TFDS now supports the Croissant 🥐 format! Read the documentation to know more.

tfds.features.Sequence

View source on GitHub

Composite FeatureConnector for a dict where each value is a list.

Inherits From: FeatureConnector 

tfds.features.Sequence(
 feature: feature_lib.FeatureConnectorArg,
 length: Optional[int] = None,
 *,
 doc: feature_lib.DocArg = None
)

Sequence correspond to sequence of tfds.features.FeatureConnector. At generation time, a list for each of the sequence element is given. The output of tf.data.Dataset will batch all the elements of the sequence together.

If the length of the sequence is static and known in advance, it should be specified in the constructor using the length param.

Note that Sequence does not support features which are of type tf.io.FixedLenSequenceFeature.

Example:

At construction time:

tfds.features.Sequence(tfds.features.Image(), length=NB_FRAME)

or:

tfds.features.Sequence({
 'frame': tfds.features.Image(shape=(64, 64, 3))
 'action': tfds.features.ClassLabel(['up', 'down', 'left', 'right'])
}, length=NB_FRAME)

During data generation:

yield {
 'frame': np.ones(shape=(NB_FRAME, 64, 64, 3)),
 'action': ['left', 'left', 'up', ...],
}

Tensor returned by .as_dataset():

{
 'frame': tf.Tensor(shape=(NB_FRAME, 64, 64, 3), dtype=tf.uint8),
 'action': tf.Tensor(shape=(NB_FRAME,), dtype=tf.int64),
}

At generation time, you can specify a list of features dict, a dict of list values or a stacked numpy array. The lists will automatically be distributed into their corresponding FeatureConnector.

Args

feature The features to wrap (any feature supported)
length int, length of the sequence if static and known in advance
doc Documentation of this feature (e.g. description).

Attributes

doc

dtype Return the dtype (or dict of dtype) of this FeatureConnector.
feature The inner feature.
flat_features

flat_sequence_ranks

flat_serialized_info

np_dtype

numpy_dtype

shape Return the shape (or dict of shape) of this FeatureConnector.
tf_dtype

tf_example_spec Returns the tf.Example proto structure.

Methods

catalog_documentation

View source 

catalog_documentation() -> List[feature_lib.CatalogFeatureDocumentation]

Returns the feature documentation to be shown in the catalog.

cls_from_name

View source 

@classmethod
cls_from_name(
 python_class_name: str
) -> Type['FeatureConnector']

Returns the feature class for the given Python class.

decode_batch_example

View source 

decode_batch_example(
 tfexample_data
)

Decode multiple features batched in a single tf.Tensor.

This function is used to decode features wrapped in tfds.features.Sequence(). By default, this function apply decode_example on each individual elements using tf.map_fn. However, for optimization, features can overwrite this method to apply a custom batch decoding.

Args
tfexample_data Same tf.Tensor inputs as decode_example, but with and additional first dimension for the sequence length.

Returns
tensor_data Tensor or dictionary of tensor, output of the tf.data.Dataset object

decode_example

View source 

decode_example(
 serialized_example, *, decoders=None
)

Decode the feature dict to TF compatible input.

Args
tfexample_data Data or dictionary of data, as read by the tf-example reader. It correspond to the tf.Tensor() (or dict of tf.Tensor()) extracted from the tf.train.Example, matching the info defined in get_serialized_info().

Returns
tensor_data Tensor or dictionary of tensor, output of the tf.data.Dataset object

decode_example_np

View source 

decode_example_np(
 serialized_example, *, decoders=None
)

Encode the feature dict into NumPy-compatible input.

Args
example_data Value to convert to NumPy.

Returns
np_data Data as NumPy-compatible type: either a Python primitive (bytes, int, etc) or a NumPy array.

decode_ragged_example

View source 

decode_ragged_example(
 tfexample_data
)

Decode nested features from a tf.RaggedTensor.

This function is used to decode features wrapped in nested tfds.features.Sequence(). By default, this function apply decode_batch_example on the flat values of the ragged tensor. For optimization, features can overwrite this method to apply a custom batch decoding.

Args
tfexample_data tf.RaggedTensor inputs containing the nested encoded examples.

Returns
tensor_data The decoded tf.RaggedTensor or dictionary of tensor, output of the tf.data.Dataset object

deserialize_example

View source 

deserialize_example(
 serialized_example: Union[tf.Tensor, bytes], *, decoders=None
) -> utils.TensorDict

Decodes the tf.train.Example data into tf.Tensor.

See serialize_example to encode the data into proto.

Args
serialized_example The tensor-like object containing the serialized tf.train.Example proto.
decoders Eventual decoders to apply (see documentation)

Returns
The decoded features tensors.

deserialize_example_np

View source 

deserialize_example_np(
 serialized_example: Union[tf.Tensor, bytes], *, decoders=None
) -> utils.NpArrayOrScalarDict

encode_example

View source 

encode_example(
 example_dict
)

Encode the feature dict into tf-example compatible input.

The input example_data can be anything that the user passed at data generation. For example:

For features:

features={
 'image': tfds.features.Image(),
 'custom_feature': tfds.features.CustomFeature(),
}

At data generation (in _generate_examples), if the user yields:

yield {
 'image': 'path/to/img.png',
 'custom_feature': [123, 'str', lambda x: x+1]
}

Then

Args
example_data Value or dictionary of values to convert into tf-example compatible data.

Returns
tfexample_data Data or dictionary of data to write as tf-example. Data can be a list or numpy array. Note that numpy arrays are flattened so it's the feature connector responsibility to reshape them in decode_example(). Note that tf.train.Example only supports int64, float32 and string so the data returned here should be integer, float or string. User type can be restored in decode_example().

from_config

View source 

@classmethod
from_config(
 root_dir: str
) -> FeatureConnector

Reconstructs the FeatureConnector from the config file.

Usage:

features = FeatureConnector.from_config('path/to/dir')

Args
root_dir Directory containing the features.json file.

Returns
The reconstructed feature instance.

from_json

View source 

@classmethod
from_json(
 value: Json
) -> FeatureConnector

FeatureConnector factory.

This function should be called from the tfds.features.FeatureConnector base class. Subclass should implement the from_json_content.

Example:

feature = tfds.features.FeatureConnector.from_json(
 {'type': 'Image', 'content': {'shape': [32, 32, 3], 'dtype': 'uint8'} }
)
assert isinstance(feature, tfds.features.Image)

Args
value dict(type=, content=) containing the feature to restore. Match dict returned by to_json.

Returns
The reconstructed FeatureConnector.

from_json_content

View source 

@classmethod
from_json_content(
 value: Union[Json, feature_pb2.Sequence]
) -> 'Sequence'

FeatureConnector factory (to overwrite).

Subclasses should overwrite this method. This method is used when importing the feature connector from the config.

This function should not be called directly. FeatureConnector.from_json should be called instead.

See existing FeatureConnectors for implementation examples.

Args
value FeatureConnector information represented as either Json or a Feature proto. The content must match what is returned by to_json_content.
doc Documentation of this feature (e.g. description).

Returns
The reconstructed FeatureConnector.

from_proto

View source 

@classmethod
from_proto(
 feature_proto: feature_pb2.Feature
) -> T

Instantiates a feature from its proto representation.

get_serialized_info

View source 

get_serialized_info()

See base class for details.

get_tensor_info

View source 

get_tensor_info()

See base class for details.

get_tensor_spec

View source 

get_tensor_spec() -> TreeDict[tf.TensorSpec]

Returns the tf.TensorSpec of this feature (not the element spec!).

Note that the output of this method may not correspond to the element spec of the dataset. For example, currently this method does not support RaggedTensorSpec.

load_metadata

View source 

load_metadata(
 *args, **kwargs
)

See base class for details.

repr_html

View source 

repr_html(
 ex: np.ndarray
) -> str

Returns the HTML str representation of the object.

repr_html_batch

View source 

repr_html_batch(
 ex: np.ndarray
) -> str

Returns the HTML str representation of the object (Sequence).

repr_html_ragged

View source 

repr_html_ragged(
 ex: np.ndarray
) -> str

Returns the HTML str representation of the object (Nested sequence).

save_config

View source 

save_config(
 root_dir: str
) -> None

Exports the FeatureConnector to a file.

Args
root_dir path/to/dir containing the features.json

save_metadata

View source 

save_metadata(
 *args, **kwargs
)

See base class for details.

serialize_example

View source 

serialize_example(
 example_data
) -> bytes

Encodes nested data values into tf.train.Example bytes.

See deserialize_example to decode the proto into tf.Tensor.

Args
example_data Example data to encode (numpy-like nested dict)

Returns
The serialized tf.train.Example.

to_json

View source 

to_json() -> Json

Exports the FeatureConnector to Json.

Each feature is serialized as a dict(type=..., content=...).

  • type: The cannonical name of the feature (module.FeatureName).
  • content: is specific to each feature connector and defined in to_json_content. Can contain nested sub-features (like for tfds.features.FeaturesDict and tfds.features.Sequence).

For example:

tfds.features.FeaturesDict({
 'input': tfds.features.Image(),
 'target': tfds.features.ClassLabel(num_classes=10),
})

Is serialized as:

{
"type":"tensorflow_datasets.core.features.features_dict.FeaturesDict",
"content":{
"input":{
"type":"tensorflow_datasets.core.features.image_feature.Image",
"content":{
"shape":[null,null,3],
"dtype":"uint8",
"encoding_format":"png"
}
},
"target":{
"type":
"tensorflow_datasets.core.features.class_label_feature.ClassLabel",
"content":{
"num_classes":10
}
}
}
}

Returns
A dict(type=, content=). Will be forwarded to from_json when reconstructing the feature.

to_json_content

View source 

to_json_content() -> feature_pb2.Sequence

FeatureConnector factory (to overwrite).

This function should be overwritten by the subclass to allow re-importing the feature connector from the config. See existing FeatureConnector for example of implementation.

Returns
The FeatureConnector metadata in either a dict, or a Feature proto. This output is used in from_json_content when reconstructing the feature.

to_proto

View source 

to_proto() -> feature_pb2.Feature

Exports the FeatureConnector to the Feature proto.

For features that have a specific schema defined in a proto, this function needs to be overriden. If there's no specific proto schema, then the feature will be represented using JSON.

Returns
The feature proto describing this feature.

__contains__

View source 

__contains__(
 key: str
) -> bool

__getitem__

View source 

__getitem__(
 key
)

Convenience method to access the underlying features.

Class Variables

ALIASES []

Except as otherwise noted, the content of this page is licensed under the Creative Commons Attribution 4.0 License, and code samples are licensed under the Apache 2.0 License. For details, see the Google Developers Site Policies. Java is a registered trademark of Oracle and/or its affiliates.

Last updated 2024年04月26日 UTC.