def set_features_and_args_from_config(self, coco_config):
"""Set features and arguments based on COCO config."""
if coco_config == KEYPOINTS:
feature_names = ['image', 'person_keypoints', 'person_boxes']
elif coco_config == BOXES:
feature_names = ['image', 'boxes', 'box_labels']
elif coco_config == MSEG:
feature_names = self.MSEG_FEATURE_NAMES + ['instance_segmentation']
elif coco_config == ALL:
feature_names = self.MSEG_FEATURE_NAMES + [
'person_keypoints', 'person_boxes', 'boxes', 'box_labels',
'instance_segmentation'
]
else:
raise ValueError(f'COCO config {coco_config} not valid!')
feature_args = {}
if 'person_keypoints' in feature_names:
feature_args['person_keypoints'] = dict(
num_keypoints=len(COCO_KEYPOINT_LABELS))
if 'box_labels' in feature_names:
feature_args['box_labels'] = dict(num_box_labels=len(COCO_BOX_LABELS))
self.feature_utils = fids_features.FeatureUtils(feature_args)
self.feature_names = feature_names
missing_features = [
feature for feature in COCO_POSSIBLY_MISSING_FEATURES
if feature in feature_names
]
if missing_features:
self.splits_with_missing_features = {
'train': missing_features,
'validation': missing_features,
}
def __init__(self,
name,
config_name,
feature_names,
splits,
splits_with_missing_features=None,
feature_args=None):
"""Constructor.
Args:
name: name of the dataset.
config_name: name of the configuration of the dataset
feature_names: list of features available in this dataset.
splits: list of splits available in this dataset.
splits_with_missing_features: Dictionary which maps each split to a list
of feature_names which may be missing. By default, a missing feature
will thow an error during dataset creation.
feature_args: Dictionary, mapping feature_names to named arguments for
for each feature. Example:
feature_args={'box_labels': {'num_box_labels': 5}} See TfdsTypes
in the fids_features module for which feature_name
consumes which arguments.
"""
self.name = name
self.config_name = config_name
self.feature_utils = fids_features.FeatureUtils(feature_args)
self.feature_names = feature_names # Validation depends on feature_utils
self.splits = splits
self.splits_with_missing_features = splits_with_missing_features
# Set to positive value during debugging to limit the number of examples
# which are created per split.
self._debug_num_examples_per_split = 0
def setUp(self):
super().setUp()
feature_args = {
'box_labels': {'num_box_labels': 5},
'person_keypoints': {'num_keypoints': 16},
'person_position': {'dim_per_position': 4},
}
self.features = fids_features.FeatureUtils(feature_args)
def testGetSpecificFakeFeature(
self, feature_name, expected, feature_args=None):
feature_utils = fids_features.FeatureUtils(feature_args)
fake_feature = feature_utils.get_fake_feature(feature_name)
if isinstance(expected, np.ndarray):
self.assertAllEqual(fake_feature, expected)
elif isinstance(expected, list):
self._assertSequenceEqual(fake_feature, expected)
else:
self.assertEqual(fake_feature, expected)
def set_features_and_splits(self):
box_feature_args = {'box_labels': {'num_box_labels': len(BOX_LABELS)}}
self.feature_utils = fids_features.FeatureUtils(box_feature_args)
mseg_segmentation_features = self.MSEG_FEATURE_NAMES.copy()
mseg_segmentation_features.remove('image')
if self.bdd_type == ALL_DENSE_3K:
self.feature_names = self.MSEG_FEATURE_NAMES + [
'boxes', 'box_labels', 'drivable', 'instance_segmentation'
]
self.splits = ['train', 'test']
self.splits_with_missing_features = {
'test':
mseg_segmentation_features + [
'boxes',
'box_labels',
'drivable',
]
}
elif self.bdd_type == MSEG:
self.feature_names = self.MSEG_FEATURE_NAMES + [
'instance_segmentation'
]
self.splits = ['train', 'validation', 'test']
self.splits_with_missing_features = {
'test': mseg_segmentation_features + ['instance_segmentation']
}
elif self.bdd_type == DETECTION_100K:
self.feature_names = ['image', 'boxes', 'box_labels', 'drivable']
self.splits = ['train', 'validation', 'test']
self.splits_with_missing_features = {
'test': ['boxes', 'box_labels', 'drivable']
}
else:
assert self.bdd_type == ALL_SPARSE_107K, (
f'Unexpected BDD type {self.bdd_type}')
self.feature_names = self.MSEG_FEATURE_NAMES + [
'boxes',
'box_labels',
'drivable',
'instance_segmentation',
]
self.splits = ['train', 'validation', 'test']
missing_features = mseg_segmentation_features + [
'boxes', 'box_labels', 'drivable', 'instance_segmentation'
]
self.splits_with_missing_features = {
'train': missing_features,
'validation': missing_features,
'test': missing_features
}
def testWrongFeatureNameInFeatureArgsRaisesAttributeError(self):
with self.assertRaises(AttributeError):
fids_features.FeatureUtils({'boxlabels': {'num_box_labels': 3}})
def testWrongNamedArgsInFeatureArgsRaisesTypeError(self, feature_args):
with self.assertRaises(TypeError):
fids_features.FeatureUtils(feature_args)