Github user jkbradley commented on a diff in the pull request:
https://github.com/apache/spark/pull/4419#discussion_r29296386
--- Diff:
mllib/src/main/scala/org/apache/spark/mllib/clustering/LDAOptimizer.scala ---
@@ -208,3 +225,224 @@ class EMLDAOptimizer extends LDAOptimizer{
new DistributedLDAModel(this, iterationTimes)
}
}
+
+
+/**
+ * :: Experimental ::
+ *
+ * An online optimizer for LDA. The Optimizer implements the Online LDA
algorithm, which
+ * processes a subset of the corpus by each call to next, and update the
term-topic
+ * distribution adaptively.
+ *
+ * References:
+ * Hoffman, Blei and Bach, "Online Learning for Latent Dirichlet
Allocation." NIPS, 2010.
+ */
+@Experimental
+class OnlineLDAOptimizer extends LDAOptimizer {
+
+ // LDA common parameters
+ private var k: Int = 0
+ private var D: Int = 0
+ private var vocabSize: Int = 0
+ private var alpha: Double = 0
+ private var eta: Double = 0
+ private var randomSeed: Long = 0
+
+ // Online LDA specific parameters
+ private var tau_0: Double = -1
+ private var kappa: Double = -1
+ private var batchSize: Int = -1
+
+ // internal data structure
+ private var docs: RDD[(Long, Vector)] = null
+ private var lambda: BDM[Double] = null
+ private var Elogbeta: BDM[Double]= null
+ private var expElogbeta: BDM[Double] = null
+
+ // count of invocation to next, used to help deciding the weight for
each iteration
+ private var iteration = 0
+
+ /**
+ * A (positive) learning parameter that downweights early iterations
+ */
+ def getTau_0: Double = {
+ if (this.tau_0 == -1) {
+ 1024
+ } else {
+ this.tau_0
+ }
+ }
+
+ /**
+ * A (positive) learning parameter that downweights early iterations
+ * Automatic setting of parameter:
+ * - default = 1024, which follows the recommendation from OnlineLDA
paper.
+ */
+ def setTau_0(tau_0: Double): this.type = {
+ require(tau_0 > 0 || tau_0 == -1.0, s"LDA tau_0 must be positive, but
was set to $tau_0")
+ this.tau_0 = tau_0
+ this
+ }
+
+ /**
+ * Learning rate: exponential decay rate
+ */
+ def getKappa: Double = {
+ if (this.kappa == -1) {
+ 0.5
+ } else {
+ this.kappa
+ }
+ }
+
+ /**
+ * Learning rate: exponential decay rate---should be between
+ * (0.5, 1.0] to guarantee asymptotic convergence.
+ * - default = 0.5, which follows the recommendation from OnlineLDA
paper.
+ */
+ def setKappa(kappa: Double): this.type = {
+ require(kappa >= 0 || kappa == -1.0,
+ s"Online LDA kappa must be nonnegative (or -1 for auto), but was set
to $kappa")
+ this.kappa = kappa
+ this
+ }
+
+ /**
+ * Mini-batch size, which controls how many documents are used in each
iteration
+ */
+ def getBatchSize: Int = {
--- End diff --
Let's use "miniBatchFraction" to match GradientDescent. Also, that way,
the parameter default value can be problem-independent.
---
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 [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
