_best.pth"),
)
copyfile(
os.path.join(log_dir, f"model_epoch_{cfg['n_epochs']}.pth"),
os.path.join(log_dir, "model_final.pth"),
)
# Final prints
logging.info(
f"[Completed] {train_counter['epoch']} epochs -- time: {t_end - t_start:.1f}s "
f"(training: {total_train_time:.1f}s, validation: {total_valid_time:.1f}s)",
)
logging.info(f"Logs and model was saved at: {log_dir}")
def _define_transforms(self):
"""Define and initialize all data transforms.
* training set images transform
* training set masks transform
* validation set images transform
* validation set masks transform
* validation set images post-transform
* test set images transform
* test set masks transform
* test set images post-transform
* prediction set images transform
* prediction set images post-transform
@return True if data transforms could be instantiated, False otherwise.
"""
if self._mask_type == MaskType.UNKNOWN:
raise Exception("The mask type is unknown. Cannot continue!")
# Depending on the mask type, we will need to adapt the Mask Loader
# and Transform. We start by initializing the most common types.
MaskLoader = LoadMask(self._mask_type)
MaskTransform = Identity
# Adapt the transform for the LABEL types
if self._mask_type == MaskType.TIFF_LABELS or self._mask_type == MaskType.NUMPY_LABELS:
MaskTransform = ToOneHot(num_classes=self._out_channels)
# The H5_ONE_HOT type requires a different loader
if self._mask_type == MaskType.H5_ONE_HOT:
# MaskLoader: still missing
raise Exception("HDF5 one-hot masks are not supported yet!")
# Define transforms for training
self._train_image_transforms = Compose([
LoadImage(image_only=True),
ScaleIntensity(),
AddChannel(),
RandSpatialCrop(self._roi_size, random_size=False),
RandRotate90(prob=0.5, spatial_axes=(0, 1)),
ToTensor()
])
self._train_mask_transforms = Compose([
MaskLoader, MaskTransform,
RandSpatialCrop(self._roi_size, random_size=False),
RandRotate90(prob=0.5, spatial_axes=(0, 1)),
ToTensor()
])
# Define transforms for validation
self._validation_image_transforms = Compose([
LoadImage(image_only=True),
ScaleIntensity(),
AddChannel(),
ToTensor()
])
self._validation_mask_transforms = Compose(
[MaskLoader, MaskTransform, ToTensor()])
# Define transforms for testing
self._test_image_transforms = Compose([
LoadImage(image_only=True),
ScaleIntensity(),
AddChannel(),
ToTensor()
])
self._test_mask_transforms = Compose(
[MaskLoader, MaskTransform, ToTensor()])
# Define transforms for prediction
self._prediction_image_transforms = Compose([
LoadImage(image_only=True),
ScaleIntensity(),
AddChannel(),
ToTensor()
])
# Post transforms
self._validation_post_transforms = Compose(
[Activations(softmax=True),
AsDiscrete(threshold_values=True)])
self._test_post_transforms = Compose(
[Activations(softmax=True),
AsDiscrete(threshold_values=True)])
self._prediction_post_transforms = Compose(
[Activations(softmax=True),
AsDiscrete(threshold_values=True)])
"/workspace/data/medical/ixi/IXI-T1/IXI584-Guys-1129-T1.nii.gz",
"/workspace/data/medical/ixi/IXI-T1/IXI253-HH-1694-T1.nii.gz",
"/workspace/data/medical/ixi/IXI-T1/IXI092-HH-1436-T1.nii.gz",
"/workspace/data/medical/ixi/IXI-T1/IXI574-IOP-1156-T1.nii.gz",
"/workspace/data/medical/ixi/IXI-T1/IXI585-Guys-1130-T1.nii.gz"
]
# 2 binary labels for gender classification: man and woman
labels = np.array([0, 0, 0, 1, 0, 0, 0, 1, 1, 0, 0, 0, 1, 0, 1, 0, 1, 0, 1, 0])
# Define transforms
train_transforms = Compose([
ScaleIntensity(),
AddChannel(),
Resize((96, 96, 96)),
RandRotate90(),
ToTensor()
])
val_transforms = Compose(
[ScaleIntensity(),
AddChannel(),
Resize((96, 96, 96)),
ToTensor()])
# Define nifti dataset, data loader
check_ds = NiftiDataset(image_files=images,
labels=labels,
transform=train_transforms)
check_loader = DataLoader(check_ds,
batch_size=2,
num_workers=2,
pin_memory=torch.cuda.is_available())
def main():
monai.config.print_config()
logging.basicConfig(stream=sys.stdout, level=logging.INFO)
# IXI dataset as a demo, downloadable from https://brain-development.org/ixi-dataset/
images = [
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI607-Guys-1097-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI175-HH-1570-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI385-HH-2078-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI344-Guys-0905-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI409-Guys-0960-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI584-Guys-1129-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI253-HH-1694-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI092-HH-1436-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI574-IOP-1156-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI585-Guys-1130-T1.nii.gz"
]),
]
# 2 binary labels for gender classification: man and woman
labels = np.array([0, 0, 1, 0, 1, 0, 1, 0, 1, 0], dtype=np.int64)
# Define transforms for image
val_transforms = Compose(
[ScaleIntensity(),
AddChannel(),
Resize((96, 96, 96)),
ToTensor()])
# Define nifti dataset
val_ds = NiftiDataset(image_files=images,
labels=labels,
transform=val_transforms,
image_only=False)
# create a validation data loader
val_loader = DataLoader(val_ds,
batch_size=2,
num_workers=4,
pin_memory=torch.cuda.is_available())
# Create DenseNet121
device = torch.device("cuda:0")
model = monai.networks.nets.densenet.densenet121(spatial_dims=3,
in_channels=1,
out_channels=2).to(device)
model.load_state_dict(torch.load("best_metric_model.pth"))
model.eval()
with torch.no_grad():
num_correct = 0.0
metric_count = 0
saver = CSVSaver(output_dir="./output")
for val_data in val_loader:
val_images, val_labels = val_data[0].to(device), val_data[1].to(
device)
val_outputs = model(val_images).argmax(dim=1)
value = torch.eq(val_outputs, val_labels)
metric_count += len(value)
num_correct += value.sum().item()
saver.save_batch(val_outputs, val_data[2])
metric = num_correct / metric_count
print("evaluation metric:", metric)
saver.finalize()
def run_test(batch_size=64, train_steps=200, device="cuda:0"):
class _TestBatch(Dataset):
def __init__(self, transforms):
self.transforms = transforms
def __getitem__(self, _unused_id):
im, seg = create_test_image_2d(128, 128, noise_max=1, num_objs=4, num_seg_classes=1)
seed = np.random.randint(2147483647)
self.transforms.set_random_state(seed=seed)
im = self.transforms(im)
self.transforms.set_random_state(seed=seed)
seg = self.transforms(seg)
return im, seg
def __len__(self):
return train_steps
net = UNet(
spatial_dims=2, in_channels=1, out_channels=1, channels=(4, 8, 16, 32), strides=(2, 2, 2), num_res_units=2
).to(device)
loss = DiceLoss(sigmoid=True)
opt = torch.optim.Adam(net.parameters(), 1e-2)
train_transforms = Compose(
[AddChannel(), ScaleIntensity(), RandSpatialCrop((96, 96), random_size=False), RandRotate90(), ToTensor()]
)
src = DataLoader(_TestBatch(train_transforms), batch_size=batch_size, shuffle=True)
net.train()
epoch_loss = 0
step = 0
for img, seg in src:
step += 1
opt.zero_grad()
output = net(img.to(device))
step_loss = loss(output, seg.to(device))
step_loss.backward()
opt.step()
epoch_loss += step_loss.item()
epoch_loss /= step
return epoch_loss, step
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
train_imtrans = Compose([
ScaleIntensity(),
AddChannel(),
RandSpatialCrop((96, 96, 96), random_size=False),
RandRotate90(prob=0.5, spatial_axes=(0, 2)),
ToTensor()
])
train_segtrans = Compose([
AddChannel(),
RandSpatialCrop((96, 96, 96), random_size=False),
RandRotate90(prob=0.5, spatial_axes=(0, 2)),
ToTensor()
])
val_imtrans = Compose([ScaleIntensity(), AddChannel(), ToTensor()])
val_segtrans = Compose([AddChannel(), ToTensor()])
# define nifti dataset, data loader
check_ds = NiftiDataset(images,
segs,
transform=train_imtrans,
seg_transform=train_segtrans)
def test_single_input(self, test_data):
result = ToTensor()(test_data)
self.assertTrue(isinstance(result, torch.Tensor))
assert_allclose(result, test_data, type_test=False)
self.assertEqual(result.ndim, 0)
def run_inference_test(root_dir, 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([
LoadImaged(keys=["img", "seg"]),
EnsureChannelFirstd(keys=["img", "seg"]),
# 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"),
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 = monai.data.DataLoader(val_ds, batch_size=1, num_workers=4)
val_post_tran = Compose([
ToTensor(),
Activations(sigmoid=True),
AsDiscrete(threshold_values=True)
])
dice_metric = DiceMetric(include_background=True,
reduction="mean",
get_not_nans=False)
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(device)
model_filename = os.path.join(root_dir, "best_metric_model.pth")
model.load_state_dict(torch.load(model_filename))
with eval_mode(model):
# 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)
# decollate prediction into a list and execute post processing for every item
val_outputs = [
val_post_tran(i) for i in decollate_batch(val_outputs)
]
# compute metrics
dice_metric(y_pred=val_outputs, y=val_labels)
saver.save_batch(val_outputs, val_data["img_meta_dict"])
return dice_metric.aggregate().item()
if ifmode == 'train': ## for train mode
print('Training start ...')
# 자유롭게 작성
images, labels = DataLoad(imdir=os.path.join(DATASET_PATH, 'train'))
images = ImagePreprocessing(images)
images = np.array(images)
labels = np.array(labels)
## Define transforms
train_transforms = Compose([
AddChannel(),
ScaleIntensity(),
RandRotate(degrees=180, prob=0.5, reshape=False),
RandFlip(spatial_axis=0, prob=0.5),
RandFlip(spatial_axis=1, prob=0.5),
ToTensor()
])
#RandZoom(min_zoom=0.9, max_zoom=1.1, prob=0.5, keep_size=True),
val_transforms = Compose([AddChannel(), ScaleIntensity(), ToTensor()])
# Split data
x_train, y_train, x_test, y_test = split_dataset(images,
labels,
valid_frac=VAL_RATIO)
# Clone training data
x_train, y_train = clone_dataset(x_train, y_train)
# dataset = PNSDataset(torch.from_numpy(images).float(), torch.from_numpy(labels).long())
# subset_size = [len(images) - int(len(images) * VAL_RATIO),int(len(images) * VAL_RATIO)]
def run_training_test(root_dir,
device="cuda:0",
cachedataset=0,
readers=(None, None)):
monai.config.print_config()
images = sorted(glob(os.path.join(root_dir, "img*.nii.gz")))
segs = sorted(glob(os.path.join(root_dir, "seg*.nii.gz")))
train_files = [{
"img": img,
"seg": seg
} for img, seg in zip(images[:20], segs[:20])]
val_files = [{
"img": img,
"seg": seg
} for img, seg in zip(images[-20:], segs[-20:])]
# define transforms for image and segmentation
train_transforms = Compose([
LoadImaged(keys=["img", "seg"], reader=readers[0]),
EnsureChannelFirstd(keys=["img", "seg"]),
# 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"),
RandCropByPosNegLabeld(keys=["img", "seg"],
label_key="seg",
spatial_size=[96, 96, 96],
pos=1,
neg=1,
num_samples=4),
RandRotate90d(keys=["img", "seg"], prob=0.8, spatial_axes=[0, 2]),
ToTensord(keys=["img", "seg"]),
])
train_transforms.set_random_state(1234)
val_transforms = Compose([
LoadImaged(keys=["img", "seg"], reader=readers[1]),
EnsureChannelFirstd(keys=["img", "seg"]),
# 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"),
ToTensord(keys=["img", "seg"]),
])
# create a training data loader
if cachedataset == 2:
train_ds = monai.data.CacheDataset(data=train_files,
transform=train_transforms,
cache_rate=0.8)
elif cachedataset == 3:
train_ds = monai.data.LMDBDataset(data=train_files,
transform=train_transforms,
cache_dir=root_dir)
else:
train_ds = monai.data.Dataset(data=train_files,
transform=train_transforms)
# use batch_size=2 to load images and use RandCropByPosNegLabeld to generate 2 x 4 images for network training
train_loader = monai.data.DataLoader(train_ds,
batch_size=2,
shuffle=True,
num_workers=4)
# create a validation data loader
val_ds = monai.data.Dataset(data=val_files, transform=val_transforms)
val_loader = monai.data.DataLoader(val_ds, batch_size=1, num_workers=4)
val_post_tran = Compose([
ToTensor(),
Activations(sigmoid=True),
AsDiscrete(threshold_values=True)
])
dice_metric = DiceMetric(include_background=True,
reduction="mean",
get_not_nans=False)
# create UNet, DiceLoss and Adam optimizer
model = monai.networks.nets.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)
loss_function = monai.losses.DiceLoss(sigmoid=True)
optimizer = torch.optim.Adam(model.parameters(), 5e-4)
# start a typical PyTorch training
val_interval = 2
best_metric, best_metric_epoch = -1, -1
epoch_loss_values = []
metric_values = []
writer = SummaryWriter(log_dir=os.path.join(root_dir, "runs"))
model_filename = os.path.join(root_dir, "best_metric_model.pth")
for epoch in range(6):
print("-" * 10)
print(f"Epoch {epoch + 1}/{6}")
model.train()
epoch_loss = 0
step = 0
for batch_data in train_loader:
step += 1
inputs, labels = batch_data["img"].to(
device), batch_data["seg"].to(device)
optimizer.zero_grad()
outputs = model(inputs)
loss = loss_function(outputs, labels)
loss.backward()
optimizer.step()
epoch_loss += loss.item()
epoch_len = len(train_ds) // train_loader.batch_size
print(f"{step}/{epoch_len}, train_loss:{loss.item():0.4f}")
writer.add_scalar("train_loss", loss.item(),
epoch_len * epoch + step)
epoch_loss /= step
epoch_loss_values.append(epoch_loss)
print(f"epoch {epoch +1} average loss:{epoch_loss:0.4f}")
if (epoch + 1) % val_interval == 0:
with eval_mode(model):
val_images = None
val_labels = None
val_outputs = None
for val_data in val_loader:
val_images, val_labels = val_data["img"].to(
device), val_data["seg"].to(device)
sw_batch_size, roi_size = 4, (96, 96, 96)
val_outputs = sliding_window_inference(
val_images, roi_size, sw_batch_size, model)
# decollate prediction into a list and execute post processing for every item
val_outputs = [
val_post_tran(i) for i in decollate_batch(val_outputs)
]
# compute metrics
dice_metric(y_pred=val_outputs, y=val_labels)
metric = dice_metric.aggregate().item()
dice_metric.reset()
metric_values.append(metric)
if metric > best_metric:
best_metric = metric
best_metric_epoch = epoch + 1
torch.save(model.state_dict(), model_filename)
print("saved new best metric model")
print(
f"current epoch {epoch +1} current mean dice: {metric:0.4f} "
f"best mean dice: {best_metric:0.4f} at epoch {best_metric_epoch}"
)
writer.add_scalar("val_mean_dice", metric, epoch + 1)
# plot the last model output as GIF image in TensorBoard with the corresponding image and label
plot_2d_or_3d_image(val_images,
epoch + 1,
writer,
index=0,
tag="image")
plot_2d_or_3d_image(val_labels,
epoch + 1,
writer,
index=0,
tag="label")
plot_2d_or_3d_image(val_outputs,
epoch + 1,
writer,
index=0,
tag="output")
print(
f"train completed, best_metric: {best_metric:0.4f} at epoch: {best_metric_epoch}"
)
writer.close()
return epoch_loss_values, best_metric, best_metric_epoch
def main():
monai.config.print_config()
logging.basicConfig(stream=sys.stdout, level=logging.INFO)
# IXI dataset as a demo, downloadable from https://brain-development.org/ixi-dataset/
images = [
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI314-IOP-0889-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI249-Guys-1072-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI609-HH-2600-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI173-HH-1590-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI020-Guys-0700-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI342-Guys-0909-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI134-Guys-0780-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI577-HH-2661-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI066-Guys-0731-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI130-HH-1528-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI607-Guys-1097-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI175-HH-1570-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI385-HH-2078-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI344-Guys-0905-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI409-Guys-0960-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI584-Guys-1129-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI253-HH-1694-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI092-HH-1436-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI574-IOP-1156-T1.nii.gz"
]),
os.sep.join([
"workspace", "data", "medical", "ixi", "IXI-T1",
"IXI585-Guys-1130-T1.nii.gz"
]),
]
# 2 binary labels for gender classification: man and woman
labels = np.array(
[0, 0, 0, 1, 0, 0, 0, 1, 1, 0, 0, 0, 1, 0, 1, 0, 1, 0, 1, 0],
dtype=np.int64)
# define transforms
train_transforms = Compose([
ScaleIntensity(),
AddChannel(),
Resize((96, 96, 96)),
RandRotate90(),
ToTensor()
])
val_transforms = Compose(
[ScaleIntensity(),
AddChannel(),
Resize((96, 96, 96)),
ToTensor()])
# define nifti dataset, data loader
check_ds = NiftiDataset(image_files=images,
labels=labels,
transform=train_transforms)
check_loader = DataLoader(check_ds,
batch_size=2,
num_workers=2,
pin_memory=torch.cuda.is_available())
im, label = monai.utils.misc.first(check_loader)
print(type(im), im.shape, label)
# create DenseNet121, CrossEntropyLoss and Adam optimizer
net = monai.networks.nets.densenet.densenet121(spatial_dims=3,
in_channels=1,
out_channels=2)
loss = torch.nn.CrossEntropyLoss()
lr = 1e-5
opt = torch.optim.Adam(net.parameters(), lr)
device = torch.device("cuda:0")
# Ignite trainer expects batch=(img, label) and returns output=loss at every iteration,
# user can add output_transform to return other values, like: y_pred, y, etc.
trainer = create_supervised_trainer(net, opt, loss, device, False)
# adding checkpoint handler to save models (network params and optimizer stats) during training
checkpoint_handler = ModelCheckpoint("./runs/",
"net",
n_saved=10,
require_empty=False)
trainer.add_event_handler(event_name=Events.EPOCH_COMPLETED,
handler=checkpoint_handler,
to_save={
"net": net,
"opt": opt
})
# StatsHandler prints loss at every iteration and print metrics at every epoch,
# we don't set metrics for trainer here, so just print loss, user can also customize print functions
# and can use output_transform to convert engine.state.output if it's not loss value
train_stats_handler = StatsHandler(name="trainer")
train_stats_handler.attach(trainer)
# TensorBoardStatsHandler plots loss at every iteration and plots metrics at every epoch, same as StatsHandler
train_tensorboard_stats_handler = TensorBoardStatsHandler()
train_tensorboard_stats_handler.attach(trainer)
# set parameters for validation
validation_every_n_epochs = 1
metric_name = "Accuracy"
# add evaluation metric to the evaluator engine
val_metrics = {metric_name: Accuracy()}
# Ignite evaluator expects batch=(img, label) and returns output=(y_pred, y) at every iteration,
# user can add output_transform to return other values
evaluator = create_supervised_evaluator(net, val_metrics, device, True)
# add stats event handler to print validation stats via evaluator
val_stats_handler = StatsHandler(
name="evaluator",
output_transform=lambda x:
None, # no need to print loss value, so disable per iteration output
global_epoch_transform=lambda x: trainer.state.epoch,
) # fetch global epoch number from trainer
val_stats_handler.attach(evaluator)
# add handler to record metrics to TensorBoard at every epoch
val_tensorboard_stats_handler = TensorBoardStatsHandler(
output_transform=lambda x:
None, # no need to plot loss value, so disable per iteration output
global_epoch_transform=lambda x: trainer.state.epoch,
) # fetch global epoch number from trainer
val_tensorboard_stats_handler.attach(evaluator)
# add early stopping handler to evaluator
early_stopper = EarlyStopping(
patience=4,
score_function=stopping_fn_from_metric(metric_name),
trainer=trainer)
evaluator.add_event_handler(event_name=Events.EPOCH_COMPLETED,
handler=early_stopper)
# create a validation data loader
val_ds = NiftiDataset(image_files=images[-10:],
labels=labels[-10:],
transform=val_transforms)
val_loader = DataLoader(val_ds,
batch_size=2,
num_workers=2,
pin_memory=torch.cuda.is_available())
@trainer.on(Events.EPOCH_COMPLETED(every=validation_every_n_epochs))
def run_validation(engine):
evaluator.run(val_loader)
# create a training data loader
train_ds = NiftiDataset(image_files=images[:10],
labels=labels[:10],
transform=train_transforms)
train_loader = DataLoader(train_ds,
batch_size=2,
shuffle=True,
num_workers=2,
pin_memory=torch.cuda.is_available())
train_epochs = 30
state = trainer.run(train_loader, train_epochs)
print(state)
def __init__(
self,
transform: InvertibleTransform,
loader: TorchDataLoader,
output_keys: Union[str, Sequence[str]] = CommonKeys.PRED,
batch_keys: Union[str, Sequence[str]] = CommonKeys.IMAGE,
meta_key_postfix: str = "meta_dict",
collate_fn: Optional[Callable] = no_collation,
postfix: str = "inverted",
nearest_interp: Union[bool, Sequence[bool]] = True,
to_tensor: Union[bool, Sequence[bool]] = True,
device: Union[Union[str, torch.device],
Sequence[Union[str, torch.device]]] = "cpu",
post_func: Union[Callable, Sequence[Callable]] = lambda x: x,
num_workers: Optional[int] = 0,
) -> None:
"""
Args:
transform: a callable data transform on input data.
loader: data loader used to run transforms and generate the batch of data.
output_keys: the key of expected data in `ignite.engine.output`, invert transforms on it.
it also can be a list of keys, will invert transform for each of them. Default to "pred".
batch_keys: the key of input data in `ignite.engine.batch`. will get the applied transforms
for this input data, then invert them for the expected data with `output_keys`.
It can also be a list of keys, each matches to the `output_keys` data. default to "image".
meta_key_postfix: use `{batch_key}_{postfix}` to to fetch the meta data according to the key data,
default is `meta_dict`, the meta data is a dictionary object.
For example, to handle key `image`, read/write affine matrices from the
metadata `image_meta_dict` dictionary's `affine` field.
collate_fn: how to collate data after inverse transformations.
default won't do any collation, so the output will be a list of size batch size.
postfix: will save the inverted result into `ignite.engine.output` with key `{output_key}_{postfix}`.
nearest_interp: whether to use `nearest` interpolation mode when inverting the spatial transforms,
default to `True`. If `False`, use the same interpolation mode as the original transform.
it also can be a list of bool, each matches to the `output_keys` data.
to_tensor: whether to convert the inverted data into PyTorch Tensor first, default to `True`.
it also can be a list of bool, each matches to the `output_keys` data.
device: if converted to Tensor, move the inverted results to target device before `post_func`,
default to "cpu", it also can be a list of string or `torch.device`,
each matches to the `output_keys` data.
post_func: post processing for the inverted data, should be a callable function.
it also can be a list of callable, each matches to the `output_keys` data.
num_workers: number of workers when run data loader for inverse transforms,
default to 0 as only run one iteration and multi-processing may be even slower.
Set to `None`, to use the `num_workers` of the input transform data loader.
"""
self.transform = transform
self.inverter = BatchInverseTransform(
transform=transform,
loader=loader,
collate_fn=collate_fn,
num_workers=num_workers,
)
self.output_keys = ensure_tuple(output_keys)
self.batch_keys = ensure_tuple_rep(batch_keys, len(self.output_keys))
self.meta_key_postfix = meta_key_postfix
self.postfix = postfix
self.nearest_interp = ensure_tuple_rep(nearest_interp,
len(self.output_keys))
self.to_tensor = ensure_tuple_rep(to_tensor, len(self.output_keys))
self.device = ensure_tuple_rep(device, len(self.output_keys))
self.post_func = ensure_tuple_rep(post_func, len(self.output_keys))
self._totensor = ToTensor()
def get_rsna_valid_aug(name=None, image_size=160):
return Compose([
ScaleIntensity(),
Resize((image_size, image_size, image_size)),
ToTensor()
])
def main():
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)
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()])
ds = NiftiDataset(images,
segs,
transform=imtrans,
seg_transform=segtrans,
image_only=False)
device = torch.device("cuda:0")
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,
)
net.to(device)
# define sliding window size and batch size for windows inference
roi_size = (96, 96, 96)
sw_batch_size = 4
def _sliding_window_processor(engine, batch):
net.eval()
with torch.no_grad():
val_images, val_labels = batch[0].to(device), batch[1].to(device)
seg_probs = sliding_window_inference(val_images, roi_size,
sw_batch_size, net)
return seg_probs, val_labels
evaluator = Engine(_sliding_window_processor)
# add evaluation metric to the evaluator engine
MeanDice(add_sigmoid=True,
to_onehot_y=False).attach(evaluator, "Mean_Dice")
# StatsHandler prints loss at every iteration and print metrics at every epoch,
# we don't need to print loss for evaluator, so just print metrics, user can also customize print functions
val_stats_handler = StatsHandler(
name="evaluator",
output_transform=lambda x:
None, # no need to print loss value, so disable per iteration output
)
val_stats_handler.attach(evaluator)
# for the array data format, assume the 3rd item of batch data is the meta_data
file_saver = SegmentationSaver(
output_dir="tempdir",
output_ext=".nii.gz",
output_postfix="seg",
name="evaluator",
batch_transform=lambda x: x[2],
output_transform=lambda output: predict_segmentation(output[0]),
)
file_saver.attach(evaluator)
# the model was trained by "unet_training_array" example
ckpt_saver = CheckpointLoader(load_path="./runs/net_checkpoint_100.pth",
load_dict={"net": net})
ckpt_saver.attach(evaluator)
# sliding window inference for one image at every iteration
loader = DataLoader(ds,
batch_size=1,
num_workers=1,
pin_memory=torch.cuda.is_available())
state = evaluator.run(loader)
print(state)
shutil.rmtree(tempdir)
def run_training_test(root_dir,
train_x,
train_y,
val_x,
val_y,
device=torch.device("cuda:0")):
monai.config.print_config()
# define transforms for image and classification
train_transforms = Compose([
LoadPNG(image_only=True),
AddChannel(),
ScaleIntensity(),
RandRotate(range_x=15, prob=0.5, keep_size=True),
RandFlip(spatial_axis=0, prob=0.5),
RandZoom(min_zoom=0.9, max_zoom=1.1, prob=0.5),
ToTensor(),
])
train_transforms.set_random_state(1234)
val_transforms = Compose(
[LoadPNG(image_only=True),
AddChannel(),
ScaleIntensity(),
ToTensor()])
# create train, val data loaders
train_ds = MedNISTDataset(train_x, train_y, train_transforms)
train_loader = DataLoader(train_ds,
batch_size=300,
shuffle=True,
num_workers=10)
val_ds = MedNISTDataset(val_x, val_y, val_transforms)
val_loader = DataLoader(val_ds, batch_size=300, num_workers=10)
model = densenet121(spatial_dims=2,
in_channels=1,
out_channels=len(np.unique(train_y))).to(device)
loss_function = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), 1e-5)
epoch_num = 4
val_interval = 1
# start training validation
best_metric = -1
best_metric_epoch = -1
epoch_loss_values = list()
metric_values = list()
model_filename = os.path.join(root_dir, "best_metric_model.pth")
for epoch in range(epoch_num):
print("-" * 10)
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[0].to(device), batch_data[1].to(device)
optimizer.zero_grad()
outputs = model(inputs)
loss = loss_function(outputs, labels)
loss.backward()
optimizer.step()
epoch_loss += loss.item()
epoch_loss /= step
epoch_loss_values.append(epoch_loss)
print(f"epoch {epoch + 1} average loss:{epoch_loss:0.4f}")
if (epoch + 1) % val_interval == 0:
model.eval()
with torch.no_grad():
y_pred = torch.tensor([], dtype=torch.float32, device=device)
y = torch.tensor([], dtype=torch.long, device=device)
for val_data in val_loader:
val_images, val_labels = val_data[0].to(
device), val_data[1].to(device)
y_pred = torch.cat([y_pred, model(val_images)], dim=0)
y = torch.cat([y, val_labels], dim=0)
auc_metric = compute_roc_auc(y_pred,
y,
to_onehot_y=True,
softmax=True)
metric_values.append(auc_metric)
acc_value = torch.eq(y_pred.argmax(dim=1), y)
acc_metric = acc_value.sum().item() / len(acc_value)
if auc_metric > best_metric:
best_metric = auc_metric
best_metric_epoch = epoch + 1
torch.save(model.state_dict(), model_filename)
print("saved new best metric model")
print(
f"current epoch {epoch +1} current AUC: {auc_metric:0.4f} "
f"current accuracy: {acc_metric:0.4f} best AUC: {best_metric:0.4f} at epoch {best_metric_epoch}"
)
print(
f"train completed, best_metric: {best_metric:0.4f} at epoch: {best_metric_epoch}"
)
return epoch_loss_values, best_metric, best_metric_epoch
def main():
monai.config.print_config()
logging.basicConfig(stream=sys.stdout, level=logging.INFO)
# IXI dataset as a demo, downloadable from https://brain-development.org/ixi-dataset/
images = [
"/workspace/data/medical/ixi/IXI-T1/IXI607-Guys-1097-T1.nii.gz",
"/workspace/data/medical/ixi/IXI-T1/IXI175-HH-1570-T1.nii.gz",
"/workspace/data/medical/ixi/IXI-T1/IXI385-HH-2078-T1.nii.gz",
"/workspace/data/medical/ixi/IXI-T1/IXI344-Guys-0905-T1.nii.gz",
"/workspace/data/medical/ixi/IXI-T1/IXI409-Guys-0960-T1.nii.gz",
"/workspace/data/medical/ixi/IXI-T1/IXI584-Guys-1129-T1.nii.gz",
"/workspace/data/medical/ixi/IXI-T1/IXI253-HH-1694-T1.nii.gz",
"/workspace/data/medical/ixi/IXI-T1/IXI092-HH-1436-T1.nii.gz",
"/workspace/data/medical/ixi/IXI-T1/IXI574-IOP-1156-T1.nii.gz",
"/workspace/data/medical/ixi/IXI-T1/IXI585-Guys-1130-T1.nii.gz",
]
# 2 binary labels for gender classification: man and woman
labels = np.array([0, 0, 1, 0, 1, 0, 1, 0, 1, 0])
# define transforms for image
val_transforms = Compose(
[ScaleIntensity(),
AddChannel(),
Resize((96, 96, 96)),
ToTensor()])
# define nifti dataset
val_ds = NiftiDataset(image_files=images,
labels=labels,
transform=val_transforms,
image_only=False)
# create DenseNet121
net = monai.networks.nets.densenet.densenet121(
spatial_dims=3,
in_channels=1,
out_channels=2,
)
device = torch.device("cuda:0")
metric_name = "Accuracy"
# add evaluation metric to the evaluator engine
val_metrics = {metric_name: Accuracy()}
def prepare_batch(batch, device=None, non_blocking=False):
return _prepare_batch((batch[0], batch[1]), device, non_blocking)
# Ignite evaluator expects batch=(img, label) and returns output=(y_pred, y) at every iteration,
# user can add output_transform to return other values
evaluator = create_supervised_evaluator(net,
val_metrics,
device,
True,
prepare_batch=prepare_batch)
# add stats event handler to print validation stats via evaluator
val_stats_handler = StatsHandler(
name="evaluator",
output_transform=lambda x:
None, # no need to print loss value, so disable per iteration output
)
val_stats_handler.attach(evaluator)
# for the array data format, assume the 3rd item of batch data is the meta_data
prediction_saver = ClassificationSaver(
output_dir="tempdir",
batch_transform=lambda batch: batch[2],
output_transform=lambda output: output[0].argmax(1),
)
prediction_saver.attach(evaluator)
# the model was trained by "densenet_training_array" example
CheckpointLoader(load_path="./runs/net_checkpoint_20.pth",
load_dict={
"net": net
}).attach(evaluator)
# create a validation data loader
val_loader = DataLoader(val_ds,
batch_size=2,
num_workers=4,
pin_memory=torch.cuda.is_available())
state = evaluator.run(val_loader)
def main():
monai.config.print_config()
logging.basicConfig(stream=sys.stdout, level=logging.INFO)
# create a temporary directory and 40 random image, mask paris
tempdir = tempfile.mkdtemp()
print(f"generating synthetic data to {tempdir} (this may take a while)")
for i in range(40):
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
train_imtrans = Compose([
ScaleIntensity(),
AddChannel(),
RandSpatialCrop((96, 96, 96), random_size=False),
ToTensor()
])
train_segtrans = Compose([
AddChannel(),
RandSpatialCrop((96, 96, 96), random_size=False),
ToTensor()
])
val_imtrans = Compose(
[ScaleIntensity(),
AddChannel(),
Resize((96, 96, 96)),
ToTensor()])
val_segtrans = Compose([AddChannel(), Resize((96, 96, 96)), ToTensor()])
# define nifti dataset, data loader
check_ds = NiftiDataset(images,
segs,
transform=train_imtrans,
seg_transform=train_segtrans)
check_loader = DataLoader(check_ds,
batch_size=10,
num_workers=2,
pin_memory=torch.cuda.is_available())
im, seg = monai.utils.misc.first(check_loader)
print(im.shape, seg.shape)
# create a training data loader
train_ds = NiftiDataset(images[:20],
segs[:20],
transform=train_imtrans,
seg_transform=train_segtrans)
train_loader = DataLoader(train_ds,
batch_size=5,
shuffle=True,
num_workers=8,
pin_memory=torch.cuda.is_available())
# create a validation data loader
val_ds = NiftiDataset(images[-20:],
segs[-20:],
transform=val_imtrans,
seg_transform=val_segtrans)
val_loader = DataLoader(val_ds,
batch_size=5,
num_workers=8,
pin_memory=torch.cuda.is_available())
# create UNet, DiceLoss and Adam optimizer
net = monai.networks.nets.UNet(
dimensions=3,
in_channels=1,
out_channels=1,
channels=(16, 32, 64, 128, 256),
strides=(2, 2, 2, 2),
num_res_units=2,
)
loss = monai.losses.DiceLoss(sigmoid=True)
lr = 1e-3
opt = torch.optim.Adam(net.parameters(), lr)
device = torch.device("cuda:0")
# Ignite trainer expects batch=(img, seg) and returns output=loss at every iteration,
# user can add output_transform to return other values, like: y_pred, y, etc.
trainer = create_supervised_trainer(net, opt, loss, device, False)
# adding checkpoint handler to save models (network params and optimizer stats) during training
checkpoint_handler = ModelCheckpoint("./runs/",
"net",
n_saved=10,
require_empty=False)
trainer.add_event_handler(event_name=Events.EPOCH_COMPLETED,
handler=checkpoint_handler,
to_save={
"net": net,
"opt": opt
})
# StatsHandler prints loss at every iteration and print metrics at every epoch,
# we don't set metrics for trainer here, so just print loss, user can also customize print functions
# and can use output_transform to convert engine.state.output if it's not a loss value
train_stats_handler = StatsHandler(name="trainer")
train_stats_handler.attach(trainer)
# TensorBoardStatsHandler plots loss at every iteration and plots metrics at every epoch, same as StatsHandler
train_tensorboard_stats_handler = TensorBoardStatsHandler()
train_tensorboard_stats_handler.attach(trainer)
validation_every_n_epochs = 1
# Set parameters for validation
metric_name = "Mean_Dice"
# add evaluation metric to the evaluator engine
val_metrics = {metric_name: MeanDice(sigmoid=True, to_onehot_y=False)}
# Ignite evaluator expects batch=(img, seg) and returns output=(y_pred, y) at every iteration,
# user can add output_transform to return other values
evaluator = create_supervised_evaluator(net, val_metrics, device, True)
@trainer.on(Events.EPOCH_COMPLETED(every=validation_every_n_epochs))
def run_validation(engine):
evaluator.run(val_loader)
# add early stopping handler to evaluator
early_stopper = EarlyStopping(
patience=4,
score_function=stopping_fn_from_metric(metric_name),
trainer=trainer)
evaluator.add_event_handler(event_name=Events.EPOCH_COMPLETED,
handler=early_stopper)
# add stats event handler to print validation stats via evaluator
val_stats_handler = StatsHandler(
name="evaluator",
output_transform=lambda x:
None, # no need to print loss value, so disable per iteration output
global_epoch_transform=lambda x: trainer.state.epoch,
) # fetch global epoch number from trainer
val_stats_handler.attach(evaluator)
# add handler to record metrics to TensorBoard at every validation epoch
val_tensorboard_stats_handler = TensorBoardStatsHandler(
output_transform=lambda x:
None, # no need to plot loss value, so disable per iteration output
global_epoch_transform=lambda x: trainer.state.epoch,
) # fetch global epoch number from trainer
val_tensorboard_stats_handler.attach(evaluator)
# add handler to draw the first image and the corresponding label and model output in the last batch
# here we draw the 3D output as GIF format along Depth axis, at every validation epoch
val_tensorboard_image_handler = TensorBoardImageHandler(
batch_transform=lambda batch: (batch[0], batch[1]),
output_transform=lambda output: predict_segmentation(output[0]),
global_iter_transform=lambda x: trainer.state.epoch,
)
evaluator.add_event_handler(event_name=Events.EPOCH_COMPLETED,
handler=val_tensorboard_image_handler)
train_epochs = 30
state = trainer.run(train_loader, train_epochs)
print(state)
shutil.rmtree(tempdir)
def main():
monai.config.print_config()
logging.basicConfig(stream=sys.stdout, level=logging.INFO)
data_dir = '/home/marafath/scratch/eu_data'
labels = np.load('eu_labels.npy')
train_images = []
train_labels = []
val_images = []
val_labels = []
n_count = 0
p_count = 0
idx = 0
for case in os.listdir(data_dir):
if p_count < 13 and labels[idx] == 1:
val_images.append(
os.path.join(data_dir, case, 'image_masked.nii.gz'))
val_labels.append(labels[idx])
p_count += 1
idx += 1
elif n_count < 11 and labels[idx] == 0:
val_images.append(
os.path.join(data_dir, case, 'image_masked.nii.gz'))
val_labels.append(labels[idx])
n_count += 1
idx += 1
else:
train_images.append(
os.path.join(data_dir, case, 'image_masked.nii.gz'))
train_labels.append(labels[idx])
idx += 1
# Define transforms
train_transforms = Compose([
ScaleIntensity(),
AddChannel(),
RandZoom(min_zoom=0.9, max_zoom=1.1, prob=0.5),
SpatialPad((256, 256, 92), mode='constant'),
Resize((256, 256, 92)),
ToTensor()
])
val_transforms = Compose([
ScaleIntensity(),
AddChannel(),
SpatialPad((256, 256, 92), mode='constant'),
Resize((256, 256, 92)),
ToTensor()
])
# create a training data loader
train_ds = NiftiDataset(image_files=train_images,
labels=train_labels,
transform=train_transforms)
train_loader = DataLoader(train_ds,
batch_size=2,
shuffle=True,
num_workers=2,
pin_memory=torch.cuda.is_available())
# create a validation data loader
val_ds = NiftiDataset(image_files=val_images,
labels=val_labels,
transform=val_transforms)
val_loader = DataLoader(val_ds,
batch_size=2,
num_workers=2,
pin_memory=torch.cuda.is_available())
# Create DenseNet121, CrossEntropyLoss and Adam optimizer
device = torch.device('cuda:0')
model = monai.networks.nets.densenet.densenet121(
spatial_dims=3,
in_channels=1,
out_channels=2,
).to(device)
loss_function = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), 1e-3)
# finetuning
#model.load_state_dict(torch.load('best_metric_model_d121.pth'))
# start a typical PyTorch training
val_interval = 1
best_metric = -1
best_metric_epoch = -1
epoch_loss_values = list()
metric_values = list()
writer = SummaryWriter()
epc = 100 # Number of epoch
for epoch in range(epc):
print('-' * 10)
print('epoch {}/{}'.format(epoch + 1, epc))
model.train()
epoch_loss = 0
step = 0
for batch_data in train_loader:
step += 1
inputs, labels = batch_data[0].to(device), batch_data[1].to(
device=device, dtype=torch.int64)
optimizer.zero_grad()
outputs = model(inputs)
loss = loss_function(outputs, labels)
loss.backward()
optimizer.step()
epoch_loss += loss.item()
epoch_len = len(train_ds) // train_loader.batch_size
print('{}/{}, train_loss: {:.4f}'.format(step, epoch_len,
loss.item()))
writer.add_scalar('train_loss', loss.item(),
epoch_len * epoch + step)
epoch_loss /= step
epoch_loss_values.append(epoch_loss)
print('epoch {} average loss: {:.4f}'.format(epoch + 1, epoch_loss))
if (epoch + 1) % val_interval == 0:
model.eval()
with torch.no_grad():
num_correct = 0.
metric_count = 0
for val_data in val_loader:
val_images, val_labels = val_data[0].to(
device), val_data[1].to(device)
val_outputs = model(val_images)
value = torch.eq(val_outputs.argmax(dim=1), val_labels)
metric_count += len(value)
num_correct += value.sum().item()
metric = num_correct / metric_count
metric_values.append(metric)
#torch.save(model.state_dict(), 'model_d121_epoch_{}.pth'.format(epoch + 1))
if metric > best_metric:
best_metric = metric
best_metric_epoch = epoch + 1
torch.save(
model.state_dict(),
'/home/marafath/scratch/saved_models/best_metric_model_d121.pth'
)
print('saved new best metric model')
print(
'current epoch: {} current accuracy: {:.4f} best accuracy: {:.4f} at epoch {}'
.format(epoch + 1, metric, best_metric, best_metric_epoch))
writer.add_scalar('val_accuracy', metric, epoch + 1)
print('train completed, best_metric: {:.4f} at epoch: {}'.format(
best_metric, best_metric_epoch))
writer.close()
