Skip to content

Commit 5596086

Browse files
dbtsaimengxr
dbtsai
authored and
mengxr
committed
[SPARK-1969][MLlib] Online summarizer APIs for mean, variance, min, and max
It basically moved the private ColumnStatisticsAggregator class from RowMatrix to public available DeveloperApi with documentation and unitests. Changes: 1) Moved the private implementation from org.apache.spark.mllib.linalg.ColumnStatisticsAggregator to org.apache.spark.mllib.stat.MultivariateOnlineSummarizer 2) When creating OnlineSummarizer object, the number of columns is not needed in the constructor. It's determined when users add the first sample. 3) Added the APIs documentation for MultivariateOnlineSummarizer. 4) Added the unittests for MultivariateOnlineSummarizer. Author: DB Tsai <[email protected]> Closes apache#955 from dbtsai/dbtsai-summarizer and squashes the following commits: b13ac90 [DB Tsai] dbtsai-summarizer
1 parent cbff187 commit 5596086

File tree

5 files changed

+458
-134
lines changed

5 files changed

+458
-134
lines changed

‎mllib/src/main/scala/org/apache/spark/mllib/linalg/distributed/RowMatrix.scala‎

Lines changed: 2 additions & 134 deletions
Original file line numberDiff line numberDiff line change
@@ -28,138 +28,7 @@ import org.apache.spark.annotation.Experimental
2828
import org.apache.spark.mllib.linalg._
2929
import org.apache.spark.rdd.RDD
3030
import org.apache.spark.Logging
31-
import org.apache.spark.mllib.stat.MultivariateStatisticalSummary
32-
33-
/**
34-
* Column statistics aggregator implementing
35-
* [[org.apache.spark.mllib.stat.MultivariateStatisticalSummary]]
36-
* together with add() and merge() function.
37-
* A numerically stable algorithm is implemented to compute sample mean and variance:
38-
* [[http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance variance-wiki]].
39-
* Zero elements (including explicit zero values) are skipped when calling add() and merge(),
40-
* to have time complexity O(nnz) instead of O(n) for each column.
41-
*/
42-
private class ColumnStatisticsAggregator(private val n: Int)
43-
extends MultivariateStatisticalSummary with Serializable {
44-
45-
private val currMean: BDV[Double] = BDV.zeros[Double](n)
46-
private val currM2n: BDV[Double] = BDV.zeros[Double](n)
47-
private var totalCnt = 0.0
48-
private val nnz: BDV[Double] = BDV.zeros[Double](n)
49-
private val currMax: BDV[Double] = BDV.fill(n)(Double.MinValue)
50-
private val currMin: BDV[Double] = BDV.fill(n)(Double.MaxValue)
51-
52-
override def mean: Vector = {
53-
val realMean = BDV.zeros[Double](n)
54-
var i = 0
55-
while (i < n) {
56-
realMean(i) = currMean(i) * nnz(i) / totalCnt
57-
i += 1
58-
}
59-
Vectors.fromBreeze(realMean)
60-
}
61-
62-
override def variance: Vector = {
63-
val realVariance = BDV.zeros[Double](n)
64-
65-
val denominator = totalCnt - 1.0
66-
67-
// Sample variance is computed, if the denominator is less than 0, the variance is just 0.
68-
if (denominator > 0.0) {
69-
val deltaMean = currMean
70-
var i = 0
71-
while (i < currM2n.size) {
72-
realVariance(i) =
73-
currM2n(i) + deltaMean(i) * deltaMean(i) * nnz(i) * (totalCnt - nnz(i)) / totalCnt
74-
realVariance(i) /= denominator
75-
i += 1
76-
}
77-
}
78-
79-
Vectors.fromBreeze(realVariance)
80-
}
81-
82-
override def count: Long = totalCnt.toLong
83-
84-
override def numNonzeros: Vector = Vectors.fromBreeze(nnz)
85-
86-
override def max: Vector = {
87-
var i = 0
88-
while (i < n) {
89-
if ((nnz(i) < totalCnt) && (currMax(i) < 0.0)) currMax(i) = 0.0
90-
i += 1
91-
}
92-
Vectors.fromBreeze(currMax)
93-
}
94-
95-
override def min: Vector = {
96-
var i = 0
97-
while (i < n) {
98-
if ((nnz(i) < totalCnt) && (currMin(i) > 0.0)) currMin(i) = 0.0
99-
i += 1
100-
}
101-
Vectors.fromBreeze(currMin)
102-
}
103-
104-
/**
105-
* Aggregates a row.
106-
*/
107-
def add(currData: BV[Double]): this.type = {
108-
currData.activeIterator.foreach {
109-
case (_, 0.0) => // Skip explicit zero elements.
110-
case (i, value) =>
111-
if (currMax(i) < value) {
112-
currMax(i) = value
113-
}
114-
if (currMin(i) > value) {
115-
currMin(i) = value
116-
}
117-
118-
val tmpPrevMean = currMean(i)
119-
currMean(i) = (currMean(i) * nnz(i) + value) / (nnz(i) + 1.0)
120-
currM2n(i) += (value - currMean(i)) * (value - tmpPrevMean)
121-
122-
nnz(i) += 1.0
123-
}
124-
125-
totalCnt += 1.0
126-
this
127-
}
128-
129-
/**
130-
* Merges another aggregator.
131-
*/
132-
def merge(other: ColumnStatisticsAggregator): this.type = {
133-
require(n == other.n, s"Dimensions mismatch. Expecting $n but got ${other.n}.")
134-
135-
totalCnt += other.totalCnt
136-
val deltaMean = currMean - other.currMean
137-
138-
var i = 0
139-
while (i < n) {
140-
// merge mean together
141-
if (other.currMean(i) != 0.0) {
142-
currMean(i) = (currMean(i) * nnz(i) + other.currMean(i) * other.nnz(i)) /
143-
(nnz(i) + other.nnz(i))
144-
}
145-
// merge m2n together
146-
if (nnz(i) + other.nnz(i) != 0.0) {
147-
currM2n(i) += other.currM2n(i) + deltaMean(i) * deltaMean(i) * nnz(i) * other.nnz(i) /
148-
(nnz(i) + other.nnz(i))
149-
}
150-
if (currMax(i) < other.currMax(i)) {
151-
currMax(i) = other.currMax(i)
152-
}
153-
if (currMin(i) > other.currMin(i)) {
154-
currMin(i) = other.currMin(i)
155-
}
156-
i += 1
157-
}
158-
159-
nnz += other.nnz
160-
this
161-
}
162-
}
31+
import org.apache.spark.mllib.stat.{MultivariateOnlineSummarizer, MultivariateStatisticalSummary}
16332

16433
/**
16534
* :: Experimental ::
@@ -478,8 +347,7 @@ class RowMatrix(
478347
* Computes column-wise summary statistics.
479348
*/
480349
def computeColumnSummaryStatistics(): MultivariateStatisticalSummary = {
481-
val zeroValue = new ColumnStatisticsAggregator(numCols().toInt)
482-
val summary = rows.map(_.toBreeze).aggregate[ColumnStatisticsAggregator](zeroValue)(
350+
val summary = rows.aggregate[MultivariateOnlineSummarizer](new MultivariateOnlineSummarizer)(
483351
(aggregator, data) => aggregator.add(data),
484352
(aggregator1, aggregator2) => aggregator1.merge(aggregator2)
485353
)

‎mllib/src/main/scala/org/apache/spark/mllib/stat/MultivariateOnlineSummarizer.scala‎

Lines changed: 201 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,201 @@
1+
/*
2+
* Licensed to the Apache Software Foundation (ASF) under one or more
3+
* contributor license agreements. See the NOTICE file distributed with
4+
* this work for additional information regarding copyright ownership.
5+
* The ASF licenses this file to You under the Apache License, Version 2.0
6+
* (the "License"); you may not use this file except in compliance with
7+
* the License. You may obtain a copy of the License at
8+
*
9+
* http://www.apache.org/licenses/LICENSE-2.0
10+
*
11+
* Unless required by applicable law or agreed to in writing, software
12+
* distributed under the License is distributed on an "AS IS" BASIS,
13+
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14+
* See the License for the specific language governing permissions and
15+
* limitations under the License.
16+
*/
17+
18+
package org.apache.spark.mllib.stat
19+
20+
import breeze.linalg.{DenseVector => BDV}
21+
22+
import org.apache.spark.annotation.DeveloperApi
23+
import org.apache.spark.mllib.linalg.{Vectors, Vector}
24+
25+
/**
26+
* :: DeveloperApi ::
27+
* MultivariateOnlineSummarizer implements [[MultivariateStatisticalSummary]] to compute the mean,
28+
* variance, minimum, maximum, counts, and nonzero counts for samples in sparse or dense vector
29+
* format in a online fashion.
30+
*
31+
* Two MultivariateOnlineSummarizer can be merged together to have a statistical summary of
32+
* the corresponding joint dataset.
33+
*
34+
* A numerically stable algorithm is implemented to compute sample mean and variance:
35+
* Reference: [[http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance variance-wiki]]
36+
* Zero elements (including explicit zero values) are skipped when calling add(),
37+
* to have time complexity O(nnz) instead of O(n) for each column.
38+
*/
39+
@DeveloperApi
40+
class MultivariateOnlineSummarizer extends MultivariateStatisticalSummary with Serializable {
41+
42+
private var n = 0
43+
private var currMean: BDV[Double] = _
44+
private var currM2n: BDV[Double] = _
45+
private var totalCnt: Long = 0
46+
private var nnz: BDV[Double] = _
47+
private var currMax: BDV[Double] = _
48+
private var currMin: BDV[Double] = _
49+
50+
/**
51+
* Add a new sample to this summarizer, and update the statistical summary.
52+
*
53+
* @param sample The sample in dense/sparse vector format to be added into this summarizer.
54+
* @return This MultivariateOnlineSummarizer object.
55+
*/
56+
def add(sample: Vector): this.type = {
57+
if (n == 0) {
58+
require(sample.toBreeze.length > 0, s"Vector should have dimension larger than zero.")
59+
n = sample.toBreeze.length
60+
61+
currMean = BDV.zeros[Double](n)
62+
currM2n = BDV.zeros[Double](n)
63+
nnz = BDV.zeros[Double](n)
64+
currMax = BDV.fill(n)(Double.MinValue)
65+
currMin = BDV.fill(n)(Double.MaxValue)
66+
}
67+
68+
require(n == sample.toBreeze.length, s"Dimensions mismatch when adding new sample." +
69+
s" Expecting $n but got ${sample.toBreeze.length}.")
70+
71+
sample.toBreeze.activeIterator.foreach {
72+
case (_, 0.0) => // Skip explicit zero elements.
73+
case (i, value) =>
74+
if (currMax(i) < value) {
75+
currMax(i) = value
76+
}
77+
if (currMin(i) > value) {
78+
currMin(i) = value
79+
}
80+
81+
val tmpPrevMean = currMean(i)
82+
currMean(i) = (currMean(i) * nnz(i) + value) / (nnz(i) + 1.0)
83+
currM2n(i) += (value - currMean(i)) * (value - tmpPrevMean)
84+
85+
nnz(i) += 1.0
86+
}
87+
88+
totalCnt += 1
89+
this
90+
}
91+
92+
/**
93+
* Merge another MultivariateOnlineSummarizer, and update the statistical summary.
94+
* (Note that it's in place merging; as a result, `this` object will be modified.)
95+
*
96+
* @param other The other MultivariateOnlineSummarizer to be merged.
97+
* @return This MultivariateOnlineSummarizer object.
98+
*/
99+
def merge(other: MultivariateOnlineSummarizer): this.type = {
100+
if (this.totalCnt != 0 && other.totalCnt != 0) {
101+
require(n == other.n, s"Dimensions mismatch when merging with another summarizer. " +
102+
s"Expecting $n but got ${other.n}.")
103+
totalCnt += other.totalCnt
104+
val deltaMean: BDV[Double] = currMean - other.currMean
105+
var i = 0
106+
while (i < n) {
107+
// merge mean together
108+
if (other.currMean(i) != 0.0) {
109+
currMean(i) = (currMean(i) * nnz(i) + other.currMean(i) * other.nnz(i)) /
110+
(nnz(i) + other.nnz(i))
111+
}
112+
// merge m2n together
113+
if (nnz(i) + other.nnz(i) != 0.0) {
114+
currM2n(i) += other.currM2n(i) + deltaMean(i) * deltaMean(i) * nnz(i) * other.nnz(i) /
115+
(nnz(i) + other.nnz(i))
116+
}
117+
if (currMax(i) < other.currMax(i)) {
118+
currMax(i) = other.currMax(i)
119+
}
120+
if (currMin(i) > other.currMin(i)) {
121+
currMin(i) = other.currMin(i)
122+
}
123+
i += 1
124+
}
125+
nnz += other.nnz
126+
} else if (totalCnt == 0 && other.totalCnt != 0) {
127+
this.n = other.n
128+
this.currMean = other.currMean.copy
129+
this.currM2n = other.currM2n.copy
130+
this.totalCnt = other.totalCnt
131+
this.nnz = other.nnz.copy
132+
this.currMax = other.currMax.copy
133+
this.currMin = other.currMin.copy
134+
}
135+
this
136+
}
137+
138+
override def mean: Vector = {
139+
require(totalCnt > 0, s"Nothing has been added to this summarizer.")
140+
141+
val realMean = BDV.zeros[Double](n)
142+
var i = 0
143+
while (i < n) {
144+
realMean(i) = currMean(i) * (nnz(i) / totalCnt)
145+
i += 1
146+
}
147+
Vectors.fromBreeze(realMean)
148+
}
149+
150+
override def variance: Vector = {
151+
require(totalCnt > 0, s"Nothing has been added to this summarizer.")
152+
153+
val realVariance = BDV.zeros[Double](n)
154+
155+
val denominator = totalCnt - 1.0
156+
157+
// Sample variance is computed, if the denominator is less than 0, the variance is just 0.
158+
if (denominator > 0.0) {
159+
val deltaMean = currMean
160+
var i = 0
161+
while (i < currM2n.size) {
162+
realVariance(i) =
163+
currM2n(i) + deltaMean(i) * deltaMean(i) * nnz(i) * (totalCnt - nnz(i)) / totalCnt
164+
realVariance(i) /= denominator
165+
i += 1
166+
}
167+
}
168+
169+
Vectors.fromBreeze(realVariance)
170+
}
171+
172+
override def count: Long = totalCnt
173+
174+
override def numNonzeros: Vector = {
175+
require(totalCnt > 0, s"Nothing has been added to this summarizer.")
176+
177+
Vectors.fromBreeze(nnz)
178+
}
179+
180+
override def max: Vector = {
181+
require(totalCnt > 0, s"Nothing has been added to this summarizer.")
182+
183+
var i = 0
184+
while (i < n) {
185+
if ((nnz(i) < totalCnt) && (currMax(i) < 0.0)) currMax(i) = 0.0
186+
i += 1
187+
}
188+
Vectors.fromBreeze(currMax)
189+
}
190+
191+
override def min: Vector = {
192+
require(totalCnt > 0, s"Nothing has been added to this summarizer.")
193+
194+
var i = 0
195+
while (i < n) {
196+
if ((nnz(i) < totalCnt) && (currMin(i) > 0.0)) currMin(i) = 0.0
197+
i += 1
198+
}
199+
Vectors.fromBreeze(currMin)
200+
}
201+
}

0 commit comments

Comments
 (0)