Loading TOC...

cntk:learning-rate-schedule-from-pairs

cntk:learning-rate-schedule-from-pairs(
   $pairs as json:array,
   [$epoch-size as xs:unsignedLong],
   [$minibatch-size as xs:unsignedLong]
) as cntk:schedule

Summary

Returns a learning rate schedule, based on a sequence of pairs of "number of epochs" and "their learning rate". Check the example for detailed explanation.

Parameters
$pairs A sequence of pairs of "number of epochs" and "their learning rate". An array of array, where each inner array has length 2, and with the first element being of type xs:unsignedLong and second being of type xs:double.
$epoch-size The epoch size. Defaults to a special sentinel value to represent the whole data sweep.
$minibatch-size The minibatch size. Defaults to a sentinel value which means ignored. This is a "best-effort" parameter: there is no guarantee that the size of every minibatch is exactly $minibatch-size.

Example

  let $pairs :=
json:to-array((json:to-array((10,0.1)),json:to-array((50,0.05)),json:to-array((1,0.01))))
  return cntk:learning-rate-schedule-from-pairs($pairs, 100)
  => a cntk:schedule, where the first 10 epochs (100*10 samples)
are using 0.1 as learning rate, the second 50 epochs (100*50 samples) using
0.05, and all subsequence samples use 0.01.

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