def encode_sample(self, sample_dict):
"""See base class for details."""
# Flatten dict matching the tf-example specs
# Use NonMutableDict to ensure there is no collision between features keys
tfexample_dict = utils.NonMutableDict()
# Iterate over sample fields
for feature_key, (feature, sample_value) in utils.zip_dict(
self._feature_dict,
sample_dict
):
# Encode the field with the associated encoder
encoded_feature = feature.encode_sample(sample_value)
# Singleton case
if not feature.specs_keys:
tfexample_dict[feature_key] = encoded_feature
# Feature contains sub features
else:
_assert_keys_match(encoded_feature.keys(), feature.specs_keys)
tfexample_dict.update({
posixpath.join(feature_key, k): encoded_feature[k]
for k in feature.specs_keys
})
return tfexample_dict
def __init__(self, split_info=None):
self._splits = utils.NonMutableDict(
error_msg="Overlap between splits. Split {key} has been added with "
"itself.")
if split_info:
self.add(SlicedSplitInfo(split_info=split_info, slice_value=None))
def get_specs(self):
"""See base class for details."""
# Flatten tf-example specs dict
# Use NonMutableDict to ensure there is no collision between features keys
specs_dict = utils.NonMutableDict()
for feature_key, feature in self._feature_dict.items():
feature_specs = feature.get_specs()
# Features can be either containers (dict of other features) or plain
# features (ex: single tensor). Plain features have a None
# feature.specs_keys
if not feature.specs_keys:
specs_dict[feature_key] = feature_specs
else:
# Sanity check which should always be True, as feature.specs_keys is
# computed using feature.get_specs()
_assert_keys_match(feature_specs.keys(), feature.specs_keys)
specs_dict.update({
posixpath.join(feature_key, k): v for k, v in feature_specs.items()
})
return specs_dict
