def get_train_transform(patch_size):
tr_transforms = []
tr_transforms.append(
SpatialTransform_2(
patch_size, [i // 2 for i in patch_size],
do_elastic_deform=True,
deformation_scale=(0, 0.05),
do_rotation=True,
angle_x=(-5 / 360. * 2 * np.pi, 5 / 360. * 2 * np.pi),
angle_y=(-5 / 360. * 2 * np.pi, 5 / 360. * 2 * np.pi),
angle_z=(-5 / 360. * 2 * np.pi, 5 / 360. * 2 * np.pi),
do_scale=True,
scale=(0.75, 1.25),
border_mode_data='constant',
border_cval_data=-2.34,
border_mode_seg='constant',
border_cval_seg=0))
tr_transforms.append(MirrorTransform(axes=(0, 1, 2)))
tr_transforms.append(
BrightnessMultiplicativeTransform((0.7, 1.5),
per_channel=True,
p_per_sample=0.15))
tr_transforms.append(
GammaTransform(gamma_range=(0.5, 2),
invert_image=True,
per_channel=True,
p_per_sample=0.15))
tr_transforms.append(
GaussianNoiseTransform(noise_variance=(0, 0.15), p_per_sample=0.15))
tr_transforms = Compose(tr_transforms)
return tr_transforms
def get_train_transform(patch_size):
tr_transforms = []
tr_transforms.append(
SpatialTransform_2(
None, [i // 2 for i in patch_size],
do_elastic_deform=False,
deformation_scale=(0, 0.25),
do_rotation=True,
angle_x=(-15 / 360. * 2 * np.pi, 15 / 360. * 2 * np.pi),
angle_y=(-15 / 360. * 2 * np.pi, 15 / 360. * 2 * np.pi),
angle_z=(-15 / 360. * 2 * np.pi, 15 / 360. * 2 * np.pi),
do_scale=False,
scale=(0.8, 1.2),
border_mode_data='constant',
border_cval_data=0,
border_mode_seg='constant',
border_cval_seg=0,
order_seg=1,
order_data=3,
random_crop=False,
p_el_per_sample=0.3,
p_rot_per_sample=0.3,
p_scale_per_sample=0.3))
tr_transforms.append(
RndTransform(MirrorTransform(axes=(0, 1, 2)), prob=0.3))
tr_transforms = Compose(transforms=tr_transforms)
return tr_transforms
def test_random_distributions_2D(self):
### test whether all 4 possible mirrorings occur in approximately equal frquencies in 2D
batch_gen = BasicDataLoader((self.x_2D, self.y_2D), self.batch_size, number_of_threads_in_multithreaded=None)
batch_gen = SingleThreadedAugmenter(batch_gen, MirrorTransform((0, 1)))
counts = np.zeros(shape=(4,))
for b in range(self.num_batches):
batch = next(batch_gen)
for ix in range(self.batch_size):
if (batch['data'][ix, :, :, :] == self.cam_left).all():
counts[0] = counts[0] + 1
elif (batch['data'][ix, :, :, :] == self.cam_updown).all():
counts[1] = counts[1] + 1
elif (batch['data'][ix, :, :, :] == self.cam_updown_left).all():
counts[2] = counts[2] + 1
elif (batch['data'][ix, :, :, :] == self.cam).all():
counts[3] = counts[3] + 1
self.assertTrue([1 if (2200 < c < 2800) else 0 for c in counts] == [1]*4, "2D Images were not mirrored along "
"all axes with equal probability. "
"This may also indicate that "
"mirroring is not working")
def test_image_pipeline_and_pin_memory(self):
'''
This just should not crash
:return:
'''
try:
import torch
except ImportError:
'''dont test if torch is not installed'''
return
tr_transforms = []
tr_transforms.append(MirrorTransform())
tr_transforms.append(
TransposeAxesTransform(transpose_any_of_these=(0, 1),
p_per_sample=0.5))
tr_transforms.append(NumpyToTensor(keys='data', cast_to='float'))
composed = Compose(tr_transforms)
dl = self.dl_images
mt = MultiThreadedAugmenter(dl, composed, 4, 1, None, True)
for _ in range(50):
res = mt.next()
assert isinstance(res['data'], torch.Tensor)
assert res['data'].is_pinned()
# let mt finish caching, otherwise it's going to print an error (which is not a problem and will not prevent
# the success of the test but it does not look pretty)
sleep(2)
def get_transforms(
patch_shape=(256, 320), other_transforms=None, random_crop=False):
"""
Initializes the transforms for training.
Args:
patch_shape:
other_transforms: List of transforms that you would like to add (optional). Defaults to None.
random_crop (boolean): whether or not you want to random crop or center crop. Currently, the Transformed3DGenerator
only supports random cropping. Transformed2DGenerator supports both random_crop = True and False.
"""
ndim = len(patch_shape)
spatial_transform = SpatialTransform(patch_shape,
do_elastic_deform=True,
alpha=(0., 1500.),
sigma=(30., 80.),
do_rotation=True,
angle_z=(0, 2 * np.pi),
do_scale=True,
scale=(0.75, 2.),
border_mode_data="nearest",
border_cval_data=0,
order_data=1,
random_crop=random_crop,
p_el_per_sample=0.1,
p_scale_per_sample=0.1,
p_rot_per_sample=0.1)
mirror_transform = MirrorTransform(axes=(0, 1))
transforms_list = [spatial_transform, mirror_transform]
if other_transforms is not None:
transforms_list = transforms_list + other_transforms
composed = Compose(transforms_list)
return composed
def get_train_transform(patch_size):
"""
data augmentation for training data, inspired by:
https://github.com/MIC-DKFZ/batchgenerators/blob/master/batchgenerators/examples/brats2017/brats2017_dataloader_3D.py
:param patch_size: shape of network's input
:return list of transformations
"""
train_transforms = []
def rad(deg):
return (-deg / 360 * 2 * np.pi, deg / 360 * 2 * np.pi)
train_transforms.append(
SpatialTransform_2(
patch_size,
(10, 10, 10),
do_elastic_deform=True,
deformation_scale=(0, 0.25),
do_rotation=True,
angle_z=rad(15),
angle_x=(0, 0),
angle_y=(0, 0),
do_scale=True,
scale=(0.75, 1.25),
border_mode_data='constant',
border_cval_data=0,
border_mode_seg='constant',
border_cval_seg=0,
order_seg=1,
random_crop=False,
p_el_per_sample=0.2,
p_rot_per_sample=0.2,
p_scale_per_sample=0.2,
))
train_transforms.append(MirrorTransform(axes=(0, 1)))
train_transforms.append(
BrightnessMultiplicativeTransform((0.7, 1.5),
per_channel=True,
p_per_sample=0.2))
train_transforms.append(
GammaTransform(gamma_range=(0.2, 1.0),
invert_image=False,
per_channel=False,
p_per_sample=0.2))
train_transforms.append(
GaussianNoiseTransform(noise_variance=(0, 0.05), p_per_sample=0.2))
train_transforms.append(
GaussianBlurTransform(blur_sigma=(0.2, 1.0),
different_sigma_per_channel=False,
p_per_channel=0.0,
p_per_sample=0.2))
return Compose(train_transforms)
def get_train_transform(patch_size):
# we now create a list of transforms. These are not necessarily the best transforms to use for BraTS, this is just
# to showcase some things
tr_transforms = []
# the first thing we want to run is the SpatialTransform. It reduces the size of our data to patch_size and thus
# also reduces the computational cost of all subsequent operations. All subsequent operations do not modify the
# shape and do not transform spatially, so no border artifacts will be introduced
# Here we use the new SpatialTransform_2 which uses a new way of parameterizing elastic_deform
# We use all spatial transformations with a probability of 0.2 per sample. This means that 1 - (1 - 0.1) ** 3 = 27%
# of samples will be augmented, the rest will just be cropped
tr_transforms.append(
SpatialTransform_2(
patch_size, [i // 2 for i in patch_size],
do_elastic_deform=True,
deformation_scale=(0, 0.25),
do_rotation=True,
angle_x=(-15 / 360. * 2 * np.pi, 15 / 360. * 2 * np.pi),
angle_y=(-15 / 360. * 2 * np.pi, 15 / 360. * 2 * np.pi),
angle_z=(-15 / 360. * 2 * np.pi, 15 / 360. * 2 * np.pi),
do_scale=True,
scale=(0.75, 1.25),
border_mode_data='constant',
border_cval_data=0,
border_mode_seg='constant',
border_cval_seg=0,
order_seg=1,
order_data=3,
random_crop=True,
p_el_per_sample=0.1,
p_rot_per_sample=0.1,
p_scale_per_sample=0.1))
# now we mirror along all axes
tr_transforms.append(MirrorTransform(axes=(0, 1, 2)))
# gamma transform. This is a nonlinear transformation of intensity values
# (https://en.wikipedia.org/wiki/Gamma_correction)
tr_transforms.append(
GammaTransform(gamma_range=(0.5, 2),
invert_image=False,
per_channel=True,
p_per_sample=0.15))
# we can also invert the image, apply the transform and then invert back
tr_transforms.append(
GammaTransform(gamma_range=(0.5, 2),
invert_image=True,
per_channel=True,
p_per_sample=0.15))
# Gaussian Noise
tr_transforms.append(
GaussianNoiseTransform(noise_variance=(0, 0.05), p_per_sample=0.15))
# now we compose these transforms together
tr_transforms = Compose(tr_transforms)
return tr_transforms
def _augment_data(self, batch_generator, type=None):
if self.Config.DATA_AUGMENTATION:
num_processes = 16 # 2D: 8 is a bit faster than 16
# num_processes = 8
else:
num_processes = 6
tfs = [] #transforms
if self.Config.NORMALIZE_DATA:
tfs.append(ZeroMeanUnitVarianceTransform(per_channel=self.Config.NORMALIZE_PER_CHANNEL))
if self.Config.DATA_AUGMENTATION:
if type == "train":
# scale: inverted: 0.5 -> bigger; 2 -> smaller
# patch_center_dist_from_border: if 144/2=72 -> always exactly centered; otherwise a bit off center (brain can get off image and will be cut then)
if self.Config.DAUG_SCALE:
center_dist_from_border = int(self.Config.INPUT_DIM[0] / 2.) - 10 # (144,144) -> 62
tfs.append(SpatialTransform(self.Config.INPUT_DIM,
patch_center_dist_from_border=center_dist_from_border,
do_elastic_deform=self.Config.DAUG_ELASTIC_DEFORM,
alpha=(90., 120.), sigma=(9., 11.),
do_rotation=self.Config.DAUG_ROTATE,
angle_x=(-0.8, 0.8), angle_y=(-0.8, 0.8), angle_z=(-0.8, 0.8),
do_scale=True, scale=(0.9, 1.5), border_mode_data='constant',
border_cval_data=0,
order_data=3,
border_mode_seg='constant', border_cval_seg=0,
order_seg=0, random_crop=True, p_el_per_sample=0.2,
p_rot_per_sample=0.2, p_scale_per_sample=0.2))
if self.Config.DAUG_RESAMPLE:
tfs.append(SimulateLowResolutionTransform(zoom_range=(0.5, 1), p_per_sample=0.2))
if self.Config.DAUG_NOISE:
tfs.append(GaussianNoiseTransform(noise_variance=(0, 0.05), p_per_sample=0.2))
if self.Config.DAUG_MIRROR:
tfs.append(MirrorTransform())
if self.Config.DAUG_FLIP_PEAKS:
tfs.append(FlipVectorAxisTransform())
tfs.append(NumpyToTensor(keys=["data", "seg"], cast_to="float"))
#num_cached_per_queue 1 or 2 does not really make a difference
batch_gen = MultiThreadedAugmenter(batch_generator, Compose(tfs), num_processes=num_processes,
num_cached_per_queue=1, seeds=None, pin_memory=True)
return batch_gen # data: (batch_size, channels, x, y), seg: (batch_size, channels, x, y)
def test_segmentations_2D(self):
### test whether segmentations are mirrored coherently with images
batch_gen = BasicDataLoader((self.x_2D, self.y_2D), self.batch_size, number_of_threads_in_multithreaded=None)
batch_gen = SingleThreadedAugmenter(batch_gen, MirrorTransform((0, 1)))
equivalent = True
for b in range(self.num_batches):
batch = next(batch_gen)
for ix in range(self.batch_size):
if (batch['data'][ix] != batch['seg'][ix]).all():
equivalent = False
self.assertTrue(equivalent, "2D images and seg were not mirrored in the same way (they should though because "
"seg needs to match the corresponding data")
def create_data_gen_pipeline(cf, cities=None, data_split='train', do_aug=True, random=True, n_batches=None):
"""
create mutli-threaded train/val/test batch generation and augmentation pipeline.
:param cities: list of strings or None
:param patient_data: dictionary containing one dictionary per patient in the train/test subset
:param test_pids: (optional) list of test patient ids, calls the test generator.
:param do_aug: (optional) whether to perform data augmentation (training) or not (validation/testing)
:param random: bool, whether to draw random batches or go through data linearly
:return: multithreaded_generator
"""
data_gen = BatchGenerator(cities=cities, batch_size=cf.batch_size, data_dir=cf.data_dir,
label_density=cf.label_density, data_split=data_split, resolution=cf.resolution,
gt_instances=cf.gt_instances, n_batches=n_batches, random=random)
my_transforms = []
if do_aug:
mirror_transform = MirrorTransform(axes=(3,))
my_transforms.append(mirror_transform)
spatial_transform = SpatialTransform(patch_size=cf.patch_size[-2:],
patch_center_dist_from_border=cf.da_kwargs['rand_crop_dist'],
do_elastic_deform=cf.da_kwargs['do_elastic_deform'],
alpha=cf.da_kwargs['alpha'], sigma=cf.da_kwargs['sigma'],
do_rotation=cf.da_kwargs['do_rotation'], angle_x=cf.da_kwargs['angle_x'],
angle_y=cf.da_kwargs['angle_y'], angle_z=cf.da_kwargs['angle_z'],
do_scale=cf.da_kwargs['do_scale'], scale=cf.da_kwargs['scale'],
random_crop=cf.da_kwargs['random_crop'],
border_mode_data=cf.da_kwargs['border_mode_data'],
border_mode_seg=cf.da_kwargs['border_mode_seg'],
border_cval_seg=cf.da_kwargs['border_cval_seg'])
my_transforms.append(spatial_transform)
else:
my_transforms.append(CenterCropTransform(crop_size=cf.patch_size[-2:]))
my_transforms.append(GammaTransform(cf.da_kwargs['gamma_range'], invert_image=False, per_channel=True,
retain_stats=cf.da_kwargs['gamma_retain_stats'],
p_per_sample=cf.da_kwargs['p_gamma']))
my_transforms.append(AddLossMask(cf.ignore_label))
if cf.label_switches is not None:
my_transforms.append(StochasticLabelSwitches(cf.name2trainId, cf.label_switches))
all_transforms = Compose(my_transforms)
multithreaded_generator = MultiThreadedAugmenter(data_gen, all_transforms, num_processes=cf.n_workers,
seeds=range(cf.n_workers))
return multithreaded_generator
def combine_transform():
'''
组合变换：对比度+镜像
:return:
'''
my_transforms = []
# 对比度变换
brightness_transform = ContrastAugmentationTransform((0.3, 3.),
preserve_range=True)
my_transforms.append(brightness_transform)
# 镜像变换
mirror_transform = MirrorTransform(axes=(2, 3))
my_transforms.append(mirror_transform)
all_transform = Compose(my_transforms)
batchgen = my_data_loader.DataLoader(camera(), 4)
multithreaded_generator = MultiThreadedAugmenter(batchgen, all_transform,
4, 2)
# 显示转换效果
my_data_loader.plot_batch(multithreaded_generator.__next__())
def test_image_pipeline(self):
'''
This just should not crash
:return:
'''
tr_transforms = []
tr_transforms.append(MirrorTransform())
tr_transforms.append(
TransposeAxesTransform(transpose_any_of_these=(0, 1),
p_per_sample=0.5))
composed = Compose(tr_transforms)
dl = self.dl_images
mt = MultiThreadedAugmenter(dl, composed, 4, 1, None, False)
for _ in range(50):
res = mt.next()
# let mt finish caching, otherwise it's going to print an error (which is not a problem and will not prevent
# the success of the test but it does not look pretty)
sleep(2)
def get_valid_transform(patch_size):
"""
data augmentation for validation data
inspired by:
https://github.com/MIC-DKFZ/batchgenerators/blob/master/batchgenerators/examples/brats2017/brats2017_dataloader_3D.py
:param patch_size: shape of network's input
:return list of transformations
"""
train_transforms = []
train_transforms.append(
SpatialTransform_2(patch_size,
patch_size,
do_elastic_deform=False,
deformation_scale=(0, 0),
do_rotation=False,
angle_x=(0, 0),
angle_y=(0, 0),
angle_z=(0, 0),
do_scale=False,
scale=(1.0, 1.0),
border_mode_data='constant',
border_cval_data=0,
border_mode_seg='constant',
border_cval_seg=0,
order_seg=1,
order_data=3,
random_crop=True,
p_el_per_sample=0.1,
p_rot_per_sample=0.1,
p_scale_per_sample=0.1))
train_transforms.append(MirrorTransform(axes=(0, 1)))
return Compose(train_transforms)
def _augment_data(self, batch_generator, type=None):
if self.Config.DATA_AUGMENTATION:
num_processes = 15 # 15 is a bit faster than 8 on cluster
# num_processes = multiprocessing.cpu_count() # on cluster: gives all cores, not only assigned cores
else:
num_processes = 6
tfs = []
if self.Config.NORMALIZE_DATA:
tfs.append(ZeroMeanUnitVarianceTransform(per_channel=self.Config.NORMALIZE_PER_CHANNEL))
if self.Config.SPATIAL_TRANSFORM == "SpatialTransformPeaks":
SpatialTransformUsed = SpatialTransformPeaks
elif self.Config.SPATIAL_TRANSFORM == "SpatialTransformCustom":
SpatialTransformUsed = SpatialTransformCustom
else:
SpatialTransformUsed = SpatialTransform
if self.Config.DATA_AUGMENTATION:
if type == "train":
# patch_center_dist_from_border:
# if 144/2=72 -> always exactly centered; otherwise a bit off center
# (brain can get off image and will be cut then)
if self.Config.DAUG_SCALE:
if self.Config.INPUT_RESCALING:
source_mm = 2 # for bb
target_mm = float(self.Config.RESOLUTION[:-2])
scale_factor = target_mm / source_mm
scale = (scale_factor, scale_factor)
else:
scale = (0.9, 1.5)
if self.Config.PAD_TO_SQUARE:
patch_size = self.Config.INPUT_DIM
else:
patch_size = None # keeps dimensions of the data
# spatial transform automatically crops/pads to correct size
center_dist_from_border = int(self.Config.INPUT_DIM[0] / 2.) - 10 # (144,144) -> 62
tfs.append(SpatialTransformUsed(patch_size,
patch_center_dist_from_border=center_dist_from_border,
do_elastic_deform=self.Config.DAUG_ELASTIC_DEFORM,
alpha=self.Config.DAUG_ALPHA, sigma=self.Config.DAUG_SIGMA,
do_rotation=self.Config.DAUG_ROTATE,
angle_x=self.Config.DAUG_ROTATE_ANGLE,
angle_y=self.Config.DAUG_ROTATE_ANGLE,
angle_z=self.Config.DAUG_ROTATE_ANGLE,
do_scale=True, scale=scale, border_mode_data='constant',
border_cval_data=0,
order_data=3,
border_mode_seg='constant', border_cval_seg=0,
order_seg=0, random_crop=True,
p_el_per_sample=self.Config.P_SAMP,
p_rot_per_sample=self.Config.P_SAMP,
p_scale_per_sample=self.Config.P_SAMP))
if self.Config.DAUG_RESAMPLE:
tfs.append(SimulateLowResolutionTransform(zoom_range=(0.5, 1), p_per_sample=0.2, per_channel=False))
if self.Config.DAUG_RESAMPLE_LEGACY:
tfs.append(ResampleTransformLegacy(zoom_range=(0.5, 1)))
if self.Config.DAUG_GAUSSIAN_BLUR:
tfs.append(GaussianBlurTransform(blur_sigma=self.Config.DAUG_BLUR_SIGMA,
different_sigma_per_channel=False,
p_per_sample=self.Config.P_SAMP))
if self.Config.DAUG_NOISE:
tfs.append(GaussianNoiseTransform(noise_variance=self.Config.DAUG_NOISE_VARIANCE,
p_per_sample=self.Config.P_SAMP))
if self.Config.DAUG_MIRROR:
tfs.append(MirrorTransform())
if self.Config.DAUG_FLIP_PEAKS:
tfs.append(FlipVectorAxisTransform())
tfs.append(NumpyToTensor(keys=["data", "seg"], cast_to="float"))
#num_cached_per_queue 1 or 2 does not really make a difference
batch_gen = MultiThreadedAugmenter(batch_generator, Compose(tfs), num_processes=num_processes,
num_cached_per_queue=1, seeds=None, pin_memory=True)
return batch_gen # data: (batch_size, channels, x, y), seg: (batch_size, channels, x, y)
def run(self, img_data, seg_data):
# Define label for segmentation for segmentation augmentation
if self.seg_augmentation: seg_label = "seg"
else: seg_label = "class"
# Create a parser for the batchgenerators module
data_generator = DataParser(img_data, seg_data, seg_label)
# Initialize empty transform list
transforms = []
# Add mirror augmentation
if self.mirror:
aug_mirror = MirrorTransform(axes=self.config_mirror_axes)
transforms.append(aug_mirror)
# Add contrast augmentation
if self.contrast:
aug_contrast = ContrastAugmentationTransform(
self.config_contrast_range,
preserve_range=self.config_contrast_preserverange,
per_channel=self.coloraug_per_channel,
p_per_sample=self.config_p_per_sample)
transforms.append(aug_contrast)
# Add brightness augmentation
if self.brightness:
aug_brightness = BrightnessMultiplicativeTransform(
self.config_brightness_range,
per_channel=self.coloraug_per_channel,
p_per_sample=self.config_p_per_sample)
transforms.append(aug_brightness)
# Add gamma augmentation
if self.gamma:
aug_gamma = GammaTransform(self.config_gamma_range,
invert_image=False,
per_channel=self.coloraug_per_channel,
retain_stats=True,
p_per_sample=self.config_p_per_sample)
transforms.append(aug_gamma)
# Add gaussian noise augmentation
if self.gaussian_noise:
aug_gaussian_noise = GaussianNoiseTransform(
self.config_gaussian_noise_range,
p_per_sample=self.config_p_per_sample)
transforms.append(aug_gaussian_noise)
# Add spatial transformations as augmentation
# (rotation, scaling, elastic deformation)
if self.rotations or self.scaling or self.elastic_deform or \
self.cropping:
# Identify patch shape (full image or cropping)
if self.cropping: patch_shape = self.cropping_patch_shape
else: patch_shape = img_data[0].shape[0:-1]
# Assembling the spatial transformation
aug_spatial_transform = SpatialTransform(
patch_shape, [i // 2 for i in patch_shape],
do_elastic_deform=self.elastic_deform,
alpha=self.config_elastic_deform_alpha,
sigma=self.config_elastic_deform_sigma,
do_rotation=self.rotations,
angle_x=self.config_rotations_angleX,
angle_y=self.config_rotations_angleY,
angle_z=self.config_rotations_angleZ,
do_scale=self.scaling,
scale=self.config_scaling_range,
border_mode_data='constant',
border_cval_data=0,
border_mode_seg='constant',
border_cval_seg=0,
order_data=3,
order_seg=0,
p_el_per_sample=self.config_p_per_sample,
p_rot_per_sample=self.config_p_per_sample,
p_scale_per_sample=self.config_p_per_sample,
random_crop=self.cropping)
# Append spatial transformation to transformation list
transforms.append(aug_spatial_transform)
# Compose the batchgenerators transforms
all_transforms = Compose(transforms)
# Assemble transforms into a augmentation generator
augmentation_generator = SingleThreadedAugmenter(
data_generator, all_transforms)
# Perform the data augmentation x times (x = cycles)
aug_img_data = None
aug_seg_data = None
for i in range(0, self.cycles):
# Run the computation process for the data augmentations
augmentation = next(augmentation_generator)
# Access augmentated data from the batchgenerators data structure
if aug_img_data is None and aug_seg_data is None:
aug_img_data = augmentation["data"]
aug_seg_data = augmentation[seg_label]
# Concatenate the new data augmentated data with the cached data
else:
aug_img_data = np.concatenate(
(augmentation["data"], aug_img_data), axis=0)
aug_seg_data = np.concatenate(
(augmentation[seg_label], aug_seg_data), axis=0)
# Transform data from channel-first back to channel-last structure
# Data structure channel-first 3D: (batch, channel, x, y, z)
# Data structure channel-last 3D: (batch, x, y, z, channel)
aug_img_data = np.moveaxis(aug_img_data, 1, -1)
aug_seg_data = np.moveaxis(aug_seg_data, 1, -1)
# Return augmentated image and segmentation data
return aug_img_data, aug_seg_data
def main(args):
########################################
# #
# DEFINE THE HYPERPARAMETERS #
# #
########################################
# load settings from config file
config_file = args.get("config_file")
config_handler = ConfigHandler()
config_dict = config_handler(config_file)
# some are changed rarely and given manually if required
train_size = args.get("train_size")
val_size = args.get("val_size")
margin = args.get("margin")
optimizer = args.get("optimizer")
if optimizer != "SGD" and optimizer != "Adam":
ValueError("Invalid optimizer")
elif optimizer == "Adam":
optimizer_cls = torch.optim.Adam
else:
optimizer_cls = torch.optim.SGD
params = Parameters(
fixed_params={
"model": config_dict["model"],
"training": {
**config_dict["training"],
"optimizer_cls": optimizer_cls,
**config_dict["optimizer"],
"criterions": {
"FocalLoss": losses.FocalLoss(),
"SmoothL1Loss": losses.SmoothL1Loss()
},
# "criterions": {"FocalMSELoss": losses.FocalMSELoss(),
# "SmoothL1Loss": losses.SmoothL1Loss()},
# "lr_sched_cls": ReduceLROnPlateauCallbackPyTorch,
# "lr_sched_params": {"verbose": True},
"lr_sched_cls": None,
"lr_sched_params": {},
"metrics": {}
}
})
########################################
# #
# DEFINE THE AUGMENTATIONS #
# #
########################################
my_transforms = []
mirror_transform = MirrorTransform(axes=(1, 2))
my_transforms.append(mirror_transform)
crop_size = config_dict["data"]["crop_size"]
img_shape = config_dict["data"]["img_shape"]
shape_limit = config_dict["data"]["shape_limit"]
if (crop_size is not None and crop_size[0] == crop_size[1]) or \
(img_shape is not None and len(img_shape) > 1
and img_shape[0] == img_shape[1]):
rot_transform = Rot90Transform(axes=(0, 1), p_per_sample=0.5)
my_transforms.append(rot_transform)
else:
rot_transform = Rot90Transform(axes=(0, 1),
num_rot=(0, 2),
p_per_sample=0.5)
my_transforms.append(rot_transform)
# apply a more extended augmentation (if desiered)
if "ext_aug" in config_dict["data"].keys() and \
config_dict["data"]["ext_aug"] is not None and \
config_dict["data"]["ext_aug"]:
if crop_size is not None:
size = [crop_size[0] + 25, crop_size[1] + 25]
elif img_shape is not None:
size = [img_shape[0] + 5, img_shape[1] + 5]
elif shape_limit is not None:
size = [shape_limit[0] + 5, shape_limit[1] + 5]
else:
raise KeyError("Crop size or image shape requried!")
if crop_size is not None:
spatial_transforms = SpatialTransform([size[0] - 25, size[1] - 25],
np.asarray(size) // 2,
do_elastic_deform=False,
do_rotation=True,
angle_x=(0, 0.01 * np.pi),
do_scale=True,
scale=(0.9, 1.1),
random_crop=True,
border_mode_data="mirror",
border_mode_seg="mirror")
my_transforms.append(spatial_transforms)
elif img_shape is not None or shape_limit is not None:
spatial_transforms = SpatialTransform(
[size[0] - 5, size[1] - 5],
np.asarray(size) // 2,
do_elastic_deform=False,
do_rotation=False,
#angle_x=(0, 0.01 * np.pi),
do_scale=True,
scale=(0.9, 1.1),
random_crop=True,
border_mode_data="constant",
border_mode_seg="nearest")
my_transforms.append(spatial_transforms)
# bbox generation
bb_transform = ConvertSegToBB(dim=2, margin=margin)
my_transforms.append(bb_transform)
transforms = Compose(my_transforms)
########################################
# #
# DEFINE THE DATASETS and MANAGER #
# #
########################################
# paths to csv files containing labels (and other information)
csv_calc_train = '/home/temp/moriz/data/' \
'calc_case_description_train_set.csv'
csv_mass_train = '/home/temp/moriz/data/' \
'mass_case_description_train_set.csv'
# path to data directory
ddsm_dir = '/home/temp/moriz/data/CBIS-DDSM/'
# path to data directory
inbreast_dir = '/images/Mammography/INbreast/AllDICOMs/'
# paths to csv files containing labels (and other information)
xls_file = '/images/Mammography/INbreast/INbreast.xls'
# determine class and load function
if config_dict["data"]["dataset_type"] == "INbreast":
dataset_cls = CacheINbreastDataset
data_dir = inbreast_dir
csv_file = None
if config_dict["data"]["level"] == "crops":
load_fn = inbreast_utils.load_pos_crops
elif config_dict["data"]["level"] == "images":
load_fn = inbreast_utils.load_sample
elif config_dict["data"]["level"] == "both":
#TODO: fix
load_fn = inbreast_utils.load_sample_and_crops
else:
raise TypeError("Level required!")
elif config_dict["data"]["dataset_type"] == "DDSM":
data_dir = ddsm_dir
if config_dict["data"]["level"] == "crops":
load_fn = ddsm_utils.load_pos_crops
elif config_dict["data"]["level"] == "images":
load_fn = ddsm_utils.load_sample
elif config_dict["data"]["level"] == "images+":
load_fn = ddsm_utils.load_sample_with_crops
else:
raise TypeError("Level required!")
if config_dict["data"]["type"] == "mass":
csv_file = csv_mass_train
elif config_dict["data"]["type"] == "calc":
csv_file = csv_calc_train
elif config_dict["data"]["type"] == "both":
raise NotImplementedError("Todo")
else:
raise TypeError("Unknown lesion type!")
if "mode" in config_dict["data"].keys():
if config_dict["data"]["mode"] == "lazy":
dataset_cls = LazyDDSMDataset
if config_dict["data"]["level"] == "crops":
load_fn = ddsm_utils.load_single_pos_crops
elif config_dict["data"]["mode"] == "cache":
dataset_cls = CacheDDSMDataset
else:
raise TypeError("Unsupported loading mode!")
else:
dataset_cls = CacheDDSMDataset
else:
raise TypeError("Dataset is not supported!")
dataset_train_dict = {
'data_path': data_dir,
'xls_file': xls_file,
'csv_file': csv_file,
'load_fn': load_fn,
'num_elements': config_dict["debug"]["n_train"],
**config_dict["data"]
}
dataset_val_dict = {
'data_path': data_dir,
'xls_file': xls_file,
'csv_file': csv_file,
'load_fn': load_fn,
'num_elements': config_dict["debug"]["n_val"],
**config_dict["data"]
}
datamgr_train_dict = {
'batch_size': params.nested_get("batch_size"),
'n_process_augmentation': 4,
'transforms': transforms,
'sampler_cls': RandomSampler,
'data_loader_cls': BaseDataLoader
}
datamgr_val_dict = {
'batch_size': params.nested_get("batch_size"),
'n_process_augmentation': 4,
'transforms': transforms,
'sampler_cls': SequentialSampler,
'data_loader_cls': BaseDataLoader
}
########################################
# #
# INITIALIZE THE ACTUAL EXPERIMENT #
# #
########################################
checkpoint_path = config_dict["checkpoint_path"]["path"]
# if "checkpoint_path" in args and args["checkpoint_path"] is not None:
# checkpoint_path = args.get("checkpoint_path")
experiment = \
RetinaNetExperiment(params,
RetinaNet,
name = config_dict["logging"]["name"],
save_path = checkpoint_path,
dataset_cls=dataset_cls,
dataset_train_kwargs=dataset_train_dict,
datamgr_train_kwargs=datamgr_train_dict,
dataset_val_kwargs=dataset_val_dict,
datamgr_val_kwargs=datamgr_val_dict,
optim_builder=create_optims_default_pytorch,
gpu_ids=list(range(args.get('gpus'))),
val_score_key="val_FocalLoss",
val_score_mode="lowest",
checkpoint_freq=2)
########################################
# #
# LOGGING DEFINITION AND CONFIGURATION #
# #
########################################
logger_kwargs = config_dict["logging"]
# setup initial logging
log_file = os.path.join(experiment.save_path, 'logger.log')
logging.basicConfig(level=logging.INFO,
handlers=[
TrixiHandler(PytorchVisdomLogger,
**config_dict["logging"]),
logging.StreamHandler(),
logging.FileHandler(log_file)
])
logger = logging.getLogger("RetinaNet Logger")
with open(experiment.save_path + "/config.yml", 'w') as file:
yaml.dump(config_dict, file)
########################################
# #
# LOAD PATHS AND EXECUTE MODEL #
# #
########################################
seed = config_dict["data"]["seed"]
if "train_size" in config_dict["data"].keys():
train_size = config_dict["data"]["train_size"]
if "val_size" in config_dict["data"].keys():
val_size = config_dict["data"]["val_size"]
if config_dict["data"]["dataset_type"] == "INbreast":
if not config_dict["kfold"]["enable"]:
train_paths, _, val_paths = \
inbreast_utils.load_single_set(inbreast_dir,
xls_file=xls_file,
train_size=train_size,
val_size=val_size,
type=config_dict["data"]["type"],
random_state=seed)
if img_shape is not None or crop_size is not None:
experiment.run(train_paths, val_paths)
else:
experiment.run(train_paths, None)
else:
paths = inbreast_utils.get_paths(inbreast_dir,
xls_file=xls_file,
type=config_dict["data"]["type"])
if "splits" in config_dict["kfold"].keys():
num_splits = config_dict["kfold"]["splits"]
else:
num_splits = 5
experiment.kfold(paths,
num_splits=num_splits,
random_seed=seed,
dataset_type="INbreast")
else:
train_paths, val_paths, _ = \
ddsm_utils.load_single_set(ddsm_dir,
csv_file=csv_file,
train_size=train_size,
val_size=None,
random_state=seed)
if img_shape is not None or crop_size is not None:
experiment.run(train_paths, val_paths)
else:
experiment.run(train_paths, None)
def get_batches(self,
batch_size=128,
type=None,
subjects=None,
num_batches=None):
data = subjects
seg = []
#6 -> >30GB RAM
if self.HP.DATA_AUGMENTATION:
num_processes = 8 # 6 is a bit faster than 16
else:
num_processes = 6
nr_of_samples = len(subjects) * self.HP.INPUT_DIM[0]
if num_batches is None:
num_batches_multithr = int(
nr_of_samples / batch_size /
num_processes) #number of batches for exactly one epoch
else:
num_batches_multithr = int(num_batches / num_processes)
if self.HP.TYPE == "combined":
# Simple with .npy -> just a little bit faster than Nifti (<10%) and f1 not better => use Nifti
# batch_gen = SlicesBatchGeneratorRandomNpyImg_fusion((data, seg), batch_size=batch_size)
batch_gen = SlicesBatchGeneratorRandomNpyImg_fusion(
(data, seg), batch_size=batch_size)
else:
batch_gen = SlicesBatchGeneratorRandomNiftiImg(
(data, seg), batch_size=batch_size)
# batch_gen = SlicesBatchGeneratorRandomNiftiImg_5slices((data, seg), batch_size=batch_size)
batch_gen.HP = self.HP
tfs = [] #transforms
if self.HP.NORMALIZE_DATA:
tfs.append(
ZeroMeanUnitVarianceTransform(
per_channel=self.HP.NORMALIZE_PER_CHANNEL))
if self.HP.DATASET == "Schizo" and self.HP.RESOLUTION == "2mm":
tfs.append(PadToMultipleTransform(16))
if self.HP.DATA_AUGMENTATION:
if type == "train":
# scale: inverted: 0.5 -> bigger; 2 -> smaller
# patch_center_dist_from_border: if 144/2=72 -> always exactly centered; otherwise a bit off center (brain can get off image and will be cut then)
if self.HP.DAUG_SCALE:
center_dist_from_border = int(
self.HP.INPUT_DIM[0] / 2.) - 10 # (144,144) -> 62
tfs.append(
SpatialTransform(
self.HP.INPUT_DIM,
patch_center_dist_from_border=
center_dist_from_border,
do_elastic_deform=self.HP.DAUG_ELASTIC_DEFORM,
alpha=(90., 120.),
sigma=(9., 11.),
do_rotation=self.HP.DAUG_ROTATE,
angle_x=(-0.8, 0.8),
angle_y=(-0.8, 0.8),
angle_z=(-0.8, 0.8),
do_scale=True,
scale=(0.9, 1.5),
border_mode_data='constant',
border_cval_data=0,
order_data=3,
border_mode_seg='constant',
border_cval_seg=0,
order_seg=0,
random_crop=True))
if self.HP.DAUG_RESAMPLE:
tfs.append(ResampleTransform(zoom_range=(0.5, 1)))
if self.HP.DAUG_NOISE:
tfs.append(GaussianNoiseTransform(noise_variance=(0,
0.05)))
if self.HP.DAUG_MIRROR:
tfs.append(MirrorTransform())
if self.HP.DAUG_FLIP_PEAKS:
tfs.append(FlipVectorAxisTransform())
#num_cached_per_queue 1 or 2 does not really make a difference
batch_gen = MultiThreadedAugmenter(batch_gen,
Compose(tfs),
num_processes=num_processes,
num_cached_per_queue=1,
seeds=None)
return batch_gen # data: (batch_size, channels, x, y), seg: (batch_size, channels, x, y)
def _make_training_transforms(self):
if self.no_data_augmentation:
print("No data augmentation will be performed during training!")
return []
patch_size = self.patch_size[::-1] # (x, y, z) order
rot_angle_x = self.training_augmentation_args.get('angle_x', 15)
rot_angle_y = self.training_augmentation_args.get('angle_y', 15)
rot_angle_z = self.training_augmentation_args.get('angle_z', 15)
p_per_sample = self.training_augmentation_args.get(
'p_per_sample', 0.15)
train_transforms = [
SpatialTransform_2(
patch_size,
patch_size // 2,
do_elastic_deform=self.training_augmentation_args.get(
'do_elastic_deform', True),
deformation_scale=self.training_augmentation_args.get(
'deformation_scale', (0, 0.25)),
do_rotation=self.training_augmentation_args.get(
'do_rotation', True),
angle_x=(-rot_angle_x / 360. * 2 * np.pi,
rot_angle_x / 360. * 2 * np.pi),
angle_y=(-rot_angle_y / 360. * 2 * np.pi,
rot_angle_y / 360. * 2 * np.pi),
angle_z=(-rot_angle_z / 360. * 2 * np.pi,
rot_angle_z / 360. * 2 * np.pi),
do_scale=self.training_augmentation_args.get('do_scale', True),
scale=self.training_augmentation_args.get(
'scale', (0.75, 1.25)),
border_mode_data='nearest',
border_cval_data=0,
order_data=3,
# border_mode_seg='nearest', border_cval_seg=0,
# order_seg=0,
random_crop=False,
p_el_per_sample=self.training_augmentation_args.get(
'p_el_per_sample', 0.5),
p_rot_per_sample=self.training_augmentation_args.get(
'p_rot_per_sample', 0.5),
p_scale_per_sample=self.training_augmentation_args.get(
'p_scale_per_sample', 0.5))
]
if self.training_augmentation_args.get("do_mirror", False):
train_transforms.append(MirrorTransform(axes=(0, 1, 2)))
train_transforms.append(
BrightnessMultiplicativeTransform(
self.training_augmentation_args.get('brightness_range',
(0.7, 1.5)),
per_channel=True,
p_per_sample=p_per_sample))
train_transforms.append(
GaussianNoiseTransform(
noise_variance=self.training_augmentation_args.get(
'gaussian_noise_variance', (0, 0.05)),
p_per_sample=p_per_sample))
train_transforms.append(
GammaTransform(gamma_range=self.training_augmentation_args.get(
'gamma_range', (0.5, 2)),
invert_image=False,
per_channel=True,
p_per_sample=p_per_sample))
print("train_transforms\n", train_transforms)
return train_transforms
train_transform = transforms.Compose([
# mt_transforms.CenterCrop2D((200, 200)),
mt_transforms.ElasticTransform(alpha_range=(28.0, 30.0),
sigma_range=(3.5, 4.0),
p=0.3),
mt_transforms.RandomAffine(degrees=4.6,
scale=(0.98, 1.02),
translate=(0.03, 0.03)),
mt_transforms.RandomTensorChannelShift((-0.10, 0.10)),
mt_transforms.ToTensor()
# mt_transforms.NormalizeInstance(),
])
gamma_t = GammaTransform(data_key="img", gamma_range=(0.1, 10))
mirror_t = MirrorTransform(data_key="img", label_key="seg")
spatial_t = SpatialTransform(patch_size=(8, 8, 8),
data_key="img",
label_key="seg")
gauss_noise_t = GaussianNoiseTransform(data_key="img", noise_variance=(0, 1))
zoom_t = ZoomTransform(zoom_factors=2, data_key="img")
def show_basic(x, gt, info=None):
if info is not None:
print("Test for " + info)
print("img size: {}, max: {}, min: {}, avg: {}.".format(
def get_default_augmentation(dataloader_train,
dataloader_val,
patch_size,
params=default_3D_augmentation_params,
border_val_seg=-1,
pin_memory=True,
seeds_train=None,
seeds_val=None,
regions=None):
assert params.get(
'mirror') is None, "old version of params, use new keyword do_mirror"
tr_transforms = []
if params.get("selected_data_channels") is not None:
tr_transforms.append(
DataChannelSelectionTransform(
params.get("selected_data_channels")))
if params.get("selected_seg_channels") is not None:
tr_transforms.append(
SegChannelSelectionTransform(params.get("selected_seg_channels")))
# don't do color augmentations while in 2d mode with 3d data because the color channel is overloaded!!
if params.get("dummy_2D") is not None and params.get("dummy_2D"):
tr_transforms.append(Convert3DTo2DTransform())
patch_size_spatial = patch_size[1:]
else:
patch_size_spatial = patch_size
# Set order_data=0 and order_seg=0 for some more speed for cascade???
tr_transforms.append(
SpatialTransform(patch_size_spatial,
patch_center_dist_from_border=None,
do_elastic_deform=params.get("do_elastic"),
alpha=params.get("elastic_deform_alpha"),
sigma=params.get("elastic_deform_sigma"),
do_rotation=params.get("do_rotation"),
angle_x=params.get("rotation_x"),
angle_y=params.get("rotation_y"),
angle_z=params.get("rotation_z"),
do_scale=params.get("do_scaling"),
scale=params.get("scale_range"),
border_mode_data=params.get("border_mode_data"),
border_cval_data=0,
order_data=3,
border_mode_seg="constant",
border_cval_seg=border_val_seg,
order_seg=1,
random_crop=params.get("random_crop"),
p_el_per_sample=params.get("p_eldef"),
p_scale_per_sample=params.get("p_scale"),
p_rot_per_sample=params.get("p_rot"),
independent_scale_for_each_axis=params.get(
"independent_scale_factor_for_each_axis")))
if params.get("dummy_2D") is not None and params.get("dummy_2D"):
tr_transforms.append(Convert2DTo3DTransform())
if params.get("do_gamma"):
tr_transforms.append(
GammaTransform(params.get("gamma_range"),
False,
True,
retain_stats=params.get("gamma_retain_stats"),
p_per_sample=params["p_gamma"]))
if params.get("do_mirror"):
tr_transforms.append(MirrorTransform(params.get("mirror_axes")))
if params.get("mask_was_used_for_normalization") is not None:
mask_was_used_for_normalization = params.get(
"mask_was_used_for_normalization")
tr_transforms.append(
MaskTransform(mask_was_used_for_normalization,
mask_idx_in_seg=0,
set_outside_to=0))
tr_transforms.append(RemoveLabelTransform(-1, 0))
if params.get("move_last_seg_chanel_to_data") is not None and params.get(
"move_last_seg_chanel_to_data"):
tr_transforms.append(
MoveSegAsOneHotToData(1, params.get("all_segmentation_labels"),
'seg', 'data'))
if params.get("cascade_do_cascade_augmentations"
) and not None and params.get(
"cascade_do_cascade_augmentations"):
# Remove the following transforms to remove cascade DA ??
tr_transforms.append(
ApplyRandomBinaryOperatorTransform(
channel_idx=list(
range(-len(params.get("all_segmentation_labels")), 0)),
p_per_sample=params.get(
"cascade_random_binary_transform_p"),
key="data",
strel_size=params.get(
"cascade_random_binary_transform_size")))
tr_transforms.append(
RemoveRandomConnectedComponentFromOneHotEncodingTransform(
channel_idx=list(
range(-len(params.get("all_segmentation_labels")), 0)),
key="data",
p_per_sample=params.get("cascade_remove_conn_comp_p"),
fill_with_other_class_p=params.get(
"cascade_remove_conn_comp_max_size_percent_threshold"),
dont_do_if_covers_more_than_X_percent=params.get(
"cascade_remove_conn_comp_fill_with_other_class_p")))
tr_transforms.append(RenameTransform('seg', 'target', True))
if regions is not None:
tr_transforms.append(
ConvertSegmentationToRegionsTransform(regions, 'target', 'target'))
tr_transforms.append(NumpyToTensor(['data', 'target'], 'float'))
tr_transforms = Compose(tr_transforms)
# from batchgenerators.dataloading import SingleThreadedAugmenter
# batchgenerator_train = SingleThreadedAugmenter(dataloader_train, tr_transforms)
# import IPython;IPython.embed()
batchgenerator_train = MultiThreadedAugmenter(
dataloader_train,
tr_transforms,
params.get('num_threads'),
params.get("num_cached_per_thread"),
seeds=seeds_train,
pin_memory=pin_memory)
val_transforms = []
val_transforms.append(RemoveLabelTransform(-1, 0))
if params.get("selected_data_channels") is not None:
val_transforms.append(
DataChannelSelectionTransform(
params.get("selected_data_channels")))
if params.get("selected_seg_channels") is not None:
val_transforms.append(
SegChannelSelectionTransform(params.get("selected_seg_channels")))
if params.get("move_last_seg_chanel_to_data") is not None and params.get(
"move_last_seg_chanel_to_data"):
val_transforms.append(
MoveSegAsOneHotToData(1, params.get("all_segmentation_labels"),
'seg', 'data'))
val_transforms.append(RenameTransform('seg', 'target', True))
if regions is not None:
val_transforms.append(
ConvertSegmentationToRegionsTransform(regions, 'target', 'target'))
val_transforms.append(NumpyToTensor(['data', 'target'], 'float'))
val_transforms = Compose(val_transforms)
# batchgenerator_val = SingleThreadedAugmenter(dataloader_val, val_transforms)
batchgenerator_val = MultiThreadedAugmenter(
dataloader_val,
val_transforms,
max(params.get('num_threads') // 2, 1),
params.get("num_cached_per_thread"),
seeds=seeds_val,
pin_memory=pin_memory)
return batchgenerator_train, batchgenerator_val
},
encode_block=ResBlockStack,
encode_kwargs_fn=encode_kwargs_fn,
decode_block=ResBlock).cuda()
patch_size = (160, 160, 80)
optimizer = optim.Adam(model.parameters(), lr=1e-4)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
factor=0.2,
patience=30)
tr_transform = Compose([
GammaTransform((0.9, 1.1)),
ContrastAugmentationTransform((0.9, 1.1)),
BrightnessMultiplicativeTransform((0.9, 1.1)),
MirrorTransform(axes=[0]),
SpatialTransform_2(
patch_size,
(90, 90, 50),
random_crop=True,
do_elastic_deform=True,
deformation_scale=(0, 0.05),
do_rotation=True,
angle_x=(-0.1 * np.pi, 0.1 * np.pi),
angle_y=(0, 0),
angle_z=(0, 0),
do_scale=True,
scale=(0.9, 1.1),
border_mode_data='constant',
),
RandomCropTransform(crop_size=patch_size),
def get_moreDA_augmentation(dataloader_train,
dataloader_val,
patch_size,
params=default_3D_augmentation_params,
border_val_seg=-1,
seeds_train=None,
seeds_val=None,
order_seg=1,
order_data=3,
deep_supervision_scales=None,
soft_ds=False,
classes=None,
pin_memory=True,
regions=None,
use_nondetMultiThreadedAugmenter: bool = False):
assert params.get(
'mirror') is None, "old version of params, use new keyword do_mirror"
tr_transforms = []
if params.get("selected_data_channels") is not None:
tr_transforms.append(
DataChannelSelectionTransform(
params.get("selected_data_channels")))
if params.get("selected_seg_channels") is not None:
tr_transforms.append(
SegChannelSelectionTransform(params.get("selected_seg_channels")))
# don't do color augmentations while in 2d mode with 3d data because the color channel is overloaded!!
if params.get("dummy_2D") is not None and params.get("dummy_2D"):
ignore_axes = (0, )
tr_transforms.append(Convert3DTo2DTransform())
patch_size_spatial = patch_size[1:]
else:
patch_size_spatial = patch_size
ignore_axes = None
tr_transforms.append(
SpatialTransform(patch_size_spatial,
patch_center_dist_from_border=None,
do_elastic_deform=params.get("do_elastic"),
alpha=params.get("elastic_deform_alpha"),
sigma=params.get("elastic_deform_sigma"),
do_rotation=params.get("do_rotation"),
angle_x=params.get("rotation_x"),
angle_y=params.get("rotation_y"),
angle_z=params.get("rotation_z"),
p_rot_per_axis=params.get("rotation_p_per_axis"),
do_scale=params.get("do_scaling"),
scale=params.get("scale_range"),
border_mode_data=params.get("border_mode_data"),
border_cval_data=0,
order_data=order_data,
border_mode_seg="constant",
border_cval_seg=border_val_seg,
order_seg=order_seg,
random_crop=params.get("random_crop"),
p_el_per_sample=params.get("p_eldef"),
p_scale_per_sample=params.get("p_scale"),
p_rot_per_sample=params.get("p_rot"),
independent_scale_for_each_axis=params.get(
"independent_scale_factor_for_each_axis")))
if params.get("dummy_2D"):
tr_transforms.append(Convert2DTo3DTransform())
# we need to put the color augmentations after the dummy 2d part (if applicable). Otherwise the overloaded color
# channel gets in the way
tr_transforms.append(GaussianNoiseTransform(p_per_sample=0.1))
tr_transforms.append(
GaussianBlurTransform((0.5, 1.),
different_sigma_per_channel=True,
p_per_sample=0.2,
p_per_channel=0.5))
tr_transforms.append(
BrightnessMultiplicativeTransform(multiplier_range=(0.75, 1.25),
p_per_sample=0.15))
if params.get("do_additive_brightness"):
tr_transforms.append(
BrightnessTransform(
params.get("additive_brightness_mu"),
params.get("additive_brightness_sigma"),
True,
p_per_sample=params.get("additive_brightness_p_per_sample"),
p_per_channel=params.get("additive_brightness_p_per_channel")))
tr_transforms.append(ContrastAugmentationTransform(p_per_sample=0.15))
tr_transforms.append(
SimulateLowResolutionTransform(zoom_range=(0.5, 1),
per_channel=True,
p_per_channel=0.5,
order_downsample=0,
order_upsample=3,
p_per_sample=0.25,
ignore_axes=ignore_axes))
tr_transforms.append(
GammaTransform(params.get("gamma_range"),
True,
True,
retain_stats=params.get("gamma_retain_stats"),
p_per_sample=0.1)) # inverted gamma
if params.get("do_gamma"):
tr_transforms.append(
GammaTransform(params.get("gamma_range"),
False,
True,
retain_stats=params.get("gamma_retain_stats"),
p_per_sample=params["p_gamma"]))
if params.get("do_mirror") or params.get("mirror"):
tr_transforms.append(MirrorTransform(params.get("mirror_axes")))
if params.get("mask_was_used_for_normalization") is not None:
mask_was_used_for_normalization = params.get(
"mask_was_used_for_normalization")
tr_transforms.append(
MaskTransform(mask_was_used_for_normalization,
mask_idx_in_seg=0,
set_outside_to=0))
tr_transforms.append(RemoveLabelTransform(-1, 0))
if params.get("move_last_seg_chanel_to_data") is not None and params.get(
"move_last_seg_chanel_to_data"):
tr_transforms.append(
MoveSegAsOneHotToData(1, params.get("all_segmentation_labels"),
'seg', 'data'))
if params.get(
"cascade_do_cascade_augmentations") is not None and params.get(
"cascade_do_cascade_augmentations"):
if params.get("cascade_random_binary_transform_p") > 0:
tr_transforms.append(
ApplyRandomBinaryOperatorTransform(
channel_idx=list(
range(-len(params.get("all_segmentation_labels")),
0)),
p_per_sample=params.get(
"cascade_random_binary_transform_p"),
key="data",
strel_size=params.get(
"cascade_random_binary_transform_size"),
p_per_label=params.get(
"cascade_random_binary_transform_p_per_label")))
if params.get("cascade_remove_conn_comp_p") > 0:
tr_transforms.append(
RemoveRandomConnectedComponentFromOneHotEncodingTransform(
channel_idx=list(
range(-len(params.get("all_segmentation_labels")),
0)),
key="data",
p_per_sample=params.get("cascade_remove_conn_comp_p"),
fill_with_other_class_p=params.get(
"cascade_remove_conn_comp_max_size_percent_threshold"
),
dont_do_if_covers_more_than_X_percent=params.get(
"cascade_remove_conn_comp_fill_with_other_class_p")
))
tr_transforms.append(RenameTransform('seg', 'target', True))
if regions is not None:
tr_transforms.append(
ConvertSegmentationToRegionsTransform(regions, 'target', 'target'))
if deep_supervision_scales is not None:
if soft_ds:
assert classes is not None
tr_transforms.append(
DownsampleSegForDSTransform3(deep_supervision_scales, 'target',
'target', classes))
else:
tr_transforms.append(
DownsampleSegForDSTransform2(deep_supervision_scales,
0,
input_key='target',
output_key='target'))
tr_transforms.append(NumpyToTensor(['data', 'target'], 'float'))
tr_transforms = Compose(tr_transforms)
if use_nondetMultiThreadedAugmenter:
if NonDetMultiThreadedAugmenter is None:
raise RuntimeError(
'NonDetMultiThreadedAugmenter is not yet available')
batchgenerator_train = NonDetMultiThreadedAugmenter(
dataloader_train,
tr_transforms,
params.get('num_threads'),
params.get("num_cached_per_thread"),
seeds=seeds_train,
pin_memory=pin_memory)
else:
batchgenerator_train = MultiThreadedAugmenter(
dataloader_train,
tr_transforms,
params.get('num_threads'),
params.get("num_cached_per_thread"),
seeds=seeds_train,
pin_memory=pin_memory)
# batchgenerator_train = SingleThreadedAugmenter(dataloader_train, tr_transforms)
# import IPython;IPython.embed()
val_transforms = []
val_transforms.append(RemoveLabelTransform(-1, 0))
if params.get("selected_data_channels") is not None:
val_transforms.append(
DataChannelSelectionTransform(
params.get("selected_data_channels")))
if params.get("selected_seg_channels") is not None:
val_transforms.append(
SegChannelSelectionTransform(params.get("selected_seg_channels")))
if params.get("move_last_seg_chanel_to_data") is not None and params.get(
"move_last_seg_chanel_to_data"):
val_transforms.append(
MoveSegAsOneHotToData(1, params.get("all_segmentation_labels"),
'seg', 'data'))
val_transforms.append(RenameTransform('seg', 'target', True))
if regions is not None:
val_transforms.append(
ConvertSegmentationToRegionsTransform(regions, 'target', 'target'))
if deep_supervision_scales is not None:
if soft_ds:
assert classes is not None
val_transforms.append(
DownsampleSegForDSTransform3(deep_supervision_scales, 'target',
'target', classes))
else:
val_transforms.append(
DownsampleSegForDSTransform2(deep_supervision_scales,
0,
input_key='target',
output_key='target'))
val_transforms.append(NumpyToTensor(['data', 'target'], 'float'))
val_transforms = Compose(val_transforms)
if use_nondetMultiThreadedAugmenter:
if NonDetMultiThreadedAugmenter is None:
raise RuntimeError(
'NonDetMultiThreadedAugmenter is not yet available')
batchgenerator_val = NonDetMultiThreadedAugmenter(
dataloader_val,
val_transforms,
max(params.get('num_threads') // 2, 1),
params.get("num_cached_per_thread"),
seeds=seeds_val,
pin_memory=pin_memory)
else:
batchgenerator_val = MultiThreadedAugmenter(
dataloader_val,
val_transforms,
max(params.get('num_threads') // 2, 1),
params.get("num_cached_per_thread"),
seeds=seeds_val,
pin_memory=pin_memory)
# batchgenerator_val = SingleThreadedAugmenter(dataloader_val, val_transforms)
return batchgenerator_train, batchgenerator_val
do_rotation=True,
angle_z=(0, 2 * np.pi), # 旋转
do_scale=True,
scale=(0.3, 3.), # 缩放
border_mode_data='constant',
border_cval_data=0,
order_data=1,
random_crop=False)
my_transforms.append(spatial_transform)
GaussianNoise = GaussianNoiseTransform() # 高斯噪声
my_transforms.append(GaussianNoise)
GaussianBlur = GaussianBlurTransform() # 高斯模糊
my_transforms.append(GaussianBlur)
Brightness = BrightnessTransform(0, 0.2) # 亮度
my_transforms.append(Brightness)
brightness_transform = ContrastAugmentationTransform(
(0.3, 3.), preserve_range=True) # 对比度
my_transforms.append(brightness_transform)
SimulateLowResolution = SimulateLowResolutionTransform() # 低分辨率
my_transforms.append(SimulateLowResolution)
Gamma = GammaTransform() # 伽马增强
my_transforms.append(Gamma)
mirror_transform = MirrorTransform(axes=(0, 1)) # 镜像
my_transforms.append(mirror_transform)
all_transforms = Compose(my_transforms)
multithreaded_generator = MultiThreadedAugmenter(batchgen, all_transforms, 1,
2)
t = multithreaded_generator.next()
plot_batch(t)
angle_z=(-5 / 360. * 2 * np.pi, 5 / 360. * 2 * np.pi),
do_scale=True,
scale=(0.9, 1.02),
border_mode_data='constant',
border_cval_data=0,
border_mode_seg='constant',
border_cval_seg=0,
order_seg=1,
order_data=3,
random_crop=False,
p_el_per_sample=0.1,
p_rot_per_sample=0.1,
p_scale_per_sample=0.1)
my_transforms.append(spatial_transform)
my_transforms.append(MirrorTransform(axes=(0, 1, 2)))
my_transforms.append(
GammaTransform(gamma_range=(0.7, 1.),
invert_image=False,
per_channel=True,
p_per_sample=0.1))
all_transforms = Compose(my_transforms)
train_loader = SingleThreadedAugmenter(
batchgen, all_transforms
) #data loader for training, applying on the fly transformation
# add other data loaders
test_loader = torch.utils.data.DataLoader(
dataset_test,
def get_train_transforms(self) -> List[AbstractTransform]:
# used for transpost and rot90
matching_axes = np.array(
[sum([i == j for j in self.patch_size]) for i in self.patch_size])
valid_axes = list(np.where(matching_axes == np.max(matching_axes))[0])
tr_transforms = []
if self.data_aug_params['selected_seg_channels'] is not None:
tr_transforms.append(
SegChannelSelectionTransform(
self.data_aug_params['selected_seg_channels']))
if self.do_dummy_2D_aug:
ignore_axes = (0, )
tr_transforms.append(Convert3DTo2DTransform())
patch_size_spatial = self.patch_size[1:]
else:
patch_size_spatial = self.patch_size
ignore_axes = None
tr_transforms.append(
SpatialTransform(
patch_size_spatial,
patch_center_dist_from_border=None,
do_elastic_deform=False,
do_rotation=True,
angle_x=self.data_aug_params["rotation_x"],
angle_y=self.data_aug_params["rotation_y"],
angle_z=self.data_aug_params["rotation_z"],
p_rot_per_axis=0.5,
do_scale=True,
scale=self.data_aug_params['scale_range'],
border_mode_data="constant",
border_cval_data=0,
order_data=3,
border_mode_seg="constant",
border_cval_seg=-1,
order_seg=1,
random_crop=False,
p_el_per_sample=0.2,
p_scale_per_sample=0.2,
p_rot_per_sample=0.4,
independent_scale_for_each_axis=True,
))
if self.do_dummy_2D_aug:
tr_transforms.append(Convert2DTo3DTransform())
if np.any(matching_axes > 1):
tr_transforms.append(
Rot90Transform((0, 1, 2, 3),
axes=valid_axes,
data_key='data',
label_key='seg',
p_per_sample=0.5), )
if np.any(matching_axes > 1):
tr_transforms.append(
TransposeAxesTransform(valid_axes,
data_key='data',
label_key='seg',
p_per_sample=0.5))
tr_transforms.append(
OneOfTransform([
MedianFilterTransform((2, 8),
same_for_each_channel=False,
p_per_sample=0.2,
p_per_channel=0.5),
GaussianBlurTransform((0.3, 1.5),
different_sigma_per_channel=True,
p_per_sample=0.2,
p_per_channel=0.5)
]))
tr_transforms.append(GaussianNoiseTransform(p_per_sample=0.1))
tr_transforms.append(
BrightnessTransform(0,
0.5,
per_channel=True,
p_per_sample=0.1,
p_per_channel=0.5))
tr_transforms.append(
OneOfTransform([
ContrastAugmentationTransform(contrast_range=(0.5, 2),
preserve_range=True,
per_channel=True,
data_key='data',
p_per_sample=0.2,
p_per_channel=0.5),
ContrastAugmentationTransform(contrast_range=(0.5, 2),
preserve_range=False,
per_channel=True,
data_key='data',
p_per_sample=0.2,
p_per_channel=0.5),
]))
tr_transforms.append(
SimulateLowResolutionTransform(zoom_range=(0.25, 1),
per_channel=True,
p_per_channel=0.5,
order_downsample=0,
order_upsample=3,
p_per_sample=0.15,
ignore_axes=ignore_axes))
tr_transforms.append(
GammaTransform((0.7, 1.5),
invert_image=True,
per_channel=True,
retain_stats=True,
p_per_sample=0.1))
tr_transforms.append(
GammaTransform((0.7, 1.5),
invert_image=True,
per_channel=True,
retain_stats=True,
p_per_sample=0.1))
if self.do_mirroring:
tr_transforms.append(MirrorTransform(self.mirror_axes))
tr_transforms.append(
BlankRectangleTransform([[max(1, p // 10), p // 3]
for p in self.patch_size],
rectangle_value=np.mean,
num_rectangles=(1, 5),
force_square=False,
p_per_sample=0.4,
p_per_channel=0.5))
tr_transforms.append(
BrightnessGradientAdditiveTransform(
lambda x, y: np.exp(
np.random.uniform(np.log(x[y] // 6), np.log(x[y]))),
(-0.5, 1.5),
max_strength=lambda x, y: np.random.uniform(-5, -1)
if np.random.uniform() < 0.5 else np.random.uniform(1, 5),
mean_centered=False,
same_for_all_channels=False,
p_per_sample=0.3,
p_per_channel=0.5))
tr_transforms.append(
LocalGammaTransform(
lambda x, y: np.exp(
np.random.uniform(np.log(x[y] // 6), np.log(x[y]))),
(-0.5, 1.5),
lambda: np.random.uniform(0.01, 0.8)
if np.random.uniform() < 0.5 else np.random.uniform(1.5, 4),
same_for_all_channels=False,
p_per_sample=0.3,
p_per_channel=0.5))
tr_transforms.append(
SharpeningTransform(strength=(0.1, 1),
same_for_each_channel=False,
p_per_sample=0.2,
p_per_channel=0.5))
if any(self.use_mask_for_norm.values()):
tr_transforms.append(
MaskTransform(self.use_mask_for_norm,
mask_idx_in_seg=0,
set_outside_to=0))
tr_transforms.append(RemoveLabelTransform(-1, 0))
if self.data_aug_params["move_last_seg_chanel_to_data"]:
all_class_labels = np.arange(1, self.num_classes)
tr_transforms.append(
MoveSegAsOneHotToData(1, all_class_labels, 'seg', 'data'))
if self.data_aug_params["cascade_do_cascade_augmentations"]:
tr_transforms.append(
ApplyRandomBinaryOperatorTransform(channel_idx=list(
range(-len(all_class_labels), 0)),
p_per_sample=0.4,
key="data",
strel_size=(1, 8),
p_per_label=1))
tr_transforms.append(
RemoveRandomConnectedComponentFromOneHotEncodingTransform(
channel_idx=list(range(-len(all_class_labels), 0)),
key="data",
p_per_sample=0.2,
fill_with_other_class_p=0.15,
dont_do_if_covers_more_than_X_percent=0))
tr_transforms.append(RenameTransform('seg', 'target', True))
if self.regions is not None:
tr_transforms.append(
ConvertSegmentationToRegionsTransform(self.regions, 'target',
'target'))
if self.deep_supervision_scales is not None:
tr_transforms.append(
DownsampleSegForDSTransform2(self.deep_supervision_scales,
0,
input_key='target',
output_key='target'))
tr_transforms.append(NumpyToTensor(['data', 'target'], 'float'))
return tr_transforms