def rotationCaeAugmentationsNoJitterNoFlip(dataset_config, given_shape):
"""
Applies augmentations for the training of color restoration autoencoders.
Augmentations: max_div_images --> RandomCropWithResize --> Random90xRotation --> clip_by_value
:param given_shape: (Array) Array containing the shape of the data from the input pipeline. E.g. [None,None,3]
:param dataset_config: (Dictionary) The config of the dataset to train and val on.
:return: dataset_augmentor_train: (DatasetAugmentationBuilder) Handles the augmentation of the training dataset.
:return: dataset_augmentor_val: (DatasetAugmentationBuilder) Handles the augmentation of the validation datasets.
"""
data_shape = dataset_config["dataShape"]
crop_shape = [1, data_shape[0], data_shape[1], data_shape[2]]
max_div_images = MaxDivNormalizer(255.0)
dataset_augmentor_train = DatasetAugmentationBuilder(
preprocessors=[
max_div_images,
RandomCropWithResize(crop_shape=crop_shape)
],
generators=[Random90xRotation()],
output_preprocessors=[ClipByValue()])
dataset_augmentor_val = dataset_augmentor_train
return dataset_augmentor_train, dataset_augmentor_val
def denoisingCaeAugmentationsNoJitter(dataset_config, given_shape):
"""
Applies augmentations for the training of denoising autoencoders.
Augmentations: max_div_images --> RandomCropWithResize --> RandomHorizontalFlip --> noise_images --> clip_by_value
|--> ClipByValue
:param given_shape: (Array) Array containing the shape of the data from the input pipeline. E.g. [None,None,3]
:param dataset_config: (Dictionary) The config of the dataset to train and val on.
:return: dataset_augmentor_train: (DatasetAugmentationBuilder) Handles the augmentation of the training dataset.
:return: dataset_augmentor_val: (DatasetAugmentationBuilder) Handles the augmentation of the validation datasets.
"""
data_shape = dataset_config["dataShape"]
crop_shape = [1, data_shape[0], data_shape[1], data_shape[2]]
max_div_images = MaxDivNormalizer(255.0)
noise_images = RandomNormalNoise(input_name="image",
output_name="noisy_image")
clip_by_value = ClipByValue(input_name="noisy_image",
output_name="noisy_image")
dataset_augmentor_train = DatasetAugmentationBuilder(
preprocessors=[
max_div_images,
RandomCropWithResize(crop_shape=crop_shape),
RandomHorizontalFlip()
],
generators=[noise_images],
output_preprocessors=[ClipByValue(), clip_by_value])
dataset_augmentor_val = dataset_augmentor_train
return dataset_augmentor_train, dataset_augmentor_val
def caeAugmentations(dataset_config, given_shape):
"""
Applies base augmentations for the training of plain autoencoders.
Augmentations: max_div_images --> RandomCropWithResize --> RandomHorizontalFlip --> RandomColorJitter --> ClipByValue
:param given_shape: (Array) Array containing the shape of the data from the input pipeline. E.g. [None,None,3]
:param dataset_config: (Dictionary) The config of the dataset to train and val on.
:return: dataset_augmentor_train: (DatasetAugmentationBuilder) Handles the augmentation of the training dataset.
:return: dataset_augmentor_val: (DatasetAugmentationBuilder) Handles the augmentation of the validation datasets.
"""
data_shape = dataset_config["dataShape"]
crop_shape = [1, data_shape[0], data_shape[1], data_shape[2]]
max_div_images = MaxDivNormalizer(255.0)
dataset_augmentor_train = DatasetAugmentationBuilder(preprocessors=[
max_div_images,
RandomCropWithResize(crop_shape=crop_shape),
RandomHorizontalFlip(),
RandomColorJitter(),
ClipByValue()
])
dataset_augmentor_val = dataset_augmentor_train
return dataset_augmentor_train, dataset_augmentor_val
def simClrAugmentations(dataset_config, given_shape):
"""
Applies augmentations for the training of contrastive learning.
Augmentations: max_div_images --> random_crop_1 --> random_flip_2 --> jitter_images_2 --> clip_2
|--> random_crop_2--> random_flip_1 --> jitter_images_1--> clip_1
:param given_shape: (Array) Array containing the shape of the data from the input pipeline. E.g. [None,None,3]
:param dataset_config: (Dictionary) The config of the dataset to train and val on.
:return: dataset_augmentor_train: (DatasetAugmentationBuilder) Handles the augmentation of the training dataset.
:return: dataset_augmentor_val: (DatasetAugmentationBuilder) Handles the augmentation of the validation datasets.
"""
data_shape = dataset_config["dataShape"]
crop_shape = [1, data_shape[0], data_shape[1], data_shape[2]]
max_div_images = MaxDivNormalizer(255.0)
# Order matters here! First image to image2!
random_crop_1 = RandomCropWithResize(input_name="image",
output_name="image2",
crop_shape=crop_shape)
random_crop_2 = RandomCropWithResize(input_name="image",
output_name="image",
crop_shape=crop_shape)
random_flip_1 = RandomHorizontalFlip(input_name="image",
output_name="image")
random_flip_2 = RandomHorizontalFlip(input_name="image2",
output_name="image2")
jitter_images_1 = RandomColorJitter(input_name="image",
output_name="image")
jitter_images_2 = RandomColorJitter(input_name="image2",
output_name="image2")
clip_1 = ClipByValue(input_name="image", output_name="image")
clip_2 = ClipByValue(input_name="image2", output_name="image2")
dataset_augmentor_train = DatasetAugmentationBuilder(
preprocessors=[max_div_images],
generators=[
random_crop_1, random_crop_2, random_flip_1, random_flip_2,
jitter_images_1, jitter_images_2
],
output_preprocessors=[clip_1, clip_2])
dataset_augmentor_val = dataset_augmentor_train
return dataset_augmentor_train, dataset_augmentor_val
def prepareDataset(provider,
dataset_config,
xfold_seed=None,
augment_data=False,
normalize_data="zeroMeanUnitVariance",
rescale_input=None):
"""
Prepares the given dataset for the experiment: If a xfold_seed is given the dataset will be shuffled.
:param provider: (ADataProvider) The DataProvider to load the dataset.
:param dataset_config: (Dictionary) The config of the dataset.
:param xfold_seed: (Integer) The seed for the shuffle operation. None by default.
:param augment_data: (Boolean) If True the data will be augmented as defined below. False by default.
:param normalize_data: (Boolean) If True the data will be normalized. True by default.
:param rescale_input: (Array) If set the input data will be rescaled. None by default.
:return: dataset: (Dictionary) The shuffled dataset in the given split.
:return: dataset_generator: (keras.preprocessing.image.ImageDataGenerator) The keras generator for the dataset.
"""
dataset_split = dataset_config["splitOfDataset"]
dataset = provider.getSplittedDatasetInNumpy()
if xfold_seed:
dataset = shuffelDatasetForXFold(dataset, xfold_seed, dataset_split)
if normalize_data == "zeroMeanUnitVariance":
dataset = ZeroMeanUnitVarianceNormalizer().preprocessingInNumpy(
dataset)
elif normalize_data == "maxDiv":
dataset = MaxDivNormalizer(255.0).preprocessingInNumpy(dataset)
if rescale_input:
dataset = rescaleImages(dataset, rescale_input)
if augment_data:
dataset_generator = ImageDataGenerator(
rotation_range=15,
width_shift_range=0.1,
height_shift_range=0.1,
horizontal_flip=True,
)
else:
dataset_generator = ImageDataGenerator()
dataset_generator.fit(dataset["x_train"])
return dataset, dataset_generator
def targetTaskAugmentations(dataset_config, given_shape):
"""
Applies base augmentations for the training of target models.
Train Augmentations: max_div_images --> RandomCropWithResize --> RandomHorizontalFlip --> RandomColorJitter --> ClipByValue
one_hot
Val Augmentations: max_div_images --> (CenterCropWithResize) --> (ClipByValue)
one_hot
:param given_shape: (Array) Array containing the shape of the data from the input pipeline. E.g. [None,None,3]
:param dataset_config: (Dictionary) The config of the dataset to train and val on.
:return: dataset_augmentor_train: (DatasetAugmentationBuilder) Handles the augmentation of the training dataset.
:return: dataset_augmentor_val: (DatasetAugmentationBuilder) Handles the augmentation of the validation datasets.
"""
data_shape = dataset_config["dataShape"]
crop_shape = [1, data_shape[0], data_shape[1], data_shape[2]]
if "labelName" in dataset_config.keys():
label_name = dataset_config["labelName"]
else:
label_name = "label"
max_div_images = MaxDivNormalizer(255.0)
one_hot = OneHot(dataset_config["numClasses"],
input_name=label_name,
output_name="label")
dataset_augmentor_train = DatasetAugmentationBuilder(preprocessors=[
one_hot, max_div_images,
RandomCropWithResize(crop_shape=crop_shape),
RandomHorizontalFlip(),
RandomColorJitter(),
ClipByValue()
])
val_preprocessors = [one_hot, max_div_images]
if list(given_shape) != list(data_shape):
val_preprocessors.append(CenterCropWithResize(crop_shape=crop_shape))
val_preprocessors.append(ClipByValue())
dataset_augmentor_val = DatasetAugmentationBuilder(
preprocessors=val_preprocessors)
return dataset_augmentor_train, dataset_augmentor_val
