View source on GitHub |
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. |