Loading TOC...

cntk.forwardBackward

cntk.forwardBackward(
   $graph as cntk.variable,
   $features as cntk.variable,
   $blank-to-token-id as (Number|String),
   $delay-constraint as cntk.integer,
   $name as String
) as cntk.function

Summary

Criterion node for training methods that rely on forward-backward Viterbi-like passes, e.g. Connectionist Temporal Classification (CTC) training The node takes as the input the graph of labels, produced by the labels_to_graph operation that determines the exact forward/backward procedure.

Parameters
$graph Labels graph.
$features Network output.
$blank-to-token-id Id of the CTC blank label.
$delay-constraint Label output delay constraint introduced during training that allows to have shorter delay during inference. This is using the original time information to enforce that CTC tokens only get aligned within a time margin. Setting this parameter smaller will result in shorted delay between label output during decoding, yet may hurt accuracy. delayConstraint=-1 means no constraint.
$name The name of the function instance in the network.

Stack Overflow iconStack Overflow: Get the most useful answers to questions from the MarkLogic community, or ask your own question.