spark.mllib源码阅读-优化算法2-Updater

版权声明：本文为博主原创文章，未经博主允许不得转载。 https://blog.csdn.net/zbc1090549839/article/details/64442103

Updater是Spark中进行机器学习时对用于更新参数的轮子，参数更新的过程是

1、第i轮的机器学习求解得到的参数wi

2、第i+1轮计算得到的梯度值

3、正则化选项

来计算第i+1轮的机器学习要求解的参数wi+1

 

Spark实现了三类Updater，SimpleUpdater、L1Updater及SquaredL2Updater，他们之间关系为

SimpleUpdater：

无正则化的Updater，直接基于梯度值来更新参数wi+1=wi - gradient*rate

实现代码如下：

class SimpleUpdater extends Updater {
  override def compute(
      weightsOld: Vector,
      gradient: Vector,
      stepSize: Double,
      iter: Int,
      regParam: Double): (Vector, Double) = {
    val thisIterStepSize = stepSize / math.sqrt(iter) //计算rate
    val brzWeights: BV[Double] = weightsOld.asBreeze.toDenseVector
    brzAxpy(-thisIterStepSize, gradient.asBreeze, brzWeights) //brzWeights - thisIterStepSize*gradient

    (Vectors.fromBreeze(brzWeights), 0)
  }
}

 L1Updater：

L1正则化的Updater，在参数更新的时候，对参数值的范围做了一定的限制，即使w更稀疏至于L1正则化的过程，原始文档讲的很详细

* If w(参数，下同) is greater than shrinkageVal(通过一个基于当前迭代次数和正则化因子的计算的值), set weight component to w-shrinkageVal.
* If w is less than -shrinkageVal, set weight component to w+shrinkageVal.
* If w is (-shrinkageVal, shrinkageVal), set weight component to 0.

实现代码也很简单：

class L1Updater extends Updater {
  override def compute(
      weightsOld: Vector,
      gradient: Vector,
      stepSize: Double,
      iter: Int,
      regParam: Double): (Vector, Double) = {
    val thisIterStepSize = stepSize / math.sqrt(iter)
    // Take gradient step
    val brzWeights: BV[Double] = weightsOld.asBreeze.toDenseVector
    brzAxpy(-thisIterStepSize, gradient.asBreeze, brzWeights)
    // Apply proximal operator (soft thresholding)
    val shrinkageVal = regParam * thisIterStepSize
    var i = 0
    val len = brzWeights.length
    while (i < len) {
      val wi = brzWeights(i)
      brzWeights(i) = signum(wi) * max(0.0, abs(wi) - shrinkageVal)
      i += 1
    }
    (Vectors.fromBreeze(brzWeights), brzNorm(brzWeights, 1.0) * regParam)
  }
}

L2正则化Updater：

L2正则化Updater，在原有损失函数的基础上加上1/2 ||w||^2，对正则化之后的损失函数求梯度得到参数的更新

w' = (1 - thisIterStepSize * regParam) * w - thisIterStepSize * gradient
class SquaredL2Updater extends Updater {
  override def compute(
      weightsOld: Vector,
      gradient: Vector,
      stepSize: Double,
      iter: Int,
      regParam: Double): (Vector, Double) = {
    val thisIterStepSize = stepSize / math.sqrt(iter)
    val brzWeights: BV[Double] = weightsOld.asBreeze.toDenseVector
    brzWeights :*= (1.0 - thisIterStepSize * regParam)
    brzAxpy(-thisIterStepSize, gradient.asBreeze, brzWeights)
    val norm = brzNorm(brzWeights, 2.0)
    (Vectors.fromBreeze(brzWeights), 0.5 * regParam * norm * norm)
  }
}

在定制自己的正则化方式时，可以继承抽象基类Updater并实现参数更新compute方法