[SPARK-30659][ML][PYSPARK] LogisticRegression blockify input vectors

[zhengruifeng]

What changes were proposed in this pull request?

1, reorg the fit method in LR to several blocks (createModel, createBounds, createOptimizer, createInitCoefWithInterceptMatrix);
2, add new param blockSize;
3, if blockSize==1, keep original behavior, code path trainOnRows;
4, if blockSize>1, standardize and stack input vectors to blocks (like ALS/MLP), code path trainOnBlocks

Why are the changes needed?

On dense dataset epsilon_normalized.t:
1, reduce RAM to persist traing dataset; (save about 40% RAM)
2, use Level-2 BLAS routines; (4x ~ 5x faster)

Does this PR introduce any user-facing change?

Yes, a new param is added

How was this patch tested?

existing and added testsuites

[zhengruifeng]

performace test on epsilon_normalized.t

code:

importorg.apache.spark.ml.classification._importorg.apache.spark.storage.StorageLevelvaldf= spark.read.option("numFeatures", "2000").format("libsvm").load("/data1/Datasets/epsilon/epsilon_normalized.t").withColumn("label", (col("label")+1)/2)
df.persist(StorageLevel.MEMORY_AND_DISK)
df.count
vallr=newLogisticRegression().setBlockSize(1).setMaxIter(10)
lr.fit(df)
valresults=Seq(1, 4, 16, 64, 256, 1024, 4096).map { size =>valstart=System.currentTimeMillis; valmodel= lr.setBlockSize(size).fit(df); valend=System.currentTimeMillis; (size, model.coefficients, end - start) }

results:

scala> results.map(_._3)
res3: Seq[Long] = List(31076, 6771, 6732, 7590, 7186, 7094, 7276)
scala> results.map(_._2).foreach(coef => println(coef.toString.take(100)))
[2.1557250220880024,-0.22767392418436572,4.569220246330072,0.04739667339597046,0.14605181933865558,-
[2.1557250220880064,-0.22767392418436597,4.56922024633007,0.04739667339596951,0.14605181933865646,-0
[2.1557250220880064,-0.2276739241843657,4.569220246330077,0.04739667339597028,0.14605181933865646,-0
[2.155725022088007,-0.2276739241843664,4.569220246330073,0.047396673395969764,0.14605181933865688,-0
[2.1557250220880038,-0.22767392418436605,4.569220246330073,0.04739667339597022,0.14605181933865458,-
[2.1557250220880033,-0.22767392418436683,4.569220246330072,0.047396673395970035,0.14605181933865727,
[2.1557250220880033,-0.22767392418436613,4.56922024633007,0.0473966733959703,0.1460518193386559,-0.0

blockSize==1

blockSize==256

[zhengruifeng]

performace test on sparse dataset: the first 10,000 instances of webspam_wc_normalized_trigram

code:

valdf= spark.read.option("numFeatures", "8289919").format("libsvm").load("/data1/Datasets/webspam/webspam_wc_normalized_trigram.svm.10k").withColumn("label", (col("label")+1)/2)
df.persist(StorageLevel.MEMORY_AND_DISK)
df.count
vallr=newLogisticRegression().setBlockSize(1).setMaxIter(10)
lr.fit(df)
valresults=Seq(1, 4, 16, 64, 256, 1024, 4096).map { size =>valstart=System.currentTimeMillis; valmodel= lr.setBlockSize(size).fit(df); valend=System.currentTimeMillis; (size, model.coefficients, end - start) }

results:

scala> results.map(_._3)
res17: Seq[Long] = List(33948, 425923, 129811, 56288, 47587, 42816, 39809)
scala> results.map(_._2).foreach(coef => println(coef.toString.take(100)))
(8289919,[549219,551719,592137,592138,592141,592154,592160,592162,592163,592164,592166,592167,592168
(8289919,[549219,551719,592137,592138,592141,592154,592160,592162,592163,592164,592166,592167,592168
(8289919,[549219,551719,592137,592138,592141,592154,592160,592162,592163,592164,592166,592167,592168
(8289919,[549219,551719,592137,592138,592141,592154,592160,592162,592163,592164,592166,592167,592168
(8289919,[549219,551719,592137,592138,592141,592154,592160,592162,592163,592164,592166,592167,592168
(8289919,[549219,551719,592137,592138,592141,592154,592160,592162,592163,592164,592166,592167,592168
(8289919,[549219,551719,592137,592138,592141,592154,592160,592162,592163,592164,592166,592167,592168
scala> results.map(_._2).foreach(coef => println(coef.toString.takeRight(100)))
87,-1188.1053920127556,335.5565308836645,-135.79302172669907,849.0515530033497,-27.040836637047736])
91,-1188.105392012755,335.55653088366444,-135.79302172669907,849.0515530033497,-27.040836637047736])
9,-1188.1053920127551,335.55653088366444,-135.79302172669904,849.0515530033495,-27.040836637047725])
94,-1188.1053920127556,335.55653088366444,-135.79302172669904,849.0515530033495,-27.04083663704773])
1,-1188.1053920127551,335.55653088366444,-135.79302172669904,849.0515530033493,-27.040836637047722])
5,-1188.1053920127556,335.55653088366444,-135.79302172669904,849.0515530033495,-27.040836637047736])
29,-1188.105392012756,335.55653088366444,-135.79302172669904,849.0515530033495,-27.040836637047736])

blockSize==1

blockSize=16

test with Master:

import org.apache.spark.ml.classification._
import org.apache.spark.storage.StorageLevel
val df = spark.read.option("numFeatures", "8289919").format("libsvm").load("/data1/Datasets/webspam/webspam_wc_normalized_trigram.svm.10k").withColumn("label", (col("label")+1)/2)
df.persist(StorageLevel.MEMORY_AND_DISK)
df.count
val lr = new LogisticRegression().setMaxIter(10)
lr.fit(df)
val start = System.currentTimeMillis; val model = lr.setMaxIter(10).fit(df); val end = System.currentTimeMillis; end - start
scala> val start = System.currentTimeMillis; val model = lr.setMaxIter(10).fit(df); val end = System.currentTimeMillis; end - start
start: Long = 1588735447883 model: org.apache.spark.ml.classification.LogisticRegressionModel = LogisticRegressionModel: uid=logreg_99d29a0ecc13, numClasses=2, numFeatures=8289919
end: Long = 1588735483170
res3: Long = 35287

In this PR, when blockSize==1, the duration is 33948, so there will be no performance regression on sparse datasets.

[SparkQA]

Test build #122334 has finished for PR 28458 at commit 563cee9.

• This patch fails to generate documentation.
• This patch merges cleanly.
• This patch adds the following public classes (experimental):
• instr.logWarning(s\"All labels belong to a single class and fitIntercept=false. It's a \" +

[SparkQA]

Test build #122338 has finished for PR 28458 at commit 0577bb2.

• This patch fails PySpark unit tests.
• This patch merges cleanly.
• This patch adds no public classes.

[zhengruifeng]

retest this please

[SparkQA]

Test build #122345 has finished for PR 28458 at commit 0577bb2.

• This patch fails due to an unknown error code, -9.
• This patch merges cleanly.
• This patch adds no public classes.

[zhengruifeng]

retest this please

[SparkQA]

Test build #122349 has finished for PR 28458 at commit 0577bb2.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

[SparkQA]

Test build #122356 has finished for PR 28458 at commit 1aca7c5.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

[SparkQA]

Test build #122357 has finished for PR 28458 at commit 8f6582a.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

[zhengruifeng]

This PR is a update of #27374, it can avoid performance regression on sparse datasets by default (with blockSize=1).
On dense datasets like epsilon, with a reasonable blockSize (100~1000), it is much faster than existing impl.

[zhengruifeng]

Merged to master

[HyukjinKwon]

@zhengruifeng, this was reviewed by nobody while the codes are near 1k lines.