tf.contrib.crf.crf_multitag_sequence_score

Computes the unnormalized score of all tag sequences matching tag_bitmap.

tf.contrib.crf.crf_multitag_sequence_score(
    inputs, tag_bitmap, sequence_lengths, transition_params
)

tag_bitmap enables more than one tag to be considered correct at each time step. This is useful when an observed output at a given time step is consistent with more than one tag, and thus the log likelihood of that observation must take into account all possible consistent tags.

Using one-hot vectors in tag_bitmap gives results identical to crf_sequence_score.

Args
inputs	A [batch_size, max_seq_len, num_tags] tensor of unary potentials to use as input to the CRF layer.
tag_bitmap	A [batch_size, max_seq_len, num_tags] boolean tensor representing all active tags at each index for which to calculate the unnormalized score.
sequence_lengths	A [batch_size] vector of true sequence lengths.
transition_params	A [num_tags, num_tags] transition matrix.

Returns
sequence_scores	A [batch_size] vector of unnormalized sequence scores.