View source on GitHub 
Performs greedy constrained sequence on a batch of examples.
text.greedy_constrained_sequence(
scores, sequence_length=None, allowed_transitions=None, transition_weights=None,
use_log_space=False, use_start_and_end_states=False, name=None
)
Constrains a set of predictions based on a set of legal transitions
and/or a set of transition weights, returning the legal sequence that
maximizes the product or sum of the state scores and the transition weights
at each step. If use_log_space
is true, the sum is used; if false, the
product is used.
This op also takes a parameter use_start_and_end_states
, which when true
will add an implicit start and end state to each sequence. These implicit
states allow the user to specify additional weights and permitted transitions
to start and end a sequence (so, for instance, if you wanted to forbid your
output from ending in a certain set of states you could do so).
Inputs to this op can take one of three forms: a single TensorFlow tensor of scores with no sequence lengths, a TensorFlow tensor of scores along with a TensorFlow tensor of sequence lengths, or a RaggedTensor. If only the scores tensor is passed, this op will assume that the sequence lengths are equal to the size of the tensor (and so use all the data provided). If a scores tensor and sequence_lengths tensor is provided, the op will only use the data in the scores tensor as specified by the sequence_lengths tensor. Finally, if a RaggedTensor is provided, the sequence_lengths will be ignored and the variable length sequences in the RaggedTensor will be used.
Args  

scores

<float32> [batch_size, num_steps, num_states]
A tensor of scores, where scores[b, t, s] is the predicted score for
transitioning to state s at step t for batch b . The num_states
dimension must correspond to the num_states attribute for this op. This
input may be ragged; if it is ragged, the ragged tensor should have the
same structure [b, t, s] and only axis 1 should be ragged.

sequence_length

<{int32, int64}>[batch_size]
A rank1 tensor representing the length of the output sequence. If None,
and the 'scores' input is not ragged, sequence lengths will be assumed
to be the length of the score tensor.

allowed_transitions

if use_start_and_end_states is TRUE:
<bool>[num_states+1, num_states+1]
if use_start_and_end_states is FALSE:
<bool>[num_states, num_states]
A rank2 tensor representing allowed transitions.

transition_weights

if use_start_and_end_states is TRUE:
<float32>[num_states+1, num_states+1]
if use_start_and_end_states is FALSE:
<float32>[num_states, num_states]
A rank2 tensor representing transition weights.

use_log_space

Whether to use log space for the calculation. If false, calculations will be done in expspace. 
use_start_and_end_states

If True, sequences will have an implicit start and end state added. 
name

The name scope within which this op should be constructed. 
Returns  

An 