tf.contrib.crf.crf_multitag_sequence_score(
inputs,
tag_bitmap,
sequence_lengths,
transition_params
)
Defined in tensorflow/contrib/crf/python/ops/crf.py.
Computes the unnormalized score of all tag sequences matching tag_bitmap.
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.