spark-reviews mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From mengxr <...@git.apache.org>
Subject [GitHub] spark pull request: [SPARK-2082] stratified sampling in PairRDDFun...
Date Wed, 09 Jul 2014 05:48:02 GMT
Github user mengxr commented on a diff in the pull request:

    https://github.com/apache/spark/pull/1025#discussion_r14694394
  
    --- Diff: core/src/main/scala/org/apache/spark/util/random/StratifiedSampler.scala ---
    @@ -0,0 +1,311 @@
    +/*
    + * 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.spark.util.random
    +
    +import scala.collection.Map
    +import scala.collection.mutable.{ArrayBuffer, HashMap, Map => MMap}
    +
    +import org.apache.commons.math3.random.RandomDataGenerator
    +import org.apache.spark.Logging
    +import org.apache.spark.rdd.RDD
    +
    +/**
    + * Auxiliary functions and data structures for the sampleByKey method in PairRDDFunctions.
    + *
    + * Essentially, when exact sample size is necessary, we make additional passes over the
RDD to
    + * compute the exact threshold value to use for each stratum to guarantee exact sample
size with
    + * high probability. This is achieved by maintaining a waitlist of size O(log(s)), where
s is the
    + * desired sample size for each stratum.
    + *
    + * Like in simple random sampling, we generate a random value for each item from the
    + * uniform  distribution [0.0, 1.0]. All items with values <= min(values of items
in the waitlist)
    + * are accepted into the sample instantly. The threshold for instant accept is designed
so that
    + * s - numAccepted = O(log(s)), where s is again the desired sample size. Thus, by maintaining
a
    + * waitlist size = O(log(s)), we will be able to create a sample of the exact size s
by adding
    + * a portion of the waitlist to the set of items that are instantly accepted. The exact
threshold
    + * is computed by sorting the values in the waitlist and picking the value at (s - numAccepted).
    + *
    + * Note that since we use the same seed for the RNG when computing the thresholds and
the actual
    + * sample, our computed thresholds are guaranteed to produce the desired sample size.
    + *
    + * For more theoretical background on the sampling techniques used here, please refer
to
    + * http://jmlr.org/proceedings/papers/v28/meng13a.html
    + */
    +
    +private[spark] object StratifiedSampler extends Logging {
    +
    +  /**
    +   * Count the number of items instantly accepted and generate the waitlist for each
stratum.
    +   *
    +   * This is only invoked when exact sample size is required.
    +   */
    +  def getCounts[K, V](rdd: RDD[(K, V)],
    +      withReplacement: Boolean,
    +      fractions: Map[K, Double],
    +      counts: Option[Map[K, Long]],
    +      seed: Long): MMap[K, Stratum] = {
    +    val combOp = getCombOp[K]
    +    val mappedPartitionRDD = rdd.mapPartitionsWithIndex({ case (partition, iter) =>
    +      val zeroU: MMap[K, Stratum] = new HashMap[K, Stratum]()
    +      val rng = new RandomDataGenerator()
    +      rng.reSeed(seed + partition)
    +      val seqOp = getSeqOp(withReplacement, fractions, rng, counts)
    +      Iterator(iter.aggregate(zeroU)(seqOp, combOp))
    +    }, preservesPartitioning=true)
    +    mappedPartitionRDD.reduce(combOp)
    +  }
    +
    +  /**
    +   * Returns the function used by aggregate to collect sampling statistics for each partition.
    +   */
    +  def getSeqOp[K, V](withReplacement: Boolean,
    +      fractions: Map[K, Double],
    +      rng: RandomDataGenerator,
    +      counts: Option[Map[K, Long]]): (MMap[K, Stratum], (K, V)) => MMap[K, Stratum]
= {
    +    val delta = 5e-5
    +    (result: MMap[K, Stratum], item: (K, V)) => {
    +      val key = item._1
    +      val fraction = fractions(key)
    +      if (!result.contains(key)) {
    +        result += (key -> new Stratum())
    +      }
    +      val stratum = result(key)
    +
    +      if (withReplacement) {
    +        // compute acceptBound and waitListBound only if they haven't been computed already
    +        // since they don't change from iteration to iteration.
    +        // TODO change this to the streaming version
    +        if (stratum.areBoundsEmpty) {
    +          val n = counts.get(key)
    +          val sampleSize = math.ceil(n * fraction).toLong
    +          val lmbd1 = PoissonBounds.getLowerBound(sampleSize)
    +          val minCount = PoissonBounds.getMinCount(lmbd1)
    +          val lmbd2 = if (lmbd1 == 0) {
    +            PoissonBounds.getUpperBound(sampleSize)
    +          } else {
    +            PoissonBounds.getUpperBound(sampleSize - minCount)
    +          }
    +          stratum.acceptBound = lmbd1 / n
    +          stratum.waitListBound = lmbd2 / n
    --- End diff --
    
    Sorry, this was my fault.


---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastructure@apache.org or file a JIRA ticket
with INFRA.
---

Mime
View raw message