package glintlda.util
import akka.util.Timeout
import glint.models.client.BigMatrix
import glintlda.LDAConfig
import scala.concurrent.{ExecutionContext, Future}
/**
* This is a specialized buffer that keeps aggregating values locally
* We wish to use this for words that are most common according to the power law distribution
*
* @param cutoff The cutoff feature. Keep this relatively small as we need cutoff * topics space to store the buffer
* @param config The LDA configuration
*/
class AggregateBuffer(val cutoff: Int, config: LDAConfig) {
val size: Int = Math.min(cutoff, config.vocabularyTerms) * config.topics
val buffer = new Array[Long](size)
/**
* Adds a value to the buffer
*
* @param feature The feature
* @param topic The topic
* @param value The value
*/
@inline
def add(feature: Long, topic: Int, value: Int): Unit = {
buffer(feature.toInt * config.topics + topic) += value
}
/**
* Flushes to given lda model
*
* @param matrix The matrix to flush to
* @param ec The execution context in which the push requests will be executed
* @param timeout The timeout for the push requests
* @return
*/
def flush(matrix: BigMatrix[Long])(implicit ec: ExecutionContext, timeout: Timeout): Future[Boolean] = {
if (cutoff > 0) {
val rows = new Array[Long](size)
val cols = new Array[Int](size)
var i = 0
while (i < size) {
rows(i) = i / config.topics
cols(i) = i % config.topics
i += 1
}
matrix.push(rows, cols, buffer)
} else {
Future { true }
}
}
}
