From 8b6b46e4ff5f19fb7befecaaa0eda63bf29a0e2c Mon Sep 17 00:00:00 2001 From: "Joseph K. Bradley" Date: Thu, 7 May 2015 01:12:14 -0700 Subject: [SPARK-7421] [MLLIB] OnlineLDA cleanups Small changes, primarily to allow us more flexibility in the future: * Rename "tau_0" to "tau0" * Mark LDAOptimizer trait sealed and DeveloperApi. * Mark LDAOptimizer subclasses as final. * Mark setOptimizer (the one taking an LDAOptimizer) and getOptimizer as DeveloperApi since we may need to change them in the future CC: hhbyyh Author: Joseph K. Bradley Closes #5956 from jkbradley/onlinelda-cleanups and squashes the following commits: f4be508 [Joseph K. Bradley] added newline f4003e4 [Joseph K. Bradley] Changes: * Rename "tau_0" to "tau0" * Mark LDAOptimizer trait sealed and DeveloperApi. * Mark LDAOptimizer subclasses as final. * Mark setOptimizer (the one taking an LDAOptimizer) and getOptimizer as DeveloperApi since we may need to change them in the future --- .../org/apache/spark/mllib/clustering/LDA.scala | 15 +++++++-- .../spark/mllib/clustering/LDAOptimizer.scala | 37 ++++++++++------------ .../spark/mllib/clustering/JavaLDASuite.java | 2 +- .../apache/spark/mllib/clustering/LDASuite.scala | 8 ++--- 4 files changed, 34 insertions(+), 28 deletions(-) (limited to 'mllib') diff --git a/mllib/src/main/scala/org/apache/spark/mllib/clustering/LDA.scala b/mllib/src/main/scala/org/apache/spark/mllib/clustering/LDA.scala index c8daa2388e..a410547a72 100644 --- a/mllib/src/main/scala/org/apache/spark/mllib/clustering/LDA.scala +++ b/mllib/src/main/scala/org/apache/spark/mllib/clustering/LDA.scala @@ -18,8 +18,9 @@ package org.apache.spark.mllib.clustering import breeze.linalg.{DenseVector => BDV} + import org.apache.spark.Logging -import org.apache.spark.annotation.Experimental +import org.apache.spark.annotation.{DeveloperApi, Experimental} import org.apache.spark.api.java.JavaPairRDD import org.apache.spark.graphx._ import org.apache.spark.mllib.linalg.Vector @@ -197,12 +198,20 @@ class LDA private ( } - /** LDAOptimizer used to perform the actual calculation */ + /** + * :: DeveloperApi :: + * + * LDAOptimizer used to perform the actual calculation + */ + @DeveloperApi def getOptimizer: LDAOptimizer = ldaOptimizer /** + * :: DeveloperApi :: + * * LDAOptimizer used to perform the actual calculation (default = EMLDAOptimizer) */ + @DeveloperApi def setOptimizer(optimizer: LDAOptimizer): this.type = { this.ldaOptimizer = optimizer this @@ -210,7 +219,7 @@ class LDA private ( /** * Set the LDAOptimizer used to perform the actual calculation by algorithm name. - * Currently "em", "online" is supported. + * Currently "em", "online" are supported. */ def setOptimizer(optimizerName: String): this.type = { this.ldaOptimizer = diff --git a/mllib/src/main/scala/org/apache/spark/mllib/clustering/LDAOptimizer.scala b/mllib/src/main/scala/org/apache/spark/mllib/clustering/LDAOptimizer.scala index 093aa0f315..6fa2fe053c 100644 --- a/mllib/src/main/scala/org/apache/spark/mllib/clustering/LDAOptimizer.scala +++ b/mllib/src/main/scala/org/apache/spark/mllib/clustering/LDAOptimizer.scala @@ -23,7 +23,7 @@ import breeze.linalg.{DenseVector => BDV, DenseMatrix => BDM, sum, normalize, kr import breeze.numerics.{digamma, exp, abs} import breeze.stats.distributions.{Gamma, RandBasis} -import org.apache.spark.annotation.Experimental +import org.apache.spark.annotation.DeveloperApi import org.apache.spark.graphx._ import org.apache.spark.graphx.impl.GraphImpl import org.apache.spark.mllib.impl.PeriodicGraphCheckpointer @@ -31,13 +31,13 @@ import org.apache.spark.mllib.linalg.{Matrices, SparseVector, DenseVector, Vecto import org.apache.spark.rdd.RDD /** - * :: Experimental :: + * :: DeveloperApi :: * * An LDAOptimizer specifies which optimization/learning/inference algorithm to use, and it can * hold optimizer-specific parameters for users to set. */ -@Experimental -trait LDAOptimizer { +@DeveloperApi +sealed trait LDAOptimizer { /* DEVELOPERS NOTE: @@ -59,7 +59,7 @@ trait LDAOptimizer { } /** - * :: Experimental :: + * :: DeveloperApi :: * * Optimizer for EM algorithm which stores data + parameter graph, plus algorithm parameters. * @@ -75,8 +75,8 @@ trait LDAOptimizer { * "On Smoothing and Inference for Topic Models." UAI, 2009. * */ -@Experimental -class EMLDAOptimizer extends LDAOptimizer { +@DeveloperApi +final class EMLDAOptimizer extends LDAOptimizer { import LDA._ @@ -211,7 +211,7 @@ class EMLDAOptimizer extends LDAOptimizer { /** - * :: Experimental :: + * :: DeveloperApi :: * * An online optimizer for LDA. The Optimizer implements the Online variational Bayes LDA * algorithm, which processes a subset of the corpus on each iteration, and updates the term-topic @@ -220,8 +220,8 @@ class EMLDAOptimizer extends LDAOptimizer { * Original Online LDA paper: * Hoffman, Blei and Bach, "Online Learning for Latent Dirichlet Allocation." NIPS, 2010. */ -@Experimental -class OnlineLDAOptimizer extends LDAOptimizer { +@DeveloperApi +final class OnlineLDAOptimizer extends LDAOptimizer { // LDA common parameters private var k: Int = 0 @@ -243,8 +243,8 @@ class OnlineLDAOptimizer extends LDAOptimizer { private var randomGenerator: java.util.Random = null // Online LDA specific parameters - // Learning rate is: (tau_0 + t)^{-kappa} - private var tau_0: Double = 1024 + // Learning rate is: (tau0 + t)^{-kappa} + private var tau0: Double = 1024 private var kappa: Double = 0.51 private var miniBatchFraction: Double = 0.05 @@ -265,16 +265,16 @@ class OnlineLDAOptimizer extends LDAOptimizer { * A (positive) learning parameter that downweights early iterations. Larger values make early * iterations count less. */ - def getTau_0: Double = this.tau_0 + def getTau0: Double = this.tau0 /** * A (positive) learning parameter that downweights early iterations. Larger values make early * iterations count less. * Default: 1024, following the original Online LDA paper. */ - def setTau_0(tau_0: Double): this.type = { - require(tau_0 > 0, s"LDA tau_0 must be positive, but was set to $tau_0") - this.tau_0 = tau_0 + def setTau0(tau0: Double): this.type = { + require(tau0 > 0, s"LDA tau0 must be positive, but was set to $tau0") + this.tau0 = tau0 this } @@ -434,11 +434,8 @@ class OnlineLDAOptimizer extends LDAOptimizer { * Update lambda based on the batch submitted. batchSize can be different for each iteration. */ private[clustering] def update(stat: BDM[Double], iter: Int, batchSize: Int): Unit = { - val tau_0 = this.getTau_0 - val kappa = this.getKappa - // weight of the mini-batch. - val weight = math.pow(tau_0 + iter, -kappa) + val weight = math.pow(getTau0 + iter, -getKappa) // Update lambda based on documents. lambda = lambda * (1 - weight) + diff --git a/mllib/src/test/java/org/apache/spark/mllib/clustering/JavaLDASuite.java b/mllib/src/test/java/org/apache/spark/mllib/clustering/JavaLDASuite.java index f394d90396..96c2da1699 100644 --- a/mllib/src/test/java/org/apache/spark/mllib/clustering/JavaLDASuite.java +++ b/mllib/src/test/java/org/apache/spark/mllib/clustering/JavaLDASuite.java @@ -117,7 +117,7 @@ public class JavaLDASuite implements Serializable { // Train a model OnlineLDAOptimizer op = new OnlineLDAOptimizer() - .setTau_0(1024) + .setTau0(1024) .setKappa(0.51) .setGammaShape(1e40) .setMiniBatchFraction(0.5); diff --git a/mllib/src/test/scala/org/apache/spark/mllib/clustering/LDASuite.scala b/mllib/src/test/scala/org/apache/spark/mllib/clustering/LDASuite.scala index 2dcc881f5a..d5b7d96335 100644 --- a/mllib/src/test/scala/org/apache/spark/mllib/clustering/LDASuite.scala +++ b/mllib/src/test/scala/org/apache/spark/mllib/clustering/LDASuite.scala @@ -138,12 +138,12 @@ class LDASuite extends FunSuite with MLlibTestSparkContext { val lda = new LDA().setK(2) val corpus = sc.parallelize(tinyCorpus, 2) val op = new OnlineLDAOptimizer().initialize(corpus, lda) - op.setKappa(0.9876).setMiniBatchFraction(0.123).setTau_0(567) + op.setKappa(0.9876).setMiniBatchFraction(0.123).setTau0(567) assert(op.getAlpha == 0.5) // default 1.0 / k assert(op.getEta == 0.5) // default 1.0 / k assert(op.getKappa == 0.9876) assert(op.getMiniBatchFraction == 0.123) - assert(op.getTau_0 == 567) + assert(op.getTau0 == 567) } test("OnlineLDAOptimizer one iteration") { @@ -159,7 +159,7 @@ class LDASuite extends FunSuite with MLlibTestSparkContext { val corpus = sc.parallelize(docs, 2) // Set GammaShape large to avoid the stochastic impact. - val op = new OnlineLDAOptimizer().setTau_0(1024).setKappa(0.51).setGammaShape(1e40) + val op = new OnlineLDAOptimizer().setTau0(1024).setKappa(0.51).setGammaShape(1e40) .setMiniBatchFraction(1) val lda = new LDA().setK(k).setMaxIterations(1).setOptimizer(op).setSeed(12345) @@ -192,7 +192,7 @@ class LDASuite extends FunSuite with MLlibTestSparkContext { ).zipWithIndex.map { case (wordCounts, docId) => (docId.toLong, wordCounts) } val docs = sc.parallelize(toydata) - val op = new OnlineLDAOptimizer().setMiniBatchFraction(1).setTau_0(1024).setKappa(0.51) + val op = new OnlineLDAOptimizer().setMiniBatchFraction(1).setTau0(1024).setKappa(0.51) .setGammaShape(1e10) val lda = new LDA().setK(2) .setDocConcentration(0.01) -- cgit v1.2.3