def __init__(self,
data_split_csv,
train_preprocessings,
valid_preprocessings,
transforms,
augments=None,
**kwargs):
super().__init__(**kwargs)
self.data_split_csv = data_split_csv
self.train_preprocessings = compose(train_preprocessings)
self.valid_preprocessings = compose(valid_preprocessings)
self.transforms = compose(transforms)
self.augments = compose(augments)
self.data_paths = []
self.input_paths = []
# Collect the data paths according to the dataset split csv.
with open(self.data_split_csv, "r") as f:
type_ = 'train' if self.type == 'train' else 'valid'
rows = csv.reader(f)
for case_name, split_type in rows:
if split_type == type_:
label_paths = sorted(
list((self.data_dir /
case_name).glob('clf_label*.npy')))
graph_paths = sorted(
list((self.data_dir / case_name).glob('graph.bin')))
self.data_paths.extend([(graph_path, label_path)
for graph_path, label_path in zip(
graph_paths, label_paths)])
def __init__(self, downscale_factor, transforms, augments=None, num_frames=5, temporal_order='last', **kwargs):
super().__init__(**kwargs)
if downscale_factor not in [2, 3, 4]:
raise ValueError(f'The downscale factor should be 2, 3, 4. Got {downscale_factor}.')
self.downscale_factor = downscale_factor
self.transforms = compose(transforms)
self.augments = compose(augments)
self.num_frames = num_frames
if temporal_order not in ['last', 'middle']:
raise ValueError(f"The temporal order should be 'last' or 'middle'. Got {temporal_order}.")
self.temporal_order = temporal_order
# Save the data paths and the target frame index for training; only need to save the data paths
# for validation to process dynamic length of the sequences.
lr_paths = sorted((self.data_dir / self.type / 'LR' / f'X{downscale_factor}').glob('**/*2d+1d*.nii.gz'))
hr_paths = sorted((self.data_dir / self.type / 'HR').glob('**/*2d+1d*.nii.gz'))
if self.type == 'train':
self.data = []
for lr_path, hr_path in zip(lr_paths, hr_paths):
T = nib.load(str(lr_path)).header.get_data_shape()[-1]
self.data.extend([(lr_path, hr_path, t) for t in range(T)])
else:
self.data = [(lr_path, hr_path) for lr_path, hr_path in zip(lr_paths, hr_paths)]
def __init__(self,
data_split_csv,
re_sample_size,
train_preprocessings,
valid_preprocessings,
transforms,
augments=None,
**kwargs):
super().__init__(**kwargs)
self.data_split_csv = data_split_csv
self.npoints = re_sample_size
self.train_preprocessings = compose(train_preprocessings)
self.valid_preprocessings = compose(valid_preprocessings)
self.transforms = compose(transforms)
self.augments = compose(augments)
self.data_paths = []
# Collect the data paths according to the dataset split csv.
with open(self.data_split_csv, "r") as f:
type_ = 'Training' if self.type == 'train' else 'Validation'
rows = csv.reader(f)
for file_name, split_type in rows:
if split_type == type_:
data_path = self.data_dir / f'{file_name}'
self.data_paths.append(data_path)
def __init__(self, data_dir, data_split_csv,
device, net, metric_fns, saved_dir=None, exported=None):
self.data_dir = data_dir
self.data_split_csv = data_split_csv
self.preprocessings = compose(preprocessings)
self.transforms = compose(transforms)
self.sample_size = sample_size
self.shift = shift
self.device = device
self.net = net
self.metric_fns = metric_fns
self.saved_dir = saved_dir
self.exported = exported
self.log = self._init_log()
def __init__(self,
data_split_csv,
train_preprocessings,
valid_preprocessings,
transforms,
augments=None,
**kwargs):
super().__init__(**kwargs)
self.data_split_csv = data_split_csv
self.train_preprocessings = compose(train_preprocessings)
self.valid_preprocessings = compose(valid_preprocessings)
self.transforms = compose(transforms)
self.augments = compose(augments)
self.data_paths = []
self.input_paths = []
# Collect the data paths according to the dataset split csv.
with open(self.data_split_csv, "r") as f:
type_ = 'train' if self.type == 'train' else 'valid'
rows = csv.reader(f)
for case_name, split_type in rows:
if split_type == type_:
image_paths = sorted(
list((self.data_dir / case_name).glob('image*.npy')))
label_paths = sorted(
list((self.data_dir / case_name).glob('label*.npy')))
gcn_label_paths = sorted(
list((self.data_dir /
case_name).glob('gcn_label*.npy')))
segments_paths = sorted(
list(
(self.data_dir / case_name).glob('segments*.npy')))
features_paths = sorted(
list(
(self.data_dir / case_name).glob('features*.npy')))
adj_paths = sorted(
list((self.data_dir / case_name).glob('adj_arr*.npy')))
self.data_paths.extend([
(image_path, label_path, gcn_label_path)
for image_path, label_path, gcn_label_path in zip(
image_paths, label_paths, gcn_label_paths)
])
self.input_paths.extend([
(s_path, f_path, a_path)
for s_path, f_path, a_path in zip(
segments_paths, features_paths, adj_paths)
])
def __init__(self, downscale_factor, transforms, augments=None, **kwargs):
super().__init__(**kwargs)
if downscale_factor not in [2, 3, 4]:
raise ValueError(
f'The downscale factor should be 2, 3, 4. Got {downscale_factor}.'
)
self.downscale_factor = downscale_factor
self.transforms = compose(transforms)
self.augments = compose(augments)
lr_paths = sorted((self.data_dir / self.type / 'LR' /
f'X{downscale_factor}').glob('**/*2d*.nii.gz'))
hr_paths = sorted(
(self.data_dir / self.type / 'HR').glob('**/*2d*.nii.gz'))
self.data = [(lr_path, hr_path)
for lr_path, hr_path in zip(lr_paths, hr_paths)]
def __init__(self, data_split_csv, positive_sampling_rate, sample_size,
preprocessings, augments, transforms, **kwargs):
super().__init__(**kwargs)
self.data_split_csv = data_split_csv
self.positive_sampling_rate = positive_sampling_rate
self.sample_size = sample_size
self.preprocessings = compose(preprocessings)
self.augments = compose(augments)
self.transforms = compose(transforms)
self.data_paths = []
# Collect the data paths according to the dataset split csv.
with open(self.data_split_csv, "r") as f:
type_ = 'Training' if self.type == 'train' else 'Validation'
rows = csv.reader(f)
for case_name, split_type in rows:
if split_type == type_:
image_path = self.data_dir / f'{case_name[:5]}-{case_name[5:]}image.nii.gz'
label_path = self.data_dir / f'{case_name[:5]}-{case_name[5:]}label_GTV1.nii.gz'
self.data_paths.append([image_path, label_path])
def test_composed_transforms(config, dummy_input):
"""Test to compose multiple augmentations
including RandomCrop, Normalize, ToTensor.
"""
cfg = config
transforms = compose(cfg.dataset.transforms)
# H, W, C
image, label = dummy_input(image_size=(512, 512, 3),
label_size=(512, 512, 1))
_image, _label = transforms(image,
label,
dtypes=[torch.float, torch.long],
elastic_deformation_orders=[3, 0])
assert _image.dtype == torch.float
assert _image.size() == (256, 256, image.shape[2])
assert _label.dtype == torch.long
assert _label.size() == (256, 256, label.shape[2])
# Test feeding only image
_image = transforms(image, dtypes=[torch.float])
assert _image.dtype == torch.float
assert _image.size() == (256, 256, image.shape[2])
def __init__(self, image, label, transforms, **kwargs):
super().__init__(**kwargs)
self.image, self.label = image, label
self.transforms = compose(transforms)