org.deeplearning4j.spark.impl.computationgraph.SparkComputationGraph Maven / Gradle / Ivy
/*
*
* * Copyright 2016 Skymind,Inc.
* *
* * Licensed 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.deeplearning4j.spark.impl.computationgraph;
import lombok.NonNull;
import org.apache.spark.SparkContext;
import org.apache.spark.api.java.JavaDoubleRDD;
import org.apache.spark.api.java.JavaPairRDD;
import org.apache.spark.api.java.JavaRDD;
import org.apache.spark.api.java.JavaSparkContext;
import org.apache.spark.api.java.function.DoubleFunction;
import org.apache.spark.broadcast.Broadcast;
import org.canova.api.records.reader.RecordReader;
import org.deeplearning4j.nn.conf.ComputationGraphConfiguration;
import org.deeplearning4j.nn.conf.graph.GraphVertex;
import org.deeplearning4j.nn.conf.graph.LayerVertex;
import org.deeplearning4j.nn.conf.layers.FeedForwardLayer;
import org.deeplearning4j.nn.gradient.Gradient;
import org.deeplearning4j.nn.graph.ComputationGraph;
import org.deeplearning4j.nn.multilayer.MultiLayerNetwork;
import org.deeplearning4j.nn.updater.graph.ComputationGraphUpdater;
import org.deeplearning4j.optimize.api.IterationListener;
import org.deeplearning4j.spark.canova.RecordReaderFunction;
import org.deeplearning4j.spark.impl.common.Adder;
import org.deeplearning4j.spark.impl.common.gradient.GradientAdder;
import org.deeplearning4j.spark.impl.common.misc.*;
import org.deeplearning4j.spark.impl.common.updater.UpdaterAggregatorCombinerCG;
import org.deeplearning4j.spark.impl.common.updater.UpdaterElementCombinerCG;
import org.deeplearning4j.spark.impl.computationgraph.dataset.DataSetToMultiDataSetFn;
import org.deeplearning4j.spark.impl.computationgraph.dataset.PairDataSetToMultiDataSetFn;
import org.deeplearning4j.spark.impl.computationgraph.gradientaccum.GradientAccumFlatMapCG;
import org.deeplearning4j.spark.impl.computationgraph.scoring.ScoreExamplesFunction;
import org.deeplearning4j.spark.impl.computationgraph.scoring.ScoreExamplesWithKeyFunction;
import org.deeplearning4j.util.ModelSerializer;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.dataset.DataSet;
import org.nd4j.linalg.dataset.api.MultiDataSet;
import org.nd4j.linalg.heartbeat.Heartbeat;
import org.nd4j.linalg.heartbeat.reports.Environment;
import org.nd4j.linalg.heartbeat.reports.Event;
import org.nd4j.linalg.heartbeat.reports.Task;
import org.nd4j.linalg.heartbeat.utils.EnvironmentUtils;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;
import scala.Tuple3;
import java.io.Serializable;
import java.util.ArrayList;
import java.util.Collection;
import java.util.List;
import java.util.concurrent.atomic.AtomicInteger;
/**Main class for training ComputationGraph networks using Spark
*
* @author Alex Black
*/
public class SparkComputationGraph implements Serializable {
public static final int DEFAULT_EVAL_SCORE_BATCH_SIZE = 50;
private transient JavaSparkContext sc;
private ComputationGraphConfiguration conf;
private ComputationGraph network;
private Broadcast params;
private Broadcast updater;
private boolean averageEachIteration = false;
private boolean initDone = false;
public final static String AVERAGE_EACH_ITERATION = "org.deeplearning4j.spark.iteration.average";
public final static String ACCUM_GRADIENT = "org.deeplearning4j.spark.iteration.accumgrad";
public final static String DIVIDE_ACCUM_GRADIENT = "org.deeplearning4j.spark.iteration.dividegrad";
private double lastScore;
private static final Logger log = LoggerFactory.getLogger(SparkComputationGraph.class);
private transient AtomicInteger iterationsCount = new AtomicInteger(0);
private List listeners = new ArrayList<>();
/**
* Instantiate a ComputationGraph instance with the given context and network.
* @param sparkContext the spark context to use
* @param network the network to use
*/
public SparkComputationGraph(SparkContext sparkContext, ComputationGraph network) {
this(new JavaSparkContext(sparkContext),network);
}
public SparkComputationGraph(JavaSparkContext javaSparkContext, ComputationGraph network){
sc = javaSparkContext;
this.conf = network.getConfiguration().clone();
this.network = network;
this.network.init();
this.updater = sc.broadcast(network.getUpdater());
this.averageEachIteration = sc.getConf().getBoolean(AVERAGE_EACH_ITERATION, false);
}
public SparkComputationGraph(SparkContext sparkContext, ComputationGraphConfiguration conf) {
this(new JavaSparkContext(sparkContext),conf);
}
public SparkComputationGraph(JavaSparkContext sparkContext, ComputationGraphConfiguration conf){
sc = sparkContext;
this.conf = conf.clone();
this.network = new ComputationGraph(conf);
this.network.init();
this.averageEachIteration = sparkContext.sc().conf().getBoolean(AVERAGE_EACH_ITERATION, false);
this.updater = sc.broadcast(network.getUpdater());
}
/**Train a ComputationGraph network based on data loaded from a text file + {@link RecordReader}.
* This method splits the data into approximately {@code examplesPerFit} sized splits, and trains on each split.
* one after the other. See {@link #fitDataSet(JavaRDD, int, int, int)} for further details.
* NOTE: This method can only be used with ComputationGraph instances that have a single input and single output
* @param path the path to the text file
* @param labelIndex the label index
* @param recordReader the record reader to parse results
* @param examplesPerFit Number of examples to fit on at each iteration
* @param totalExamples total number of examples
* @param numPartitions Number of partitions. Usually set to number of executors
* @return {@link MultiLayerNetwork}
*/
public ComputationGraph fit(String path,int labelIndex, RecordReader recordReader, int examplesPerFit, int totalExamples, int numPartitions ) {
if(network.getNumInputArrays() != 1 || network.getNumOutputArrays() != 1){
throw new UnsupportedOperationException("Cannot train ComputationGraph with multiple inputs/outputs from text file + record reader");
}
JavaRDD points = loadFromTextFile(path, labelIndex, recordReader).map(new DataSetToMultiDataSetFn());
return fitMultiDataSet(points, examplesPerFit, totalExamples, numPartitions);
}
private JavaRDD loadFromTextFile(String path, int labelIndex, RecordReader recordReader ){
JavaRDD lines = sc.textFile(path);
int nOut = ((FeedForwardLayer)network.getOutputLayer(0)).getNOut();
return lines.map(new RecordReaderFunction(recordReader, labelIndex, nOut));
}
public ComputationGraph getNetwork() {
return network;
}
public void setNetwork(ComputationGraph network) {
this.network = network;
}
/**Fit the data, splitting into smaller data subsets if necessary. This allows large {@code JavaRDD}s)
* to be trained as a set of smaller steps instead of all together.
* Using this method, training progresses as follows:
* train on {@code examplesPerFit} examples -> average parameters -> train on {@code examplesPerFit} -> average
* parameters etc until entire data set has been processed
* Note: The actual number of splits for the input data is based on rounding up.
* Suppose {@code examplesPerFit}=1000, with {@code rdd.count()}=1200. Then, we round up to 2000 examples, and the
* network will then be fit in two steps (as 2000/1000=2), with 1200/2=600 examples at each step. These 600 examples
* will then be distributed approximately equally (no guarantees) amongst each executor/core for training.
*
* @param rdd Data to train on
* @param examplesPerFit Number of examples to learn on (between averaging) across all executors. For example, if set to
* 1000 and rdd.count() == 10k, then we do 10 sets of learning, each on 1000 examples.
* To use all examples, set maxExamplesPerFit to Integer.MAX_VALUE
* @param totalExamples total number of examples in the data RDD
* @param numPartitions number of partitions to divide the data in to
* @return Trained network
*/
public ComputationGraph fitMultiDataSet(JavaRDD rdd, int examplesPerFit, int totalExamples, int numPartitions ){
int nSplits;
if(examplesPerFit == Integer.MAX_VALUE || examplesPerFit >= totalExamples ) nSplits = 1;
else {
if(totalExamples%examplesPerFit==0){
nSplits = (totalExamples / examplesPerFit);
} else {
nSplits = (totalExamples/ examplesPerFit) + 1;
}
}
if(nSplits == 1){
fitDataSet(rdd);
} else {
double[] splitWeights = new double[nSplits];
for( int i=0; i[] subsets = rdd.randomSplit(splitWeights);
for( int i=0; i next = subsets[i].repartition(numPartitions);
fitDataSet(next);
}
}
return network;
}
/**DataSet version of {@link #fitMultiDataSet(JavaRDD, int, int, int)}.
* Handles conversion from DataSet to MultiDataSet internally.
*/
public ComputationGraph fitDataSet(JavaRDD rdd, int examplesPerFit, int totalExamples, int numPartitions) {
if(network.getNumInputArrays() != 1 || network.getNumOutputArrays() != 1){
throw new UnsupportedOperationException("Cannot train ComputationGraph with multiple inputs/outputs from DataSet");
}
JavaRDD mds = rdd.map(new DataSetToMultiDataSetFn());
return fitMultiDataSet(mds, examplesPerFit, totalExamples, numPartitions);
}
/**
* Fit the dataset rdd
* @param rdd the rdd to fitDataSet
* @return the ComputationGraph after parameter averaging
*/
public ComputationGraph fitDataSet(JavaRDD rdd) {
int iterations = network.getLayer(0).conf().getNumIterations();
log.info("Running distributed training: (averaging each iteration = " + averageEachIteration + "), (iterations = " +
iterations + "), (num partions = " + rdd.partitions().size() + ")");
if(!averageEachIteration) {
//Do multiple iterations and average once at the end
runIteration(rdd);
} else {
//Temporarily set numIterations = 1. Control numIterations externally here so we can average between iterations
for(GraphVertex gv : conf.getVertices().values()) {
if(gv instanceof LayerVertex){
((LayerVertex)gv).getLayerConf().setNumIterations(1); //TODO - do this more elegantly...
}
}
//Run learning, and average at each iteration
for(int i = 0; i < iterations; i++) {
runIteration(rdd);
}
//Reset number of iterations in config
if(iterations > 1 ){
for(GraphVertex gv : conf.getVertices().values()) {
if(gv instanceof LayerVertex){
((LayerVertex)gv).getLayerConf().setNumIterations(iterations); //TODO - do this more elegantly...
}
}
}
}
return network;
}
protected void runIteration(JavaRDD rdd) {
log.info("Broadcasting initial parameters of length " + network.numParams(false));
int maxRep = 0;
long maxSm = 0;
INDArray valToBroadcast = network.params(false);
this.params = sc.broadcast(valToBroadcast);
ComputationGraphUpdater updater = network.getUpdater();
if(updater == null) {
network.setUpdater(new ComputationGraphUpdater(network));
log.warn("Unable to propagate null updater");
updater = network.getUpdater();
}
this.updater = sc.broadcast(updater);
int paramsLength = network.numParams(true);
boolean accumGrad = sc.getConf().getBoolean(ACCUM_GRADIENT, false);
if(accumGrad) {
//Learning via averaging gradients
JavaRDD> results = rdd.mapPartitions(new GradientAccumFlatMapCG(conf.toJson(),
this.params, this.updater),true).cache();
JavaRDD resultsGradient = results.map(new GradientFromTupleFunctionCG());
log.info("Ran iterative reduce... averaging gradients now.");
GradientAdder a = new GradientAdder(paramsLength);
resultsGradient.foreach(a);
INDArray accumulatedGradient = a.getAccumulator().value();
boolean divideGrad = sc.getConf().getBoolean(DIVIDE_ACCUM_GRADIENT,false);
if(divideGrad) {
maxRep = results.partitions().size();
accumulatedGradient.divi(maxRep);
}
log.info("Accumulated parameters");
log.info("Summed gradients.");
network.setParams(network.params(false).addi(accumulatedGradient));
log.info("Set parameters");
log.info("Processing updaters");
JavaRDD resultsUpdater = results.map(new UpdaterFromGradientTupleFunctionCG());
// implements DoubleFunction>
JavaDoubleRDD scores = results.mapToDouble(new ScoreMappingG());
if (!initDone) {
JavaDoubleRDD sm = results.mapToDouble(new SMappingG());
maxSm = sm.mean().longValue();
}
lastScore = scores.mean();
ComputationGraphUpdater.Aggregator aggregator = resultsUpdater.aggregate(
null,
new UpdaterElementCombinerCG(),
new UpdaterAggregatorCombinerCG()
);
ComputationGraphUpdater combinedUpdater = aggregator.getUpdater();
network.setUpdater(combinedUpdater);
log.info("Set updater");
}
else {
//Standard parameter averaging
JavaRDD> results = rdd.mapPartitions(new IterativeReduceFlatMapCG(conf.toJson(),
this.params, this.updater),true).cache();
JavaRDD resultsParams = results.map(new INDArrayFromTupleFunctionCG());
log.info("Running iterative reduce and averaging parameters");
Adder a = new Adder(paramsLength,sc.accumulator(0));
resultsParams.foreach(a);
INDArray newParams = a.getAccumulator().value();
maxRep = a.getCounter().value();
newParams.divi(maxRep);
network.setParams(newParams);
log.info("Accumulated and set parameters");
JavaRDD resultsUpdater = results.map(new UpdaterFromTupleFunctionCG());
JavaDoubleRDD scores = results.mapToDouble(new ScoreMapping());
if (!initDone) {
JavaDoubleRDD sm = results.mapToDouble(new SMapping());
maxSm = sm.mean().longValue();
}
lastScore = scores.mean();
ComputationGraphUpdater.Aggregator aggregator = resultsUpdater.aggregate(
null,
new UpdaterElementCombinerCG(),
new UpdaterAggregatorCombinerCG()
);
ComputationGraphUpdater combinedUpdater = aggregator.getUpdater();
network.setUpdater(combinedUpdater);
log.info("Processed and set updater");
}
if (listeners.size() > 0) {
network.setScore(lastScore);
invokeListeners(network, iterationsCount.incrementAndGet());
}
if (!initDone) {
initDone = true;
update(maxRep, maxSm);
}
}
/**
* This method allows you to specify IterationListeners for this model.
*
* PLEASE NOTE:
* 1. These iteration listeners should be configured to use remote UiServer
* 2. Remote UiServer should be accessible via network from Spark master node.
*
* @param listeners
*/
public void setListeners(@NonNull Collection listeners) {
this.listeners.clear();
this.listeners.addAll(listeners);
}
protected void invokeListeners(ComputationGraph network, int iteration) {
for (IterationListener listener: listeners) {
try {
listener.iterationDone(network, iteration);
} catch (Exception e) {
log.error("Exception caught at IterationListener invocation" + e.getMessage());
e.printStackTrace();
}
}
}
/** Gets the last (average) minibatch score from calling fit */
public double getScore(){
return lastScore;
}
public double calculateScoreDataSet(JavaRDD data, boolean average){
long n = data.count();
JavaRDD scores = data.mapPartitions(new ScoreFlatMapFunctionCGDataSet(conf.toJson(), sc.broadcast(network.params(false))));
List scoresList = scores.collect();
double sum = 0.0;
for(Double d : scoresList) sum += d;
if(average) return sum / n;
return sum;
}
public double calculateScore(JavaRDD data, boolean average){
long n = data.count();
JavaRDD scores = data.mapPartitions(new ScoreFlatMapFunctionCGMultiDataSet(conf.toJson(), sc.broadcast(network.params(false))));
List scoresList = scores.collect();
double sum = 0.0;
for(Double d : scoresList) sum += d;
if(average) return sum / n;
return sum;
}
/** DataSet version of {@link #scoreExamples(JavaRDD, boolean)}
*/
public JavaDoubleRDD scoreExamplesDataSet(JavaRDD data, boolean includeRegularizationTerms) {
return scoreExamples(data.map(new DataSetToMultiDataSetFn()),includeRegularizationTerms);
}
/**DataSet version of {@link #scoreExamples(JavaPairRDD, boolean, int)}
*/
public JavaDoubleRDD scoreExamplesDataSet(JavaRDD data, boolean includeRegularizationTerms, int batchSize) {
return scoreExamples(data.map(new DataSetToMultiDataSetFn()), includeRegularizationTerms, batchSize);
}
/**DataSet version of {@link #scoreExamples(JavaPairRDD, boolean)}
*/
public JavaPairRDD scoreExamplesDataSet(JavaPairRDD data, boolean includeRegularizationTerms){
return scoreExamples(data.mapToPair(new PairDataSetToMultiDataSetFn()),includeRegularizationTerms,DEFAULT_EVAL_SCORE_BATCH_SIZE);
}
/**DataSet version of {@link #scoreExamples(JavaPairRDD, boolean,int)}
*/
public JavaPairRDD scoreExamplesDataSet(JavaPairRDD data, boolean includeRegularizationTerms, int batchSize){
return scoreExamples(data.mapToPair(new PairDataSetToMultiDataSetFn()),includeRegularizationTerms,batchSize);
}
/** Score the examples individually, using the default batch size {@link #DEFAULT_EVAL_SCORE_BATCH_SIZE}. Unlike {@link #calculateScore(JavaRDD, boolean)},
* this method returns a score for each example separately. If scoring is needed for specific examples use either
* {@link #scoreExamples(JavaPairRDD, boolean)} or {@link #scoreExamples(JavaPairRDD, boolean, int)} which can have
* a key for each example.
* @param data Data to score
* @param includeRegularizationTerms If true: include the l1/l2 regularization terms with the score (if any)
* @return A JavaDoubleRDD containing the scores of each example
* @see ComputationGraph#scoreExamples(MultiDataSet, boolean)
*/
public JavaDoubleRDD scoreExamples(JavaRDD data, boolean includeRegularizationTerms) {
return scoreExamples(data,includeRegularizationTerms,DEFAULT_EVAL_SCORE_BATCH_SIZE);
}
/** Score the examples individually, using a specified batch size. Unlike {@link #calculateScore(JavaRDD, boolean)},
* this method returns a score for each example separately. If scoring is needed for specific examples use either
* {@link #scoreExamples(JavaPairRDD, boolean)} or {@link #scoreExamples(JavaPairRDD, boolean, int)} which can have
* a key for each example.
* @param data Data to score
* @param includeRegularizationTerms If true: include the l1/l2 regularization terms with the score (if any)
* @param batchSize Batch size to use when doing scoring
* @return A JavaDoubleRDD containing the scores of each example
* @see ComputationGraph#scoreExamples(MultiDataSet, boolean)
*/
public JavaDoubleRDD scoreExamples(JavaRDD data, boolean includeRegularizationTerms, int batchSize) {
return data.mapPartitionsToDouble(new ScoreExamplesFunction(sc.broadcast(network.params()), sc.broadcast(conf.toJson()),
includeRegularizationTerms, batchSize));
}
/** Score the examples individually, using the default batch size {@link #DEFAULT_EVAL_SCORE_BATCH_SIZE}. Unlike {@link #calculateScore(JavaRDD, boolean)},
* this method returns a score for each example separately
* Note: The provided JavaPairRDD has a key that is associated with each example and returned score.
* Note: The DataSet objects passed in must have exactly one example in them (otherwise: can't have a 1:1 association
* between keys and data sets to score)
* @param data Data to score
* @param includeRegularizationTerms If true: include the l1/l2 regularization terms with the score (if any)
* @param Key type
* @return A {@code JavaPairRDD} containing the scores of each example
* @see MultiLayerNetwork#scoreExamples(DataSet, boolean)
*/
public JavaPairRDD scoreExamples(JavaPairRDD data, boolean includeRegularizationTerms){
return scoreExamples(data,includeRegularizationTerms,DEFAULT_EVAL_SCORE_BATCH_SIZE);
}
private void update(int mr, long mg) {
Environment env = EnvironmentUtils.buildEnvironment();
env.setNumCores(mr);
env.setAvailableMemory(mg);
Task task = ModelSerializer.taskByModel(network);
Heartbeat.getInstance().reportEvent(Event.SPARK, env, task);
}
/** Score the examples individually, using a specified batch size. Unlike {@link #calculateScore(JavaRDD, boolean)},
* this method returns a score for each example separately
* Note: The provided JavaPairRDD has a key that is associated with each example and returned score.
* Note: The DataSet objects passed in must have exactly one example in them (otherwise: can't have a 1:1 association
* between keys and data sets to score)
* @param data Data to score
* @param includeRegularizationTerms If true: include the l1/l2 regularization terms with the score (if any)
* @param Key type
* @return A {@code JavaPairRDD} containing the scores of each example
* @see MultiLayerNetwork#scoreExamples(DataSet, boolean)
*/
public JavaPairRDD scoreExamples(JavaPairRDD data, boolean includeRegularizationTerms, int batchSize ){
return data.mapPartitionsToPair(new ScoreExamplesWithKeyFunction(sc.broadcast(network.params()), sc.broadcast(conf.toJson()),
includeRegularizationTerms, batchSize));
}
private static class ScoreMapping implements DoubleFunction> {
@Override
public double call(Tuple3 t3) throws Exception {
return t3._3().getS();
}
}
private static class ScoreMappingG implements DoubleFunction> {
@Override
public double call(Tuple3 t3) throws Exception {
return t3._3().getS();
}
}
private static class SMapping implements DoubleFunction> {
@Override
public double call(Tuple3 t3) throws Exception {
return t3._3().getM();
}
}
private static class SMappingG implements DoubleFunction> {
@Override
public double call(Tuple3 t3) throws Exception {
return t3._3().getM();
}
}
}
© 2015 - 2025 Weber Informatics LLC | Privacy Policy