Apache Spark源码走读之23 -- Spark MLLib中拟牛顿法L-BFGS的源码实现

about云腾讯认证空间

本帖最后由 pig2 于 2015-1-6 14:18 编辑
问题导读
１．牛顿法有哪些优点体现？
２．L-BFGS算法中使用到的正则化方法是什么？

概要

本文就拟牛顿法L-BFGS的由来做一个简要的回顾，然后就其在spark mllib中的实现进行源码走读。

拟牛顿法数学原理

代码实现

L-BFGS算法中使用到的正则化方法是SquaredL2Updater。

算法实现上使用到了由scalanlp的成员项目breeze库中的BreezeLBFGS函数,mllib中自定义了BreezeLBFGS所需要的DiffFunctions.

runLBFGS函数的源码实现如下

def
 runLBFGS(
      data: RDD[(Double, Vector)],
      gradient: Gradient,
      updater: Updater,
      numCorrections: Int,
      convergenceTol: Double,
      maxNumIterations: Int,
      regParam: Double,
      initialWeights: Vector): (Vector, Array[Double]) = {

    
val lossHistory = new ArrayBuffer[Double](maxNumIterations)

    val numExamples = data.count()

    val costFun =
      new CostFun(data, gradient, updater, regParam, numExamples)

    val lbfgs = new BreezeLBFGS[BDV[Double]](maxNumIterations, numCorrections, convergenceTol)

    val states =
      lbfgs.iterations(new CachedDiffFunction(costFun), initialWeights.toBreeze.toDenseVector)

    /**
     * NOTE: lossSum and loss is computed using the weights from the previous iteration
     * and regVal is the regularization value computed in the previous iteration as well.
     */
    var state = states.next()
    while(states.hasNext) {
      lossHistory.append(state.value)
      state = states.next()
    }
    lossHistory.append(state.value)
    val weights = Vectors.fromBreeze(state.x)

    logInfo("LBFGS.runLBFGS finished. Last 10 losses %s".format(
      lossHistory.takeRight(10).mkString(", ")))

    (weights, lossHistory.toArray)
  }
复制代码

costFun函数是算法实现中的重点

private
 
class CostFun(
    data: RDD[(Double, Vector)],
    gradient: Gradient,
    updater: Updater,
    regParam: Double,
    numExamples: Long) extends DiffFunction[BDV[Double]] {

    private var i = 0

    override def calculate(weights: BDV[Double]) = {
      // Have a local copy to avoid the serialization of CostFun object which is not serializable.
      val localData = data
      val localGradient = gradient

      val (gradientSum, lossSum) = localData.aggregate((BDV.zeros[Double](weights.size), 0.0))(
          seqOp = (c, v) => (c, v) match { case ((grad, loss), (label, features)) =>
            val l = localGradient.compute(
              features, label, Vectors.fromBreeze(weights), Vectors.fromBreeze(grad))
            (grad, loss + l)
          },
          combOp = (c1, c2) => (c1, c2) match { case ((grad1, loss1), (grad2, loss2)) =>
            (grad1 += grad2, loss1 + loss2)
          })

      /**
       * regVal is sum of weight squares if it's L2 updater;
       * for other updater, the same logic is followed.
       */
      val regVal = updater.compute(
        Vectors.fromBreeze(weights),
        Vectors.dense(new Array[Double](weights.size)), 0, 1, regParam)._2

      val loss = lossSum / numExamples + regVal
      /**
       * It will return the gradient part of regularization using updater.
       *
       * Given the input parameters, the updater basically does the following,
       *
       * w' = w - thisIterStepSize * (gradient + regGradient(w))
       * Note that regGradient is function of w
       *
       * If we set gradient = 0, thisIterStepSize = 1, then
       *
       * regGradient(w) = w - w'
       *
       * TODO: We need to clean it up by separating the logic of regularization out
       *       from updater to regularizer.
       */
      // The following gradientTotal is actually the regularization part of gradient.
      // Will add the gradientSum computed from the data with weights in the next step.
      val gradientTotal = weights - updater.compute(
        Vectors.fromBreeze(weights),
        Vectors.dense(new Array[Double](weights.size)), 1, 1, regParam)._1.toBreeze

      // gradientTotal = gradientSum / numExamples + gradientTotal
      axpy(1.0 / numExamples, gradientSum, gradientTotal)

      i += 1

      (loss, gradientTotal)
    }
  }

}
复制代码

355815741 · 发表于 2015-1-5 09:50:14

学习了，谢谢lz分享~

图文精华

Apache Spark源码走读之23 -- Spark MLLib中拟牛顿法L-BFGS的源码实现

已有(1)人评论

灌水之王

推荐 /2