View source on GitHub
|
View source on GitHub
|
A sequence of categorical terms where ids are set by hashing.
tf.feature_column.sequence_categorical_column_with_hash_bucket(
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.