Missed TensorFlow Dev Summit? Check out the video playlist. Watch recordings


TensorFlow 1 version View source on GitHub

A sequence of categorical terms where ids are set by hashing.

    key, hash_bucket_size, dtype=tf.dtypes.string

Pass this to embedding_column or indicator_column to convert sequence categorical data into dense representation for input to sequence NN, such as RNN.


tokens = sequence_categorical_column_with_hash_bucket(
    'tokens', hash_bucket_size=1000)
tokens_embedding = embedding_column(tokens, dimension=10)
columns = [tokens_embedding]

features = tf.io.parse_example(..., features=make_parse_example_spec(columns))
sequence_feature_layer = SequenceFeatures(columns)
sequence_input, sequence_length = sequence_feature_layer(features)
sequence_length_mask = tf.sequence_mask(sequence_length)

rnn_cell = tf.keras.layers.SimpleRNNCell(hidden_size)
rnn_layer = tf.keras.layers.RNN(rnn_cell)
outputs, state = rnn_layer(sequence_input, mask=sequence_length_mask)


  • key: A unique string identifying the input feature.
  • hash_bucket_size: An int > 1. The number of buckets.
  • dtype: The type of features. Only string and integer types are supported.


A SequenceCategoricalColumn.


  • ValueError: hash_bucket_size is not greater than 1.
  • ValueError: dtype is neither string nor integer.