tf.feature_column.sequence_categorical_column_with_identity

View source on GitHub

Returns a feature column that represents sequences of integers.

tf.feature_column.sequence_categorical_column_with_identity(
    key, num_buckets, default_value=None
)

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

Example:

watches = sequence_categorical_column_with_identity(
    'watches', num_buckets=1000)
watches_embedding = embedding_column(watches, dimension=10)
columns = [watches_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)

Args:

Returns:

A SequenceCategoricalColumn.

Raises: