Github user thunterdb commented on a diff in the pull request:
https://github.com/apache/spark/pull/17419#discussion_r109063248
--- Diff: mllib/src/main/scala/org/apache/spark/ml/stat/Summarizer.scala ---
@@ -0,0 +1,746 @@
+/*
+ * 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 breeze.{linalg => la}
+import breeze.linalg.{Vector => BV}
+import breeze.numerics
+
+import org.apache.spark.SparkException
+import org.apache.spark.annotation.Since
+import org.apache.spark.internal.Logging
+import org.apache.spark.ml.linalg.{DenseVector, SparseVector, 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, UnsafeProjection, UnsafeRow}
+import
org.apache.spark.sql.catalyst.expressions.aggregate.{AggregateExpression,
Complete, TypedImperativeAggregate}
+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 column a column that contains Vector object.
+ * @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(column: Column): Column
+}
+
+/**
+ * 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(column: Column): Column = {
+ val start = SummaryBuilderImpl.Buffer.fromMetrics(requestedCompMetrics)
+ val agg = SummaryBuilderImpl.MetricsAggregate(
+ requestedMetrics,
+ start,
+ column.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)),
+ ("min", Min, arrayDType, Seq(ComputeMin)),
+ ("normL2", NormL2, arrayDType, Seq(ComputeM2)),
+ ("normL1", NormL1, arrayDType, Seq(ComputeL1))
+ )
+
+ /**
+ * The metrics that are currently implemented.
+ */
+ sealed trait Metrics
+ 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
+ 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
+
+ /**
+ * The buffer that contains all the summary statistics. If the value is
null, it is considered
+ * to be not required.
+ *
+ * If it is required but the size of the vectors (n) is not yet know, it
is initialized to
+ * an empty array.
+ */
+ case class Buffer private (
+ var n: Int = -1, // 0
+ var mean: Array[Double] = null, // 1
+ var m2n: Array[Double] = null, // 2
+ var m2: Array[Double] = null, // 3
+ var l1: Array[Double] = null, // 4
+ var totalCount: Long = 0, // 5
+ var totalWeightSum: Double = 0.0, // 6
+ var totalWeightSquareSum: Double = 0.0, // 7
+ var weightSum: Array[Double] = null, // 8
+ var nnz: Array[Long] = null, // 9
+ var max: Array[Double] = null, // 10
+ var min: Array[Double] = null // 11
+ ) {
+ override def toString: String = {
+ def v(x: Array[Double]) = if (x==null) "null" else
x.toSeq.mkString("[", " ", "]")
+ def vl(x: Array[Long]) = if (x==null) "null" else
x.toSeq.mkString("[", " ", "]")
+
+ s"Buffer(n=$n mean=${v(mean)} m2n=${v(m2n)} m2=${v(m2)}
l1=${v(l1)}" +
+ s" totalCount=$totalCount totalWeightSum=$totalWeightSum" +
+ s" totalWeightSquareSum=$totalWeightSquareSum
weightSum=${v(weightSum)} nnz=${vl(nnz)}" +
+ s" max=${v(max)} min=${v(min)})"
+ }
+ }
+
+ object Buffer extends Logging {
+ // Recursive function, but the number of cases is really small.
+ def fromMetrics(requested: Seq[ComputeMetrics]): Buffer = {
+ if (requested.isEmpty) {
+ new Buffer()
+ } else {
+ val b = fromMetrics(requested.tail)
+ requested.head match {
+ case ComputeMean => b.copy(mean = Array.empty)
+ case ComputeM2n => b.copy(m2n = Array.empty)
+ case ComputeM2 => b.copy(m2 = Array.empty)
+ case ComputeL1 => b.copy(l1 = Array.empty)
+ case ComputeWeightSum => b.copy(weightSum = Array.empty)
+ case ComputeNNZ => b.copy(nnz = Array.empty)
+ case ComputeMax => b.copy(max = Array.empty)
+ case ComputeMin => b.copy(min = Array.empty)
+ case _ => b // These cases are already being computed
+ }
+ }
+ }
+
+ /**
+ * (testing only). Makes a buffer with all the metrics enabled.
+ */
+ def allMetrics(): Buffer = {
+ fromMetrics(Seq(ComputeMean, ComputeM2n, ComputeM2, ComputeL1,
+ ComputeWeightSum, ComputeNNZ, ComputeMax,
+ ComputeMin))
+ }
+
+ val bufferSchema: StructType = {
+ val fields = Seq(
+ "n" -> IntegerType,
+ "mean" -> arrayDType,
+ "m2n" -> arrayDType,
+ "m2" -> arrayDType,
+ "l1" -> arrayDType,
+ "totalCount" -> LongType,
+ "totalWeightSum" -> DoubleType,
+ "totalWeightSquareSum" -> DoubleType,
+ "weightSum" -> arrayDType,
+ "nnz" -> arrayLType,
+ "max" -> arrayDType,
+ "min" -> arrayDType
+ )
+ StructType(fields.map { case (name, t) => StructField(name, t,
nullable = true)})
+ }
+
+ val numFields = bufferSchema.fields.length
+
+ def updateInPlace(buffer: Buffer, v: Vector, w: Double): Unit = {
+ val startN = buffer.n
+ if (startN == -1) {
+ // The buffer was not initialized, we initialize it with the
incoming row.
+ fillBufferWithRow(buffer, v, w)
+ return
+ } else {
+ require(startN == v.size,
+ s"Trying to insert a vector of size $v into a buffer that " +
+ s"has been sized with $startN")
+ }
+ val n = buffer.n
+ assert(n > 0, n)
+ // Always update the following fields.
+ buffer.totalWeightSum += w
+ buffer.totalCount += 1
+ buffer.totalWeightSquareSum += w * w
+ // All the fields that we compute on demand:
+ // TODO: the most common case is dense vectors. In that case we
should
+ // directly use BLAS instructions instead of iterating through a
scala iterator.
+ v.foreachActive { (index, value) =>
--- End diff --
Oh yes it does not. Note that the benchmark below is works with vectors of
size 1, so as to analyze the overhead of dataframes vs RDDs. I will put a more
realistic benchmark later.
---
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 infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org
