tf.keras.utils.FeatureSpace

One-stop utility for preprocessing and encoding structured data.

Inherits From: Layer, Operation

tf.keras.utils.FeatureSpace(
    features,
    output_mode='concat',
    crosses=None,
    crossing_dim=32,
    hashing_dim=32,
    num_discretization_bins=32,
    name=None
)

feature_names Dict mapping the names of your features to their type specification, e.g. {"my_feature": "integer_categorical"} or {"my_feature": FeatureSpace.integer_categorical()}. For a complete list of all supported types, see "Available feature types" paragraph below. output_mode One of "concat" or "dict". In concat mode, all features get concatenated together into a single vector. In dict mode, the FeatureSpace returns a dict of individually encoded features (with the same keys as the input dict keys). crosses List of features to be crossed together, e.g. crosses=[("feature_1", "feature_2")]. The features will be "crossed" by hashing their combined value into a fixed-length vector. crossing_dim Default vector size for hashing crossed features. Defaults to 32. hashing_dim Default vector size for hashing features of type "integer_hashed" and "string_hashed". Defaults to 32. num_discretization_bins Default number of bins to be used for discretizing features of type "float_discretized". Defaults to 32.

Available feature types:

Note that all features can be referred to by their string name, e.g. "integer_categorical". When using the string name, the default argument values are used.

# Plain float values.
FeatureSpace.float(name=None)

# Float values to be preprocessed via featurewise standardization
# (i.e. via a `keras.layers.Normalization` layer).
FeatureSpace.float_normalized(name=None)

# Float values to be preprocessed via linear rescaling
# (i.e. via a `keras.layers.Rescaling` layer).
FeatureSpace.float_rescaled(scale=1., offset=0., name=None)

# Float values to be discretized. By default, the discrete
# representation will then be one-hot encoded.
FeatureSpace.float_discretized(
    num_bins, bin_boundaries=None, output_mode="one_hot", name=None)

# Integer values to be indexed. By default, the discrete
# representation will then be one-hot encoded.
FeatureSpace.integer_categorical(
    max_tokens=None, num_oov_indices=1, output_mode="one_hot", name=None)

# String values to be indexed. By default, the discrete
# representation will then be one-hot encoded.
FeatureSpace.string_categorical(
    max_tokens=None, num_oov_indices=1, output_mode="one_hot", name=None)

# Integer values to be hashed into a fixed number of bins.
# By default, the discrete representation will then be one-hot encoded.
FeatureSpace.integer_hashed(num_bins, output_mode="one_hot", name=None)

# String values to be hashed into a fixed number of bins.
# By default, the discrete representation will then be one-hot encoded.
FeatureSpace.string_hashed(num_bins, output_mode="one_hot", name=None)

Examples:

Basic usage with a dict of input data:

raw_data = {
    "float_values": [0.0, 0.1, 0.2, 0.3],
    "string_values": ["zero", "one", "two", "three"],
    "int_values": [0, 1, 2, 3],
}
dataset = tf.data.Dataset.from_tensor_slices(raw_data)

feature_space = FeatureSpace(
    features={
        "float_values": "float_normalized",
        "string_values": "string_categorical",
        "int_values": "integer_categorical",
    },
    crosses=[("string_values", "int_values")],
    output_mode="concat",
)
# Before you start using the FeatureSpace,
# you must `adapt()` it on some data.
feature_space.adapt(dataset)

# You can call the FeatureSpace on a dict of data (batched or unbatched).
output_vector = feature_space(raw_data)

Basic usage with tf.data:

# Unlabeled data
preprocessed_ds = unlabeled_dataset.map(feature_space)

# Labeled data
preprocessed_ds = labeled_dataset.map(lambda x, y: (feature_space(x), y))

Basic usage with the Keras Functional API:

# Retrieve a dict Keras Input objects
inputs = feature_space.get_inputs()
# Retrieve the corresponding encoded Keras tensors
encoded_features = feature_space.get_encoded_features()
# Build a Functional model
outputs = keras.layers.Dense(1, activation="sigmoid")(encoded_features)
model = keras.Model(inputs, outputs)

Customizing each feature or feature cross:

feature_space = FeatureSpace(
    features={
        "float_values": FeatureSpace.float_normalized(),
        "string_values": FeatureSpace.string_categorical(max_tokens=10),
        "int_values": FeatureSpace.integer_categorical(max_tokens=10),
    },
    crosses=[
        FeatureSpace.cross(("string_values", "int_values"), crossing_dim=32)
    ],
    output_mode="concat",
)

Returning a dict of integer-encoded features:

feature_space = FeatureSpace(
    features={
        "string_values": FeatureSpace.string_categorical(output_mode="int"),
        "int_values": FeatureSpace.integer_categorical(output_mode="int"),
    },
    crosses=[
        FeatureSpace.cross(
            feature_names=("string_values", "int_values"),
            crossing_dim=32,
            output_mode="int",
        )
    ],
    output_mode="dict",
)

Specifying your own Keras preprocessing layer:

# Let's say that one of the features is a short text paragraph that
# we want to encode as a vector (one vector per paragraph) via TF-IDF.
data = {
    "text": ["1st string", "2nd string", "3rd string"],
}

# There's a Keras layer for this: TextVectorization.
custom_layer = layers.TextVectorization(output_mode="tf_idf")

# We can use FeatureSpace.feature to create a custom feature
# that will use our preprocessing layer.
feature_space = FeatureSpace(
    features={
        "text": FeatureSpace.feature(
            preprocessor=custom_layer, dtype="string", output_mode="float"
        ),
    },
    output_mode="concat",
)
feature_space.adapt(tf.data.Dataset.from_tensor_slices(data))
output_vector = feature_space(data)

Retrieving the underlying Keras preprocessing layers:

# The preprocessing layer of each feature is available in `.preprocessors`.
preprocessing_layer = feature_space.preprocessors["feature1"]

# The crossing layer of each feature cross is available in `.crossers`.
# It's an instance of keras.layers.HashedCrossing.
crossing_layer = feature_space.crossers["feature1_X_feature2"]

Saving and reloading a FeatureSpace:

feature_space.save("featurespace.keras")
reloaded_feature_space = keras.models.load_model("featurespace.keras")

input Retrieves the input tensor(s) of a symbolic operation.

Only returns the tensor(s) corresponding to the first time the operation was called. output Retrieves the output tensor(s) of a layer.

Only returns the tensor(s) corresponding to the first time the operation was called.

Methods

adapt

View source

adapt(
    dataset
)

cross

View source

@classmethod
cross(
    feature_names, crossing_dim, output_mode='one_hot'
)

feature

View source

@classmethod
feature(
    dtype, preprocessor, output_mode
)

float

View source

@classmethod
float(
    name=None
)

float_discretized

View source

@classmethod
float_discretized(
    num_bins, bin_boundaries=None, output_mode='one_hot', name=None
)

float_normalized

View source

@classmethod
float_normalized(
    name=None
)

float_rescaled

View source

@classmethod
float_rescaled(
    scale=1.0, offset=0.0, name=None
)

from_config

View source

@classmethod
from_config(
    config
)

Creates a layer from its config.

This method is the reverse of get_config, capable of instantiating the same layer from the config dictionary. It does not handle layer connectivity (handled by Network), nor weights (handled by set_weights).

Args

config A Python dictionary, typically the output of get_config.

Returns
A layer instance.

get_encoded_features

View source

get_encoded_features()

get_inputs

View source

get_inputs()

integer_categorical

View source

@classmethod
integer_categorical(
    max_tokens=None,
    num_oov_indices=1,
    output_mode='one_hot',
    name=None
)

integer_hashed

View source

@classmethod
integer_hashed(
    num_bins, output_mode='one_hot', name=None
)

save

View source

save(
    filepath
)

Save the FeatureSpace instance to a .keras file.

You can reload it via keras.models.load_model():

feature_space.save("featurespace.keras")
reloaded_fs = keras.models.load_model("featurespace.keras")

string_categorical

View source

@classmethod
string_categorical(
    max_tokens=None,
    num_oov_indices=1,
    output_mode='one_hot',
    name=None
)

string_hashed

View source

@classmethod
string_hashed(
    num_bins, output_mode='one_hot', name=None
)

symbolic_call

View source

symbolic_call(
    *args, **kwargs
)