def dataset_val(self):
"""
Returns the validation dataset for videos. No random augmentation is performed.
:return: The validation dataset.
"""
iterator_val = IdListIterator(self.val_id_list_file_name,
random=False,
keys=['video_id', 'frame_id'])
sources = self.datasources()
generator_sources = self.data_generator_sources()
generators_val = self.data_generators(None)
image_transformation = self.spatial_transformation()
image_key = 'merged' if self.pad_image or self.crop_image_size is not None else 'image'
dataset_val = ReferenceTransformationDataset(
dim=self.dim,
reference_datasource_keys={'image': image_key},
reference_transformation=image_transformation,
datasources=sources,
data_generators=generators_val,
data_generator_sources=generator_sources,
iterator=iterator_val,
all_generators_post_processing=lambda x: self.
all_generators_post_processing(x, False),
debug_image_folder=os.path.join(self.debug_folder_prefix,
'debug_val')
if self.save_debug_images else None,
use_only_first_reference_datasource_entry=True)
return dataset_val
def dataset_val_single_frame(self):
"""
Returns the validation dataset for single frames. No random augmentation is performed.
:return: The validation dataset.
"""
sources = self.datasources_single_frame()
generator_sources = self.data_generator_sources()
generators_val = self.data_generators_single_frame(None)
image_transformation = self.spatial_transformation()
image_key = 'merged' if self.pad_image or self.crop_image_size is not None else 'image'
dataset_val = ReferenceTransformationDataset(
dim=self.dim,
reference_datasource_keys={'image': image_key},
reference_transformation=image_transformation,
datasources=sources,
data_generators=generators_val,
data_generator_sources=generator_sources,
iterator=None,
all_generators_post_processing=lambda x: self.
all_generators_post_processing(x, False),
debug_image_folder=os.path.join(self.debug_folder_prefix,
'debug_val')
if self.save_debug_images else None)
return dataset_val
def dataset_train_single_frame(self):
"""
Returns the training dataset for single frames. Random augmentation is performed.
:return: The training dataset.
"""
iterator_train = IdListIterator(self.train_id_list_file_name,
random=True,
keys=['video_id', 'frame_id'])
sources = self.datasources_single_frame()
generator_sources = self.data_generator_sources()
generators_train = self.data_generators_single_frame(
self.postprocessing_random)
image_transformation = self.spatial_transformation_augmented()
image_key = 'merged' if self.pad_image or self.crop_image_size is not None else 'image'
dataset_train = ReferenceTransformationDataset(
dim=self.dim,
reference_datasource_keys={'image': image_key},
reference_transformation=image_transformation,
datasources=sources,
data_generators=generators_train,
data_generator_sources=generator_sources,
iterator=iterator_train,
all_generators_post_processing=lambda x: self.
all_generators_post_processing(x, True),
debug_image_folder=os.path.join(self.debug_folder_prefix,
'debug_train')
if self.save_debug_images else None)
return dataset_train
def dataset_val(self):
"""
Returns the validation dataset. No random augmentation is performed.
:return: The validation dataset.
"""
iterator = IdListIterator(self.test_file,
random=False,
keys=['image_id'])
sources = self.datasources()
generator_sources = self.data_generator_sources()
generators = self.data_generators(self.postprocessing,
self.split_labels)
reference_transformation = self.spatial_transformation()
if self.cv == 0:
del sources['mask']
del generator_sources['mask']
del generators['mask']
return ReferenceTransformationDataset(
dim=self.dim,
reference_datasource_keys={
'image': 'image',
'landmarks': 'landmarks'
},
reference_transformation=reference_transformation,
datasources=sources,
data_generators=generators,
data_generator_sources=generator_sources,
iterator=iterator,
debug_image_folder='debug_val' if self.save_debug_images else None)
def dataset_train(self):
"""
Returns the training dataset. Random augmentation is performed.
:return: The training dataset.
"""
iterator = IdListIterator(self.train_file,
random=True,
keys=['image_id'])
sources = self.datasources()
generator_sources = self.data_generator_sources()
generators = self.data_generators(self.postprocessing_random,
self.split_labels)
reference_transformation = self.spatial_transformation_augmented()
return ReferenceTransformationDataset(
dim=self.dim,
reference_datasource_keys={
'image': 'image',
'landmarks': 'landmarks'
},
reference_transformation=reference_transformation,
datasources=sources,
data_generators=generators,
data_generator_sources=generator_sources,
iterator=iterator,
debug_image_folder='debug_train'
if self.save_debug_images else None)
def dataset_val_single_frame(self):
sources = self.datasources()
generator_sources = self.data_generator_sources()
generators_val = self.data_generators()
image_transformation = self.image_transformation_val()
dataset_val = ReferenceTransformationDataset(
dim=self.dim,
reference_datasource_keys={'image': 'image'},
reference_transformation=image_transformation,
datasources=sources,
data_generators=generators_val,
data_generator_sources=generator_sources,
iterator=None,
all_generators_post_processing=None,
debug_image_folder=os.path.join(self.debug_folder_prefix,
'debug_val')
if self.save_debug_images else None)
return dataset_val
def dataset_train(self):
"""
Returns the training dataset for videos. Random augmentation is performed.
:return: The training dataset.
"""
full_video_frame_list_image = VideoFrameList(
self.video_frame_list_file_name,
int(self.num_frames / 2),
int(self.num_frames / 2) - 1,
border_mode='valid',
random_start=True,
random_skip_probability=self.random_skip_probability)
iterator_train = IdListIterator(
self.train_id_list_file_name,
random=True,
keys=['video_id', 'frame_id'],
postprocessing=lambda x: full_video_frame_list_image.
get_id_dict_list(x['video_id'], x['frame_id']))
sources = self.datasources()
generator_sources = self.data_generator_sources()
generators_train = self.data_generators(self.postprocessing_random)
image_transformation = self.spatial_transformation_augmented()
image_key = 'merged' if self.pad_image or self.crop_image_size is not None else 'image'
dataset_train = ReferenceTransformationDataset(
dim=self.dim,
reference_datasource_keys={'image': image_key},
reference_transformation=image_transformation,
datasources=sources,
data_generators=generators_train,
data_generator_sources=generator_sources,
iterator=iterator_train,
all_generators_post_processing=lambda x: self.
all_generators_post_processing(x, True),
debug_image_folder=os.path.join(self.debug_folder_prefix,
'debug_train')
if self.save_debug_images else None,
use_only_first_reference_datasource_entry=True)
return dataset_train
def dataset_val(self):
"""
Returns the validation dataset. No random augmentation is performed.
:return: The validation dataset.
"""
data_sources = self.data_sources(False)
data_generator_sources = self.data_generator_sources()
data_generators = self.data_generators(self.intensity_postprocessing)
image_transformation = self.spatial_transformation()
iterator = IdListIterator(self.val_id_list_file_name,
random=False,
keys=['image_id'])
dataset = ReferenceTransformationDataset(dim=self.dim,
reference_datasource_keys={'image': 'image_datasource'},
reference_transformation=image_transformation,
datasources=data_sources,
data_generators=data_generators,
data_generator_sources=data_generator_sources,
iterator=iterator,
debug_image_folder='debug_val' if self.save_debug_images else None)
return dataset