spark-reviews mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From WeichenXu123 <...@git.apache.org>
Subject [GitHub] spark pull request #18798: [SPARK-19634][ML] Multivariate summarizer - dataf...
Date Tue, 01 Aug 2017 22:38:18 GMT
Github user WeichenXu123 commented on a diff in the pull request:

    https://github.com/apache/spark/pull/18798#discussion_r130746893
  
    --- Diff: mllib/src/main/scala/org/apache/spark/ml/stat/Summarizer.scala ---
    @@ -0,0 +1,633 @@
    +/*
    + * 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.ml.stat
    +
    +import java.io._
    +
    +import org.apache.spark.annotation.Since
    +import org.apache.spark.internal.Logging
    +import org.apache.spark.ml.linalg.{Vector, Vectors, VectorUDT}
    +import org.apache.spark.sql.Column
    +import org.apache.spark.sql.catalyst.InternalRow
    +import org.apache.spark.sql.catalyst.expressions.{Expression, UnsafeArrayData}
    +import org.apache.spark.sql.catalyst.expressions.aggregate.{AggregateExpression, Complete,
TypedImperativeAggregate}
    +import org.apache.spark.sql.catalyst.util.ArrayData
    +import org.apache.spark.sql.functions.lit
    +import org.apache.spark.sql.types._
    +
    +/**
    + * A builder object that provides summary statistics about a given column.
    + *
    + * Users should not directly create such builders, but instead use one of the methods
in
    + * [[Summarizer]].
    + */
    +@Since("2.2.0")
    +abstract class SummaryBuilder {
    +  /**
    +   * Returns an aggregate object that contains the summary of the column with the requested
metrics.
    +   * @param featuresCol a column that contains features Vector object.
    +   * @param weightCol a column that contains weight value.
    +   * @return an aggregate column that contains the statistics. The exact content of this
    +   *         structure is determined during the creation of the builder.
    +   */
    +  @Since("2.2.0")
    +  def summary(featuresCol: Column, weightCol: Column): Column
    +
    +  @Since("2.2.0")
    +  def summary(featuresCol: Column): Column = summary(featuresCol, lit(1.0))
    +}
    +
    +/**
    + * Tools for vectorized statistics on MLlib Vectors.
    + *
    + * The methods in this package provide various statistics for Vectors contained inside
DataFrames.
    + *
    + * This class lets users pick the statistics they would like to extract for a given column.
Here is
    + * an example in Scala:
    + * {{{
    + *   val dataframe = ... // Some dataframe containing a feature column
    + *   val allStats = dataframe.select(Summarizer.metrics("min", "max").summary($"features"))
    + *   val Row(min_, max_) = allStats.first()
    + * }}}
    + *
    + * If one wants to get a single metric, shortcuts are also available:
    + * {{{
    + *   val meanDF = dataframe.select(Summarizer.mean($"features"))
    + *   val Row(mean_) = meanDF.first()
    + * }}}
    + */
    +@Since("2.2.0")
    +object Summarizer extends Logging {
    +
    +  import SummaryBuilderImpl._
    +
    +  /**
    +   * Given a list of metrics, provides a builder that it turns computes metrics from
a column.
    +   *
    +   * See the documentation of [[Summarizer]] for an example.
    +   *
    +   * The following metrics are accepted (case sensitive):
    +   *  - mean: a vector that contains the coefficient-wise mean.
    +   *  - variance: a vector tha contains the coefficient-wise variance.
    +   *  - count: the count of all vectors seen.
    +   *  - numNonzeros: a vector with the number of non-zeros for each coefficients
    +   *  - max: the maximum for each coefficient.
    +   *  - min: the minimum for each coefficient.
    +   *  - normL2: the Euclidian norm for each coefficient.
    +   *  - normL1: the L1 norm of each coefficient (sum of the absolute values).
    +   * @param firstMetric the metric being provided
    +   * @param metrics additional metrics that can be provided.
    +   * @return a builder.
    +   * @throws IllegalArgumentException if one of the metric names is not understood.
    +   */
    +  @Since("2.2.0")
    +  def metrics(firstMetric: String, metrics: String*): SummaryBuilder = {
    +    val (typedMetrics, computeMetrics) = getRelevantMetrics(Seq(firstMetric) ++ metrics)
    +    new SummaryBuilderImpl(typedMetrics, computeMetrics)
    +  }
    +
    +  def mean(col: Column): Column = getSingleMetric(col, "mean")
    +
    +  def variance(col: Column): Column = getSingleMetric(col, "variance")
    +
    +  def count(col: Column): Column = getSingleMetric(col, "count")
    +
    +  def numNonZeros(col: Column): Column = getSingleMetric(col, "numNonZeros")
    +
    +  def max(col: Column): Column = getSingleMetric(col, "max")
    +
    +  def min(col: Column): Column = getSingleMetric(col, "min")
    +
    +  def normL1(col: Column): Column = getSingleMetric(col, "normL1")
    +
    +  def normL2(col: Column): Column = getSingleMetric(col, "normL2")
    +
    +  private def getSingleMetric(col: Column, metric: String): Column = {
    +    val c1 = metrics(metric).summary(col)
    +    c1.getField(metric).as(s"$metric($col)")
    +  }
    +}
    +
    +private[ml] class SummaryBuilderImpl(
    +    requestedMetrics: Seq[SummaryBuilderImpl.Metrics],
    +    requestedCompMetrics: Seq[SummaryBuilderImpl.ComputeMetrics]
    +  ) extends SummaryBuilder {
    +
    +  override def summary(featuresCol: Column, weightCol: Column): Column = {
    +
    +    val agg = SummaryBuilderImpl.MetricsAggregate(
    +      requestedMetrics,
    +      requestedCompMetrics,
    +      featuresCol.expr,
    +      weightCol.expr,
    +      mutableAggBufferOffset = 0,
    +      inputAggBufferOffset = 0)
    +
    +    new Column(AggregateExpression(agg, mode = Complete, isDistinct = false))
    +  }
    +}
    +
    +private[ml]
    +object SummaryBuilderImpl extends Logging {
    +
    +  def implementedMetrics: Seq[String] = allMetrics.map(_._1).sorted
    +
    +  @throws[IllegalArgumentException]("When the list is empty or not a subset of known
metrics")
    +  def getRelevantMetrics(requested: Seq[String]): (Seq[Metrics], Seq[ComputeMetrics])
= {
    +    val all = requested.map { req =>
    +      val (_, metric, _, deps) = allMetrics.find(tup => tup._1 == req).getOrElse {
    +        throw new IllegalArgumentException(s"Metric $req cannot be found." +
    +          s" Valid metrics are $implementedMetrics")
    +      }
    +      metric -> deps
    +    }
    +    // Do not sort, otherwise the user has to look the schema to see the order that it
    +    // is going to be given in.
    +    val metrics = all.map(_._1)
    +    val computeMetrics = all.flatMap(_._2).distinct.sortBy(_.toString)
    +    metrics -> computeMetrics
    +  }
    +
    +  def structureForMetrics(metrics: Seq[Metrics]): StructType = {
    +    val dct = allMetrics.map { case (n, m, dt, _) => (m, (n, dt)) }.toMap
    +    val fields = metrics.map(dct.apply).map { case (n, dt) =>
    +        StructField(n, dt, nullable = false)
    +    }
    +    StructType(fields)
    +  }
    +
    +  private val arrayDType = ArrayType(DoubleType, containsNull = false)
    +  private val arrayLType = ArrayType(LongType, containsNull = false)
    +
    +  /**
    +   * All the metrics that can be currently computed by Spark for vectors.
    +   *
    +   * This list associates the user name, the internal (typed) name, and the list of computation
    +   * metrics that need to de computed internally to get the final result.
    +   */
    +  private val allMetrics: Seq[(String, Metrics, DataType, Seq[ComputeMetrics])] = Seq(
    +    ("mean", Mean, arrayDType, Seq(ComputeMean, ComputeWeightSum)),
    +    ("variance", Variance, arrayDType, Seq(ComputeWeightSum, ComputeMean, ComputeM2n)),
    +    ("count", Count, LongType, Seq()),
    +    ("numNonZeros", NumNonZeros, arrayLType, Seq(ComputeNNZ)),
    +    ("max", Max, arrayDType, Seq(ComputeMax, ComputeNNZ)),
    +    ("min", Min, arrayDType, Seq(ComputeMin, ComputeNNZ)),
    +    ("normL2", NormL2, arrayDType, Seq(ComputeM2)),
    +    ("normL1", NormL1, arrayDType, Seq(ComputeL1))
    +  )
    +
    +  /**
    +   * The metrics that are currently implemented.
    +   */
    +  sealed trait Metrics extends Serializable
    +  case object Mean extends Metrics
    +  case object Variance extends Metrics
    +  case object Count extends Metrics
    +  case object NumNonZeros extends Metrics
    +  case object Max extends Metrics
    +  case object Min extends Metrics
    +  case object NormL2 extends Metrics
    +  case object NormL1 extends Metrics
    +
    +  /**
    +   * The running metrics that are going to be computed.
    +   *
    +   * There is a bipartite graph between the metrics and the computed metrics.
    +   */
    +  sealed trait ComputeMetrics extends Serializable
    +  case object ComputeMean extends ComputeMetrics
    +  case object ComputeM2n extends ComputeMetrics
    +  case object ComputeM2 extends ComputeMetrics
    +  case object ComputeL1 extends ComputeMetrics
    +  case object ComputeWeightSum extends ComputeMetrics
    +  case object ComputeNNZ extends ComputeMetrics
    +  case object ComputeMax extends ComputeMetrics
    +  case object ComputeMin extends ComputeMetrics
    +
    +  class SummarizerBuffer(
    +      requestedMetrics: Seq[Metrics],
    +      requestedCompMetrics: Seq[ComputeMetrics]
    +  ) extends Serializable {
    +
    +    private var n = 0
    +    private var currMean: Array[Double] = null
    +    private var currM2n: Array[Double] = null
    +    private var currM2: Array[Double] = null
    +    private var currL1: Array[Double] = null
    +    private var totalCnt: Long = 0
    +    private var totalWeightSum: Double = 0.0
    +    private var weightSquareSum: Double = 0.0
    +    private var weightSum: Array[Double] = null
    +    private var nnz: Array[Long] = null
    +    private var currMax: Array[Double] = null
    +    private var currMin: Array[Double] = null
    +
    +    def this() {
    +      this(
    +        Seq(Mean, Variance, Count, NumNonZeros, Max, Min, NormL2, NormL1),
    +        Seq(ComputeMean, ComputeM2n, ComputeM2, ComputeL1,
    +          ComputeWeightSum, ComputeNNZ, ComputeMax, ComputeMin)
    +      )
    +    }
    +
    +    /**
    +     * Add a new sample to this summarizer, and update the statistical summary.
    +     */
    +    def addRaw(instance: TraversableIndexedSeq, weight: Double): this.type = {
    +      require(weight >= 0.0, s"sample weight, ${weight} has to be >= 0.0")
    +      if (weight == 0.0) return this
    +
    +      if (n == 0) {
    +        require(instance.size > 0, s"Vector should have dimension larger than zero.")
    +        n = instance.size
    +
    +        if (requestedCompMetrics.contains(ComputeMean)) { currMean = Array.ofDim[Double](n)
}
    +        if (requestedCompMetrics.contains(ComputeM2n)) { currM2n = Array.ofDim[Double](n)
}
    +        if (requestedCompMetrics.contains(ComputeM2)) { currM2 = Array.ofDim[Double](n)
}
    +        if (requestedCompMetrics.contains(ComputeL1)) { currL1 = Array.ofDim[Double](n)
}
    +        if (requestedCompMetrics.contains(ComputeWeightSum)) { weightSum = Array.ofDim[Double](n)
}
    +        if (requestedCompMetrics.contains(ComputeNNZ)) { nnz = Array.ofDim[Long](n) }
    +        if (requestedCompMetrics.contains(ComputeMax)) {
    +          currMax = Array.fill[Double](n)(Double.MinValue)
    +        }
    +        if (requestedCompMetrics.contains(ComputeMin)) {
    +          currMin = Array.fill[Double](n)(Double.MaxValue)
    +        }
    +      }
    +
    +      require(n == instance.size, s"Dimensions mismatch when adding new sample." +
    +        s" Expecting $n but got ${instance.size}.")
    +
    +      val localCurrMean = currMean
    +      val localCurrM2n = currM2n
    +      val localCurrM2 = currM2
    +      val localCurrL1 = currL1
    +      val localWeightSum = weightSum
    +      val localNumNonzeros = nnz
    +      val localCurrMax = currMax
    +      val localCurrMin = currMin
    +      instance.foreachActive { (index, value) =>
    +        if (value != 0.0) {
    +          if (localCurrMax != null && localCurrMax(index) < value) {
    +            localCurrMax(index) = value
    +          }
    +          if (localCurrMin != null && localCurrMin(index) > value) {
    +            localCurrMin(index) = value
    +          }
    +
    +          if (localWeightSum != null) {
    +            if (localCurrMean != null) {
    +              val prevMean = localCurrMean(index)
    +              val diff = value - prevMean
    +              localCurrMean(index) = prevMean + weight * diff / (localWeightSum(index)
+ weight)
    +
    +              if (localCurrM2n != null) {
    +                localCurrM2n(index) += weight * (value - localCurrMean(index)) * diff
    +              }
    +            }
    +            localWeightSum(index) += weight
    +          }
    +
    +          if (localCurrM2 != null) {
    +            localCurrM2(index) += weight * value * value
    +          }
    +          if (localCurrL1 != null) {
    +            localCurrL1(index) += weight * math.abs(value)
    +          }
    +
    +          if (localNumNonzeros != null) {
    +            localNumNonzeros(index) += 1
    +          }
    +        }
    +      }
    +
    +      totalWeightSum += weight
    +      weightSquareSum += weight * weight
    +      totalCnt += 1
    +      this
    +    }
    +
    +    def addRaw(instance: TraversableIndexedSeq): this.type = addRaw(instance, 1.0)
    +
    +    // For test
    +    def add(sample: Vector, weight: Double): this.type = {
    +      val v = new TraversableIndexedSeq {
    --- End diff --
    
    Yes... this method is only used in testsuite (let writing testing more convenient). 


---
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.
---

---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscribe@spark.apache.org
For additional commands, e-mail: reviews-help@spark.apache.org


Mime
View raw message