[
https://issues.apache.org/jira/browse/FLINK-2259?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15253911#comment-15253911
]
ASF GitHub Bot commented on FLINK-2259:
---------------------------------------
Github user tillrohrmann commented on a diff in the pull request:
https://github.com/apache/flink/pull/1898#discussion_r60737947
--- Diff:
flink-libraries/flink-ml/src/main/scala/org/apache/flink/ml/preprocessing/Splitter.scala
---
@@ -0,0 +1,215 @@
+/*
+ * 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.preprocessing
+
+import org.apache.flink.api.common.typeinfo.{TypeInformation,
BasicTypeInfo}
+import org.apache.flink.api.java.Utils
+import org.apache.flink.api.scala. DataSet
+import org.apache.flink.api.scala.utils._
+
+import org.apache.flink.ml.common.{FlinkMLTools, ParameterMap,
WithParameters}
+import _root_.scala.reflect.ClassTag
+
+object Splitter {
+
+ case class TrainTestDataSet[T: TypeInformation : ClassTag](training:
DataSet[T],
+ testing:
DataSet[T])
+
+ case class TrainTestHoldoutDataSet[T: TypeInformation :
ClassTag](training: DataSet[T],
+
testing: DataSet[T],
+
holdout: DataSet[T])
+ //
--------------------------------------------------------------------------------------------
+ // randomSplit
+ //
--------------------------------------------------------------------------------------------
+ /**
+ * Split a DataSet by the probability fraction of each element.
+ *
+ * @param input DataSet to be split
+ * @param fraction Probability that each element is chosen,
should be [0,1] without
+ * replacement, and [0, ∞) with replacement.
While fraction is larger
+ * than 1, the elements are expected to be
selected multi times into
+ * sample on average. This fraction refers to the
first element in the
+ * resulting array.
+ * @param precise Sampling by default is random and can result
in slightly lop-sided
+ * sample sets. When precise is true, equal
sample set size are forced,
+ * however this is somewhat less efficient.
+ * @param seed Random number generator seed.
+ * @return An array of two datasets
+ */
+
+ def randomSplit[T: TypeInformation : ClassTag]( input: DataSet[T],
+ fraction: Double,
+ precise: Boolean = false,
+ seed: Long =
Utils.RNG.nextLong())
+ : Array[DataSet[T]] = {
+ import org.apache.flink.api.scala._
+
+ val indexedInput: DataSet[(Long, T)] = input.zipWithIndex
+
+ val leftSplit: DataSet[(Long, T)] = precise match {
+ case false => indexedInput.sample(false, fraction, seed)
+ case true => {
+ val count = indexedInput.count()
+ val numOfSamples = math.round(fraction * count).toInt
+ indexedInput.sampleWithSize(false, numOfSamples, seed)
+ }
+ }
+
+ val rightSplit: DataSet[(Long, T)] = indexedInput.leftOuterJoin[(Long,
T)](leftSplit)
+ .where(0)
+ .equalTo(0) {
+ (full: (Long,T) , left: (Long, T)) => (if (left == null) full
else null)
+ }
+ .filter( o => o != null )
+ Array(leftSplit.map(o => o._2), rightSplit.map(o => o._2))
+ }
+
+ //
--------------------------------------------------------------------------------------------
+ // multiRandomSplit
+ //
--------------------------------------------------------------------------------------------
+ /**
+ * Split a DataSet by the probability fraction of each element of a
vector.
+ *
+ * @param input DataSet to be split
+ * @param fracArray An array of PROPORTIONS for splitting the
DataSet. Unlike the
+ * randomSplit function, number greater than 1 do
not lead to over
+ * sampling. The number of splits is dictated by
the length of this array.
+ * The number are normalized, eg. Array(1.0, 2.0)
would yield
+ * two data sets with a 33/66% split.
+ * @param precise Sampling by default is random and can result
in slightly lop-sided
+ * sample sets. When precise is true, equal
sample set size are forced,
+ * however this is somewhat less efficient.
+ * @param seed Random number generator seed.
+ * @return An array of DataSets whose length is equal to the length of
fracArray
+ */
+ def multiRandomSplit[T: TypeInformation : ClassTag](input: DataSet[T],
+ fracArray: Array[Double],
+ precise: Boolean = false,
+ seed: Long = Utils.RNG.nextLong())
+ : Array[DataSet[T]] = {
+ val splits = fracArray.length
+ val output = new Array[DataSet[T]](splits)
+ val aggs = fracArray.scanRight((0.0))( _ + _ )
+ val fracs = fracArray.zip(aggs).map( o => o._1 / o._2)
+
+ ////
+ var tempDS = input
+ for (k <- 0 to splits-2){
+ println( (splits -k))
+ var temp = Splitter.randomSplit(tempDS, fracs(k), true)
+ output(k) = temp(0)
+ tempDS = temp(1)
+ }
--- End diff --
I'm not so sure, whether this scales so well. The reason is that we're
constructing some really long pipelines with this for loop where each iteration
contains an outer join and is the input for the next iteration. I'm wondering
whether we cannot simply achieve the assignment of elements to splits in one
iteration. We could for example generate for each element a random number and
see in which bin it falls (indicated by the normalized fraction array). Then we
could apply a number of filters to the result with the assigned splits in order
to obtain the different splits.
> Support training Estimators using a (train, validation, test) split of the
> available data
> -----------------------------------------------------------------------------------------
>
> Key: FLINK-2259
> URL: https://issues.apache.org/jira/browse/FLINK-2259
> Project: Flink
> Issue Type: New Feature
> Components: Machine Learning Library
> Reporter: Theodore Vasiloudis
> Assignee: Trevor Grant
> Priority: Minor
> Labels: ML
>
> When there is an abundance of data available, a good way to train models is
> to split the available data into 3 parts: Train, Validation and Test.
> We use the Train data to train the model, the Validation part is used to
> estimate the test error and select hyperparameters, and the Test is used to
> evaluate the performance of the model, and assess its generalization [1]
> This is a common approach when training Artificial Neural Networks, and a
> good strategy to choose in data-rich environments. Therefore we should have
> some support of this data-analysis process in our Estimators.
> [1] Friedman, Jerome, Trevor Hastie, and Robert Tibshirani. The elements of
> statistical learning. Vol. 1. Springer, Berlin: Springer series in
> statistics, 2001.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
