[
https://issues.apache.org/jira/browse/FLINK-1723?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14633184#comment-14633184
]
ASF GitHub Bot commented on FLINK-1723:
---------------------------------------
Github user tillrohrmann commented on a diff in the pull request:
https://github.com/apache/flink/pull/891#discussion_r34976620
--- 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 --
The test shouldn't fail. Maybe there is an error then.
What one could do to have different sequences on each node is to xor the
subtask id with the seed. But IMO this does not change the statistical
properties of the sample because we don't know the underlying order of the
elements. E.g. the underlying order of the element could be that way that we
obtain the same sample set as with an identical seed and a different order.
> 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)
