/// <summary>Compute dot product between two vectors.</summary>
public static double Dot(SimpleMatrix vector1, SimpleMatrix vector2)
{
if (vector1.NumRows() == 1)
{
// vector1: row vector, assume that vector2 is a row vector too
return(vector1.Mult(vector2.Transpose()).Get(0));
}
else
{
if (vector1.NumCols() == 1)
{
// vector1: col vector, assume that vector2 is also a column vector.
return(vector1.Transpose().Mult(vector2).Get(0));
}
else
{
throw new AssertionError("Error in neural.Utils.dot: vector1 is a matrix " + vector1.NumRows() + " x " + vector1.NumCols());
}
}
}
/// <summary>
/// Returns a column vector where each entry is the nth bilinear
/// product of the nth slices of the two tensors.
/// </summary>
public virtual SimpleMatrix BilinearProducts(SimpleMatrix @in)
{
if (@in.NumCols() != 1)
{
throw new AssertionError("Expected a column vector");
}
if (@in.NumRows() != numCols)
{
throw new AssertionError("Number of rows in the input does not match number of columns in tensor");
}
if (numRows != numCols)
{
throw new AssertionError("Can only perform this operation on a SimpleTensor with square slices");
}
SimpleMatrix inT = @in.Transpose();
SimpleMatrix @out = new SimpleMatrix(numSlices, 1);
for (int slice = 0; slice < numSlices; ++slice)
{
double result = inT.Mult(slices[slice]).Mult(@in).Get(0);
@out.Set(slice, result);
}
return(@out);
}
public virtual SimpleMatrix GetAntecedentEmbedding(SimpleMatrix mentionEmbedding)
{
return(antecedentMatrix.Mult(mentionEmbedding));
}
public virtual SimpleMatrix GetAnaphorEmbedding(SimpleMatrix mentionEmbedding)
{
return(anaphorMatrix.Mult(mentionEmbedding));
}
public virtual double GetPairwiseScore(SimpleMatrix antecedentEmbedding, SimpleMatrix anaphorEmbedding, SimpleMatrix pairFeatures)
{
SimpleMatrix firstLayerOutput = NeuralUtils.ElementwiseApplyReLU(antecedentEmbedding.Plus(anaphorEmbedding).Plus(pairFeaturesMatrix.Mult(pairFeatures)).Plus(pairwiseFirstLayerBias));
return(Score(firstLayerOutput, pairwiseModel));
}
/// <summary>
/// This is the method to call for assigning labels and node vectors
/// to the Tree.
/// </summary>
/// <remarks>
/// This is the method to call for assigning labels and node vectors
/// to the Tree. After calling this, each of the non-leaf nodes will
/// have the node vector and the predictions of their classes
/// assigned to that subtree's node. The annotations filled in are
/// the RNNCoreAnnotations.NodeVector, Predictions, and
/// PredictedClass. In general, PredictedClass will be the most
/// useful annotation except when training.
/// </remarks>
public virtual void ForwardPropagateTree(Tree tree)
{
SimpleMatrix nodeVector;
// initialized below or Exception thrown // = null;
SimpleMatrix classification;
// initialized below or Exception thrown // = null;
if (tree.IsLeaf())
{
// We do nothing for the leaves. The preterminals will
// calculate the classification for this word/tag. In fact, the
// recursion should not have gotten here (unless there are
// degenerate trees of just one leaf)
log.Info("SentimentCostAndGradient: warning: We reached leaves in forwardPropagate: " + tree);
throw new AssertionError("We should not have reached leaves in forwardPropagate");
}
else
{
if (tree.IsPreTerminal())
{
classification = model.GetUnaryClassification(tree.Label().Value());
string word = tree.Children()[0].Label().Value();
SimpleMatrix wordVector = model.GetWordVector(word);
nodeVector = NeuralUtils.ElementwiseApplyTanh(wordVector);
}
else
{
if (tree.Children().Length == 1)
{
log.Info("SentimentCostAndGradient: warning: Non-preterminal nodes of size 1: " + tree);
throw new AssertionError("Non-preterminal nodes of size 1 should have already been collapsed");
}
else
{
if (tree.Children().Length == 2)
{
ForwardPropagateTree(tree.Children()[0]);
ForwardPropagateTree(tree.Children()[1]);
string leftCategory = tree.Children()[0].Label().Value();
string rightCategory = tree.Children()[1].Label().Value();
SimpleMatrix W = model.GetBinaryTransform(leftCategory, rightCategory);
classification = model.GetBinaryClassification(leftCategory, rightCategory);
SimpleMatrix leftVector = RNNCoreAnnotations.GetNodeVector(tree.Children()[0]);
SimpleMatrix rightVector = RNNCoreAnnotations.GetNodeVector(tree.Children()[1]);
SimpleMatrix childrenVector = NeuralUtils.ConcatenateWithBias(leftVector, rightVector);
if (model.op.useTensors)
{
SimpleTensor tensor = model.GetBinaryTensor(leftCategory, rightCategory);
SimpleMatrix tensorIn = NeuralUtils.Concatenate(leftVector, rightVector);
SimpleMatrix tensorOut = tensor.BilinearProducts(tensorIn);
nodeVector = NeuralUtils.ElementwiseApplyTanh(W.Mult(childrenVector).Plus(tensorOut));
}
else
{
nodeVector = NeuralUtils.ElementwiseApplyTanh(W.Mult(childrenVector));
}
}
else
{
log.Info("SentimentCostAndGradient: warning: Tree not correctly binarized: " + tree);
throw new AssertionError("Tree not correctly binarized");
}
}
}
}
SimpleMatrix predictions = NeuralUtils.Softmax(classification.Mult(NeuralUtils.ConcatenateWithBias(nodeVector)));
int index = GetPredictedClass(predictions);
if (!(tree.Label() is CoreLabel))
{
log.Info("SentimentCostAndGradient: warning: No CoreLabels in nodes: " + tree);
throw new AssertionError("Expected CoreLabels in the nodes");
}
CoreLabel label = (CoreLabel)tree.Label();
label.Set(typeof(RNNCoreAnnotations.Predictions), predictions);
label.Set(typeof(RNNCoreAnnotations.PredictedClass), index);
label.Set(typeof(RNNCoreAnnotations.NodeVector), nodeVector);
}
private void BackpropDerivativesAndError(Tree tree, TwoDimensionalMap <string, string, SimpleMatrix> binaryTD, TwoDimensionalMap <string, string, SimpleMatrix> binaryCD, TwoDimensionalMap <string, string, SimpleTensor> binaryTensorTD, IDictionary
<string, SimpleMatrix> unaryCD, IDictionary <string, SimpleMatrix> wordVectorD, SimpleMatrix deltaUp)
{
if (tree.IsLeaf())
{
return;
}
SimpleMatrix currentVector = RNNCoreAnnotations.GetNodeVector(tree);
string category = tree.Label().Value();
category = model.BasicCategory(category);
// Build a vector that looks like 0,0,1,0,0 with an indicator for the correct class
SimpleMatrix goldLabel = new SimpleMatrix(model.numClasses, 1);
int goldClass = RNNCoreAnnotations.GetGoldClass(tree);
if (goldClass >= 0)
{
goldLabel.Set(goldClass, 1.0);
}
double nodeWeight = model.op.trainOptions.GetClassWeight(goldClass);
SimpleMatrix predictions = RNNCoreAnnotations.GetPredictions(tree);
// 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
SimpleMatrix deltaClass = goldClass >= 0 ? predictions.Minus(goldLabel).Scale(nodeWeight) : new SimpleMatrix(predictions.NumRows(), predictions.NumCols());
SimpleMatrix localCD = deltaClass.Mult(NeuralUtils.ConcatenateWithBias(currentVector).Transpose());
double error = -(NeuralUtils.ElementwiseApplyLog(predictions).ElementMult(goldLabel).ElementSum());
error = error * nodeWeight;
RNNCoreAnnotations.SetPredictionError(tree, error);
if (tree.IsPreTerminal())
{
// below us is a word vector
unaryCD[category] = unaryCD[category].Plus(localCD);
string word = tree.Children()[0].Label().Value();
word = model.GetVocabWord(word);
//SimpleMatrix currentVectorDerivative = NeuralUtils.elementwiseApplyTanhDerivative(currentVector);
//SimpleMatrix deltaFromClass = model.getUnaryClassification(category).transpose().mult(deltaClass);
//SimpleMatrix deltaFull = deltaFromClass.extractMatrix(0, model.op.numHid, 0, 1).plus(deltaUp);
//SimpleMatrix wordDerivative = deltaFull.elementMult(currentVectorDerivative);
//wordVectorD.put(word, wordVectorD.get(word).plus(wordDerivative));
SimpleMatrix currentVectorDerivative = NeuralUtils.ElementwiseApplyTanhDerivative(currentVector);
SimpleMatrix deltaFromClass = model.GetUnaryClassification(category).Transpose().Mult(deltaClass);
deltaFromClass = deltaFromClass.ExtractMatrix(0, model.op.numHid, 0, 1).ElementMult(currentVectorDerivative);
SimpleMatrix deltaFull = deltaFromClass.Plus(deltaUp);
SimpleMatrix oldWordVectorD = wordVectorD[word];
if (oldWordVectorD == null)
{
wordVectorD[word] = deltaFull;
}
else
{
wordVectorD[word] = oldWordVectorD.Plus(deltaFull);
}
}
else
{
// Otherwise, this must be a binary node
string leftCategory = model.BasicCategory(tree.Children()[0].Label().Value());
string rightCategory = model.BasicCategory(tree.Children()[1].Label().Value());
if (model.op.combineClassification)
{
unaryCD[string.Empty] = unaryCD[string.Empty].Plus(localCD);
}
else
{
binaryCD.Put(leftCategory, rightCategory, binaryCD.Get(leftCategory, rightCategory).Plus(localCD));
}
SimpleMatrix currentVectorDerivative = NeuralUtils.ElementwiseApplyTanhDerivative(currentVector);
SimpleMatrix deltaFromClass = model.GetBinaryClassification(leftCategory, rightCategory).Transpose().Mult(deltaClass);
deltaFromClass = deltaFromClass.ExtractMatrix(0, model.op.numHid, 0, 1).ElementMult(currentVectorDerivative);
SimpleMatrix deltaFull = deltaFromClass.Plus(deltaUp);
SimpleMatrix leftVector = RNNCoreAnnotations.GetNodeVector(tree.Children()[0]);
SimpleMatrix rightVector = RNNCoreAnnotations.GetNodeVector(tree.Children()[1]);
SimpleMatrix childrenVector = NeuralUtils.ConcatenateWithBias(leftVector, rightVector);
SimpleMatrix W_df = deltaFull.Mult(childrenVector.Transpose());
binaryTD.Put(leftCategory, rightCategory, binaryTD.Get(leftCategory, rightCategory).Plus(W_df));
SimpleMatrix deltaDown;
if (model.op.useTensors)
{
SimpleTensor Wt_df = GetTensorGradient(deltaFull, leftVector, rightVector);
binaryTensorTD.Put(leftCategory, rightCategory, binaryTensorTD.Get(leftCategory, rightCategory).Plus(Wt_df));
deltaDown = ComputeTensorDeltaDown(deltaFull, leftVector, rightVector, model.GetBinaryTransform(leftCategory, rightCategory), model.GetBinaryTensor(leftCategory, rightCategory));
}
else
{
deltaDown = model.GetBinaryTransform(leftCategory, rightCategory).Transpose().Mult(deltaFull);
}
SimpleMatrix leftDerivative = NeuralUtils.ElementwiseApplyTanhDerivative(leftVector);
SimpleMatrix rightDerivative = NeuralUtils.ElementwiseApplyTanhDerivative(rightVector);
SimpleMatrix leftDeltaDown = deltaDown.ExtractMatrix(0, deltaFull.NumRows(), 0, 1);
SimpleMatrix rightDeltaDown = deltaDown.ExtractMatrix(deltaFull.NumRows(), deltaFull.NumRows() * 2, 0, 1);
BackpropDerivativesAndError(tree.Children()[0], binaryTD, binaryCD, binaryTensorTD, unaryCD, wordVectorD, leftDerivative.ElementMult(leftDeltaDown));
BackpropDerivativesAndError(tree.Children()[1], binaryTD, binaryCD, binaryTensorTD, unaryCD, wordVectorD, rightDerivative.ElementMult(rightDeltaDown));
}
}
public virtual void BackpropDerivative(Tree tree, IList <string> words, IdentityHashMap <Tree, SimpleMatrix> nodeVectors, TwoDimensionalMap <string, string, SimpleMatrix> binaryW_dfs, IDictionary <string, SimpleMatrix> unaryW_dfs, TwoDimensionalMap
<string, string, SimpleMatrix> binaryScoreDerivatives, IDictionary <string, SimpleMatrix> unaryScoreDerivatives, IDictionary <string, SimpleMatrix> wordVectorDerivatives, SimpleMatrix deltaUp)
{
if (tree.IsLeaf())
{
return;
}
if (tree.IsPreTerminal())
{
if (op.trainOptions.trainWordVectors)
{
string word = tree.Children()[0].Label().Value();
word = dvModel.GetVocabWord(word);
// SimpleMatrix currentVector = nodeVectors.get(tree);
// SimpleMatrix currentVectorDerivative = nonlinearityVectorToDerivative(currentVector);
// SimpleMatrix derivative = deltaUp.elementMult(currentVectorDerivative);
SimpleMatrix derivative = deltaUp;
wordVectorDerivatives[word] = wordVectorDerivatives[word].Plus(derivative);
}
return;
}
SimpleMatrix currentVector = nodeVectors[tree];
SimpleMatrix currentVectorDerivative = NeuralUtils.ElementwiseApplyTanhDerivative(currentVector);
SimpleMatrix scoreW = dvModel.GetScoreWForNode(tree);
currentVectorDerivative = currentVectorDerivative.ElementMult(scoreW.Transpose());
// the delta that is used at the current nodes
SimpleMatrix deltaCurrent = deltaUp.Plus(currentVectorDerivative);
SimpleMatrix W = dvModel.GetWForNode(tree);
SimpleMatrix WTdelta = W.Transpose().Mult(deltaCurrent);
if (tree.Children().Length == 2)
{
//TODO: RS: Change to the nice "getWForNode" setup?
string leftLabel = dvModel.BasicCategory(tree.Children()[0].Label().Value());
string rightLabel = dvModel.BasicCategory(tree.Children()[1].Label().Value());
binaryScoreDerivatives.Put(leftLabel, rightLabel, binaryScoreDerivatives.Get(leftLabel, rightLabel).Plus(currentVector.Transpose()));
SimpleMatrix leftVector = nodeVectors[tree.Children()[0]];
SimpleMatrix rightVector = nodeVectors[tree.Children()[1]];
SimpleMatrix childrenVector = NeuralUtils.ConcatenateWithBias(leftVector, rightVector);
if (op.trainOptions.useContextWords)
{
childrenVector = ConcatenateContextWords(childrenVector, tree.GetSpan(), words);
}
SimpleMatrix W_df = deltaCurrent.Mult(childrenVector.Transpose());
binaryW_dfs.Put(leftLabel, rightLabel, binaryW_dfs.Get(leftLabel, rightLabel).Plus(W_df));
// and then recurse
SimpleMatrix leftDerivative = NeuralUtils.ElementwiseApplyTanhDerivative(leftVector);
SimpleMatrix rightDerivative = NeuralUtils.ElementwiseApplyTanhDerivative(rightVector);
SimpleMatrix leftWTDelta = WTdelta.ExtractMatrix(0, deltaCurrent.NumRows(), 0, 1);
SimpleMatrix rightWTDelta = WTdelta.ExtractMatrix(deltaCurrent.NumRows(), deltaCurrent.NumRows() * 2, 0, 1);
BackpropDerivative(tree.Children()[0], words, nodeVectors, binaryW_dfs, unaryW_dfs, binaryScoreDerivatives, unaryScoreDerivatives, wordVectorDerivatives, leftDerivative.ElementMult(leftWTDelta));
BackpropDerivative(tree.Children()[1], words, nodeVectors, binaryW_dfs, unaryW_dfs, binaryScoreDerivatives, unaryScoreDerivatives, wordVectorDerivatives, rightDerivative.ElementMult(rightWTDelta));
}
else
{
if (tree.Children().Length == 1)
{
string childLabel = dvModel.BasicCategory(tree.Children()[0].Label().Value());
unaryScoreDerivatives[childLabel] = unaryScoreDerivatives[childLabel].Plus(currentVector.Transpose());
SimpleMatrix childVector = nodeVectors[tree.Children()[0]];
SimpleMatrix childVectorWithBias = NeuralUtils.ConcatenateWithBias(childVector);
if (op.trainOptions.useContextWords)
{
childVectorWithBias = ConcatenateContextWords(childVectorWithBias, tree.GetSpan(), words);
}
SimpleMatrix W_df = deltaCurrent.Mult(childVectorWithBias.Transpose());
// System.out.println("unary backprop derivative for " + childLabel);
// System.out.println("Old transform:");
// System.out.println(unaryW_dfs.get(childLabel));
// System.out.println(" Delta:");
// System.out.println(W_df.scale(scale));
unaryW_dfs[childLabel] = unaryW_dfs[childLabel].Plus(W_df);
// and then recurse
SimpleMatrix childDerivative = NeuralUtils.ElementwiseApplyTanhDerivative(childVector);
//SimpleMatrix childDerivative = childVector;
SimpleMatrix childWTDelta = WTdelta.ExtractMatrix(0, deltaCurrent.NumRows(), 0, 1);
BackpropDerivative(tree.Children()[0], words, nodeVectors, binaryW_dfs, unaryW_dfs, binaryScoreDerivatives, unaryScoreDerivatives, wordVectorDerivatives, childDerivative.ElementMult(childWTDelta));
}
}
}
private void ForwardPropagateTree(Tree tree, IList <string> words, IdentityHashMap <Tree, SimpleMatrix> nodeVectors, IdentityHashMap <Tree, double> scores)
{
if (tree.IsLeaf())
{
return;
}
if (tree.IsPreTerminal())
{
Tree wordNode = tree.Children()[0];
string word = wordNode.Label().Value();
SimpleMatrix wordVector = dvModel.GetWordVector(word);
wordVector = NeuralUtils.ElementwiseApplyTanh(wordVector);
nodeVectors[tree] = wordVector;
return;
}
foreach (Tree child in tree.Children())
{
ForwardPropagateTree(child, words, nodeVectors, scores);
}
// at this point, nodeVectors contains the vectors for all of
// the children of tree
SimpleMatrix childVec;
if (tree.Children().Length == 2)
{
childVec = NeuralUtils.ConcatenateWithBias(nodeVectors[tree.Children()[0]], nodeVectors[tree.Children()[1]]);
}
else
{
childVec = NeuralUtils.ConcatenateWithBias(nodeVectors[tree.Children()[0]]);
}
if (op.trainOptions.useContextWords)
{
childVec = ConcatenateContextWords(childVec, tree.GetSpan(), words);
}
SimpleMatrix W = dvModel.GetWForNode(tree);
if (W == null)
{
string error = "Could not find W for tree " + tree;
if (op.testOptions.verbose)
{
log.Info(error);
}
throw new NoSuchParseException(error);
}
SimpleMatrix currentVector = W.Mult(childVec);
currentVector = NeuralUtils.ElementwiseApplyTanh(currentVector);
nodeVectors[tree] = currentVector;
SimpleMatrix scoreW = dvModel.GetScoreWForNode(tree);
if (scoreW == null)
{
string error = "Could not find scoreW for tree " + tree;
if (op.testOptions.verbose)
{
log.Info(error);
}
throw new NoSuchParseException(error);
}
double score = scoreW.Dot(currentVector);
//score = NeuralUtils.sigmoid(score);
scores[tree] = score;
}
