Github user akopich commented on a diff in the pull request:
https://github.com/apache/spark/pull/19565#discussion_r146572407
--- Diff:
mllib/src/main/scala/org/apache/spark/mllib/clustering/LDAOptimizer.scala ---
@@ -497,40 +495,38 @@ final class OnlineLDAOptimizer extends LDAOptimizer
with Logging {
(u._1, u._2, u._3 + v._3)
}
- val (statsSum: BDM[Double], logphatOption: Option[BDV[Double]],
nonEmptyDocsN: Long) = stats
+ val (statsSum: BDM[Double], logphatOption: Option[BDV[Double]],
batchSize: Long) = stats
.treeAggregate((BDM.zeros[Double](k, vocabSize),
logphatPartOptionBase(), 0L))(
elementWiseSum, elementWiseSum
)
expElogbetaBc.destroy(false)
- if (nonEmptyDocsN == 0) {
+ if (batchSize == 0) {
logWarning("No non-empty documents were submitted in the batch.")
// Therefore, there is no need to update any of the model parameters
return this
}
val batchResult = statsSum *:* expElogbeta.t
- // Note that this is an optimization to avoid batch.count
- val batchSize = (miniBatchFraction * corpusSize).ceil.toInt
updateLambda(batchResult, batchSize)
- logphatOption.foreach(_ /= nonEmptyDocsN.toDouble)
- logphatOption.foreach(updateAlpha(_, nonEmptyDocsN))
+ logphatOption.foreach(_ /= batchSize.toDouble)
+ logphatOption.foreach(updateAlpha(_, batchSize))
this
}
/**
* Update lambda based on the batch submitted. batchSize can be
different for each iteration.
*/
- private def updateLambda(stat: BDM[Double], batchSize: Int): Unit = {
+ private def updateLambda(stat: BDM[Double], batchSize: Double): Unit = {
--- End diff --
I've done this just in order to achieve consistency with `updateAlpha`. Is
it a bad idea?
---
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org
