zhipeng93 commented on code in PR #139: URL: https://github.com/apache/flink-ml/pull/139#discussion_r943058613
########## flink-ml-lib/src/main/java/org/apache/flink/ml/feature/kbinsdiscretizer/KBinsDiscretizer.java: ########## @@ -0,0 +1,340 @@ +/* + * 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.feature.kbinsdiscretizer; + +import org.apache.flink.api.common.functions.MapFunction; +import org.apache.flink.api.common.functions.MapPartitionFunction; +import org.apache.flink.ml.api.Estimator; +import org.apache.flink.ml.common.datastream.DataStreamUtils; +import org.apache.flink.ml.feature.minmaxscaler.MinMaxScaler.MinMaxReduceFunctionOperator; +import org.apache.flink.ml.linalg.DenseVector; +import org.apache.flink.ml.linalg.Vector; +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.streaming.api.datastream.DataStream; +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.slf4j.Logger; +import org.slf4j.LoggerFactory; + +import java.io.IOException; +import java.util.ArrayList; +import java.util.Arrays; +import java.util.HashMap; +import java.util.HashSet; +import java.util.List; +import java.util.Map; +import java.util.Set; + +/** + * An Estimator which implements discretization (also known as quantization or binning), which + * transforms continuous features into discrete ones. The output values are in [0, numBins). + * + * <p>KBinsDiscretizer implements three different binning strategies, and it can be set by {@link + * KBinsDiscretizerParams#STRATEGY}. If the strategy is set as {@link KBinsDiscretizerParams#KMEANS} + * or {@link KBinsDiscretizerParams#QUANTILE}, users should further set {@link + * KBinsDiscretizerParams#SUB_SAMPLES} for better performance. + * + * <p>There are several cornel cases for different inputs as listed below: + * + * <ul> + * <li>When the input values of one column are all the same, then they should be mapped to the + * same bin (i.e., the zero-th bin). Thus the corresponding bin edges are {Double.MIN_VALUE, + * Double.MAX_VALUE}. + * <li>When the number of distinct values of one column is less than the specified number of bins + * and the {@link KBinsDiscretizerParams#STRATEGY} is set as {@link + * KBinsDiscretizerParams#KMEANS}, we switch to {@link KBinsDiscretizerParams#UNIFORM}. + * <li>When the width of one output bin is zero, i.e., the left edge equals to the right edge of + * the bin, we remove it. + * </ul> + */ +public class KBinsDiscretizer + implements Estimator<KBinsDiscretizer, KBinsDiscretizerModel>, + KBinsDiscretizerParams<KBinsDiscretizer> { + private static final Logger LOG = LoggerFactory.getLogger(KBinsDiscretizer.class); + private final Map<Param<?>, Object> paramMap = new HashMap<>(); + + public KBinsDiscretizer() { + ParamUtils.initializeMapWithDefaultValues(paramMap, this); + } + + @Override + public KBinsDiscretizerModel fit(Table... inputs) { + Preconditions.checkArgument(inputs.length == 1); + StreamTableEnvironment tEnv = + (StreamTableEnvironment) ((TableImpl) inputs[0]).getTableEnvironment(); + + String inputCol = getInputCol(); + String strategy = getStrategy(); + int numBins = getNumBins(); + + DataStream<DenseVector> inputData = + tEnv.toDataStream(inputs[0]) + .map( + (MapFunction<Row, DenseVector>) + value -> ((Vector) value.getField(inputCol)).toDense()); + + DataStream<DenseVector> preprocessedData; + if (strategy.equals(UNIFORM)) { + preprocessedData = + inputData + .transform( + "reduceInEachPartition", + inputData.getType(), + new MinMaxReduceFunctionOperator()) + .transform( + "reduceInFinalPartition", + inputData.getType(), + new MinMaxReduceFunctionOperator()) + .setParallelism(1); + } else { + preprocessedData = DataStreamUtils.sample(inputData, getSubSamples(), 2022L); Review Comment: As I understand, `quantitle` and `kmeans` are similar in the following aspects: - They are both used to estimate/compute the bin edges. - For accurate implementation, they both need to access/store all of the input data. Now that sampling is used for `quantitle` to accelerate the process of computing bins, with a compromise on precision. Then it seems unclear why do we do different stuff to `kmeans`. If we do not support sampling for `kmeans`, it may only work on small datasets. Also, users can also set `numSamples` as `Integer.MAX_VALUE` to do `kmeans` with no sampling. -- 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: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
