flink-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "ASF GitHub Bot (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (FLINK-1723) Add cross validation for model evaluation
Date Mon, 20 Jul 2015 08:34:04 GMT

    [ https://issues.apache.org/jira/browse/FLINK-1723?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14633165#comment-14633165
] 

ASF GitHub Bot commented on FLINK-1723:
---------------------------------------

Github user sachingoel0101 commented on a diff in the pull request:

    https://github.com/apache/flink/pull/891#discussion_r34975724
  
    --- Diff: flink-staging/flink-ml/src/main/scala/org/apache/flink/ml/evaluation/CrossValidation.scala
---
    @@ -0,0 +1,97 @@
    +/*
    + * Licensed to the Apache Software Foundation (ASF) under one
    + * or more contributor license agreements.  See the NOTICE file
    + * distributed with this work for additional information
    + * regarding copyright ownership.  The ASF licenses this file
    + * to you under the Apache License, Version 2.0 (the
    + * "License"); you may not use this file except in compliance
    + * with the License.  You may obtain a copy of the License at
    + *
    + *     http://www.apache.org/licenses/LICENSE-2.0
    + *
    + * Unless required by applicable law or agreed to in writing, software
    + * distributed under the License is distributed on an "AS IS" BASIS,
    + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
    + * See the License for the specific language governing permissions and
    + * limitations under the License.
    + */
    +package org.apache.flink.ml.evaluation
    +
    +import org.apache.flink.api.scala._
    +import org.apache.flink.ml.RichDataSet
    +import java.util.Random
    +
    +import org.apache.flink.ml.pipeline.{EvaluateDataSetOperation, FitOperation, Predictor}
    +
    +object CrossValidation {
    +  def crossValScore[P <: Predictor[P], T](
    +      predictor: P,
    +      data: DataSet[T],
    +      scorerOption: Option[Scorer] = None,
    +      cv: FoldGenerator = KFold(),
    +      seed: Long = new Random().nextLong())(implicit fitOperation: FitOperation[P, T],
    +      evaluateDataSetOperation: EvaluateDataSetOperation[P, T, Double]): Array[DataSet[Double]]
= {
    +    val folds = cv.folds(data, 1)
    +
    +    val scores = folds.map {
    +      case (training: DataSet[T], testing: DataSet[T]) =>
    +        predictor.fit(training)
    +        if (scorerOption.isEmpty) {
    +          predictor.score(testing)
    +        } else {
    +          val s = scorerOption.get
    +          s.evaluate(testing, predictor)
    +        }
    +    }
    +    // TODO: Undecided on the return type: Array[DS[Double]] or DS[Double] i.e. reduce->union?
    +    // Or: Return mean and std?
    +    scores//.reduce((right: DataSet[Double], left: DataSet[Double]) => left.union(right)).mean()
    +  }
    +}
    +
    +abstract class FoldGenerator {
    +
    +  /** Takes a DataSet as input and creates splits (folds) of the data into
    +    * (training, testing) pairs.
    +    *
    +    * @param input The DataSet that will be split into folds
    +    * @param seed Seed for replicable splitting of the data
    +    * @tparam T The type of the DataSet
    +    * @return An Array containing K (training, testing) tuples, where training and testing
are
    +    *         DataSets
    +    */
    +  def folds[T](
    +      input: DataSet[T],
    +      seed: Long = new Random().nextLong()): Array[(DataSet[T], DataSet[T])]
    +}
    +
    +class KFold(numFolds: Int) extends FoldGenerator{
    +
    +  /** Takes a DataSet as input and creates K splits (folds) of the data into non-overlapping
    +    * (training, testing) pairs.
    +    *
    +    * Code based on Apache Spark implementation
    +    * @param input The DataSet that will be split into folds
    +    * @param seed Seed for replicable splitting of the data
    +    * @tparam T The type of the DataSet
    +    * @return An Array containing K (training, testing) tuples, where training and testing
are
    +    *         DataSets
    +    */
    +  override def folds[T](
    +      input: DataSet[T],
    +      seed: Long = new Random().nextLong()): Array[(DataSet[T], DataSet[T])] = {
    +    val numFoldsF = numFolds.toFloat
    +    (1 to numFolds).map { fold =>
    +      val lb = (fold - 1) / numFoldsF
    +      val ub = fold / numFoldsF
    +      val validation = input.sampleBounded(lb, ub, complement = false, seed = seed)
    +      val training = input.sampleBounded(lb, ub, complement = true, seed = seed)
    +      (training, validation)
    --- End diff --
    
    However, in case the parallelism of data is more than one, this can lead to problem. The
random number sequence generated on every node would be the same, wouldn't it?
    I printed all the random numbers generated and it looks like this: https://gist.github.com/sachingoel0101/ecde269af996fba7a39a
    
    Further, for a parallelism of 2, the test itself fails.


> Add cross validation for model evaluation
> -----------------------------------------
>
>                 Key: FLINK-1723
>                 URL: https://issues.apache.org/jira/browse/FLINK-1723
>             Project: Flink
>          Issue Type: New Feature
>          Components: Machine Learning Library
>            Reporter: Till Rohrmann
>            Assignee: Theodore Vasiloudis
>              Labels: ML
>
> Cross validation [1] is a standard tool to estimate the test error for a model. As such
it is a crucial tool for every machine learning library.
> The cross validation should work with arbitrary Estimators and error metrics. A first
cross validation strategy it should support is the k-fold cross validation.
> Resources:
> [1] [http://en.wikipedia.org/wiki/Cross-validation]



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Mime
View raw message