org.deeplearning4j.spark.data.BatchAndExportMultiDataSetsFunction Maven / Gradle / Ivy
The newest version!
/*
* ******************************************************************************
* *
* *
* * This program and the accompanying materials are made available under the
* * terms of the Apache License, Version 2.0 which is available at
* * https://www.apache.org/licenses/LICENSE-2.0.
* *
* * See the NOTICE file distributed with this work for additional
* * information regarding copyright ownership.
* * 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.
* *
* * SPDX-License-Identifier: Apache-2.0
* *****************************************************************************
*/
package org.deeplearning4j.spark.data;
import org.apache.hadoop.conf.Configuration;
import org.apache.hadoop.fs.FSDataOutputStream;
import org.apache.hadoop.fs.FileSystem;
import org.apache.hadoop.fs.Path;
import org.apache.spark.api.java.function.Function2;
import org.apache.spark.broadcast.Broadcast;
import org.datavec.spark.util.DefaultHadoopConfig;
import org.datavec.spark.util.SerializableHadoopConfig;
import org.deeplearning4j.core.util.UIDProvider;
import org.nd4j.linalg.dataset.api.MultiDataSet;
import org.nd4j.common.primitives.Pair;
import java.net.URI;
import java.util.*;
public class BatchAndExportMultiDataSetsFunction
implements Function2, Iterator> {
private final int minibatchSize;
private final String exportBaseDirectory;
private final String jvmuid;
private final Broadcast conf;
/**
* @param minibatchSize Minibatch size to combine examples to (if necessary)
* @param exportBaseDirectory Base directory for exporting
*/
public BatchAndExportMultiDataSetsFunction(int minibatchSize, String exportBaseDirectory) {
this(minibatchSize, exportBaseDirectory, null);
}
/**
* @param minibatchSize Minibatch size to combine examples to (if necessary)
* @param exportBaseDirectory Base directory for exporting
* @param configuration Hadoop Configuration
*/
public BatchAndExportMultiDataSetsFunction(int minibatchSize, String exportBaseDirectory, Broadcast configuration) {
this.minibatchSize = minibatchSize;
this.exportBaseDirectory = exportBaseDirectory;
String fullUID = UIDProvider.getJVMUID();
this.jvmuid = (fullUID.length() <= 8 ? fullUID : fullUID.substring(0, 8));
this.conf = configuration;
}
@Override
public Iterator call(Integer partitionIdx, Iterator iterator) throws Exception {
List outputPaths = new ArrayList<>();
LinkedList tempList = new LinkedList<>();
int count = 0;
while (iterator.hasNext()) {
MultiDataSet next = iterator.next();
if (next.getFeatures(0).size(0) == minibatchSize) {
outputPaths.add(export(next, partitionIdx, count++));
continue;
}
//DataSet must be either smaller or larger than minibatch size...
tempList.add(next);
Pair> countAndPaths = processList(tempList, partitionIdx, count, false);
if (countAndPaths.getSecond() != null && !countAndPaths.getSecond().isEmpty()) {
outputPaths.addAll(countAndPaths.getSecond());
}
count = countAndPaths.getFirst();
}
//We might have some left-over examples...
Pair> countAndPaths = processList(tempList, partitionIdx, count, true);
if (countAndPaths.getSecond() != null && !countAndPaths.getSecond().isEmpty()) {
outputPaths.addAll(countAndPaths.getSecond());
}
return outputPaths.iterator();
}
private Pair> processList(LinkedList tempList, int partitionIdx,
int countBefore, boolean finalExport) throws Exception {
//Go through the list. If we have enough examples: remove the DataSet objects, merge and export them. Otherwise: do nothing
int numExamples = 0;
for (MultiDataSet ds : tempList) {
numExamples += ds.getFeatures(0).size(0);
}
if (tempList.isEmpty() || (numExamples < minibatchSize && !finalExport)) {
//No op
return new Pair<>(countBefore, Collections.emptyList());
}
List exportPaths = new ArrayList<>();
int countAfter = countBefore;
//Batch the required number together
int countSoFar = 0;
List tempToMerge = new ArrayList<>();
while (!tempList.isEmpty() && countSoFar != minibatchSize) {
MultiDataSet next = tempList.removeFirst();
if (countSoFar + next.getFeatures(0).size(0) <= minibatchSize) {
//Add the entire DataSet object
tempToMerge.add(next);
countSoFar += next.getFeatures(0).size(0);
} else {
//Split the DataSet
List examples = next.asList();
for (MultiDataSet ds : examples) {
tempList.addFirst(ds);
}
}
}
//At this point: we should have the required number of examples in tempToMerge (unless it's a final export)
MultiDataSet toExport = org.nd4j.linalg.dataset.MultiDataSet.merge(tempToMerge);
exportPaths.add(export(toExport, partitionIdx, countAfter++));
return new Pair<>(countAfter, exportPaths);
}
private String export(MultiDataSet dataSet, int partitionIdx, int outputCount) throws Exception {
String filename = "mds_" + partitionIdx + jvmuid + "_" + outputCount + ".bin";
URI uri = new URI(exportBaseDirectory
+ (exportBaseDirectory.endsWith("/") || exportBaseDirectory.endsWith("\\") ? "" : "/")
+ filename);
Configuration c = conf == null ? DefaultHadoopConfig.get() : conf.getValue().getConfiguration();
FileSystem file = FileSystem.get(uri, c);
try (FSDataOutputStream out = file.create(new Path(uri))) {
dataSet.save(out);
}
return uri.toString();
}
}