gistfile1.txt

package com.intel.analytics.bigdl.example

import com.intel.analytics.bigdl._
import com.intel.analytics.bigdl.nn.Module
import com.intel.analytics.bigdl.nn.abstractnn.{AbstractModule, Activity}
import com.intel.analytics.bigdl.tensor.Tensor
import com.intel.analytics.bigdl.utils.Engine
import org.apache.hadoop.fs.Path
import org.apache.spark.SparkContext
import org.apache.spark.sql.SparkSession

import scala.collection.mutable.ArrayBuffer

object DeepSpeechPerf {

  var warmup = 10
  var iterations = 5

  def main(args: Array[String]): Unit = {
    val modelPath = "../model" // args(0)
    val uttLen = 3000
    val batchSize = 1
    val input = Tensor[Float](batchSize, 1, 13, uttLen).rand()
    val labels = Tensor[Float](batchSize, uttLen / 3).fill(1.0f)

    val conf = Engine.createSparkConf()
      .setAppName("test deepspeech 2 on text")
      .set("spark.task.maxFailures", "1")
    val sc = new SparkContext(conf)
    Engine.init

    val model = DeepSpeech2ModelLoader.loadModel(sc, modelPath)

    println(model)
    run(model, input)
  }

  def getTopTimes(times: Array[(AbstractModule[_ <: Activity, _ <: Activity, Float],
    Long, Long)], allSum: Long): Unit = {
    var forwardSum = 0L
    var backwardSum = 0L
    times.foreach(x => {
      forwardSum += x._2
      backwardSum += x._3
    })
    println(s"forwardSum = ${forwardSum}", s"backwardSum = ${backwardSum}")

    val timeBuffer = new ArrayBuffer[(AbstractModule[_ <: Activity,
      _ <: Activity, Float], Long, Long, Double)]
    var i = 0
    while (i < times.length) {
      val rate = times(i)._2.toDouble/ allSum
      timeBuffer.append((times(i)._1, times(i)._2, times(i)._3, rate))
      i += 1
    }
    val sortData = timeBuffer.sortBy(a => a._4)
    println("111111111111111111        ")
    sortData.foreach(println)
  }

  def run(model: Module[Float], input: Tensor[Float]): Unit = {
    println("start warm up")
    for (i <- 0 to warmup) {
      println(i)
      model.forward(input)
    }

    println("start run iterations")
    val s1 = System.nanoTime()
    for (i <- 0 to iterations) {
      val s1 = System.nanoTime()
      model.forward(input)
      val end1 = System.nanoTime() - s1
      val tmp = model.getTimes()
      getTopTimes(tmp, end1)
      model.resetTimes()
    }
//    for (i <- 0 to iterations) {
//      model.forward(input)
//    }
    val end1 = System.nanoTime() - s1
    println("time " + end1/1e9 + " s")
  }

}

val realOutputs = outputs.map(out => {
val m = out.element.asInstanceOf[MklDnnLayer]
val reorder = m._outputFormats.map(outMemData => {
val shape = outMemData.shape
val layout = outMemData.layout
require(shape.length == 2 || shape.length == 4, s"output shape should be 2 or 4 dims," +
s"but get ${shape.length}, and this IRElement name is ${m.getName()}")
val realOut = if (shape.length == 2) {
val realOutFormat = new HeapData(shape, Memory.Format.nc)
val layer = ReorderMemory(inputFormat = null, outputFormat = realOutFormat,
gradInputFormat = null, gradOutputFomat = realOutFormat)
layer
} else if (shape.length == 4) {
val realOutFormat = new HeapData(shape, Memory.Format.nchw)
val layer = ReorderMemory(inputFormat = null, outputFormat = realOutFormat,
gradInputFormat = null, gradOutputFomat = realOutFormat)
layer
}
val node = new NodeFloat
out.add(node)
node
})
})

<!-- Another sample, using keys to authenticate.
<server>
  <id>siteServer</id>
  <privateKey>/path/to/private/key</privateKey>
  <passphrase>optional; leave empty if not used.</passphrase>
</server>
-->

alimaven aliyun maven http://maven.aliyun.com/nexus/content/groups/public/ central

<!-- Another sample, using keys to authenticate.
<server>
  <id>siteServer</id>
  <privateKey>/path/to/private/key</privateKey>
  <passphrase>optional; leave empty if not used.</passphrase>
</server>
-->

alimaven aliyun maven http://maven.aliyun.com/nexus/content/groups/public/ central