public FloatMatrix getValueGradient(int iterations) {
// We use TreeMap for each of these so that they stay in a
// canonical sorted order
// TODO: factor out the initialization routines
// binaryTD stands for Transform Derivatives
final MultiDimensionalMap<String, String, FloatMatrix> binaryTD =
MultiDimensionalMap.newTreeBackedMap();
// the derivatives of the FloatTensors for the binary nodes
final MultiDimensionalMap<String, String, FloatTensor> binaryFloatTensorTD =
MultiDimensionalMap.newTreeBackedMap();
// binaryCD stands for Classification Derivatives
final MultiDimensionalMap<String, String, FloatMatrix> binaryCD =
MultiDimensionalMap.newTreeBackedMap();
// unaryCD stands for Classification Derivatives
final Map<String, FloatMatrix> unaryCD = new TreeMap<>();
// word vector derivatives
final Map<String, FloatMatrix> wordVectorD = new TreeMap<>();
for (MultiDimensionalMap.Entry<String, String, FloatMatrix> entry :
binaryTransform.entrySet()) {
int numRows = entry.getValue().rows;
int numCols = entry.getValue().columns;
binaryTD.put(entry.getFirstKey(), entry.getSecondKey(), new FloatMatrix(numRows, numCols));
}
if (!combineClassification) {
for (MultiDimensionalMap.Entry<String, String, FloatMatrix> entry :
binaryClassification.entrySet()) {
int numRows = entry.getValue().rows;
int numCols = entry.getValue().columns;
binaryCD.put(entry.getFirstKey(), entry.getSecondKey(), new FloatMatrix(numRows, numCols));
}
}
if (useFloatTensors) {
for (MultiDimensionalMap.Entry<String, String, FloatTensor> entry :
binaryFloatTensors.entrySet()) {
int numRows = entry.getValue().rows();
int numCols = entry.getValue().columns;
int numSlices = entry.getValue().slices();
binaryFloatTensorTD.put(
entry.getFirstKey(),
entry.getSecondKey(),
new FloatTensor(numRows, numCols, numSlices));
}
}
for (Map.Entry<String, FloatMatrix> entry : unaryClassification.entrySet()) {
int numRows = entry.getValue().rows;
int numCols = entry.getValue().columns;
unaryCD.put(entry.getKey(), new FloatMatrix(numRows, numCols));
}
for (Map.Entry<String, FloatMatrix> entry : featureVectors.entrySet()) {
int numRows = entry.getValue().rows;
int numCols = entry.getValue().columns;
wordVectorD.put(entry.getKey(), new FloatMatrix(numRows, numCols));
}
final List<Tree> forwardPropTrees = new CopyOnWriteArrayList<>();
Parallelization.iterateInParallel(
trainingTrees,
new Parallelization.RunnableWithParams<Tree>() {
public void run(Tree currentItem, Object[] args) {
Tree trainingTree = new Tree(currentItem);
trainingTree.connect(new ArrayList<>(currentItem.children()));
// this will attach the error vectors and the node vectors
// to each node in the tree
forwardPropagateTree(trainingTree);
forwardPropTrees.add(trainingTree);
}
},
rnTnActorSystem);
// TODO: we may find a big speedup by separating the derivatives and then summing
final AtomicDouble error = new AtomicDouble(0);
Parallelization.iterateInParallel(
forwardPropTrees,
new Parallelization.RunnableWithParams<Tree>() {
public void run(Tree currentItem, Object[] args) {
backpropDerivativesAndError(
currentItem, binaryTD, binaryCD, binaryFloatTensorTD, unaryCD, wordVectorD);
error.addAndGet(currentItem.errorSum());
}
},
new Parallelization.RunnableWithParams<Tree>() {
public void run(Tree currentItem, Object[] args) {}
},
rnTnActorSystem,
new Object[] {binaryTD, binaryCD, binaryFloatTensorTD, unaryCD, wordVectorD});
// scale the error by the number of sentences so that the
// regularization isn't drowned out for large training batchs
float scale = (1.0f / trainingTrees.size());
value = error.floatValue() * scale;
value += scaleAndRegularize(binaryTD, binaryTransform, scale, regTransformMatrix);
value += scaleAndRegularize(binaryCD, binaryClassification, scale, regClassification);
value +=
scaleAndRegularizeFloatTensor(
binaryFloatTensorTD, binaryFloatTensors, scale, regTransformFloatTensor);
value += scaleAndRegularize(unaryCD, unaryClassification, scale, regClassification);
value += scaleAndRegularize(wordVectorD, featureVectors, scale, regWordVector);
FloatMatrix derivative =
MatrixUtil.toFlattenedFloat(
getNumParameters(),
binaryTD.values().iterator(),
binaryCD.values().iterator(),
binaryFloatTensorTD.values().iterator(),
unaryCD.values().iterator(),
wordVectorD.values().iterator());
if (paramAdaGrad == null) paramAdaGrad = new AdaGradFloat(1, derivative.columns);
derivative.muli(paramAdaGrad.getLearningRates(derivative));
return derivative;
}
private void backpropDerivativesAndError(
Tree tree,
MultiDimensionalMap<String, String, FloatMatrix> binaryTD,
MultiDimensionalMap<String, String, FloatMatrix> binaryCD,
MultiDimensionalMap<String, String, FloatTensor> binaryFloatTensorTD,
Map<String, FloatMatrix> unaryCD,
Map<String, FloatMatrix> wordVectorD,
FloatMatrix deltaUp) {
if (tree.isLeaf()) {
return;
}
FloatMatrix currentVector = tree.vector();
String category = tree.label();
category = basicCategory(category);
// Build a vector that looks like 0,0,1,0,0 with an indicator for the correct class
FloatMatrix goldLabel = new FloatMatrix(numOuts, 1);
int goldClass = tree.goldLabel();
if (goldClass >= 0) {
goldLabel.put(goldClass, 1.0f);
}
Float nodeWeight = classWeights.get(goldClass);
if (nodeWeight == null) nodeWeight = 1.0f;
FloatMatrix predictions = tree.prediction();
// If this is an unlabeled class, set deltaClass to 0. We could
// make this more efficient by eliminating various of the below
// calculations, but this would be the easiest way to handle the
// unlabeled class
FloatMatrix deltaClass =
goldClass >= 0
? SimpleBlas.scal(nodeWeight, predictions.sub(goldLabel))
: new FloatMatrix(predictions.rows, predictions.columns);
FloatMatrix localCD = deltaClass.mmul(appendBias(currentVector).transpose());
float error = -(MatrixFunctions.log(predictions).muli(goldLabel).sum());
error = error * nodeWeight;
tree.setError(error);
if (tree.isPreTerminal()) { // below us is a word vector
unaryCD.put(category, unaryCD.get(category).add(localCD));
String word = tree.children().get(0).label();
word = getVocabWord(word);
FloatMatrix currentVectorDerivative = activationFunction.apply(currentVector);
FloatMatrix deltaFromClass = getUnaryClassification(category).transpose().mmul(deltaClass);
deltaFromClass =
deltaFromClass.get(interval(0, numHidden), interval(0, 1)).mul(currentVectorDerivative);
FloatMatrix deltaFull = deltaFromClass.add(deltaUp);
wordVectorD.put(word, wordVectorD.get(word).add(deltaFull));
} else {
// Otherwise, this must be a binary node
String leftCategory = basicCategory(tree.children().get(0).label());
String rightCategory = basicCategory(tree.children().get(1).label());
if (combineClassification) {
unaryCD.put("", unaryCD.get("").add(localCD));
} else {
binaryCD.put(
leftCategory, rightCategory, binaryCD.get(leftCategory, rightCategory).add(localCD));
}
FloatMatrix currentVectorDerivative = activationFunction.applyDerivative(currentVector);
FloatMatrix deltaFromClass =
getBinaryClassification(leftCategory, rightCategory).transpose().mmul(deltaClass);
FloatMatrix mult = deltaFromClass.get(interval(0, numHidden), interval(0, 1));
deltaFromClass = mult.muli(currentVectorDerivative);
FloatMatrix deltaFull = deltaFromClass.add(deltaUp);
FloatMatrix leftVector = tree.children().get(0).vector();
FloatMatrix rightVector = tree.children().get(1).vector();
FloatMatrix childrenVector = appendBias(leftVector, rightVector);
// deltaFull 50 x 1, childrenVector: 50 x 2
FloatMatrix add = binaryTD.get(leftCategory, rightCategory);
FloatMatrix W_df = deltaFromClass.mmul(childrenVector.transpose());
binaryTD.put(leftCategory, rightCategory, add.add(W_df));
FloatMatrix deltaDown;
if (useFloatTensors) {
FloatTensor Wt_df = getFloatTensorGradient(deltaFull, leftVector, rightVector);
binaryFloatTensorTD.put(
leftCategory,
rightCategory,
binaryFloatTensorTD.get(leftCategory, rightCategory).add(Wt_df));
deltaDown =
computeFloatTensorDeltaDown(
deltaFull,
leftVector,
rightVector,
getBinaryTransform(leftCategory, rightCategory),
getBinaryFloatTensor(leftCategory, rightCategory));
} else {
deltaDown = getBinaryTransform(leftCategory, rightCategory).transpose().mmul(deltaFull);
}
FloatMatrix leftDerivative = activationFunction.apply(leftVector);
FloatMatrix rightDerivative = activationFunction.apply(rightVector);
FloatMatrix leftDeltaDown = deltaDown.get(interval(0, deltaFull.rows), interval(0, 1));
FloatMatrix rightDeltaDown =
deltaDown.get(interval(deltaFull.rows, deltaFull.rows * 2), interval(0, 1));
backpropDerivativesAndError(
tree.children().get(0),
binaryTD,
binaryCD,
binaryFloatTensorTD,
unaryCD,
wordVectorD,
leftDerivative.mul(leftDeltaDown));
backpropDerivativesAndError(
tree.children().get(1),
binaryTD,
binaryCD,
binaryFloatTensorTD,
unaryCD,
wordVectorD,
rightDerivative.mul(rightDeltaDown));
}
}
