lindong28 commented on a change in pull request #70: URL: https://github.com/apache/flink-ml/pull/70#discussion_r836010275
########## File path: flink-ml-lib/src/main/java/org/apache/flink/ml/clustering/kmeans/OnlineKMeans.java ########## @@ -0,0 +1,407 @@ +/* + * 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.flink.ml.clustering.kmeans; + +import org.apache.flink.api.common.functions.FlatMapFunction; +import org.apache.flink.api.common.functions.MapFunction; +import org.apache.flink.api.common.functions.ReduceFunction; +import org.apache.flink.api.common.state.ListState; +import org.apache.flink.api.common.state.ListStateDescriptor; +import org.apache.flink.api.common.typeinfo.TypeInformation; +import org.apache.flink.api.java.typeutils.ObjectArrayTypeInfo; +import org.apache.flink.iteration.DataStreamList; +import org.apache.flink.iteration.IterationBody; +import org.apache.flink.iteration.IterationBodyResult; +import org.apache.flink.iteration.Iterations; +import org.apache.flink.iteration.operator.OperatorStateUtils; +import org.apache.flink.ml.api.Estimator; +import org.apache.flink.ml.common.distance.DistanceMeasure; +import org.apache.flink.ml.linalg.BLAS; +import org.apache.flink.ml.linalg.DenseVector; +import org.apache.flink.ml.linalg.typeinfo.DenseVectorTypeInfo; +import org.apache.flink.ml.param.Param; +import org.apache.flink.ml.util.ParamUtils; +import org.apache.flink.ml.util.ReadWriteUtils; +import org.apache.flink.runtime.state.StateInitializationContext; +import org.apache.flink.streaming.api.datastream.DataStream; +import org.apache.flink.streaming.api.environment.StreamExecutionEnvironment; +import org.apache.flink.streaming.api.functions.windowing.AllWindowFunction; +import org.apache.flink.streaming.api.operators.AbstractStreamOperator; +import org.apache.flink.streaming.api.operators.TwoInputStreamOperator; +import org.apache.flink.streaming.api.windowing.windows.GlobalWindow; +import org.apache.flink.streaming.runtime.streamrecord.StreamRecord; +import org.apache.flink.table.api.Table; +import org.apache.flink.table.api.bridge.java.StreamTableEnvironment; +import org.apache.flink.table.api.internal.TableImpl; +import org.apache.flink.types.Row; +import org.apache.flink.util.Collector; +import org.apache.flink.util.Preconditions; + +import org.apache.commons.collections.IteratorUtils; + +import java.io.IOException; +import java.nio.file.Files; +import java.nio.file.Paths; +import java.util.Arrays; +import java.util.HashMap; +import java.util.List; +import java.util.Map; +import java.util.function.Supplier; + +/** + * OnlineKMeans extends the function of {@link KMeans}, supporting to train a K-Means model + * continuously according to an unbounded stream of train data. + * + * <p>OnlineKMeans makes updates with the "mini-batch" KMeans rule, generalized to incorporate + * forgetfulness (i.e. decay). After the centroids estimated on the current batch are acquired, + * OnlineKMeans computes the new centroids from the weighted average between the original and the + * estimated centroids. The weight of the estimated centroids is the number of points assigned to + * them. The weight of the original centroids is also the number of points, but additionally + * multiplying with the decay factor. + * + * <p>The decay factor scales the contribution of the clusters as estimated thus far. If the decay + * factor is 1, all batches are weighted equally. If the decay factor is 0, new centroids are + * determined entirely by recent data. Lower values correspond to more forgetting. + */ +public class OnlineKMeans + implements Estimator<OnlineKMeans, OnlineKMeansModel>, OnlineKMeansParams<OnlineKMeans> { + private final Map<Param<?>, Object> paramMap = new HashMap<>(); + private Table initModelDataTable; + + public OnlineKMeans() { + ParamUtils.initializeMapWithDefaultValues(paramMap, this); + } + + @Override + public OnlineKMeansModel fit(Table... inputs) { + Preconditions.checkArgument(inputs.length == 1); + + StreamTableEnvironment tEnv = + (StreamTableEnvironment) ((TableImpl) inputs[0]).getTableEnvironment(); + + DataStream<DenseVector> points = + tEnv.toDataStream(inputs[0]).map(new FeaturesExtractor(getFeaturesCol())); + + DataStream<KMeansModelData> initModelData = + KMeansModelData.getModelDataStream(initModelDataTable); + initModelData.getTransformation().setParallelism(1); + + IterationBody body = + new OnlineKMeansIterationBody( + DistanceMeasure.getInstance(getDistanceMeasure()), + getK(), + getDecayFactor(), + getGlobalBatchSize()); + + DataStream<KMeansModelData> onlineModelData = + Iterations.iterateUnboundedStreams( + DataStreamList.of(initModelData), DataStreamList.of(points), body) + .get(0); + + Table onlineModelDataTable = tEnv.fromDataStream(onlineModelData); + OnlineKMeansModel model = new OnlineKMeansModel().setModelData(onlineModelDataTable); + ReadWriteUtils.updateExistingParams(model, paramMap); + return model; + } + + /** Saves the metadata AND bounded model data table (if exists) to the given path. */ + @Override + public void save(String path) throws IOException { + if (initModelDataTable != null) { + ReadWriteUtils.saveModelData( + KMeansModelData.getModelDataStream(initModelDataTable), + path, + new KMeansModelData.ModelDataEncoder()); + } + + ReadWriteUtils.saveMetadata(this, path); + } + + public static OnlineKMeans load(StreamExecutionEnvironment env, String path) + throws IOException { + OnlineKMeans onlineKMeans = ReadWriteUtils.loadStageParam(path); + + String initModelDataPath = ReadWriteUtils.getDataPath(path); + if (Files.exists(Paths.get(initModelDataPath))) { + StreamTableEnvironment tEnv = StreamTableEnvironment.create(env); + + DataStream<KMeansModelData> initModelDataStream = + ReadWriteUtils.loadModelData(env, path, new KMeansModelData.ModelDataDecoder()); + + onlineKMeans.initModelDataTable = tEnv.fromDataStream(initModelDataStream); + } + + return onlineKMeans; + } + + @Override + public Map<Param<?>, Object> getParamMap() { + return paramMap; + } + + private static class OnlineKMeansIterationBody implements IterationBody { + private final DistanceMeasure distanceMeasure; + private final int k; + private final double decayFactor; + private final int batchSize; + + public OnlineKMeansIterationBody( + DistanceMeasure distanceMeasure, int k, double decayFactor, int batchSize) { + this.distanceMeasure = distanceMeasure; + this.k = k; + this.decayFactor = decayFactor; + this.batchSize = batchSize; + } + + @Override + public IterationBodyResult process( + DataStreamList variableStreams, DataStreamList dataStreams) { + DataStream<KMeansModelData> modelData = variableStreams.get(0); + DataStream<DenseVector> points = dataStreams.get(0); + + int parallelism = points.getParallelism(); Review comment: Yes, I think the reason is related to performance. If `parallelism > batchSize`, it effectively means some slot (with its CPU resource) it definitely wasted. Is there any reason user would want to do this? If not, it means user must have chosen this setup by mistake. Would it be more user friendly to alert user of this issue by throwing an exception? -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: issues-unsubscr...@flink.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org
