deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/embeddings/learning/impl/elements/SkipGram.java
/*
* ******************************************************************************
* *
* *
* * 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.models.embeddings.learning.impl.elements;
import lombok.*;
import lombok.extern.slf4j.Slf4j;
import org.apache.commons.lang3.RandomUtils;
import org.deeplearning4j.config.DL4JSystemProperties;
import org.deeplearning4j.models.embeddings.WeightLookupTable;
import org.deeplearning4j.models.embeddings.inmemory.InMemoryLookupTable;
import org.deeplearning4j.models.embeddings.learning.ElementsLearningAlgorithm;
import org.deeplearning4j.models.embeddings.loader.VectorsConfiguration;
import org.deeplearning4j.models.sequencevectors.interfaces.SequenceIterator;
import org.deeplearning4j.models.sequencevectors.sequence.Sequence;
import org.deeplearning4j.models.sequencevectors.sequence.SequenceElement;
import org.deeplearning4j.models.word2vec.wordstore.VocabCache;
import org.nd4j.linalg.api.buffer.DataType;
import org.nd4j.linalg.api.memory.MemoryWorkspace;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.api.ops.impl.nlp.SkipGramInference;
import org.nd4j.linalg.api.ops.impl.nlp.SkipGramRound;
import org.nd4j.linalg.factory.Nd4j;
import org.nd4j.linalg.util.DeviceLocalNDArray;
import org.nd4j.shade.guava.cache.Cache;
import org.nd4j.shade.guava.cache.CacheBuilder;
import org.nd4j.shade.guava.cache.Weigher;
import java.time.Duration;
import java.util.ArrayList;
import java.util.List;
import java.util.NoSuchElementException;
import java.util.Queue;
import java.util.concurrent.ConcurrentLinkedQueue;
import java.util.concurrent.atomic.AtomicLong;
@Slf4j
public class SkipGram<T extends SequenceElement> implements ElementsLearningAlgorithm<T> {
protected VocabCache<T> vocabCache;
protected WeightLookupTable<T> lookupTable;
protected VectorsConfiguration configuration;
protected int window;
protected boolean useAdaGrad;
protected double negative;
protected double sampling;
protected int[] variableWindows;
protected int vectorLength;
protected int maxQueueSize = Integer.parseInt(System.getProperty(DL4JSystemProperties.NLP_QUEUE_SIZE,"1000"));
private Cache<IterationArraysKey, Queue<IterationArrays>> iterationArrays = CacheBuilder.newBuilder()
.maximumSize(Integer.parseInt(System.getProperty(DL4JSystemProperties.NLP_CACHE_SIZE,"1000")))
.weakKeys()
.expireAfterWrite(Duration.ofMinutes(5))
.build();
protected int workers = Runtime.getRuntime().availableProcessors();
public int getWorkers() {
return workers;
}
public void setWorkers(int workers) {
this.workers = workers;
}
@Getter
@Setter
protected DeviceLocalNDArray syn0, syn1, syn1Neg, table, expTable;
protected ThreadLocal<List<BatchItem<T>>> batches = new ThreadLocal<>();
/**
* Dummy construction is required for reflection
*/
public SkipGram() {
}
public List<BatchItem<T>> getBatch() {
if (batches.get() == null)
batches.set(new ArrayList<>());
return batches.get();
}
/**
* Returns implementation code name
*
* @return
*/
@Override
public String getCodeName() {
return "SkipGram";
}
/**
* SkipGram initialization over given vocabulary and WeightLookupTable
*
* @param vocabCache
* @param lookupTable
* @param configuration
*/
@Override
public void configure(@NonNull VocabCache<T> vocabCache, @NonNull WeightLookupTable<T> lookupTable,
@NonNull VectorsConfiguration configuration) {
this.vocabCache = vocabCache;
this.lookupTable = lookupTable;
this.configuration = configuration;
if (configuration.getNegative() > 0) {
if (((InMemoryLookupTable<T>) lookupTable).getSyn1Neg() == null) {
log.info("Initializing syn1Neg...");
((InMemoryLookupTable<T>) lookupTable).setUseHS(configuration.isUseHierarchicSoftmax());
((InMemoryLookupTable<T>) lookupTable).setNegative(configuration.getNegative());
lookupTable.resetWeights(false);
}
}
this.syn0 = new DeviceLocalNDArray(((InMemoryLookupTable<T>) lookupTable).getSyn0());
this.syn1 = new DeviceLocalNDArray(((InMemoryLookupTable<T>) lookupTable).getSyn1());
this.syn1Neg = new DeviceLocalNDArray(((InMemoryLookupTable<T>) lookupTable).getSyn1Neg());
this.expTable = new DeviceLocalNDArray(Nd4j.create(((InMemoryLookupTable<T>) lookupTable).getExpTable(),
new long[]{((InMemoryLookupTable<T>) lookupTable).getExpTable().length}, syn0.get() == null ? DataType.DOUBLE
: syn0.get().dataType()));
this.table = new DeviceLocalNDArray(((InMemoryLookupTable<T>) lookupTable).getTable());
this.window = configuration.getWindow();
this.useAdaGrad = configuration.isUseAdaGrad();
this.negative = configuration.getNegative();
this.sampling = configuration.getSampling();
this.variableWindows = configuration.getVariableWindows();
this.workers = configuration.getWorkers();
this.vectorLength = configuration.getLayersSize();
}
/**
* SkipGram doesn't involve any pretraining
*
* @param iterator
*/
@Override
public void pretrain(SequenceIterator<T> iterator) {
// no-op
}
public Sequence<T> applySubsampling(@NonNull Sequence<T> sequence, @NonNull AtomicLong nextRandom) {
Sequence<T> result = new Sequence<>();
// subsampling implementation, if subsampling threshold met, just continue to next element
if (sampling > 0) {
result.setSequenceId(sequence.getSequenceId());
if (sequence.getSequenceLabels() != null)
result.setSequenceLabels(sequence.getSequenceLabels());
if (sequence.getSequenceLabel() != null)
result.setSequenceLabel(sequence.getSequenceLabel());
for (T element : sequence.getElements()) {
double numWords = vocabCache.totalWordOccurrences();
double ran = (Math.sqrt(element.getElementFrequency() / (sampling * numWords)) + 1)
* (sampling * numWords) / element.getElementFrequency();
nextRandom.set(Math.abs(nextRandom.get() * 25214903917L + 11));
if (ran < (nextRandom.get() & 0xFFFF) / (double) 65536) {
continue;
}
result.addElement(element);
}
return result;
} else
return sequence;
}
/**
* Learns sequence using SkipGram algorithm
*
* @param sequence
* @param nextRandom
* @param learningRate
*/
@Override
public double learnSequence(@NonNull Sequence<T> sequence, @NonNull AtomicLong nextRandom, double learningRate) {
Sequence<T> tempSequence = sequence;
if (sampling > 0)
tempSequence = applySubsampling(sequence, nextRandom);
double score = 0.0;
int currentWindow = window;
if (variableWindows != null && variableWindows.length != 0) {
currentWindow = variableWindows[RandomUtils.nextInt(0, variableWindows.length)];
}
for (int i = 0; i < tempSequence.getElements().size(); i++) {
nextRandom.set(Math.abs(nextRandom.get() * 25214903917L + 11));
score = skipGram(i, tempSequence.getElements(), (int) nextRandom.get() % currentWindow, nextRandom,
learningRate, currentWindow);
}
if (getBatch() != null && getBatch().size() >= configuration.getBatchSize()) {
doExec(getBatch(),null);
getBatch().clear();
}
return score;
}
public void clearBatch() {
getBatch().clear();
}
@Override
public void finish() {
if (batches != null && batches.get() != null && !batches.get().isEmpty()) {
iterateSample(null);
clearBatch();
}
}
@Override
public void finish(INDArray inferenceVector) {
if (batches != null && batches.get() != null && !batches.get().isEmpty()) {
iterateSample(null);
clearBatch();
}
}
/**
* SkipGram has no reasons for early termination ever.
*
* @return
*/
@Override
public boolean isEarlyTerminationHit() {
return false;
}
public void addBatchItem(BatchItem<T> batchItem) {
getBatch().add(batchItem);
}
private double skipGram(int i, List<T> sentence, int b, AtomicLong nextRandom, double alpha, int currentWindow) {
final T word = sentence.get(i);
if (word == null || sentence.isEmpty() || word.isLocked())
return 0.0;
double score = 0.0;
int end = currentWindow * 2 + 1 - b;
for (int a = b; a < end; a++) {
if (a != currentWindow) {
int c = i - currentWindow + a;
if (c >= 0 && c < sentence.size()) {
T lastWord = sentence.get(c);
nextRandom.set(Math.abs(nextRandom.get() * 25214903917L + 11));
BatchItem<T> batchItem = new BatchItem<>(word, lastWord, nextRandom.get(), alpha);
addBatchItem(batchItem);
}
}
}
return score;
}
public double iterateSample(BatchItem<T> item) {
double score = 0.0;
List<BatchItem<T>> items = getBatch();
if(item != null) {
items.add(item);
if(items.size() >= configuration.getBatchSize()) {
score = doExec(items, null);
}
} else if(item == null && !items.isEmpty()) {
if(items.size() >= configuration.getBatchSize()) {
score = doExec(items, null);
}
}
return score;
}
public Double doExec(List<BatchItem<T>> items,INDArray inferenceVector) {
try(MemoryWorkspace workspace = Nd4j.getWorkspaceManager().scopeOutOfWorkspaces()) {
if (items.size() > 1) {
INDArray targetArray = null;
INDArray ngStarterArray = null;
INDArray alphasArray = null;
INDArray randomValuesArr = null;
int maxCols = 1;
for (int i = 0; i < items.size(); i++) {
int curr = items.get(i).getWord().getCodeLength();
if (curr > maxCols)
maxCols = curr;
}
IterationArraysKey key = IterationArraysKey.builder()
.itemSize(items.size())
.maxCols(maxCols).build();
Queue<IterationArrays> iterationArraysQueue = iterationArrays.getIfPresent(key);
IterationArrays iterationArrays1;
if(iterationArraysQueue == null) {
iterationArraysQueue = new ConcurrentLinkedQueue<>();
iterationArrays.put(key,iterationArraysQueue);
iterationArrays1 = new IterationArrays(items.size(),maxCols);
} else {
if(iterationArraysQueue.isEmpty()) {
iterationArrays1 = new IterationArrays(items.size(),maxCols);
}else {
try {
iterationArrays1 = iterationArraysQueue.remove();
iterationArrays1.initCodes();
}catch(NoSuchElementException e) {
iterationArrays1 = new IterationArrays(items.size(),maxCols);
}
}
}
int[][] indicesArr = iterationArrays1.indicesArr;
int[][] codesArr = iterationArrays1.codesArr;
//use -1 as padding for codes that are not actually valid for a given row
INDArray codes = null;
INDArray indices = null;
long[] randomValues = iterationArrays1.randomValues;
double[] alphas = iterationArrays1.alphas;
int[] targets = iterationArrays1.targets;
int[] ngStarters = iterationArrays1.ngStarters;
for (int cnt = 0; cnt < items.size(); cnt++) {
T w1 = items.get(cnt).getWord();
T lastWord = items.get(cnt).getLastWord();
randomValues[cnt] = items.get(cnt).getRandomValue();
double alpha = items.get(cnt).getAlpha();
if (w1 == null || lastWord == null || (lastWord.getIndex() < 0 && inferenceVector == null)
|| w1.getIndex() == lastWord.getIndex() || w1.getLabel().equals("STOP")
|| lastWord.getLabel().equals("STOP") || w1.getLabel().equals("UNK")
|| lastWord.getLabel().equals("UNK")) {
continue;
}
int target = lastWord.getIndex();
int ngStarter = w1.getIndex();
targets[cnt] = target;
ngStarters[cnt] = ngStarter;
alphas[cnt] = alpha;
if (configuration.isUseHierarchicSoftmax()) {
for (int i = 0; i < w1.getCodeLength(); i++) {
int code = w1.getCodes().get(i);
int point = w1.getPoints().get(i);
if (point >= vocabCache.numWords() || point < 0)
continue;
codesArr[cnt][i] = code;
indicesArr[cnt][i] = point;
}
}
//negative sampling
if (negative > 0) {
if (syn1Neg == null) {
((InMemoryLookupTable<T>) lookupTable).initNegative();
syn1Neg = new DeviceLocalNDArray(((InMemoryLookupTable<T>) lookupTable).getSyn1Neg());
}
}
}
alphasArray = Nd4j.createFromArray(alphas);
if(negative > 0)
ngStarterArray = Nd4j.createFromArray(ngStarters);
randomValuesArr = Nd4j.createFromArray(randomValues);
targetArray = Nd4j.createFromArray(targets);
if(configuration.isUseHierarchicSoftmax())
codes = Nd4j.createFromArray(codesArr);
if(configuration.isUseHierarchicSoftmax())
indices = Nd4j.createFromArray(indicesArr);
SkipGramRound sg = SkipGramRound.builder()
.target(targetArray)
.expTable(expTable.get())
.ngStarter((negative > 0) ? ngStarterArray : Nd4j.empty(DataType.INT32))
.syn0(syn0.get())
.syn1(configuration.isUseHierarchicSoftmax() ? syn1.get() : Nd4j.empty(syn0.get().dataType()))
.syn1Neg((negative > 0) ? syn1Neg.get() : Nd4j.empty(syn0.get().dataType()))
.negTable((negative > 0) ? table.get() : Nd4j.empty(syn0.get().dataType()))
.indices(configuration.isUseHierarchicSoftmax() ? indices : Nd4j.empty(DataType.INT32))
.codes(configuration.isUseHierarchicSoftmax() ? codes: Nd4j.empty(DataType.INT8))
.alpha(alphasArray)
.randomValue(randomValuesArr)
.inferenceVector(inferenceVector != null ? inferenceVector : Nd4j.empty(syn0.get().dataType()))
.preciseMode(configuration.isPreciseMode())
.numWorkers(workers)
.iterations(inferenceVector != null ? configuration.getIterations() * configuration.getEpochs() : 1)
.build();
Nd4j.getExecutioner().exec(sg);
items.clear();
sg.inputArguments().clear();
Nd4j.close(targetArray,codes,indices,alphasArray,ngStarterArray,randomValuesArr);
if(iterationArraysQueue.size() < maxQueueSize)
iterationArraysQueue.add(iterationArrays1);
} else {
int cnt = 0;
T w1 = items.get(cnt).getWord();
T lastWord = items.get(cnt).getLastWord();
byte[] codes = new byte[w1.getCodeLength()];
int[] indices = new int[w1.getCodeLength()];
double alpha = items.get(cnt).getAlpha();
if (w1 == null || lastWord == null || (lastWord.getIndex() < 0 && inferenceVector == null)
|| w1.getIndex() == lastWord.getIndex() || w1.getLabel().equals("STOP")
|| lastWord.getLabel().equals("STOP") || w1.getLabel().equals("UNK")
|| lastWord.getLabel().equals("UNK")) {
return 0.0;
}
int target = lastWord.getIndex();
int ngStarter = w1.getIndex();
if (configuration.isUseHierarchicSoftmax()) {
for (int i = 0; i < w1.getCodeLength(); i++) {
int code = w1.getCodes().get(i);
int point = w1.getPoints().get(i);
if (point >= vocabCache.numWords() || point < 0)
continue;
if (i < w1.getCodeLength()) {
codes[i] = (byte) code;
indices[i] = point;
}
}
}
//negative sampling
if (negative > 0) {
if (syn1Neg == null) {
((InMemoryLookupTable<T>) lookupTable).initNegative();
syn1Neg = new DeviceLocalNDArray(((InMemoryLookupTable<T>) lookupTable).getSyn1Neg());
}
}
SkipGramInference sg = SkipGramInference.builder()
.inferenceVector(inferenceVector != null ? inferenceVector : Nd4j.empty(syn0.get().dataType()))
.randomValue((int) items.get(0).getRandomValue())
.syn0(syn0.get())
.negTable((negative > 0) ? table.get() : Nd4j.empty(syn0.get().dataType()))
.expTable(expTable.get())
.syn1(configuration.isUseHierarchicSoftmax() ? syn1.get() : Nd4j.empty(syn0.get().dataType()))
.syn1Neg((negative > 0) ? syn1Neg.get() : Nd4j.empty(syn0.get().dataType()))
.negTable((negative > 0) ? table.get() : Nd4j.empty(syn0.get().dataType()))
.alpha(new double[]{alpha})
.iteration(1)
.ngStarter(ngStarter)
.indices(indices)
.target(target)
.codes(codes)
.preciseMode(configuration.getPreciseMode())
.numWorkers(configuration.getWorkers())
.build();
Nd4j.getExecutioner().exec(sg);
items.clear();
}
return 0.0;
}
}
}