/**
* Discrete fourier transform 2d
*
* @param input the input to transform
* @param rows the number of rows in the transformed output matrix
* @param cols the number of columns in the transformed output matrix
* @return the discrete fourier transform of the input
*/
public static ComplexFloatMatrix complexDisceteFourierTransform(
FloatMatrix input, int rows, int cols) {
ComplexFloatMatrix base;
// pad
if (input.rows < rows || input.columns < cols)
base = MatrixUtil.complexPadWithZeros(input, rows, cols);
// truncation
else if (input.rows > rows || input.columns > cols) {
base = new ComplexFloatMatrix(input);
base =
base.get(
MatrixUtil.toIndices(RangeUtils.interval(0, rows)),
MatrixUtil.toIndices(RangeUtils.interval(0, cols)));
} else base = new ComplexFloatMatrix(input);
ComplexFloatMatrix temp = new ComplexFloatMatrix(base.rows, base.columns);
ComplexFloatMatrix ret = new ComplexFloatMatrix(base.rows, base.columns);
for (int i = 0; i < base.columns; i++) {
ComplexFloatMatrix column = base.getColumn(i);
temp.putColumn(i, complexDiscreteFourierTransform1d(column));
}
for (int i = 0; i < ret.rows; i++) {
ComplexFloatMatrix row = temp.getRow(i);
ret.putRow(i, complexDiscreteFourierTransform1d(row));
}
return ret;
}
/**
* 1d inverse discrete fourier transform see matlab's fft2 for more examples. Note that this will
* throw an exception if the input isn't a vector
*
* @param inputC the input to transform
* @return the inverse fourier transform of the passed in input
*/
public static ComplexNDArray complexInverseDisceteFourierTransform1d(ComplexNDArray inputC) {
if (inputC.shape().length != 1)
throw new IllegalArgumentException("Illegal input: Must be a vector");
double len = MatrixUtil.length(inputC);
ComplexDouble c2 = new ComplexDouble(0, -2).muli(FastMath.PI).divi(len);
ComplexDoubleMatrix range = MatrixUtil.complexRangeVector(0, len);
ComplexDoubleMatrix div2 = range.transpose().mul(c2);
ComplexDoubleMatrix div3 = range.mmul(div2).negi();
ComplexDoubleMatrix matrix = exp(div3).div(len);
ComplexDoubleMatrix complexRet = matrix.mmul(inputC);
return ComplexNDArray.wrap(inputC, complexRet);
}
/**
* 1d inverse discrete fourier transform see matlab's fft2 for more examples. Note that this will
* throw an exception if the input isn't a vector
*
* @param inputC the input to transform
* @return the inverse fourier transform of the passed in input
*/
public static ComplexFloatMatrix complexInverseDisceteFourierTransform1d(
ComplexFloatMatrix inputC) {
if (inputC.rows != 1 && inputC.columns != 1)
throw new IllegalArgumentException("Illegal input: Must be a vector");
float len = MatrixUtil.length(inputC);
ComplexFloat c2 = new ComplexFloat(0, -2).muli((float) FastMath.PI).divi(len);
ComplexFloatMatrix range = MatrixUtil.complexRangeVector(0, (int) len);
ComplexFloatMatrix div2 = range.transpose().mul(c2);
ComplexFloatMatrix div3 = range.mmul(div2).negi();
ComplexFloatMatrix matrix = exp(div3).div(len);
ComplexFloatMatrix complexRet =
inputC.isRowVector() ? matrix.mmul(inputC) : inputC.mmul(matrix);
return complexRet;
}
/**
* Discrete fourier transform 2d
*
* @param input the input to transform
* @param shape the shape of the output matrix
* @return the discrete fourier transform of the input
*/
public static ComplexDoubleMatrix complexDisceteFourierTransform(NDArray input, int[] shape) {
ComplexNDArray base;
// pad
if (ArrayUtil.anyLess(input.shape(), shape))
base = MatrixUtil.complexPadWithZeros(input, shape);
// truncation
else if (ArrayUtil.anyMore(input.shape(), shape)) {
base = new ComplexNDArray(shape);
for (int i = 0; i < ArrayUtil.prod(shape); i++) base.put(i, input.get(i));
} else base = new ComplexNDArray(input);
ComplexNDArray temp = new ComplexNDArray(shape);
ComplexNDArray ret = new ComplexNDArray(shape);
for (int i = 0; i < base.columns; i++) {
ComplexDoubleMatrix column = base.getColumn(i);
temp.putColumn(i, complexDiscreteFourierTransform1d(column));
}
for (int i = 0; i < ret.rows; i++) {
ComplexDoubleMatrix row = temp.getRow(i);
ret.putRow(i, complexDiscreteFourierTransform1d(row));
}
return ret;
}
/**
* Strips the dataset down to the specified labels and remaps them
*
* @param labels the labels to strip down to
*/
public void filterAndStrip(int[] labels) {
FloatDataSet filtered = filterBy(labels);
List<Integer> newLabels = new ArrayList<>();
// map new labels to index according to passed in labels
Map<Integer, Integer> labelMap = new HashMap<>();
for (int i = 0; i < labels.length; i++) labelMap.put(labels[i], i);
// map examples
for (int i = 0; i < filtered.numExamples(); i++) {
int o2 = filtered.get(i).outcome();
int outcome = labelMap.get(o2);
newLabels.add(outcome);
}
FloatMatrix newLabelMatrix = new FloatMatrix(filtered.numExamples(), labels.length);
if (newLabelMatrix.rows != newLabels.size())
throw new IllegalStateException("Inconsistent label sizes");
for (int i = 0; i < newLabelMatrix.rows; i++) {
Integer i2 = newLabels.get(i);
if (i2 == null) throw new IllegalStateException("Label not found on row " + i);
FloatMatrix newRow = MatrixUtil.toOutcomeVectorFloat(i2, labels.length);
newLabelMatrix.putRow(i, newRow);
}
setFirst(filtered.getFirst());
setSecond(newLabelMatrix);
}
/**
* Sets the outcome of a particular example
*
* @param example the example to applyTransformToDestination
* @param label the label of the outcome
*/
public void setOutcome(int example, int label) {
if (example > numExamples()) throw new IllegalArgumentException("No example at " + example);
if (label > numOutcomes() || label < 0) throw new IllegalArgumentException("Illegal label");
FloatMatrix outcome = MatrixUtil.toOutcomeVectorFloat(label, numOutcomes());
getSecond().putRow(example, outcome);
}
/** Returns matrices of the right size for either binary or unary (terminal) classification */
FloatMatrix randomClassificationMatrix() {
// Leave the bias column with 0 values
float range = 1.0f / (float) (Math.sqrt((float) numHidden));
FloatMatrix ret = FloatMatrix.zeros(numOuts, numHidden + 1);
FloatMatrix insert = MatrixUtil.rand(numOuts, numHidden, -range, range, rng);
ret.put(interval(0, numOuts), interval(0, numHidden), insert);
return SimpleBlas.scal(scalingForInit, ret);
}
public void normalizeZeroMeanZeroUnitVariance() {
FloatMatrix columnMeans = getFirst().columnMeans();
FloatMatrix columnStds = MatrixUtil.columnStdDeviation(getFirst());
setFirst(getFirst().subiRowVector(columnMeans));
columnStds.addi(1e-6f);
setFirst(getFirst().diviRowVector(columnStds));
}
public FloatMatrix getParameters() {
return MatrixUtil.toFlattenedFloat(
getNumParameters(),
binaryTransform.values().iterator(),
binaryClassification.values().iterator(),
binaryFloatTensors.values().iterator(),
unaryClassification.values().iterator(),
featureVectors.values().iterator());
}
/**
* 1d discrete fourier transform, note that this will throw an exception if the passed in input
* isn't a vector. See matlab's fft2 for more information
*
* @param inputC the input to transform
* @return the the discrete fourier transform of the passed in input
*/
public static ComplexFloatMatrix complexDiscreteFourierTransform1d(ComplexFloatMatrix inputC) {
if (inputC.rows != 1 && inputC.columns != 1)
throw new IllegalArgumentException("Illegal input: Must be a vector");
float len = Math.max(inputC.rows, inputC.columns);
ComplexFloat c2 = new ComplexFloat(0, -2).muli((float) FastMath.PI).divi(len);
ComplexFloatMatrix range = MatrixUtil.complexRangeVector(0, len);
ComplexFloatMatrix matrix = exp(range.mmul(range.transpose().mul(c2)));
ComplexFloatMatrix complexRet =
inputC.isRowVector() ? matrix.mmul(inputC) : inputC.mmul(matrix);
return complexRet;
}
/**
* Vectorizes the passed in text treating it as one document
*
* @param text the text to vectorize
* @param label the label of the text
* @return a dataset with a applyTransformToDestination of weights(relative to impl; could be word
* counts or tfidf scores)
*/
@Override
public DataSet vectorize(String text, String label) {
Tokenizer tokenizer = tokenizerFactory.create(text);
List<String> tokens = tokenizer.getTokens();
DoubleMatrix input = new DoubleMatrix(1, vocab.size());
for (int i = 0; i < tokens.size(); i++) {
int idx = vocab.indexOf(tokens.get(i));
if (vocab.indexOf(tokens.get(i)) >= 0) input.put(idx, wordCounts.getCount(tokens.get(i)));
}
DoubleMatrix labelMatrix = MatrixUtil.toOutcomeVector(labels.indexOf(label), labels.size());
return new DataSet(input, labelMatrix);
}
static FloatMatrix randomWordVector(int size, RandomGenerator rng) {
return MatrixUtil.uniformFloat(rng, 1, size);
}
FloatMatrix randomTransformBlock() {
float range = 1.0f / (float) (Math.sqrt((float) numHidden) * 2.0f);
FloatMatrix ret = MatrixUtil.rand(numHidden, numHidden, -range, range, rng).add(identity);
return ret;
}
/** Divides the input data applyTransformToDestination by the max number in each row */
public void scale() {
MatrixUtil.scaleByMax(getFirst());
}
public void normalize() {
MatrixUtil.normalizeMatrix(getFirst());
}
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;
}
public void roundInputToTheNearest(int numDecimalPlaces) {
setFirst(MatrixUtil.roundToTheNearest(getFirst(), numDecimalPlaces));
}
