tf.keras.layers.experimental.preprocessing.TextVectorization

Text vectorization layer.

Inherits From: PreprocessingLayer, Layer, Module

This layer has basic options for managing text in a Keras model. It transforms a batch of strings (one sample = one string) into either a list of token indices (one sample = 1D tensor of integer token indices) or a dense representation (one sample = 1D tensor of float values representing data about the sample's tokens).

If desired, the user can call this layer's adapt() method on a dataset. When this layer is adapted, it will analyze the dataset, determine the frequency of individual string values, and create a 'vocabulary' from them. This vocabulary can have unlimited size or be capped, depending on the configuration options for this layer; if there are more unique values in the input than the maximum vocabulary size, the most frequent terms will be used to create the vocabulary.

The processing of each sample contains the following steps:

  1. standardize each sample (usually lowercasing + punctuation stripping)
  2. split each sample into substrings (usually words)
  3. recombine substrings into tokens (usually ngrams)
  4. index tokens (associate a unique int value with each token)
  5. transform each sample using this index, either into a vector of ints or a dense float vector.

Some notes on passing Callables to customize splitting and normalization for this layer:

  1. Any callable can be passed to this Layer, but if you want to serialize this object you should only pass functions that are registered Keras serializables (see tf.keras.utils.register_keras_serializable for more details).
  2. When using a custom callable for standardize, the data received by the callable will be exactly as passed to this layer. The callable should return a tensor of the same shape as the input.
  3. When using a custom callable for split, the data received by the callable will have the 1st dimension squeezed out - instead of [["string to split"], ["another string to split"]], the Callable will see ["string to split", "another string to split"]. The callable should return a Tensor with the first dimension containing the split tokens - in this example, we should see something like [["string", "to", "split], ["another", "string", "to", "split"]]. This makes the callable site natively compatible with tf.strings.split().

Example:

This example instantiates a TextVectorization layer that lowercases text, splits on whitespace, strips punctuation, and outputs integer vocab indices.

text_dataset = tf.data.Dataset.from_tensor_slices(["foo", "bar", "baz"])
max_features = 5000  # Maximum vocab size.
max_len = 4  # Sequence length to pad the outputs to.
embedding_dims = 2

# Create the layer.
vectorize_layer = TextVectorization(
 max_tokens=max_features,
 output_mode='int',
 output_sequence_length=max_len)

# Now that the vocab layer has been created, call `adapt` on the text-only
# dataset to create the vocabulary. You don't have to batch, but for large
# datasets this means we're not keeping spare copies of the dataset.
vectorize_layer.adapt(text_dataset.batch(64))

# Create the model that uses the vectorize text layer
model = tf.keras.models.Sequential()

# Start by creating an explicit input layer. It needs to have a shape of
# (1,) (because we need to guarantee that there is exactly one string
# input per batch), and the dtype needs to be 'string'.
model.add(tf.keras.Input(shape=(1,), dtype=tf.string))

# The first layer in our model is the vectorization layer. After this
# layer, we have a tensor of shape (batch_size, max_len) containing vocab
# indices.
model.add(vectorize_layer)

# Now, the model can map strings to integers, and you can add an embedding
# layer to map these integers to learned embeddings.
input_data = [["foo qux bar"], ["qux baz"]]
model.predict(input_data)
array([[2, 1, 4, 0],
       [1, 3, 0, 0]])

Example:

This example instantiates a TextVectorization layer by passing a list of vocabulary terms to the layer's init method.

input_array = np.array([["earth", "wind", "and", "fire"], ["fire", "and", "earth", "michigan"]]) expected_output = [[2, 3, 4, 5], [5, 4, 2, 1]]

input_data = keras.Input(shape=(None,), dtype=dtypes.string) layer = get_layer_class()( max_tokens=None, standardize=None, split=None, output_mode=text_vectorization.INT, vocabulary=vocab_data) int_data = layer(input_data) model = keras.Model(inputs=input_data, outputs=int_data)

output_dataset = model.predict(input_array)

>>> vocab_data = ["earth", "wind", "and", "fire"]
>>> max_len = 4  # Sequence length to pad the outputs to.
>>>
>>> # Create the layer, passing the vocab directly. You can also pass the
>>> # vocabulary arg a path to a file containing one vocabulary word per
>>> # line.
>>> vectorize_layer = TextVectorization(
...  max_tokens=max_features,
...  output_mode='int',
...  output_sequence_length=max_len,
...  vocabulary=vocab_data)
>>>
>>> # Because we've passed the vocabulary directly, we don't need to adapt
>>> # the layer - the vocabulary is already set. The vocabulary contains the
>>> # padding token ('') and OOV token ('[UNK]') as well as the passed tokens.
>>> vectorize_layer.get_vocabulary()
['', '[UNK]', 'earth', 'wind', 'and', 'fire']

max_tokens The maximum size of the vocabulary for this layer. If None, there is no cap on the size of the vocabulary. Note that this vocabulary contains 1 OOV token, so the effective number of tokens is (max_tokens - 1 - (1 if output == "int" else 0)).
standardize Optional specification for standardization to apply to the input text. Values can be None (no standardization), 'lower_and_strip_punctuation' (lowercase and remove punctuation) or a Callable. Default is 'lower_and_strip_punctuation'.
split Optional specification for splitting the input text. Values can be None (no splitting), 'whitespace' (split on ASCII whitespace), or a Callable. The default is 'whitespace'.
ngrams Optional specification for ngrams to create from the possibly-split input text. Values can be None, an integer or tuple of integers; passing an integer will create ngrams up to that integer, and passing a tuple of integers will create ngrams for the specified values in the tuple. Passing None means that no ngrams will be created.
output_mode Optional specification for the output of the layer. Values can be "int", "binary", "count" or "tf-idf", configuring the layer as follows: "int": Outputs integer indices, one integer index per split string token. When output == "int", 0 is reserved for masked locations; this reduces the vocab size to max_tokens-2 instead of max_tokens-1 "binary": Outputs a single int array per batch, of either vocab_size or max_tokens size, containing 1s in all elements where the token mapped to that index exists at least once in the batch item. "count": As "binary", but the int array contains a count of the number of times the token at that index appeared in the batch item. "tf-idf": As "binary", but the TF-IDF algorithm is applied to find the value in each token slot.
output_sequence_length Only valid in INT mode. If set, the output will have its time dimension padded or truncated to exactly output_sequence_length values, resulting in a tensor of shape [batch_size, output_sequence_length] regardless of how many tokens resulted from the splitting step. Defaults to None.
pad_to_max_tokens Only valid in "binary", "count", and "tf-idf" modes. If True, the output will have its feature axis padded to max_tokens even if the number of unique tokens in the vocabulary is less than max_tokens, resulting in a tensor of shape [batch_size, max_tokens] regardless of vocabulary size. Defaults to True.
vocabulary An optional list of vocabulary terms, or a path to a text file containing a vocabulary to load into this layer. The file should contain one token per line. If the list or file contains the same token multiple times, an error will be thrown.

Methods

adapt

View source

Fits the state of the preprocessing layer to the dataset.

Overrides the default adapt method to apply relevant preprocessing to the inputs before passing to the combiner.

Arguments
data The data to train on. It can be passed either as a tf.data Dataset, as a NumPy array, a string tensor, or as a list of texts.
reset_state Optional argument specifying whether to clear the state of the layer at the start of the call to adapt. This must be True for this layer, which does not support repeated calls to adapt.

get_vocabulary

View source

set_vocabulary

View source

Sets vocabulary (and optionally document frequency) data for this layer.

This method sets the vocabulary and DF data for this layer directly, instead of analyzing a dataset through 'adapt'. It should be used whenever the vocab (and optionally document frequency) information is already known. If vocabulary data is already present in the layer, this method will replace it.

Arguments
vocab An array of string tokens.
df_data An array of document frequency data. Only necessary if the layer output_mode is TFIDF.
oov_df_value The document frequency of the OOV token. Only necessary if output_mode is TFIDF.

Raises
ValueError If there are too many inputs, the inputs do not match, or input data is missing.
RuntimeError If the vocabulary cannot be set when this function is called. This happens when "binary", "count", and "tfidf" modes, if "pad_to_max_tokens" is False and the layer itself has already been called.