You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@flink.apache.org by GitBox <gi...@apache.org> on 2022/03/17 10:07:10 UTC
[GitHub] [flink-ml] yunfengzhou-hub commented on a change in pull request #54: [FLINK-25552] Add Estimator and Transformer for MinMaxScaler in FlinkML
yunfengzhou-hub commented on a change in pull request #54:
URL: https://github.com/apache/flink-ml/pull/54#discussion_r828943448
##########
File path: flink-ml-lib/src/main/java/org/apache/flink/ml/feature/minmaxscaler/MinMaxScaler.java
##########
@@ -0,0 +1,205 @@
+/*
+ * 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.minmaxscaler;
+
+import org.apache.flink.api.common.functions.MapFunction;
+import org.apache.flink.api.common.functions.RichMapPartitionFunction;
+import org.apache.flink.api.common.state.ListState;
+import org.apache.flink.api.common.state.ListStateDescriptor;
+import org.apache.flink.ml.api.Estimator;
+import org.apache.flink.ml.common.datastream.DataStreamUtils;
+import org.apache.flink.ml.linalg.DenseVector;
+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.runtime.state.StateSnapshotContext;
+import org.apache.flink.streaming.api.datastream.DataStream;
+import org.apache.flink.streaming.api.environment.StreamExecutionEnvironment;
+import org.apache.flink.streaming.api.operators.AbstractStreamOperator;
+import org.apache.flink.streaming.api.operators.BoundedOneInput;
+import org.apache.flink.streaming.api.operators.OneInputStreamOperator;
+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 java.io.IOException;
+import java.util.HashMap;
+import java.util.Iterator;
+import java.util.Map;
+
+/**
+ * An Estimator which implements the MinMaxScaler algorithm.
+ *
+ * <p>See https://en.wikipedia.org/wiki/Feature_scaling#Rescaling_(min-max_normalization).
+ */
+public class MinMaxScaler
+ implements Estimator<MinMaxScaler, MinMaxScalerModel>, MinMaxScalerParams<MinMaxScaler> {
+ private final Map<Param<?>, Object> paramMap = new HashMap<>();
+
+ public MinMaxScaler() {
+ ParamUtils.initializeMapWithDefaultValues(paramMap, this);
+ }
+
+ @Override
+ public MinMaxScalerModel fit(Table... inputs) {
+ Preconditions.checkArgument(inputs.length == 1);
+ final String featureCol = getFeaturesCol();
+ StreamTableEnvironment tEnv =
+ (StreamTableEnvironment) ((TableImpl) inputs[0]).getTableEnvironment();
+ DataStream<DenseVector> features =
+ tEnv.toDataStream(inputs[0])
+ .map(
+ (MapFunction<Row, DenseVector>)
+ value -> (DenseVector) value.getField(featureCol));
+ DataStream<DenseVector> minMaxValues =
+ features.transform(
+ "reduceInEachPartition",
+ features.getType(),
+ new MinMaxReduceFunctionOperator())
+ .transform(
+ "reduceInFinalPartition",
+ features.getType(),
+ new MinMaxReduceFunctionOperator())
+ .setParallelism(1);
+ DataStream<MinMaxScalerModelData> modelData =
+ DataStreamUtils.mapPartition(
+ minMaxValues,
+ new RichMapPartitionFunction<DenseVector, MinMaxScalerModelData>() {
+ @Override
+ public void mapPartition(
+ Iterable<DenseVector> values,
+ Collector<MinMaxScalerModelData> out) {
+ Iterator<DenseVector> iter = values.iterator();
+ DenseVector minVector = iter.next();
+ DenseVector maxVector = iter.next();
+ out.collect(new MinMaxScalerModelData(minVector, maxVector));
+ }
+ });
+
+ MinMaxScalerModel model =
+ new MinMaxScalerModel().setModelData(tEnv.fromDataStream(modelData));
+ ReadWriteUtils.updateExistingParams(model, getParamMap());
+ return model;
+ }
+
+ /**
+ * A stream operator to compute the min and max values in each partition of the input bounded
+ * data stream.
+ */
+ private static class MinMaxReduceFunctionOperator extends AbstractStreamOperator<DenseVector>
+ implements OneInputStreamOperator<DenseVector, DenseVector>, BoundedOneInput {
+ private ListState<DenseVector> minState;
+ private ListState<DenseVector> maxState;
+
+ private DenseVector minVector;
+ private DenseVector maxVector;
+
+ @Override
+ public void endInput() {
+ if (minVector != null) {
+ output.collect(new StreamRecord<>(minVector));
+ }
+ if (maxVector != null) {
+ output.collect(new StreamRecord<>(maxVector));
+ }
+ }
+
+ @Override
+ public void processElement(StreamRecord<DenseVector> streamRecord) {
+ DenseVector currentValue = streamRecord.getValue();
+ if (minVector == null) {
+ int vecSize = currentValue.size();
+ minVector = new DenseVector(vecSize);
+ maxVector = new DenseVector(vecSize);
+ System.arraycopy(currentValue.values, 0, minVector.values, 0, vecSize);
+ System.arraycopy(currentValue.values, 0, maxVector.values, 0, vecSize);
+
+ } else {
+ for (int i = 0; i < currentValue.size(); ++i) {
+ minVector.values[i] = Math.min(minVector.values[i], currentValue.values[i]);
Review comment:
Could you please help check whether there are methods in `BLAS` that could help to find the max/min value between two double arrays, instead of using a for loop? I think using for loop on vectors could be an expensive operation.
If there is no direct method for this, maybe we can achieve this function by composing other BLAS operations. For example, the following methods would also help, if exists.
- compares two double arrays and return an array containing 1/-1 values, showing which one has bigger value.
- computes the absolute value of values in a double array.
##########
File path: flink-ml-lib/src/main/java/org/apache/flink/ml/feature/minmaxscaler/MinMaxScaler.java
##########
@@ -0,0 +1,205 @@
+/*
+ * 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.minmaxscaler;
+
+import org.apache.flink.api.common.functions.MapFunction;
+import org.apache.flink.api.common.functions.RichMapPartitionFunction;
+import org.apache.flink.api.common.state.ListState;
+import org.apache.flink.api.common.state.ListStateDescriptor;
+import org.apache.flink.ml.api.Estimator;
+import org.apache.flink.ml.common.datastream.DataStreamUtils;
+import org.apache.flink.ml.linalg.DenseVector;
+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.runtime.state.StateSnapshotContext;
+import org.apache.flink.streaming.api.datastream.DataStream;
+import org.apache.flink.streaming.api.environment.StreamExecutionEnvironment;
+import org.apache.flink.streaming.api.operators.AbstractStreamOperator;
+import org.apache.flink.streaming.api.operators.BoundedOneInput;
+import org.apache.flink.streaming.api.operators.OneInputStreamOperator;
+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 java.io.IOException;
+import java.util.HashMap;
+import java.util.Iterator;
+import java.util.Map;
+
+/**
+ * An Estimator which implements the MinMaxScaler algorithm.
+ *
+ * <p>See https://en.wikipedia.org/wiki/Feature_scaling#Rescaling_(min-max_normalization).
+ */
+public class MinMaxScaler
+ implements Estimator<MinMaxScaler, MinMaxScalerModel>, MinMaxScalerParams<MinMaxScaler> {
+ private final Map<Param<?>, Object> paramMap = new HashMap<>();
+
+ public MinMaxScaler() {
+ ParamUtils.initializeMapWithDefaultValues(paramMap, this);
+ }
+
+ @Override
+ public MinMaxScalerModel fit(Table... inputs) {
+ Preconditions.checkArgument(inputs.length == 1);
+ final String featureCol = getFeaturesCol();
+ StreamTableEnvironment tEnv =
+ (StreamTableEnvironment) ((TableImpl) inputs[0]).getTableEnvironment();
+ DataStream<DenseVector> features =
+ tEnv.toDataStream(inputs[0])
+ .map(
+ (MapFunction<Row, DenseVector>)
+ value -> (DenseVector) value.getField(featureCol));
+ DataStream<DenseVector> minMaxValues =
+ features.transform(
+ "reduceInEachPartition",
+ features.getType(),
+ new MinMaxReduceFunctionOperator())
+ .transform(
+ "reduceInFinalPartition",
+ features.getType(),
+ new MinMaxReduceFunctionOperator())
+ .setParallelism(1);
+ DataStream<MinMaxScalerModelData> modelData =
+ DataStreamUtils.mapPartition(
+ minMaxValues,
+ new RichMapPartitionFunction<DenseVector, MinMaxScalerModelData>() {
+ @Override
+ public void mapPartition(
+ Iterable<DenseVector> values,
+ Collector<MinMaxScalerModelData> out) {
+ Iterator<DenseVector> iter = values.iterator();
+ DenseVector minVector = iter.next();
+ DenseVector maxVector = iter.next();
+ out.collect(new MinMaxScalerModelData(minVector, maxVector));
+ }
+ });
+
+ MinMaxScalerModel model =
+ new MinMaxScalerModel().setModelData(tEnv.fromDataStream(modelData));
+ ReadWriteUtils.updateExistingParams(model, getParamMap());
+ return model;
+ }
+
+ /**
+ * A stream operator to compute the min and max values in each partition of the input bounded
+ * data stream.
+ */
+ private static class MinMaxReduceFunctionOperator extends AbstractStreamOperator<DenseVector>
+ implements OneInputStreamOperator<DenseVector, DenseVector>, BoundedOneInput {
+ private ListState<DenseVector> minState;
+ private ListState<DenseVector> maxState;
+
+ private DenseVector minVector;
+ private DenseVector maxVector;
+
+ @Override
+ public void endInput() {
+ if (minVector != null) {
+ output.collect(new StreamRecord<>(minVector));
+ }
+ if (maxVector != null) {
+ output.collect(new StreamRecord<>(maxVector));
+ }
+ }
+
+ @Override
+ public void processElement(StreamRecord<DenseVector> streamRecord) {
+ DenseVector currentValue = streamRecord.getValue();
+ if (minVector == null) {
+ int vecSize = currentValue.size();
+ minVector = new DenseVector(vecSize);
+ maxVector = new DenseVector(vecSize);
+ System.arraycopy(currentValue.values, 0, minVector.values, 0, vecSize);
+ System.arraycopy(currentValue.values, 0, maxVector.values, 0, vecSize);
+
+ } else {
+ for (int i = 0; i < currentValue.size(); ++i) {
+ minVector.values[i] = Math.min(minVector.values[i], currentValue.values[i]);
+ maxVector.values[i] = Math.max(maxVector.values[i], currentValue.values[i]);
+ }
+ }
+ }
+
+ @Override
+ @SuppressWarnings("unchecked")
+ public void initializeState(StateInitializationContext context) throws Exception {
+ super.initializeState(context);
+ minState =
+ context.getOperatorStateStore()
+ .getListState(
+ new ListStateDescriptor<>(
+ "minState",
+ getOperatorConfig()
+ .getTypeSerializerIn(
+ 0, getClass().getClassLoader())));
+ maxState =
+ context.getOperatorStateStore()
+ .getListState(
+ new ListStateDescriptor<>(
+ "maxState",
+ getOperatorConfig()
+ .getTypeSerializerIn(
+ 0, getClass().getClassLoader())));
+ Iterator<DenseVector> minIterator = minState.get().iterator();
+ Iterator<DenseVector> maxIterator = maxState.get().iterator();
+ if (minIterator.hasNext()) {
Review comment:
Have we checked the states or used `getUniqueElement` here? I saw zhipeng93's comment has been resolved here.
##########
File path: flink-ml-lib/src/main/java/org/apache/flink/ml/feature/minmaxscaler/MinMaxScalerModel.java
##########
@@ -0,0 +1,181 @@
+/*
+ * 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.minmaxscaler;
+
+import org.apache.flink.api.common.functions.RichMapFunction;
+import org.apache.flink.api.java.typeutils.RowTypeInfo;
+import org.apache.flink.ml.api.Model;
+import org.apache.flink.ml.common.broadcast.BroadcastUtils;
+import org.apache.flink.ml.common.datastream.TableUtils;
+import org.apache.flink.ml.linalg.DenseVector;
+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.streaming.api.environment.StreamExecutionEnvironment;
+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.table.runtime.typeutils.ExternalTypeInfo;
+import org.apache.flink.types.Row;
+import org.apache.flink.util.Preconditions;
+
+import org.apache.commons.lang3.ArrayUtils;
+
+import java.io.IOException;
+import java.util.Collections;
+import java.util.HashMap;
+import java.util.Map;
+
+/**
+ * A Model which do a minMax scaler operation using the model data computed by {@link MinMaxScaler}.
+ */
+public class MinMaxScalerModel
+ implements Model<MinMaxScalerModel>, MinMaxScalerParams<MinMaxScalerModel> {
+ private final Map<Param<?>, Object> paramMap = new HashMap<>();
+ private Table modelDataTable;
+
+ public MinMaxScalerModel() {
+ ParamUtils.initializeMapWithDefaultValues(paramMap, this);
+ }
+
+ @Override
+ public MinMaxScalerModel setModelData(Table... inputs) {
+ modelDataTable = inputs[0];
+ return this;
+ }
+
+ @Override
+ public Table[] getModelData() {
+ return new Table[] {modelDataTable};
+ }
+
+ @Override
+ @SuppressWarnings("unchecked")
+ public Table[] transform(Table... inputs) {
+ Preconditions.checkArgument(inputs.length == 1);
+ StreamTableEnvironment tEnv =
+ (StreamTableEnvironment) ((TableImpl) inputs[0]).getTableEnvironment();
+ DataStream<Row> data = tEnv.toDataStream(inputs[0]);
+ DataStream<MinMaxScalerModelData> minMaxScalerModel =
+ MinMaxScalerModelData.getModelDataStream(modelDataTable);
+ final String broadcastModelKey = "broadcastModelKey";
+ RowTypeInfo inputTypeInfo = TableUtils.getRowTypeInfo(inputs[0].getResolvedSchema());
+ RowTypeInfo outputTypeInfo =
+ new RowTypeInfo(
+ ArrayUtils.addAll(
+ inputTypeInfo.getFieldTypes(),
+ ExternalTypeInfo.of(DenseVector.class)),
+ ArrayUtils.addAll(inputTypeInfo.getFieldNames(), getPredictionCol()));
+ DataStream<Row> output =
+ BroadcastUtils.withBroadcastStream(
+ Collections.singletonList(data),
+ Collections.singletonMap(broadcastModelKey, minMaxScalerModel),
+ inputList -> {
+ DataStream input = inputList.get(0);
+ return input.map(
+ new PredictOutputFunction(
+ broadcastModelKey,
+ getMax(),
+ getMin(),
+ getFeaturesCol()),
+ outputTypeInfo);
+ });
+ return new Table[] {tEnv.fromDataStream(output)};
+ }
+
+ @Override
+ public Map<Param<?>, Object> getParamMap() {
+ return paramMap;
+ }
+
+ @Override
+ public void save(String path) throws IOException {
+ ReadWriteUtils.saveMetadata(this, path);
+ ReadWriteUtils.saveModelData(
+ MinMaxScalerModelData.getModelDataStream(modelDataTable),
+ path,
+ new MinMaxScalerModelData.ModelDataEncoder());
+ }
+
+ /**
+ * Loads model data from path.
+ *
+ * @param env Stream execution environment.
+ * @param path Model path.
+ * @return MinMaxScalerModel model.
+ */
+ public static MinMaxScalerModel load(StreamExecutionEnvironment env, String path)
+ throws IOException {
+ StreamTableEnvironment tEnv = StreamTableEnvironment.create(env);
+ MinMaxScalerModel model = ReadWriteUtils.loadStageParam(path);
+ DataStream<MinMaxScalerModelData> modelData =
+ ReadWriteUtils.loadModelData(
+ env, path, new MinMaxScalerModelData.ModelDataDecoder());
+ return model.setModelData(tEnv.fromDataStream(modelData));
+ }
+
+ /** This operator loads model data and predicts result. */
+ private static class PredictOutputFunction extends RichMapFunction<Row, Row> {
+ private final String featureCol;
+ private MinMaxScalerModelData minMaxScalerModelData;
+ private final double upperBound;
+ private final double lowerBound;
+ private final String broadcastKey;
+ private DenseVector maxVector;
+ private DenseVector minVector;
+
+ public PredictOutputFunction(
+ String broadcastKey, double upperBound, double lowerBound, String featureCol) {
+ this.upperBound = upperBound;
+ this.lowerBound = lowerBound;
+ this.broadcastKey = broadcastKey;
+ this.featureCol = featureCol;
+ }
+
+ @Override
+ public Row map(Row row) {
+ if (minMaxScalerModelData == null) {
+ minMaxScalerModelData =
+ (MinMaxScalerModelData)
+ getRuntimeContext().getBroadcastVariable(broadcastKey).get(0);
+ maxVector = minMaxScalerModelData.maxVector;
+ minVector = minMaxScalerModelData.minVector;
+ }
+ DenseVector feature = (DenseVector) row.getField(featureCol);
+ DenseVector outputVector = new DenseVector(maxVector.size());
+ if (feature != null) {
+ for (int i = 0; i < maxVector.size(); ++i) {
Review comment:
This for loop might also be replaced by BLAS operations.
##########
File path: flink-ml-lib/src/main/java/org/apache/flink/ml/feature/minmaxscaler/MinMaxScalerModel.java
##########
@@ -0,0 +1,181 @@
+/*
+ * 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.minmaxscaler;
+
+import org.apache.flink.api.common.functions.RichMapFunction;
+import org.apache.flink.api.java.typeutils.RowTypeInfo;
+import org.apache.flink.ml.api.Model;
+import org.apache.flink.ml.common.broadcast.BroadcastUtils;
+import org.apache.flink.ml.common.datastream.TableUtils;
+import org.apache.flink.ml.linalg.DenseVector;
+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.streaming.api.environment.StreamExecutionEnvironment;
+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.table.runtime.typeutils.ExternalTypeInfo;
+import org.apache.flink.types.Row;
+import org.apache.flink.util.Preconditions;
+
+import org.apache.commons.lang3.ArrayUtils;
+
+import java.io.IOException;
+import java.util.Collections;
+import java.util.HashMap;
+import java.util.Map;
+
+/**
+ * A Model which do a minMax scaler operation using the model data computed by {@link MinMaxScaler}.
+ */
+public class MinMaxScalerModel
+ implements Model<MinMaxScalerModel>, MinMaxScalerParams<MinMaxScalerModel> {
+ private final Map<Param<?>, Object> paramMap = new HashMap<>();
+ private Table modelDataTable;
+
+ public MinMaxScalerModel() {
+ ParamUtils.initializeMapWithDefaultValues(paramMap, this);
+ }
+
+ @Override
+ public MinMaxScalerModel setModelData(Table... inputs) {
+ modelDataTable = inputs[0];
+ return this;
+ }
+
+ @Override
+ public Table[] getModelData() {
+ return new Table[] {modelDataTable};
+ }
+
+ @Override
+ @SuppressWarnings("unchecked")
+ public Table[] transform(Table... inputs) {
+ Preconditions.checkArgument(inputs.length == 1);
+ StreamTableEnvironment tEnv =
+ (StreamTableEnvironment) ((TableImpl) inputs[0]).getTableEnvironment();
+ DataStream<Row> data = tEnv.toDataStream(inputs[0]);
+ DataStream<MinMaxScalerModelData> minMaxScalerModel =
+ MinMaxScalerModelData.getModelDataStream(modelDataTable);
+ final String broadcastModelKey = "broadcastModelKey";
+ RowTypeInfo inputTypeInfo = TableUtils.getRowTypeInfo(inputs[0].getResolvedSchema());
+ RowTypeInfo outputTypeInfo =
+ new RowTypeInfo(
+ ArrayUtils.addAll(
+ inputTypeInfo.getFieldTypes(),
+ ExternalTypeInfo.of(DenseVector.class)),
+ ArrayUtils.addAll(inputTypeInfo.getFieldNames(), getPredictionCol()));
+ DataStream<Row> output =
+ BroadcastUtils.withBroadcastStream(
+ Collections.singletonList(data),
+ Collections.singletonMap(broadcastModelKey, minMaxScalerModel),
+ inputList -> {
+ DataStream input = inputList.get(0);
+ return input.map(
+ new PredictOutputFunction(
+ broadcastModelKey,
+ getMax(),
+ getMin(),
+ getFeaturesCol()),
+ outputTypeInfo);
+ });
+ return new Table[] {tEnv.fromDataStream(output)};
+ }
+
+ @Override
+ public Map<Param<?>, Object> getParamMap() {
+ return paramMap;
+ }
+
+ @Override
+ public void save(String path) throws IOException {
+ ReadWriteUtils.saveMetadata(this, path);
+ ReadWriteUtils.saveModelData(
+ MinMaxScalerModelData.getModelDataStream(modelDataTable),
+ path,
+ new MinMaxScalerModelData.ModelDataEncoder());
+ }
+
+ /**
+ * Loads model data from path.
+ *
+ * @param env Stream execution environment.
+ * @param path Model path.
+ * @return MinMaxScalerModel model.
+ */
+ public static MinMaxScalerModel load(StreamExecutionEnvironment env, String path)
+ throws IOException {
+ StreamTableEnvironment tEnv = StreamTableEnvironment.create(env);
+ MinMaxScalerModel model = ReadWriteUtils.loadStageParam(path);
+ DataStream<MinMaxScalerModelData> modelData =
+ ReadWriteUtils.loadModelData(
+ env, path, new MinMaxScalerModelData.ModelDataDecoder());
+ return model.setModelData(tEnv.fromDataStream(modelData));
+ }
+
+ /** This operator loads model data and predicts result. */
+ private static class PredictOutputFunction extends RichMapFunction<Row, Row> {
+ private final String featureCol;
+ private MinMaxScalerModelData minMaxScalerModelData;
+ private final double upperBound;
+ private final double lowerBound;
+ private final String broadcastKey;
+ private DenseVector maxVector;
+ private DenseVector minVector;
+
+ public PredictOutputFunction(
+ String broadcastKey, double upperBound, double lowerBound, String featureCol) {
+ this.upperBound = upperBound;
+ this.lowerBound = lowerBound;
+ this.broadcastKey = broadcastKey;
+ this.featureCol = featureCol;
+ }
+
+ @Override
+ public Row map(Row row) {
+ if (minMaxScalerModelData == null) {
+ minMaxScalerModelData =
+ (MinMaxScalerModelData)
+ getRuntimeContext().getBroadcastVariable(broadcastKey).get(0);
+ maxVector = minMaxScalerModelData.maxVector;
+ minVector = minMaxScalerModelData.minVector;
+ }
+ DenseVector feature = (DenseVector) row.getField(featureCol);
+ DenseVector outputVector = new DenseVector(maxVector.size());
+ if (feature != null) {
Review comment:
Maybe we can make MinMaxScaler implements `HasHandleInvalid`, as the behavior to throw exception corresponds to `HasHandleInvalid.ERROR_INVALID`.
Besides, if the feature vector exists, but its dimension is different from that of maxVector, or its max/min value exceeds maxVector/minVector's range, then maybe we should also throw exception.
##########
File path: flink-ml-lib/src/main/java/org/apache/flink/ml/feature/minmaxscaler/MinMaxScalerModel.java
##########
@@ -0,0 +1,181 @@
+/*
+ * 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.minmaxscaler;
+
+import org.apache.flink.api.common.functions.RichMapFunction;
+import org.apache.flink.api.java.typeutils.RowTypeInfo;
+import org.apache.flink.ml.api.Model;
+import org.apache.flink.ml.common.broadcast.BroadcastUtils;
+import org.apache.flink.ml.common.datastream.TableUtils;
+import org.apache.flink.ml.linalg.DenseVector;
+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.streaming.api.environment.StreamExecutionEnvironment;
+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.table.runtime.typeutils.ExternalTypeInfo;
+import org.apache.flink.types.Row;
+import org.apache.flink.util.Preconditions;
+
+import org.apache.commons.lang3.ArrayUtils;
+
+import java.io.IOException;
+import java.util.Collections;
+import java.util.HashMap;
+import java.util.Map;
+
+/**
+ * A Model which do a minMax scaler operation using the model data computed by {@link MinMaxScaler}.
+ */
+public class MinMaxScalerModel
+ implements Model<MinMaxScalerModel>, MinMaxScalerParams<MinMaxScalerModel> {
+ private final Map<Param<?>, Object> paramMap = new HashMap<>();
+ private Table modelDataTable;
+
+ public MinMaxScalerModel() {
+ ParamUtils.initializeMapWithDefaultValues(paramMap, this);
+ }
+
+ @Override
+ public MinMaxScalerModel setModelData(Table... inputs) {
+ modelDataTable = inputs[0];
+ return this;
+ }
+
+ @Override
+ public Table[] getModelData() {
+ return new Table[] {modelDataTable};
+ }
+
+ @Override
+ @SuppressWarnings("unchecked")
+ public Table[] transform(Table... inputs) {
+ Preconditions.checkArgument(inputs.length == 1);
+ StreamTableEnvironment tEnv =
+ (StreamTableEnvironment) ((TableImpl) inputs[0]).getTableEnvironment();
+ DataStream<Row> data = tEnv.toDataStream(inputs[0]);
+ DataStream<MinMaxScalerModelData> minMaxScalerModel =
+ MinMaxScalerModelData.getModelDataStream(modelDataTable);
+ final String broadcastModelKey = "broadcastModelKey";
+ RowTypeInfo inputTypeInfo = TableUtils.getRowTypeInfo(inputs[0].getResolvedSchema());
+ RowTypeInfo outputTypeInfo =
+ new RowTypeInfo(
+ ArrayUtils.addAll(
+ inputTypeInfo.getFieldTypes(),
+ ExternalTypeInfo.of(DenseVector.class)),
+ ArrayUtils.addAll(inputTypeInfo.getFieldNames(), getPredictionCol()));
+ DataStream<Row> output =
+ BroadcastUtils.withBroadcastStream(
+ Collections.singletonList(data),
+ Collections.singletonMap(broadcastModelKey, minMaxScalerModel),
+ inputList -> {
+ DataStream input = inputList.get(0);
+ return input.map(
+ new PredictOutputFunction(
+ broadcastModelKey,
+ getMax(),
+ getMin(),
+ getFeaturesCol()),
+ outputTypeInfo);
+ });
+ return new Table[] {tEnv.fromDataStream(output)};
+ }
+
+ @Override
+ public Map<Param<?>, Object> getParamMap() {
+ return paramMap;
+ }
+
+ @Override
+ public void save(String path) throws IOException {
+ ReadWriteUtils.saveMetadata(this, path);
+ ReadWriteUtils.saveModelData(
+ MinMaxScalerModelData.getModelDataStream(modelDataTable),
+ path,
+ new MinMaxScalerModelData.ModelDataEncoder());
+ }
+
+ /**
+ * Loads model data from path.
+ *
+ * @param env Stream execution environment.
+ * @param path Model path.
+ * @return MinMaxScalerModel model.
+ */
+ public static MinMaxScalerModel load(StreamExecutionEnvironment env, String path)
+ throws IOException {
+ StreamTableEnvironment tEnv = StreamTableEnvironment.create(env);
+ MinMaxScalerModel model = ReadWriteUtils.loadStageParam(path);
+ DataStream<MinMaxScalerModelData> modelData =
+ ReadWriteUtils.loadModelData(
+ env, path, new MinMaxScalerModelData.ModelDataDecoder());
+ return model.setModelData(tEnv.fromDataStream(modelData));
+ }
+
+ /** This operator loads model data and predicts result. */
+ private static class PredictOutputFunction extends RichMapFunction<Row, Row> {
+ private final String featureCol;
+ private MinMaxScalerModelData minMaxScalerModelData;
+ private final double upperBound;
+ private final double lowerBound;
+ private final String broadcastKey;
+ private DenseVector maxVector;
+ private DenseVector minVector;
+
+ public PredictOutputFunction(
+ String broadcastKey, double upperBound, double lowerBound, String featureCol) {
+ this.upperBound = upperBound;
+ this.lowerBound = lowerBound;
+ this.broadcastKey = broadcastKey;
+ this.featureCol = featureCol;
+ }
+
+ @Override
+ public Row map(Row row) {
+ if (minMaxScalerModelData == null) {
+ minMaxScalerModelData =
+ (MinMaxScalerModelData)
+ getRuntimeContext().getBroadcastVariable(broadcastKey).get(0);
+ maxVector = minMaxScalerModelData.maxVector;
+ minVector = minMaxScalerModelData.minVector;
+ }
+ DenseVector feature = (DenseVector) row.getField(featureCol);
+ DenseVector outputVector = new DenseVector(maxVector.size());
+ if (feature != null) {
+ for (int i = 0; i < maxVector.size(); ++i) {
+ if ((minVector.values[i] - maxVector.values[i]) != 0.0) {
+ outputVector.values[i] =
+ (feature.values[i] - minVector.values[i])
+ / (maxVector.values[i] - minVector.values[i])
Review comment:
It seems that there is no direct usage of `maxVector`. Instead, we only use `maxVector.values[i] - minVector.values[i]`. So maybe we can just keep something like `spanVector` instead of `maxVector` in these operators and in model data.
##########
File path: flink-ml-lib/src/test/java/org/apache/flink/ml/feature/MinMaxScalerTest.java
##########
@@ -0,0 +1,208 @@
+/*
+ * 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;
+
+import org.apache.flink.api.common.functions.MapFunction;
+import org.apache.flink.api.common.restartstrategy.RestartStrategies;
+import org.apache.flink.configuration.Configuration;
+import org.apache.flink.ml.feature.minmaxscaler.MinMaxScaler;
+import org.apache.flink.ml.feature.minmaxscaler.MinMaxScalerModel;
+import org.apache.flink.ml.feature.minmaxscaler.MinMaxScalerModelData;
+import org.apache.flink.ml.linalg.DenseVector;
+import org.apache.flink.ml.linalg.Vectors;
+import org.apache.flink.ml.util.ReadWriteUtils;
+import org.apache.flink.ml.util.StageTestUtils;
+import org.apache.flink.streaming.api.datastream.DataStream;
+import org.apache.flink.streaming.api.environment.ExecutionCheckpointingOptions;
+import org.apache.flink.streaming.api.environment.StreamExecutionEnvironment;
+import org.apache.flink.table.api.DataTypes;
+import org.apache.flink.table.api.Schema;
+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.commons.collections.IteratorUtils;
+import org.junit.Assert;
+import org.junit.Before;
+import org.junit.Rule;
+import org.junit.Test;
+import org.junit.rules.TemporaryFolder;
+
+import java.util.ArrayList;
+import java.util.Arrays;
+import java.util.Collections;
+import java.util.List;
+
+import static org.junit.Assert.assertEquals;
+
+/** Tests {@link MinMaxScaler} and {@link MinMaxScalerModel}. */
+public class MinMaxScalerTest {
+ @Rule public final TemporaryFolder tempFolder = new TemporaryFolder();
+ private StreamExecutionEnvironment env;
+ private StreamTableEnvironment tEnv;
+ private Table trainDataTable;
+ private Table predictDataTable;
+ private static final List<Row> trainData =
+ new ArrayList<>(
+ Arrays.asList(
+ Row.of(Vectors.dense(0.0, 3.0)),
+ Row.of(Vectors.dense(2.1, 0.0)),
+ Row.of(Vectors.dense(4.1, 5.1)),
+ Row.of(Vectors.dense(6.1, 8.1)),
+ Row.of(Vectors.dense(200, 300))));
+ private static final List<Row> predictRows =
+ new ArrayList<>(Collections.singletonList(Row.of(Vectors.dense(150.0, 90.0))));
+
+ @Before
+ public void before() {
+ Configuration config = new Configuration();
+ config.set(ExecutionCheckpointingOptions.ENABLE_CHECKPOINTS_AFTER_TASKS_FINISH, true);
+ env = StreamExecutionEnvironment.getExecutionEnvironment(config);
+ env.setParallelism(4);
+ env.enableCheckpointing(100);
+ env.setRestartStrategy(RestartStrategies.noRestart());
+ tEnv = StreamTableEnvironment.create(env);
+ Schema schema = Schema.newBuilder().column("f0", DataTypes.of(DenseVector.class)).build();
+ DataStream<Row> dataStream = env.fromCollection(trainData);
+ trainDataTable = tEnv.fromDataStream(dataStream, schema).as("features");
+ DataStream<Row> predDataStream = env.fromCollection(predictRows);
+ predictDataTable = tEnv.fromDataStream(predDataStream, schema).as("features");
+ }
+
+ private static void verifyPredictionResult(Table output, String outputCol, DenseVector expected)
+ throws Exception {
+ StreamTableEnvironment tEnv =
+ (StreamTableEnvironment) ((TableImpl) output).getTableEnvironment();
+ DataStream<DenseVector> stream =
+ tEnv.toDataStream(output)
+ .map(
+ (MapFunction<Row, DenseVector>)
+ row -> (DenseVector) row.getField(outputCol));
+ List<DenseVector> result = IteratorUtils.toList(stream.executeAndCollect());
+ assertEquals(1, result.size());
+ assertEquals(expected, result.get(0));
+ }
+
+ @Test
+ public void testParam() {
+ MinMaxScaler minMaxScaler = new MinMaxScaler();
+ assertEquals("features", minMaxScaler.getFeaturesCol());
+ assertEquals(1.0, minMaxScaler.getMax(), 0.0001);
+ assertEquals(0.0, minMaxScaler.getMin(), 0.0001);
+ assertEquals("prediction", minMaxScaler.getPredictionCol());
+ minMaxScaler
+ .setFeaturesCol("test_features")
+ .setMax(4.0)
+ .setMin(1.0)
+ .setPredictionCol("test_output");
+ assertEquals("test_features", minMaxScaler.getFeaturesCol());
+ assertEquals(1.0, minMaxScaler.getMin(), 0.0001);
+ assertEquals(4.0, minMaxScaler.getMax(), 0.0001);
+ assertEquals("test_output", minMaxScaler.getPredictionCol());
+ }
+
+ @Test
+ public void testFeaturePredictionParam() {
+ MinMaxScaler minMaxScaler =
+ new MinMaxScaler()
+ .setMin(1.0)
+ .setMax(4.0)
+ .setFeaturesCol("test_features")
+ .setPredictionCol("test_output");
+ MinMaxScalerModel model = minMaxScaler.fit(trainDataTable.as("test_features"));
+ Table output = model.transform(predictDataTable.as("test_features"))[0];
+ assertEquals(
+ Arrays.asList("test_features", "test_output"),
+ output.getResolvedSchema().getColumnNames());
+ }
+
+ @Test
+ public void testFewerDistinctPointsThanCluster() throws Exception {
Review comment:
Maybe we can refine this naming, as there is no concept of "cluster" in MinMaxScaler.
--
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-unsubscribe@flink.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org