Github user yu-iskw commented on a diff in the pull request:
https://github.com/apache/spark/pull/6880#discussion_r43590450
--- Diff:
mllib/src/main/scala/org/apache/spark/mllib/clustering/DpMeans.scala ---
@@ -0,0 +1,279 @@
+/*
+ * 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.mllib.clustering
+
+import scala.collection.mutable.ArrayBuffer
+
+import org.apache.spark.Logging
+import org.apache.spark.annotation.Experimental
+import org.apache.spark.mllib.linalg.{Vector, Vectors}
+import org.apache.spark.mllib.linalg.BLAS.{axpy, scal}
+import org.apache.spark.mllib.util.MLUtils
+import org.apache.spark.rdd.RDD
+import org.apache.spark.storage.StorageLevel
+
+/**
+ * :: Experimental ::
+ *
+ * The Dirichlet process (DP) is a popular non-parametric Bayesian mixture
+ * model that allows for flexible clustering of data without having to
+ * determine the number of clusters in advance.
+ *
+ * Given a set of data points, this class performs cluster creation
process,
+ * based on DP-means algorithm, iterating until the maximum number of
iterations
+ * is reached or the convergence criteria is satisfied. With the current
+ * global set of centers, it locally creates a new cluster centered at `x`
+ * whenever it encounters an uncovered data point `x`. In a similar manner,
+ * a local cluster center is promoted to a global center whenever an
uncovered
+ * local cluster center is found. A data point is said to be "covered" by
+ * a cluster `c` if the distance from the point to the cluster center of
`c`
+ * is less than a given lambda value.
+ *
+ * The original paper is "Revisiting k-means: New Algorithms via Bayesian
Nonparametrics"
+ * by Brian Kulis, Michael I. Jordan. This implementation is based on
"MLbase: Distributed
+ * Machine Learning Made Easy" by Xinghao Pan, Evan R. Sparks, Andre
Wibisono
+ *
+ * @param lambdaValue distance value that controls cluster creation.
+ * @param convergenceTol The threshold value at which convergence is
considered to have occurred.
+ * @param maxIterations The maximum number of iterations to perform.
+ * // TODO
+ * @param maxClusterCount The maximum expected number of clusters.
+ */
+
+@Experimental
+class DpMeans private (
+ private var lambdaValue: Double,
+ private var convergenceTol: Double,
+ private var maxIterations: Int,
+ private var maxClusterCount: Int) extends Serializable with Logging {
+
+ /**
+ * Constructs a default instance with default parameters: {lambdaValue:
1,
+ * convergenceTol: 0.01, maxIterations: 20, maxClusterCount: 1000}.
+ */
+ def this() = this(1, 0.01, 20, 1000)
+
+ /** Sets the value for the lambda parameter, which controls the cluster
creation. Default: 1 */
+ def setLambdaValue(lambdaValue: Double): this.type = {
+ this.lambdaValue = lambdaValue
+ this
+ }
+
+ /** Returns the lambda value */
+ def getLambdaValue: Double = lambdaValue
+
+ /** Sets the threshold value at which convergence is considered to have
occurred. Default: 0.01 */
+ def setConvergenceTol(convergenceTol: Double): this.type = {
+ this.convergenceTol = convergenceTol
+ this
+ }
+
+ /** Returns the convergence threshold value . */
+ def getConvergenceTol: Double = convergenceTol
+
+ /** Sets the maximum number of iterations. Default: 20 */
+ def setMaxIterations(maxIterations: Int): this.type = {
+ this.maxIterations = maxIterations
+ this
+ }
+
+ /** Returns the maximum number of iterations. */
+ def getMaxIterations: Int = maxIterations
+
+ /** Sets the maximum number of clusters expected. Default: 1000 */
+ def setMaxClusterCount(maxClusterCount: Int): this.type = {
+ this.maxClusterCount = maxClusterCount
+ this
+ }
+
+ /** Returns the maximum number of clusters expected. */
+ def getMaxClusterCount: Int = maxClusterCount
+
+ /**
+ * Perform DP-means clustering
+ */
+ def run(data: RDD[Vector]): DpMeansModel = {
+ if (data.getStorageLevel == StorageLevel.NONE) {
+ logWarning("The input data is not directly cached, which may hurt
performance if its"
+ + " parent RDDs are also uncached.")
+ }
+
+ // Compute squared norms and cache them.
+ val norms = data.map(Vectors.norm(_, 2.0))
+ norms.persist()
+ val zippedData = data.zip(norms).map {
+ case (v, norm) => new VectorWithNorm(v, norm)
+ }
+
+ // Implementation of DP-means algorithm.
+ var iteration = 0
+ var previousK = 1
+ var covered = false
+ var converged = false
+ var localCenters = Array.empty[VectorWithNorm]
+ val globalCenters = ArrayBuffer.empty[VectorWithNorm]
+
+ // Execute clustering until the maximum number of iterations is reached
+ // or the cluster centers have converged.
+ while (iteration < maxIterations && !converged) {
+
+ // This loop simulates the cluster assignment step by iterating
through the data
+ // points until no new clusters are generated. A new cluster is
formed whenever
+ // the distance from a data point to all the existing cluster
centroids is greater
+ // than lambdaValue. The cluster centers returned by
generateNewCenters(), called
+ // localCenters, are sent to the master. The cluster creation
process is performed
+ // locally on these localCenters and the resulting cluster centers
are added to
+ // the existing set of global centers.
+ while (!covered) {
+ val findLocalCenters = zippedData
+ .mapPartitions(h =>
+ DpMeans.generateNewCenters(h, globalCenters, lambdaValue)
+ ).persist()
+ val localCentersCount = findLocalCenters.count()
+ logInfo(" Local centers count :: " + localCentersCount)
+ if (localCentersCount > getMaxClusterCount) {
+ logInfo(" Local centers count " + localCentersCount + " exceeds
user expectation")
+ sys.exit()
+ }
+ localCenters = findLocalCenters.collect()
+ if (localCenters.isEmpty) {
+ covered = true
+ }
+ // Promote a local cluster center to a global center
+ else {
+ val newGlobalCenters =
DpMeans.generateNewCenters(localCenters.iterator,
+ ArrayBuffer.empty[VectorWithNorm], lambdaValue)
+ globalCenters ++= newGlobalCenters
+ logInfo(" Number of global centers :: " + globalCenters.length)
+ }
+ }
+
+ // Find the sum and count of points belonging to each cluster
+ case class WeightedPoint(vector: Vector, count: Long)
+ val mergeClusters = (x: WeightedPoint, y: WeightedPoint) => {
+ axpy(1.0, y.vector, x.vector)
+ WeightedPoint(x.vector, x.count + y.count)
+ }
+
+ val dims = globalCenters(0).vector.size
+ val clusterStat = zippedData.mapPartitions { points =>
--- End diff --
I think we could also implement this with a single `aggregateByKey`. It
makes this more simple and more efficient.
---
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]
