use of org.nd4j.parameterserver.distributed.messages.Frame in project deeplearning4j by deeplearning4j.
the class PartitionTrainingFunction method trainAllAtOnce.
protected void trainAllAtOnce(List<Sequence<ShallowSequenceElement>> sequences) {
Frame bigFrame = new Frame(BasicSequenceProvider.getInstance().getNextValue());
for (Sequence<ShallowSequenceElement> sequence : sequences) {
Frame frame = elementsLearningAlgorithm.frameSequence(sequence, new AtomicLong(119L), 25e-3f);
bigFrame.stackMessages(frame.getMessages());
}
if (bigFrame.size() > 0)
paramServer.execDistributed(bigFrame);
}
use of org.nd4j.parameterserver.distributed.messages.Frame in project deeplearning4j by deeplearning4j.
the class SparkCBOW method frameSequence.
@Override
public Frame<? extends TrainingMessage> frameSequence(Sequence<ShallowSequenceElement> sequence, AtomicLong nextRandom, double learningRate) {
// FIXME: totalElementsCount should have real value
if (vectorsConfiguration.getSampling() > 0)
sequence = BaseSparkLearningAlgorithm.applySubsampling(sequence, nextRandom, 10L, vectorsConfiguration.getSampling());
int currentWindow = vectorsConfiguration.getWindow();
if (vectorsConfiguration.getVariableWindows() != null && vectorsConfiguration.getVariableWindows().length != 0) {
currentWindow = vectorsConfiguration.getVariableWindows()[RandomUtils.nextInt(vectorsConfiguration.getVariableWindows().length)];
}
if (frame == null)
synchronized (this) {
if (frame == null)
frame = new ThreadLocal<>();
}
if (frame.get() == null)
frame.set(new Frame<CbowRequestMessage>(BasicSequenceProvider.getInstance().getNextValue()));
for (int i = 0; i < sequence.getElements().size(); i++) {
nextRandom.set(Math.abs(nextRandom.get() * 25214903917L + 11));
int b = (int) nextRandom.get() % currentWindow;
int end = currentWindow * 2 + 1 - b;
ShallowSequenceElement currentWord = sequence.getElementByIndex(i);
List<Integer> intsList = new ArrayList<>();
for (int a = b; a < end; a++) {
if (a != currentWindow) {
int c = i - currentWindow + a;
if (c >= 0 && c < sequence.size()) {
ShallowSequenceElement lastWord = sequence.getElementByIndex(c);
intsList.add(lastWord.getIndex());
}
}
}
// just converting values to int
int[] windowWords = new int[intsList.size()];
for (int x = 0; x < windowWords.length; x++) {
windowWords[x] = intsList.get(x);
}
if (windowWords.length < 1)
continue;
iterateSample(currentWord, windowWords, nextRandom, learningRate, false, 0, true, null);
}
Frame<CbowRequestMessage> currentFrame = frame.get();
frame.set(new Frame<CbowRequestMessage>(BasicSequenceProvider.getInstance().getNextValue()));
return currentFrame;
}
use of org.nd4j.parameterserver.distributed.messages.Frame in project deeplearning4j by deeplearning4j.
the class SparkSkipGram method frameSequence.
@Override
public Frame<? extends TrainingMessage> frameSequence(Sequence<ShallowSequenceElement> sequence, AtomicLong nextRandom, double learningRate) {
// FIXME: totalElementsCount should have real value
if (vectorsConfiguration.getSampling() > 0)
sequence = BaseSparkLearningAlgorithm.applySubsampling(sequence, nextRandom, 10L, vectorsConfiguration.getSampling());
int currentWindow = vectorsConfiguration.getWindow();
if (vectorsConfiguration.getVariableWindows() != null && vectorsConfiguration.getVariableWindows().length != 0) {
currentWindow = vectorsConfiguration.getVariableWindows()[RandomUtils.nextInt(vectorsConfiguration.getVariableWindows().length)];
}
if (frame == null)
synchronized (this) {
if (frame == null)
frame = new ThreadLocal<>();
}
if (frame.get() == null)
frame.set(new Frame<SkipGramRequestMessage>(BasicSequenceProvider.getInstance().getNextValue()));
for (int i = 0; i < sequence.size(); i++) {
nextRandom.set(Math.abs(nextRandom.get() * 25214903917L + 11));
ShallowSequenceElement word = sequence.getElementByIndex(i);
if (word == null)
continue;
int b = (int) (nextRandom.get() % currentWindow);
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 < sequence.size()) {
ShallowSequenceElement lastWord = sequence.getElementByIndex(c);
iterateSample(word, lastWord, nextRandom, learningRate);
nextRandom.set(Math.abs(nextRandom.get() * 25214903917L + 11));
}
}
}
}
// at this moment we should have something in ThreadLocal Frame, so we'll send it to VoidParameterServer for processing
Frame<SkipGramRequestMessage> currentFrame = frame.get();
frame.set(new Frame<SkipGramRequestMessage>(BasicSequenceProvider.getInstance().getNextValue()));
return currentFrame;
}
use of org.nd4j.parameterserver.distributed.messages.Frame in project deeplearning4j by deeplearning4j.
the class SparkDM method frameSequence.
@Override
public Frame<? extends TrainingMessage> frameSequence(Sequence<ShallowSequenceElement> sequence, AtomicLong nextRandom, double learningRate) {
if (vectorsConfiguration.getSampling() > 0)
sequence = BaseSparkLearningAlgorithm.applySubsampling(sequence, nextRandom, 10L, vectorsConfiguration.getSampling());
int currentWindow = vectorsConfiguration.getWindow();
if (vectorsConfiguration.getVariableWindows() != null && vectorsConfiguration.getVariableWindows().length != 0) {
currentWindow = vectorsConfiguration.getVariableWindows()[RandomUtils.nextInt(vectorsConfiguration.getVariableWindows().length)];
}
if (frame == null)
synchronized (this) {
if (frame == null)
frame = new ThreadLocal<>();
}
if (frame.get() == null)
frame.set(new Frame<CbowRequestMessage>(BasicSequenceProvider.getInstance().getNextValue()));
for (int i = 0; i < sequence.getElements().size(); i++) {
nextRandom.set(Math.abs(nextRandom.get() * 25214903917L + 11));
int b = (int) nextRandom.get() % currentWindow;
int end = currentWindow * 2 + 1 - b;
ShallowSequenceElement currentWord = sequence.getElementByIndex(i);
List<Integer> intsList = new ArrayList<>();
for (int a = b; a < end; a++) {
if (a != currentWindow) {
int c = i - currentWindow + a;
if (c >= 0 && c < sequence.size()) {
ShallowSequenceElement lastWord = sequence.getElementByIndex(c);
intsList.add(lastWord.getIndex());
}
}
}
// basically it's the same as CBOW, we just add labels here
if (sequence.getSequenceLabels() != null) {
for (ShallowSequenceElement label : sequence.getSequenceLabels()) {
intsList.add(label.getIndex());
}
} else
// FIXME: we probably should throw this exception earlier?
throw new DL4JInvalidInputException("Sequence passed via RDD has no labels within, nothing to learn here");
// just converting values to int
int[] windowWords = new int[intsList.size()];
for (int x = 0; x < windowWords.length; x++) {
windowWords[x] = intsList.get(x);
}
if (windowWords.length < 1)
continue;
iterateSample(currentWord, windowWords, nextRandom, learningRate, false, 0, true, null);
}
Frame<CbowRequestMessage> currentFrame = frame.get();
frame.set(new Frame<CbowRequestMessage>(BasicSequenceProvider.getInstance().getNextValue()));
return currentFrame;
}
use of org.nd4j.parameterserver.distributed.messages.Frame in project deeplearning4j by deeplearning4j.
the class SparkDBOW method frameSequence.
@Override
public Frame<? extends TrainingMessage> frameSequence(Sequence<ShallowSequenceElement> sequence, AtomicLong nextRandom, double learningRate) {
if (vectorsConfiguration.getSampling() > 0)
sequence = BaseSparkLearningAlgorithm.applySubsampling(sequence, nextRandom, 10L, vectorsConfiguration.getSampling());
int currentWindow = vectorsConfiguration.getWindow();
if (vectorsConfiguration.getVariableWindows() != null && vectorsConfiguration.getVariableWindows().length != 0) {
currentWindow = vectorsConfiguration.getVariableWindows()[RandomUtils.nextInt(vectorsConfiguration.getVariableWindows().length)];
}
if (frame == null)
synchronized (this) {
if (frame == null)
frame = new ThreadLocal<>();
}
if (frame.get() == null)
frame.set(new Frame<SkipGramRequestMessage>(BasicSequenceProvider.getInstance().getNextValue()));
for (ShallowSequenceElement lastWord : sequence.getSequenceLabels()) {
for (ShallowSequenceElement word : sequence.getElements()) {
iterateSample(word, lastWord, nextRandom, learningRate);
nextRandom.set(Math.abs(nextRandom.get() * 25214903917L + 11));
}
}
// at this moment we should have something in ThreadLocal Frame, so we'll send it to VoidParameterServer for processing
Frame<SkipGramRequestMessage> currentFrame = frame.get();
frame.set(new Frame<SkipGramRequestMessage>(BasicSequenceProvider.getInstance().getNextValue()));
return currentFrame;
}
Aggregations