有什么方法可以序列化Spark ML Pipeline中的自定义Transformer

Question

我将ML管道与各种基于UDF的定制转换器一起使用。我正在寻找一种序列化/反序列化此管道的方法。

我使用以下方法序列化PipelineModel

ObjectOutputStream.write() 

但是，每当我尝试反序列化管道时，我都会遇到：

java.lang.ClassNotFoundException: org.sparkexample.DateTransformer

DateTransformer在哪里是我的自定义转换器。是否有实现适当序列化的方法/接口？

我发现那里

MLWritable

我的课程可能实现的接口（DateTransformer扩展了Transfrormer），但是找不到有用的示例。

Answer 1

简而言之，你不能，至少不容易。

开发人员竭尽全力让添加新的变压器/估算器变得尽可能困难。基本上，中的所有内容org.apache.spark.ml.util.ReadWrite都是私有的（除了MLWritable和MLReadable），因此无法使用其中的任何实用程序方法/类/对象。还有（我相信您已经发现）绝对没有关于如何完成此操作的文档，但是嘿，好的代码文档本身对吧？！

深入研究 中的代码org.apache.spark.ml.util.ReadWrite，org.apache.spark.ml.feature.HashingTF似乎您需要覆盖MLWritable.write和MLReadable.read。和似乎包含实际的保存/加载实现，正在保存和加载一堆元数据DefaultParamsWriter：DefaultParamsReader

• 班级
• 时间戳
• 火花版本
• uid
• 参数映射表
• （可选，额外的元数据）

因此任何实现至少需要涵盖这些，并且不需要学习任何模型的变压器可能会摆脱这一点。需要拟合的模型还需要在其实现中保存该数据save/write- 例如，这就是LocalLDAModel（https://github.com/apache/spark/blob/v1.6.3/mllib/src/main /scala/org/apache/spark/ml/clustering/LDA.scala#L523）所以学习的模型只是保存为镶木地板文件（看起来）

val data = sqlContext.read.parquet(dataPath)
        .select("vocabSize", "topicsMatrix", "docConcentration", "topicConcentration",
          "gammaShape")
        .head()

作为测试，我复制了org.apache.spark.ml.util.ReadWrite似乎需要的所有内容，并测试了以下变压器，但它没有做任何有用的事情。

警告：这几乎肯定是错误的做法，并且很可能在将来出错。 我真诚地希望我误解了一些东西，并且有人会纠正我如何实际创建一个可以序列化/反序列化的变压器

这是适用于 Spark 1.6.3 的，如果您使用的是 2.x，则可能已经损坏

import org.apache.spark.sql.types._
import org.apache.spark.ml.param._
import org.apache.hadoop.fs.Path
import org.apache.spark.SparkContext
import org.apache.spark.ml.Transformer
import org.apache.spark.ml.util.{Identifiable, MLReadable, MLReader, MLWritable, MLWriter}
import org.apache.spark.sql.{SQLContext, DataFrame}
import org.apache.spark.mllib.linalg._

import org.json4s._
import org.json4s.JsonDSL._
import org.json4s.jackson.JsonMethods._

object CustomTransform extends DefaultParamsReadable[CustomTransform] {
  /* Companion object for deserialisation */
  override def load(path: String): CustomTransform = super.load(path)
}

class CustomTransform(override val uid: String)
  extends Transformer with DefaultParamsWritable {

  def this() = this(Identifiable.randomUID("customThing"))

  def setInputCol(value: String): this.type = set(inputCol, value)
  def setOutputCol(value: String): this.type = set(outputCol, value)
  def getOutputCol(): String = getOrDefault(outputCol)

  val inputCol = new Param[String](this, "inputCol", "input column")
  val outputCol = new Param[String](this, "outputCol", "output column")

  override def transform(dataset: DataFrame): DataFrame = {
    val sqlContext = SQLContext.getOrCreate(SparkContext.getOrCreate())
    import sqlContext.implicits._

    val outCol = extractParamMap.getOrElse(outputCol, "output")
    val inCol = extractParamMap.getOrElse(inputCol, "input")
    val transformUDF = udf({ vector: SparseVector =>
      vector.values.map( _ * 10 )
      // WHAT EVER YOUR TRANSFORMER NEEDS TO DO GOES HERE
    })

    dataset.withColumn(outCol, transformUDF(col(inCol)))
  }

  override def copy(extra: ParamMap): Transformer = defaultCopy(extra)

  override def transformSchema(schema: StructType): StructType = {
    val outputFields = schema.fields :+ StructField(extractParamMap.getOrElse(outputCol, "filtered"), new VectorUDT, nullable = false)
    StructType(outputFields)
  }
}

然后我们需要https://github.com/apache/spark/blob/v1.6.3/mllib/src/main/scala/org/apache/spark/ml/util/ReadWrite.scala中的所有实用程序org.apache.spark.ml.util.ReadWrite

trait DefaultParamsWritable extends MLWritable { self: Params =>
  override def write: MLWriter = new DefaultParamsWriter(this)
}

trait DefaultParamsReadable[T] extends MLReadable[T] {
  override def read: MLReader[T] = new DefaultParamsReader
}

class DefaultParamsWriter(instance: Params) extends MLWriter {
  override protected def saveImpl(path: String): Unit = {
    DefaultParamsWriter.saveMetadata(instance, path, sc)
  }
}

object DefaultParamsWriter {

  /**
    * Saves metadata + Params to: path + "/metadata"
    *  - class
    *  - timestamp
    *  - sparkVersion
    *  - uid
    *  - paramMap
    *  - (optionally, extra metadata)
    * @param extraMetadata  Extra metadata to be saved at same level as uid, paramMap, etc.
    * @param paramMap  If given, this is saved in the "paramMap" field.
    *                  Otherwise, all [[org.apache.spark.ml.param.Param]]s are encoded using
    *                  [[org.apache.spark.ml.param.Param.jsonEncode()]].
    */
  def saveMetadata(
  instance: Params,
  path: String,
  sc: SparkContext,
  extraMetadata: Option[JObject] = None,
  paramMap: Option[JValue] = None): Unit = {
    val uid = instance.uid
    val cls = instance.getClass.getName
    val params = instance.extractParamMap().toSeq.asInstanceOf[Seq[ParamPair[Any]]]
    val jsonParams = paramMap.getOrElse(render(params.map { case ParamPair(p, v) =>
      p.name -> parse(p.jsonEncode(v))
    }.toList))
    val basicMetadata = ("class" -> cls) ~
    ("timestamp" -> System.currentTimeMillis()) ~
    ("sparkVersion" -> sc.version) ~
    ("uid" -> uid) ~
    ("paramMap" -> jsonParams)
    val metadata = extraMetadata match {
      case Some(jObject) =>
        basicMetadata ~ jObject
      case None =>
        basicMetadata
    }
    val metadataPath = new Path(path, "metadata").toString
    val metadataJson = compact(render(metadata))
    sc.parallelize(Seq(metadataJson), 1).saveAsTextFile(metadataPath)
  }
}

class DefaultParamsReader[T] extends MLReader[T] {
  override def load(path: String): T = {
    val metadata = DefaultParamsReader.loadMetadata(path, sc)
    val cls = Class.forName(metadata.className, true, Option(Thread.currentThread().getContextClassLoader).getOrElse(getClass.getClassLoader))
    val instance =
    cls.getConstructor(classOf[String]).newInstance(metadata.uid).asInstanceOf[Params]
    DefaultParamsReader.getAndSetParams(instance, metadata)
    instance.asInstanceOf[T]
  }
}

object DefaultParamsReader {

  /**
    * All info from metadata file.
    *
    * @param params       paramMap, as a [[JValue]]
    * @param metadata     All metadata, including the other fields
    * @param metadataJson Full metadata file String (for debugging)
    */
  case class Metadata(
                       className: String,
                       uid: String,
                       timestamp: Long,
                       sparkVersion: String,
                       params: JValue,
                       metadata: JValue,
                       metadataJson: String)

  /**
    * Load metadata from file.
    *
    * @param expectedClassName If non empty, this is checked against the loaded metadata.
    * @throws IllegalArgumentException if expectedClassName is specified and does not match metadata
    */
  def loadMetadata(path: String, sc: SparkContext, expectedClassName: String = ""): Metadata = {
    val metadataPath = new Path(path, "metadata").toString
    val metadataStr = sc.textFile(metadataPath, 1).first()
    val metadata = parse(metadataStr)

    implicit val format = DefaultFormats
    val className = (metadata \ "class").extract[String]
    val uid = (metadata \ "uid").extract[String]
    val timestamp = (metadata \ "timestamp").extract[Long]
    val sparkVersion = (metadata \ "sparkVersion").extract[String]
    val params = metadata \ "paramMap"
    if (expectedClassName.nonEmpty) {
      require(className == expectedClassName, s"Error loading metadata: Expected class name" +
        s" $expectedClassName but found class name $className")
    }

    Metadata(className, uid, timestamp, sparkVersion, params, metadata, metadataStr)
  }

  /**
    * Extract Params from metadata, and set them in the instance.
    * This works if all Params implement [[org.apache.spark.ml.param.Param.jsonDecode()]].
    */
  def getAndSetParams(instance: Params, metadata: Metadata): Unit = {
    implicit val format = DefaultFormats
    metadata.params match {
      case JObject(pairs) =>
        pairs.foreach { case (paramName, jsonValue) =>
          val param = instance.getParam(paramName)
          val value = param.jsonDecode(compact(render(jsonValue)))
          instance.set(param, value)
        }
      case _ =>
        throw new IllegalArgumentException(
          s"Cannot recognize JSON metadata: ${metadata.metadataJson}.")
    }
  }

  /**
    * Load a [[Params]] instance from the given path, and return it.
    * This assumes the instance implements [[MLReadable]].
    */
  def loadParamsInstance[T](path: String, sc: SparkContext): T = {
    val metadata = DefaultParamsReader.loadMetadata(path, sc)
    val cls = Class.forName(metadata.className, true, Option(Thread.currentThread().getContextClassLoader).getOrElse(getClass.getClassLoader))
    cls.getMethod("read").invoke(null).asInstanceOf[MLReader[T]].load(path)
  }
}

完成后，您可以使用CustomTransformerin aPipeline并保存/加载管道。我在 Spark shell 中很快进行了测试，它似乎可以工作，但肯定不太漂亮。