Spark MLlib Deep Learning Deep Belief Network (深度学习-深度信念网络)2.1，mllibbelief

Spark MLlib Deep Learning Deep Belief Network (深度学习-深度信念网络)2.1

http://blog.csdn.net/sunbow0

Spark MLlib Deep Learning工具箱，是根据现有深度学习教程《UFLDL教程》中的算法，在SparkMLlib中的实现。具体Spark MLlib Deep Learning(深度学习)目录结构：

第一章Neural Net(NN)

1、源码

2、源码解析

3、实例

第二章Deep Belief Nets(DBNs)

1、源码

2、源码解析

3、实例

第三章Convolution Neural Network(CNN)

第四章 Stacked Auto-Encoders(SAE)

第五章CAE

第二章Deep Belief Network (深度信念网络)

1源码

目前Spark MLlib Deep Learning工具箱源码的github地址为：

https://github.com/sunbow1/SparkMLlibDeepLearn

1.1 DBN代码 

package DBN

import org.apache.spark._
import org.apache.spark.SparkContext._
import org.apache.spark.rdd.RDD
import org.apache.spark.Logging
import org.apache.spark.mllib.regression.LabeledPoint
import org.apache.spark.mllib.linalg._
import org.apache.spark.mllib.linalg.distributed.RowMatrix

import breeze.linalg.{
  Matrix => BM,
  CSCMatrix => BSM,
  DenseMatrix => BDM,
  Vector => BV,
  DenseVector => BDV,
  SparseVector => BSV,
  axpy => brzAxpy,
  svd => brzSvd
}
import breeze.numerics.{
  exp => Bexp,
  tanh => Btanh
}

import scala.collection.mutable.ArrayBuffer
import java.util.Random
import scala.math._

/**
 * W：权重
 * b：偏置
 * c：偏置
 */
case class DBNweight(
  W: BDM[Double],
  vW: BDM[Double],
  b: BDM[Double],
  vb: BDM[Double],
  c: BDM[Double],
  vc: BDM[Double]) extends Serializable

/**
 * 配置参数
 */
case class DBNConfig(
  size: Array[Int],
  layer: Int,
  momentum: Double,
  alpha: Double) extends Serializable

/**
 * DBN(Deep Belief Network)
 */

class DBN(
  private var size: Array[Int],
  private var layer: Int,
  private var momentum: Double,
  private var alpha: Double) extends Serializable with Logging {
  //                  var size=Array(5, 10, 10)
  //                  var layer=3
  //                  var momentum=0.0
  //                  var alpha=1.0
  /**
   * size = architecture;         网络结构
   * layer = numel(nn.size);      网络层数
   * momentum = 0.0;              Momentum
   * alpha = 1.0;                 alpha
   */
  def this() = this(DBN.Architecture, 3, 0.0, 1.0)

  /** 设置神经网络结构. Default: [10, 5, 1]. */
  def setSize(size: Array[Int]): this.type = {
    this.size = size
    this
  }

  /** 设置神经网络层数据. Default: 3. */
  def setLayer(layer: Int): this.type = {
    this.layer = layer
    this
  }

  /** 设置Momentum. Default: 0.0. */
  def setMomentum(momentum: Double): this.type = {
    this.momentum = momentum
    this
  }

  /** 设置alpha. Default: 1. */
  def setAlpha(alpha: Double): this.type = {
    this.alpha = alpha
    this
  }

  /**
   * 深度信念网络（Deep Belief Network）
   * 运行训练DBNtrain
   */
  def DBNtrain(train_d: RDD[(BDM[Double], BDM[Double])], opts: Array[Double]): DBNModel = {
    // 参数配置 广播配置
    val sc = train_d.sparkContext
    val dbnconfig = DBNConfig(size, layer, momentum, alpha)
    // 初始化权重
    var dbn_W = DBN.InitialW(size)
    var dbn_vW = DBN.InitialvW(size)
    var dbn_b = DBN.Initialb(size)
    var dbn_vb = DBN.Initialvb(size)
    var dbn_c = DBN.Initialc(size)
    var dbn_vc = DBN.Initialvc(size)
    // 训练第1层
    printf("Training Level: %d.\n", 1)
    val weight0 = new DBNweight(dbn_W(0), dbn_vW(0), dbn_b(0), dbn_vb(0), dbn_c(0), dbn_vc(0))
    val weight1 = RBMtrain(train_d, opts, dbnconfig, weight0)
    dbn_W(0) = weight1.W
    dbn_vW(0) = weight1.vW
    dbn_b(0) = weight1.b
    dbn_vb(0) = weight1.vb
    dbn_c(0) = weight1.c
    dbn_vc(0) = weight1.vc
    // 打印权重
    printf("dbn_W%d.\n", 1)
    val tmpw0 = dbn_W(0)
    for (i <- 0 to tmpw0.rows - 1) {
      for (j <- 0 to tmpw0.cols - 1) {
        print(tmpw0(i, j) + "\t")
      }
      println()
    }

    // 训练第2层 至 n层
    for (i <- 2 to dbnconfig.layer - 1) {
      // 前向计算x
      //  x = sigm(repmat(rbm.c', size(x, 1), 1) + x * rbm.W');
      printf("Training Level: %d.\n", i)
      val tmp_bc_w = sc.broadcast(dbn_W(i - 2))
      val tmp_bc_c = sc.broadcast(dbn_c(i - 2))
      val train_d2 = train_d.map { f =>
        val lable = f._1
        val x = f._2
        val x2 = DBN.sigm(x * tmp_bc_w.value.t + tmp_bc_c.value.t)
        (lable, x2)
      }
      // 训练第i层
      val weighti = new DBNweight(dbn_W(i - 1), dbn_vW(i - 1), dbn_b(i - 1), dbn_vb(i - 1), dbn_c(i - 1), dbn_vc(i - 1))
      val weight2 = RBMtrain(train_d2, opts, dbnconfig, weighti)
      dbn_W(i - 1) = weight2.W
      dbn_vW(i - 1) = weight2.vW
      dbn_b(i - 1) = weight2.b
      dbn_vb(i - 1) = weight2.vb
      dbn_c(i - 1) = weight2.c
      dbn_vc(i - 1) = weight2.vc
      // 打印权重
      printf("dbn_W%d.\n", i)
      val tmpw1 = dbn_W(i - 1)
      for (i <- 0 to tmpw1.rows - 1) {
        for (j <- 0 to tmpw1.cols - 1) {
          print(tmpw1(i, j) + "\t")
        }
        println()
      }
    }
    new DBNModel(dbnconfig, dbn_W, dbn_b, dbn_c)
  }

  /**
   * 深度信念网络（Deep Belief Network）
   * 每一层神经网络进行训练rbmtrain
   */
  def RBMtrain(train_t: RDD[(BDM[Double], BDM[Double])],
    opts: Array[Double],
    dbnconfig: DBNConfig,
    weight: DBNweight): DBNweight = {
    val sc = train_t.sparkContext
    var StartTime = System.currentTimeMillis()
    var EndTime = System.currentTimeMillis()
    // 权重参数变量
    var rbm_W = weight.W
    var rbm_vW = weight.vW
    var rbm_b = weight.b
    var rbm_vb = weight.vb
    var rbm_c = weight.c
    var rbm_vc = weight.vc
    // 广播参数
    val bc_config = sc.broadcast(dbnconfig)
    // 训练样本数量
    val m = train_t.count
    // 计算batch的数量
    val batchsize = opts(0).toInt
    val numepochs = opts(1).toInt
    val numbatches = (m / batchsize).toInt
    // numepochs是循环的次数 
    for (i <- 1 to numepochs) {
      StartTime = System.currentTimeMillis()
      val splitW2 = Array.fill(numbatches)(1.0 / numbatches)
      var err = 0.0
      // 根据分组权重，随机划分每组样本数据  
      for (l <- 1 to numbatches) {
        // 1 广播权重参数
        val bc_rbm_W = sc.broadcast(rbm_W)
        val bc_rbm_vW = sc.broadcast(rbm_vW)
        val bc_rbm_b = sc.broadcast(rbm_b)
        val bc_rbm_vb = sc.broadcast(rbm_vb)
        val bc_rbm_c = sc.broadcast(rbm_c)
        val bc_rbm_vc = sc.broadcast(rbm_vc)

        //        // 打印权重
        //        println(i + "\t" + l)
        //        val tmpw0 = bc_rbm_W.value
        //        for (i <- 0 to tmpw0.rows - 1) {
        //          for (j <- 0 to tmpw0.cols - 1) {