def test_correct_results(self, degrees, keep_size, mode, padding_mode, align_corners):
rotate_fn = RandRotate(
range_x=degrees,
prob=1.0,
keep_size=keep_size,
mode=mode,
padding_mode=padding_mode,
align_corners=align_corners,
)
rotate_fn.set_random_state(243)
rotated = rotate_fn(self.imt[0])
_order = 0 if mode == "nearest" else 1
if mode == "border":
_mode = "nearest"
elif mode == "reflection":
_mode = "reflect"
else:
_mode = "constant"
angle = rotate_fn.x
expected = scipy.ndimage.rotate(
self.imt[0, 0], -np.rad2deg(angle), (0, 1), not keep_size, order=_order, mode=_mode, prefilter=False
)
expected = np.stack(expected).astype(np.float32)
good = np.sum(np.isclose(expected, rotated[0], atol=1e-3))
self.assertLessEqual(np.abs(good - expected.size), 5, "diff at most 5 pixels")
def test_correct_results(self, degrees, keep_size, order, mode,
align_corners):
rotate_fn = RandRotate(range_x=degrees,
prob=1.0,
keep_size=keep_size,
interp_order=order,
mode=mode,
align_corners=align_corners)
rotate_fn.set_random_state(243)
rotated = rotate_fn(self.imt[0])
_order = 0 if order == "nearest" else 1
if mode == "border":
_mode = "nearest"
elif mode == "reflection":
_mode = "reflect"
else:
_mode = "constant"
angle = rotate_fn.x
expected = scipy.ndimage.rotate(self.imt[0, 0],
-angle, (0, 1),
not keep_size,
order=_order,
mode=_mode,
prefilter=False)
expected = np.stack(expected).astype(np.float32)
np.testing.assert_allclose(expected, rotated[0])
def test_correct_results(self, degrees, spatial_axes, reshape, order, mode,
cval, prefilter):
rotate_fn = RandRotate(
degrees,
prob=1.0,
spatial_axes=spatial_axes,
reshape=reshape,
interp_order=order,
mode=mode,
cval=cval,
prefilter=prefilter,
)
rotate_fn.set_random_state(243)
rotated = rotate_fn(self.imt[0])
angle = rotate_fn.angle
expected = list()
for channel in self.imt[0]:
expected.append(
scipy.ndimage.rotate(channel,
angle,
spatial_axes,
reshape,
order=order,
mode=mode,
cval=cval,
prefilter=prefilter))
expected = np.stack(expected).astype(np.float32)
self.assertTrue(np.allclose(expected, rotated))
def test_correct_results(self, x, y, z, keep_size, mode, padding_mode, align_corners, expected):
rotate_fn = RandRotate(
range_x=x,
range_y=y,
range_z=z,
prob=1.0,
keep_size=keep_size,
mode=mode,
padding_mode=padding_mode,
align_corners=align_corners,
)
rotate_fn.set_random_state(243)
rotated = rotate_fn(self.imt[0])
np.testing.assert_allclose(rotated.shape, expected)
def test_correct_results(self, im_type, x, y, z, keep_size, mode, padding_mode, align_corners, expected):
rotate_fn = RandRotate(
range_x=x,
range_y=y,
range_z=z,
prob=1.0,
keep_size=keep_size,
mode=mode,
padding_mode=padding_mode,
align_corners=align_corners,
dtype=np.float64,
)
rotate_fn.set_random_state(243)
rotated = rotate_fn(im_type(self.imt[0]))
torch.testing.assert_allclose(rotated.shape, expected, rtol=1e-7, atol=0)
def test_correct_results(self, im_type, x, y, z, keep_size, mode, padding_mode, align_corners, expected):
rotate_fn = RandRotate(
range_x=x,
range_y=y,
range_z=z,
prob=1.0,
keep_size=keep_size,
mode=mode,
padding_mode=padding_mode,
align_corners=align_corners,
dtype=np.float64,
)
rotate_fn.set_random_state(243)
im = im_type(self.imt[0])
rotated = rotate_fn(im)
torch.testing.assert_allclose(rotated.shape, expected, rtol=1e-7, atol=0)
test_local_inversion(rotate_fn, rotated, im)
set_track_meta(False)
rotated = rotate_fn(im)
self.assertNotIsInstance(rotated, MetaTensor)
self.assertIsInstance(rotated, torch.Tensor)
set_track_meta(True)
def test_invert(self):
set_determinism(seed=0)
im_fname = make_nifti_image(create_test_image_3d(101, 100, 107, noise_max=100)[1]) # label image, discrete
data = [im_fname for _ in range(12)]
transform = Compose(
[
LoadImage(image_only=True),
EnsureChannelFirst(),
Orientation("RPS"),
Spacing(pixdim=(1.2, 1.01, 0.9), mode="bilinear", dtype=np.float32),
RandFlip(prob=0.5, spatial_axis=[1, 2]),
RandAxisFlip(prob=0.5),
RandRotate90(prob=0, spatial_axes=(1, 2)),
RandZoom(prob=0.5, min_zoom=0.5, max_zoom=1.1, keep_size=True),
RandRotate(prob=0.5, range_x=np.pi, mode="bilinear", align_corners=True, dtype=np.float64),
RandAffine(prob=0.5, rotate_range=np.pi, mode="nearest"),
ResizeWithPadOrCrop(100),
CastToType(dtype=torch.uint8),
]
)
# num workers = 0 for mac or gpu transforms
num_workers = 0 if sys.platform != "linux" or torch.cuda.is_available() else 2
dataset = Dataset(data, transform=transform)
self.assertIsInstance(transform.inverse(dataset[0]), MetaTensor)
loader = DataLoader(dataset, num_workers=num_workers, batch_size=1)
inverter = Invert(transform=transform, nearest_interp=True, device="cpu")
for d in loader:
d = decollate_batch(d)
for item in d:
orig = deepcopy(item)
i = inverter(item)
self.assertTupleEqual(orig.shape[1:], (100, 100, 100))
# check the nearest interpolation mode
torch.testing.assert_allclose(i.to(torch.uint8).to(torch.float), i.to(torch.float))
self.assertTupleEqual(i.shape[1:], (100, 101, 107))
# check labels match
reverted = i.detach().cpu().numpy().astype(np.int32)
original = LoadImage(image_only=True)(data[-1])
n_good = np.sum(np.isclose(reverted, original.numpy(), atol=1e-3))
reverted_name = i.meta["filename_or_obj"]
original_name = original.meta["filename_or_obj"]
self.assertEqual(reverted_name, original_name)
print("invert diff", reverted.size - n_good)
self.assertTrue((reverted.size - n_good) < 300000, f"diff. {reverted.size - n_good}")
set_determinism(seed=None)
prob=1,
min_zoom=1.1,
max_zoom=2.0,
keep_size=False)))
TESTS.append((dict, pad_collate,
Compose([
RandRotate90d("image", prob=1, max_k=3),
RandRotate90d("image", prob=1, max_k=4)
])))
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.transforms = transforms
for i, elem in enumerate(image_files_list):
if elem in list_all_images:
image_files_list_updated.append(elem)
image_class_list.append(image_class[i])
"""
transforms
"""
train_transforms = Compose(
[
LoadPNG(image_only=True),
AddChannel(),
ScaleIntensity(),
RandRotate(range_x=15, prob=0.1, keep_size=True), # low probability for rotation
RandFlip(spatial_axis=0, prob=0.5),# left right flip
RandFlip(spatial_axis=1, prob=0.5), # horizontal flip
RandZoom(min_zoom=0.9, max_zoom=1.1, prob=0.5),
ToTensor(),
Lambda(lambda x: torch.cat([x, x, x], 0)),
Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
]
)
val_transforms = Compose(
[
LoadPNG(image_only=True),
# Resize((480,640)),
AddChannel(),
ScaleIntensity(),
nsml.paused(scope=locals())
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)
roi_size=[8, 7],
random_size=True)))
TESTS.append((dict, RandRotated("image",
prob=1,
range_x=np.pi,
keep_size=False)))
TESTS.append((dict,
RandZoomd("image",
prob=1,
min_zoom=1.1,
max_zoom=2.0,
keep_size=False)))
TESTS.append((dict, RandRotate90d("image", prob=1, max_k=2)))
TESTS.append((list, RandSpatialCrop(roi_size=[8, 7], random_size=True)))
TESTS.append((list, RandRotate(prob=1, range_x=np.pi, keep_size=False)))
TESTS.append(
(list, RandZoom(prob=1, min_zoom=1.1, max_zoom=2.0, keep_size=False)))
TESTS.append((list, RandRotate90(prob=1, max_k=2)))
class _Dataset(torch.utils.data.Dataset):
def __init__(self, images, labels, transforms):
self.images = images
self.labels = labels
self.transforms = transforms
def __len__(self):
return len(self.images)
def __getitem__(self, index):
class Loader():
"""Loader for different image datasets with built in split function and download if needed.
Functions:
load_IXIT1: Loads the IXIT1 3D brain MRI dataset.
load_MedNIST: Loads the MedNIST 2D image dataset.
"""
ixi_train_transforms = Compose([ScaleIntensity(), AddChannel(), Resize((96, 96, 96)), RandRotate90()])
ixi_test_transforms = Compose([ScaleIntensity(), AddChannel(), Resize((96, 96, 96))])
mednist_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)])
mednist_test_transforms = Compose([LoadImage(image_only=True), AddChannel(), ScaleIntensity()])
@staticmethod
def load_IXIT1(download: bool = False, train_transforms: object = ixi_train_transforms,
test_transforms: object = ixi_test_transforms, test_size: float = 0.2,
val_size: float = 0.0, sample_size: float = 0.01, shuffle: bool = True):
"""Loads the IXIT1 3D Brain MRI dataset.
Consists of ~566 images of 3D Brain MRI scans and labels (0) for male and (1) for female.
Args:
download (bool): If true, then data is downloaded before loading it as dataset.
train_transforms (Compose): Specify the transformations to be applied to the training dataset.
test_transforms (Compose): Specify the transformations to be applied to the test dataset.
sample_size (float): Percentage of available images to be used.
test_size (float): Precantage of sample to be used as test data.
val_size (float): Percentage of sample to be used as validation data.
shuffle (bool): Whether or not the data should be shuffled after loading.
"""
# Download data if needed
if download:
data_url = 'http://biomedic.doc.ic.ac.uk/brain-development/downloads/IXI/IXI-T1.tar'
compressed_file = os.sep.join(['Data', 'IXI-T1.tar'])
data_dir = os.sep.join(['Data', 'IXI-T1'])
# Data download
monai.apps.download_and_extract(data_url, compressed_file, './Data/IXI-T1')
# Labels document download
labels_url = 'http://biomedic.doc.ic.ac.uk/brain-development/downloads/IXI/IXI.xls'
monai.apps.download_url(labels_url, './Data/IXI.xls')
# Get all the images and corresponding Labels
images = [impath for impath in os.listdir('./Data/IXI-T1')]
df = pd.read_excel('./Data/IXI.xls')
data = []
labels = []
for i in images:
ixi_id = int(i[3:6])
row = df.loc[df['IXI_ID'] == ixi_id]
if not row.empty:
data.append(os.sep.join(['Data', 'IXI-T1', i]))
labels.append(int(row.iat[0, 1] - 1)) # Sex labels are 1/2 but need to be 0/1
data, labels = data[:int(len(data) * sample_size)], labels[:int(len(data) * sample_size)]
# Make train test validation split
train_data, train_labels, test_data, test_labels, val_data, val_labels = _split(data, labels,
test_size, val_size)
# Construct and return Datasets
train_ds = IXIT1Dataset(train_data, train_labels, train_transforms, shuffle)
test_ds = IXIT1Dataset(test_data, test_labels, test_transforms, shuffle)
if val_size == 0:
return train_ds, test_ds
else:
val_ds = IXIT1Dataset(val_data, val_labels, test_transforms, shuffle)
return train_ds, test_ds, val_ds
@staticmethod
def load_MedNIST(download: bool = False, train_transforms: object = mednist_train_transforms,
test_transforms: object = mednist_test_transforms, test_size: float = 0.2,
val_size: float = 0.0, sample_size: float = 0.01, shuffle: bool = True):
"""Loads the MedNIST 2D image dataset.
Consists of ~60.000 2D images from 6 classes: AbdomenCT, BreastMRI, ChestCT, CXR, Hand, HeadCT.
Args:
download (bool): If true, then data is downloaded before loading it as dataset.
train_transforms (Compose): Specify the transformations to be applied to the training dataset.
test_transforms (Compose): Specify the transformations to be applied to the test dataset.
sample_size (float): Percentage of available images to be used.
test_size (float): Precantage of sample to be used as test data.
val_size (float): Percentage of sample to be used as validation data.
shuffle (bool): Whether or not the data should be shuffled after loading.
"""
root_dir = './Data'
resource = "https://www.dropbox.com/s/5wwskxctvcxiuea/MedNIST.tar.gz?dl=1"
md5 = "0bc7306e7427e00ad1c5526a6677552d"
compressed_file = os.path.join(root_dir, "MedNIST.tar.gz")
data_dir = os.path.join(root_dir, "MedNIST")
if download:
monai.apps.download_and_extract(resource, compressed_file, root_dir, md5)
# Reading image filenames from dataset folders and assigning labels
class_names = sorted(x for x in os.listdir(data_dir)
if os.path.isdir(os.path.join(data_dir, x)))
num_class = len(class_names)
image_files = [
[
os.path.join(data_dir, class_names[i], x)
for x in os.listdir(os.path.join(data_dir, class_names[i]))
]
for i in range(num_class)
]
image_files = [images[:int(len(images) * sample_size)] for images in image_files]
# Constructing data and labels
num_each = [len(image_files[i]) for i in range(num_class)]
data = []
labels = []
for i in range(num_class):
data.extend(image_files[i])
labels.extend([int(i)] * num_each[i])
if shuffle:
np.random.seed(42)
indicies = np.arange(len(data))
np.random.shuffle(indicies)
data = [data[i] for i in indicies]
labels = [labels[i] for i in indicies]
# Make train test validation split
train_data, train_labels, test_data, test_labels, val_data, val_labels = _split(data, labels,
test_size, val_size)
# Construct and return datasets
train_ds = MedNISTDataset(train_data, train_labels, train_transforms, shuffle)
test_ds = MedNISTDataset(test_data, test_labels, test_transforms, shuffle)
if val_size == 0:
return train_ds, test_ds
else:
val_ds = MedNISTDataset(val_data, val_labels, test_transforms, shuffle)
return train_ds, test_ds, val_ds
data_all_ch = list(zip(*all_data))
train_split, val_split = split_train_val(data_all_ch,
N_valid_per_magn=4,
is_val_split=is_val_split)
# data preprocessing/augmentation
trans_train = MozartTheComposer([
#ScaleIntensity(),
# AddChannel(),
# RandSpatialCrop(roi_size=256, random_size=False),
#CenterSpatialCrop(roi_size=2154), # 2154
# RandScaleIntensity(factors=0.25, prob=aug_prob),
RandRotate(range_x=15,
prob=aug_prob,
keep_size=True,
padding_mode="reflection"),
RandRotate90(prob=aug_prob, spatial_axes=(1, 2)),
RandFlip(spatial_axis=(1, 2), prob=aug_prob),
ToTensor()
])
trans_val = MozartTheComposer([
# LoadImage(PILReader(), image_only=True),
#ScaleIntensity(),
# AddChannel(),
# RandSpatialCrop(roi_size=256, random_size=False),
#CenterSpatialCrop(roi_size=2154),
ToTensor()
])
monai.config.print_config()
train_data, val_data, test_data, raw_data = get_train_val_test_data(
params['data_dir'], params['test_val_split'])
size_n = 3
sample_data = train_data.sample(size_n**2)
#show_sample_dataframe(sample_data, size_n=size_n, title="Training samples")
train_transforms = Compose([
LoadPNG(image_only=True),
AddChannel(),
ScaleIntensity(),
RandRotate(range_x=params['rotate_range_x'],
prob=params['rotate_prob'],
keep_size=True),
# RandFlip(spatial_axis=0, prob=0.5),
RandZoom(min_zoom=params['min_zoom'],
max_zoom=params['max_zoom'],
prob=params['zoom_prob'],
keep_size=True),
ToTensor()
])
val_transforms = Compose(
[LoadPNG(image_only=True),
AddChannel(),
ScaleIntensity(),
ToTensor()])
train_ds = LabeledImageDataset(train_data, train_transforms)
TESTS.append((dict, pad_collate,
RandRotated("image", prob=1, range_x=np.pi,
keep_size=False)))
TESTS.append((dict, pad_collate,
RandZoomd("image",
prob=1,
min_zoom=1.1,
max_zoom=2.0,
keep_size=False)))
TESTS.append((dict, pad_collate, RandRotate90d("image", prob=1, max_k=2)))
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)))
TESTS.append((list, pad_collate,
RandZoom(prob=1, min_zoom=1.1, max_zoom=2.0,
keep_size=False)))
TESTS.append((list, pad_collate, RandRotate90(prob=1, max_k=2)))
class _Dataset(torch.utils.data.Dataset):
def __init__(self, images, labels, transforms):
self.images = images
self.labels = labels
self.transforms = transforms
def __len__(self):
return len(self.images)
def run_training_test(root_dir,
train_x,
train_y,
val_x,
val_y,
device="cuda:0",
num_workers=10):
monai.config.print_config()
# define transforms for image and classification
train_transforms = Compose([
LoadPNG(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(),
])
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=num_workers)
val_ds = MedNISTDataset(val_x, val_y, val_transforms)
val_loader = DataLoader(val_ds, batch_size=300, num_workers=num_workers)
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
val_indices = indices[test_split:val_split]
train_indices = indices[val_split:]
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)