我正在使用Spark，我想培养机器学习模型 .
由于结果不好，我想在训练的每个时期（火车和测试数据集）上显示模型所犯的错误 .
然后，我将使用此信息来确定我的模型是否过度拟合或过度拟合数据 .
Question: 如何用火花绘制模型的学习曲线？
在下面的示例中，我实现了自己的求值程序并覆盖evaluate方法以打印我需要的度量标准，但只显示了两个值（maxIter = 1000） .
MinimalRunnableCode.scala：
import org.apache.spark.SparkConf
import org.apache.spark.ml.linalg.Vectors
import org.apache.spark.ml.regression.LinearRegression
import org.apache.spark.ml.tuning.{ParamGridBuilder, TrainValidationSplit}
import org.apache.spark.sql.SparkSession
object Min extends App {
// Open spark session.
val conf = new SparkConf()
.setMaster("local")
.set("spark.network.timeout", "800")
val ss = SparkSession.builder
.config(conf)
.getOrCreate
// Load data.
val data = ss.createDataFrame(ss.sparkContext.parallelize(
List(
(Vectors.dense(1, 2), 1),
(Vectors.dense(1, 3), 2),
(Vectors.dense(1, 2), 1),
(Vectors.dense(1, 3), 2),
(Vectors.dense(1, 2), 1),
(Vectors.dense(1, 3), 2),
(Vectors.dense(1, 2), 1),
(Vectors.dense(1, 3), 2),
(Vectors.dense(1, 2), 1),
(Vectors.dense(1, 3), 2),
(Vectors.dense(1, 4), 3)
)
))
.withColumnRenamed("_1", "features")
.withColumnRenamed("_2", "label")
val Array(training, test) = data.randomSplit(Array(0.8, 0.2), seed = 42)
// Create model of linear regression.
val lr = new LinearRegression().setMaxIter(1000)
// Create parameters grid that will be used to train different version of the linear model.
val paramGrid = new ParamGridBuilder()
.addGrid(lr.regParam, Array(0.001))
.addGrid(lr.fitIntercept)
.addGrid(lr.elasticNetParam, Array(0.5))
.build()
// Create trainer using validation split to evaluate which set of parameters performs the best.
val trainValidationSplit = new TrainValidationSplit()
.setEstimator(lr)
.setEvaluator(new CustomRegressionEvaluator)
.setEstimatorParamMaps(paramGrid)
.setTrainRatio(0.8) // 80% of the data will be used for training and the remaining 20% for validation.
// Run train validation split, and choose the best set of parameters.
var model = trainValidationSplit.fit(training)
// Close spark session.
ss.stop()
}
CustomRegressionEvaluator.scala：
import org.apache.spark.ml.evaluation.{Evaluator, RegressionEvaluator}
import org.apache.spark.ml.param.{Param, ParamMap, Params}
import org.apache.spark.ml.util.{DefaultParamsReadable, DefaultParamsWritable, Identifiable}
import org.apache.spark.mllib.evaluation.RegressionMetrics
import org.apache.spark.sql.{Dataset, Row}
import org.apache.spark.sql.functions._
import org.apache.spark.sql.types._
final class CustomRegressionEvaluator (override val uid: String) extends Evaluator with HasPredictionCol with HasLabelCol with DefaultParamsWritable {
def this() = this(Identifiable.randomUID("regEval"))
def checkNumericType(
schema: StructType,
colName: String,
msg: String = ""): Unit = {
val actualDataType = schema(colName).dataType
val message = if (msg != null && msg.trim.length > 0) " " + msg else ""
require(actualDataType.isInstanceOf[NumericType], s"Column $colName must be of type " +
s"NumericType but was actually of type $actualDataType.$message")
}
def checkColumnTypes(
schema: StructType,
colName: String,
dataTypes: Seq[DataType],
msg: String = ""): Unit = {
val actualDataType = schema(colName).dataType
val message = if (msg != null && msg.trim.length > 0) " " + msg else ""
require(dataTypes.exists(actualDataType.equals),
s"Column $colName must be of type equal to one of the following types: " +
s"${dataTypes.mkString("[", ", ", "]")} but was actually of type $actualDataType.$message")
}
var i = 0 // count the number of time the evaluate method is called
override def evaluate(dataset: Dataset[_]): Double = {
val schema = dataset.schema
checkColumnTypes(schema, $(predictionCol), Seq(DoubleType, FloatType))
checkNumericType(schema, $(labelCol))
val predictionAndLabels = dataset
.select(col($(predictionCol)).cast(DoubleType), col($(labelCol)).cast(DoubleType))
.rdd
.map { case Row(prediction: Double, label: Double) => (prediction, label) }
val metrics = new RegressionMetrics(predictionAndLabels)
val metric = "mae" match {
case "rmse" => metrics.rootMeanSquaredError
case "mse" => metrics.meanSquaredError
case "r2" => metrics.r2
case "mae" => metrics.meanAbsoluteError
}
println(s"$i $metric") // Print the metrics
i = i + 1 // Update counter
metric
}
override def copy(extra: ParamMap): RegressionEvaluator = defaultCopy(extra)
}
object RegressionEvaluator extends DefaultParamsReadable[RegressionEvaluator] {
override def load(path: String): RegressionEvaluator = super.load(path)
}
private[ml] trait HasPredictionCol extends Params {
/**
* Param for prediction column name.
* @group param
*/
final val predictionCol: Param[String] = new Param[String](this, "predictionCol", "prediction column name")
setDefault(predictionCol, "prediction")
/** @group getParam */
final def getPredictionCol: String = $(predictionCol)
}
private[ml] trait HasLabelCol extends Params {
/**
* Param for label column name.
* @group param
*/
final val labelCol: Param[String] = new Param[String](this, "labelCol", "label column name")
setDefault(labelCol, "label")
/** @group getParam */
final def getLabelCol: String = $(labelCol)
}
1 回答
这是针对
LinearRegression的特定情况以及支持客观历史的任何其他算法的可能解决方案（在这种情况下，AndLinearRegressionTrainingSummary完成工作） .让我们首先创建一个最小的可验证的完整示例：
正如您所注意到的，当您想为
spark-mllib或spark-ml的测试目的生成数据时，建议使用数据生成器 .现在，让我们训练一个线性回归量：
现在，由于我们的模型经过培训，我们所需要的只是获得客观的历史 .
以下部分需要在模型和子模型对象参数之间进行一些体操 .
如果您有一个
Pipeline左右，则需要修改此代码，因此请谨慎使用 . 这只是一个例子：我们现在可以检查这些指标：
您可以注意到，训练过程实际上在4次迭代后停止 .
如果只需要迭代次数，则可以执行以下操作：
为了演示，我已经更改了生成器中的功能数量（
nfeatures = 100）：