scala实现超限学习机（ELM）手写数字分类

介绍：超限学习机（Extreme Learning Machine）又叫极限学习机是黄广斌等于2006年提出，属于单隐层前馈神经网络SLFN。其输入权值和隐层阈值随机初始化，通过最小二乘法计算得出输出权值，具有很好的分类和回归性能，与其它机器学习算法相比最大的优势是计算速度快，因为ELM不需要进行迭代。

现将实现ELM的scala代码贴出，亲测运行成功，由于内存限制，在本地最高只能运行到900个隐层结点。若有疑问，欢迎交流。

import org.apache.log4j.{Level, Logger}import org.apache.spark.{SparkConf, SparkContext}import breeze.linalg.{DenseMatrix=>BDM,_}//为避免与mllib自带的DenseMatrix混淆，将其赋一个别名import breeze.numerics._import org.apache.spark.rdd.RDDobject ELM {  def manOf[T: Manifest](t: T): Manifest[T] = manifest[T]//返回类型  def main(args: Array[String]) {    // 屏蔽不必要的日志显示终端上    Logger.getLogger("org.apache.spark").setLevel(Level.ERROR)    Logger.getLogger("org.eclipse.jetty.server").setLevel(Level.OFF)    // 设置运行环境    val conf = new SparkConf().setAppName("ELM").setMaster("local")    //val conf = new SparkConf().setAppName("ELM").setMaster("spark://ec2-54-89-63-159.compute-1.amazonaws.com:7077")    val sc = new SparkContext(conf)    var data = sc.textFile("C:\\Users\\Administrator\\Desktop\\DeepLearnToolbox-master\\data\\stdtrainx.txt")    //var data = sc.textFile("hdfs://192.168.203.132:9000/stdtrainx.txt")    //var data = sc.textFile("file:///root/spark/file/stdtrainx.md")    println("训练数据x读取完毕")    //隐层结点数    val hiddenNumber = 100    var temp = readData(data)    var sampleNumber = temp.count().toInt//训练集样本数    val inputNumber = temp.first().length    var inputWeight = BDM.rand(inputNumber,hiddenNumber)    var inputBias = BDM.rand(1,hiddenNumber)    val s1 =(t:Double)=>t*2-1    val s2 = (line:Array[Double]) =>{      val t : BDM[Double] = BDM(line) * inputWeight+inputBias      val s = t.map(i=>1/(1+exp(-i)))      s.toArray    }    inputWeight = inputWeight.map(i=>i*2-1)//随机权值[-1,1]    inputBias = inputBias.map(i=>i*2-1)//随机阈值[-1,1]    var temp_x = temp.map(s2).collect()    var H = new BDM(hiddenNumber,sampleNumber,temp_x.flatten).t    data = sc.textFile("C:\\Users\\Administrator\\Desktop\\DeepLearnToolbox-master\\data\\stdtrainy.txt")    //data = sc.textFile("hdfs://192.168.203.132:9000/stdtrainy.txt")    //data = sc.textFile("file:///root/spark/file/stdtrainy.txt")    println("训练数据y读取完毕")    var f2 = readData(data).collect()    val train_y = array2mat(f2)    println("输出权值计算开始")    val outputWeight = inv(H.t*H)*H.t*train_y    println("输出权值计算完毕")    var actual_output = H * outputWeight    var errornum = calError(actual_output,train_y)    printf("训练误差:%.2f%%\n",errornum*1.0/sampleNumber*100)    data = sc.textFile("C:\\Users\\Administrator\\Desktop\\DeepLearnToolbox-master\\data\\stdtestx.txt")    //data = sc.textFile("hdfs://192.168.203.132:9000/stdtestx.txt")    //data = sc.textFile("file:///root/spark/file/stdtestx.txt")    temp = readData(data)    sampleNumber = temp.count().toInt//测试集样本数    temp_x = temp.map(s2).collect()    var T = new BDM(hiddenNumber,sampleNumber,temp_x.flatten).t    data = sc.textFile("C:\\Users\\Administrator\\Desktop\\DeepLearnToolbox-master\\data\\stdtesty.txt")    //data = sc.textFile("hdfs://192.168.203.132:9000/stdtesty.txt")    //data = sc.textFile("file:///root/spark/file/stdtesty.txt")    f2 = readData(data).collect()    val test_y = array2mat(f2)    actual_output = T * outputWeight    errornum = calError(actual_output,test_y)    printf("测试误差:%.2f%%\n",errornum*1.0/sampleNumber*100)    sc.stop()  }  def array2mat(array: Array[Array[Double]]) : BDM[Double] = {    val rows = array.length    val cols = array(0).length    var mat = new BDM[Double](cols,rows,array.flatten).t    return mat  }  def readData( data : RDD[String] ) : RDD[Array[Double]] = {    val s3 = (line:String) => line.split(',').map(_.toDouble)    return data.map{s3}  }  def calError( mat1 : BDM[Double],mat2 : BDM[Double]) : Int = {    var errornum = 0    for( i <- 0 until mat1.rows){      val x = argmax(mat1(i,::))      val y = argmax(mat2(i,::))      if(x!=y){        errornum += 1      }    }    return errornum  }}

刚开始写scala的时候，疯狂的用for循环，可能是被其他编程语言带动的缘故吧，还没有体会到scala的思想精髓，后来发现，大多数的for循环都能借助RDD完成，这就是函数式编程的特点之一。