bool KMeansFeatures::train_(MatrixDouble &trainingData){
if( !initialized ){
errorLog << "train_(MatrixDouble &trainingData) - The quantizer has not been initialized!" << endl;
return false;
}
//Reset any previous model
featureDataReady = false;
const UINT M = trainingData.getNumRows();
const UINT N = trainingData.getNumCols();
numInputDimensions = N;
numOutputDimensions = numClustersPerLayer[ numClustersPerLayer.size()-1 ];
//Scale the input data if needed
ranges = trainingData.getRanges();
if( useScaling ){
for(UINT i=0; i<M; i++){
for(UINT j=0; j<N; j++){
trainingData[i][j] = scale(trainingData[i][j],ranges[j].minValue,ranges[j].maxValue,0,1.0);
}
}
}
//Train the KMeans model at each layer
const UINT K = (UINT)numClustersPerLayer.size();
for(UINT k=0; k<K; k++){
KMeans kmeans;
kmeans.setNumClusters( numClustersPerLayer[k] );
kmeans.setComputeTheta( true );
kmeans.setMinChange( minChange );
kmeans.setMinNumEpochs( minNumEpochs );
kmeans.setMaxNumEpochs( maxNumEpochs );
trainingLog << "Layer " << k+1 << "/" << K << " NumClusters: " << numClustersPerLayer[k] << endl;
if( !kmeans.train_( trainingData ) ){
errorLog << "train_(MatrixDouble &trainingData) - Failed to train kmeans model at layer: " << k << endl;
return false;
}
//Save the clusters
clusters.push_back( kmeans.getClusters() );
//Project the data through the current layer to use as training data for the next layer
if( k+1 != K ){
MatrixDouble data( M, numClustersPerLayer[k] );
VectorDouble input( trainingData.getNumCols() );
VectorDouble output( data.getNumCols() );
for(UINT i=0; i<M; i++){
//Copy the data into the sample
for(UINT j=0; j<input.size(); j++){
input[j] = trainingData[i][j];
}
//Project the sample through the current layer
if( !projectDataThroughLayer( input, output, k ) ){
errorLog << "train_(MatrixDouble &trainingData) - Failed to project sample through layer: " << k << endl;
return false;
}
//Copy the result into the training data for the next layer
for(UINT j=0; j<output.size(); j++){
data[i][j] = output[j];
}
}
//Swap the data for the next layer
trainingData = data;
}
}
//Flag that the kmeans model has been trained
trained = true;
featureVector.resize( numOutputDimensions, 0 );
return true;
}
int main (int argc, const char * argv[])
{
//Create an empty matrix double
MatrixDouble matrix;
//Resize the matrix
matrix.resize( 100, 2 );
//Set all the values in the matrix to zero
matrix.setAllValues( 0 );
//Loop over the data and set the values to random values
UINT counter = 0;
for(UINT i=0; i<matrix.getNumRows(); i++){
for(UINT j=0; j<matrix.getNumCols(); j++){
matrix[i][j] = counter++;
}
}
//Add a new row at the very end of the matrix
VectorDouble newRow(2);
newRow[0] = 1000;
newRow[1] = 2000;
matrix.push_back( newRow );
//Print the values
cout << "Matrix Data: \n";
for(UINT i=0; i<matrix.getNumRows(); i++){
for(UINT j=0; j<matrix.getNumCols(); j++){
cout << matrix[i][j] << "\t";
}
cout << endl;
}
cout << endl;
//Get the second row as a vector
VectorDouble rowVector = matrix.getRowVector( 1 );
cout << "Row Vector Data: \n";
for(UINT i=0; i<rowVector.size(); i++){
cout << rowVector[i] << "\t";
}
cout << endl;
//Get the second column as a vector
VectorDouble colVector = matrix.getColVector( 1 );
cout << "Column Vector Data: \n";
for(UINT i=0; i<colVector.size(); i++){
cout << colVector[i] << "\n";
}
cout << endl;
//Get the mean of each column
VectorDouble mean = matrix.getMean();
cout << "Mean: \n";
for(UINT i=0; i<mean.size(); i++){
cout << mean[i] << "\n";
}
cout << endl;
//Get the Standard Deviation of each column
VectorDouble stdDev = matrix.getStdDev();
cout << "StdDev: \n";
for(UINT i=0; i<stdDev.size(); i++){
cout << stdDev[i] << "\n";
}
cout << endl;
//Get the covariance matrix
MatrixDouble cov = matrix.getCovarianceMatrix();
cout << "Covariance Matrix: \n";
for(UINT i=0; i<cov.getNumRows(); i++){
for(UINT j=0; j<cov.getNumCols(); j++){
cout << cov[i][j] << "\t";
}
cout << endl;
}
cout << endl;
vector< MinMax > ranges = matrix.getRanges();
cout << "Ranges: \n";
for(UINT i=0; i<ranges.size(); i++){
cout << "i: " << i << "\tMinValue: " << ranges[i].minValue << "\tMaxValue:" << ranges[i].maxValue << "\n";
}
cout << endl;
//Save the matrix data to a csv file
matrix.save( "data.csv" );
//load the matrix data from a csv file
matrix.load( "data.csv" );
return EXIT_SUCCESS;
}
bool GaussianMixtureModels::train_(MatrixDouble &data){
trained = false;
//Clear any previous training results
det.clear();
invSigma.clear();
numTrainingIterationsToConverge = 0;
if( data.getNumRows() == 0 ){
errorLog << "train_(MatrixDouble &data) - Training Failed! Training data is empty!" << endl;
return false;
}
//Resize the variables
numTrainingSamples = data.getNumRows();
numInputDimensions = data.getNumCols();
//Resize mu and resp
mu.resize(numClusters,numInputDimensions);
resp.resize(numTrainingSamples,numClusters);
//Resize sigma
sigma.resize(numClusters);
for(UINT k=0; k<numClusters; k++){
sigma[k].resize(numInputDimensions,numInputDimensions);
}
//Resize frac and lndets
frac.resize(numClusters);
lndets.resize(numClusters);
//Scale the data if needed
ranges = data.getRanges();
if( useScaling ){
for(UINT i=0; i<numTrainingSamples; i++){
for(UINT j=0; j<numInputDimensions; j++){
data[i][j] = scale(data[i][j],ranges[j].minValue,ranges[j].maxValue,0,1);
}
}
}
//Pick K random starting points for the inital guesses of Mu
Random random;
vector< UINT > randomIndexs(numTrainingSamples);
for(UINT i=0; i<numTrainingSamples; i++) randomIndexs[i] = i;
for(UINT i=0; i<numClusters; i++){
SWAP(randomIndexs[ i ],randomIndexs[ random.getRandomNumberInt(0,numTrainingSamples) ]);
}
for(UINT k=0; k<numClusters; k++){
for(UINT n=0; n<numInputDimensions; n++){
mu[k][n] = data[ randomIndexs[k] ][n];
}
}
//Setup sigma and the uniform prior on P(k)
for(UINT k=0; k<numClusters; k++){
frac[k] = 1.0/double(numClusters);
for(UINT i=0; i<numInputDimensions; i++){
for(UINT j=0; j<numInputDimensions; j++) sigma[k][i][j] = 0;
sigma[k][i][i] = 1.0e-2; //Set the diagonal to a small number
}
}
loglike = 0;
bool keepGoing = true;
double change = 99.9e99;
UINT numIterationsNoChange = 0;
VectorDouble u(numInputDimensions);
VectorDouble v(numInputDimensions);
while( keepGoing ){
//Run the estep
if( estep( data, u, v, change ) ){
//Run the mstep
mstep( data );
//Check for convergance
if( fabs( change ) < minChange ){
if( ++numIterationsNoChange >= minNumEpochs ){
keepGoing = false;
}
}else numIterationsNoChange = 0;
if( ++numTrainingIterationsToConverge >= maxNumEpochs ) keepGoing = false;
}else{
errorLog << "train_(MatrixDouble &data) - Estep failed at iteration " << numTrainingIterationsToConverge << endl;
return false;
}
}
//Compute the inverse of sigma and the determinants for prediction
if( !computeInvAndDet() ){
det.clear();
invSigma.clear();
errorLog << "train_(MatrixDouble &data) - Failed to compute inverse and determinat!" << endl;
return false;
}
//Flag that the model was trained
trained = true;
return true;
}
bool KMeans::trainModel(MatrixDouble &data){
if( numClusters == 0 ){
errorLog << "trainModel(MatrixDouble &data) - Failed to train model. NumClusters is zero!" << endl;
return false;
}
if( clusters.getNumRows() != numClusters ){
errorLog << "trainModel(MatrixDouble &data) - Failed to train model. The number of rows in the cluster matrix does not match the number of clusters! You should need to initalize the clusters matrix first before calling this function!" << endl;
return false;
}
if( clusters.getNumCols() != numInputDimensions ){
errorLog << "trainModel(MatrixDouble &data) - Failed to train model. The number of columns in the cluster matrix does not match the number of input dimensions! You should need to initalize the clusters matrix first before calling this function!" << endl;
return false;
}
Timer timer;
UINT currentIter = 0;
UINT numChanged = 0;
bool keepTraining = true;
double theta = 0;
double lastTheta = 0;
double delta = 0;
double startTime = 0;
thetaTracker.clear();
finalTheta = 0;
numTrainingIterationsToConverge = 0;
trained = false;
converged = false;
//Scale the data if needed
ranges = data.getRanges();
if( useScaling ){
data.scale(0,1);
}
//Init the assign and count vectors
//Assign is set to K+1 so that the nChanged values in the eStep at the first iteration will be updated correctly
for(UINT m=0; m<numTrainingSamples; m++) assign[m] = numClusters+1;
for(UINT k=0; k<numClusters; k++) count[k] = 0;
//Run the training loop
timer.start();
while( keepTraining ){
startTime = timer.getMilliSeconds();
//Compute the E step
numChanged = estep( data );
//Compute the M step
mstep( data );
//Update the iteration counter
currentIter++;
//Compute theta if needed
if( computeTheta ){
theta = calculateTheta(data);
delta = lastTheta - theta;
lastTheta = theta;
}else theta = delta = 0;
//Check convergance
if( numChanged == 0 && currentIter > minNumEpochs ){ converged = true; keepTraining = false; }
if( currentIter >= maxNumEpochs ){ keepTraining = false; }
if( fabs( delta ) < minChange && computeTheta && currentIter > minNumEpochs ){ converged = true; keepTraining = false; }
if( computeTheta ) thetaTracker.push_back( theta );
trainingLog << "Epoch: " << currentIter << "/" << maxNumEpochs;
trainingLog << " Epoch time: " << (timer.getMilliSeconds()-startTime)/1000.0 << " seconds";
trainingLog << " Theta: " << theta << " Delta: " << delta << endl;
}
trainingLog << "Model Trained at epoch: " << currentIter << " with a theta value of: " << theta << endl;
finalTheta = theta;
numTrainingIterationsToConverge = currentIter;
trained = true;
return true;
}
bool BernoulliRBM::train_(MatrixDouble &data){
const UINT numTrainingSamples = data.getNumRows();
numInputDimensions = data.getNumCols();
numOutputDimensions = numHiddenUnits;
numVisibleUnits = numInputDimensions;
trainingLog << "NumInputDimensions: " << numInputDimensions << endl;
trainingLog << "NumOutputDimensions: " << numOutputDimensions << endl;
if( randomizeWeightsForTraining ){
//Init the weights matrix
weightsMatrix.resize(numHiddenUnits, numVisibleUnits);
double a = 1.0 / numVisibleUnits;
for(UINT i=0; i<numHiddenUnits; i++) {
for(UINT j=0; j<numVisibleUnits; j++) {
weightsMatrix[i][j] = rand.getRandomNumberUniform(-a, a);
}
}
//Init the bias units
visibleLayerBias.resize( numVisibleUnits );
hiddenLayerBias.resize( numHiddenUnits );
std::fill(visibleLayerBias.begin(),visibleLayerBias.end(),0);
std::fill(hiddenLayerBias.begin(),hiddenLayerBias.end(),0);
}else{
if( weightsMatrix.getNumRows() != numHiddenUnits ){
errorLog << "train_(MatrixDouble &data) - Weights matrix row size does not match the number of hidden units!" << endl;
return false;
}
if( weightsMatrix.getNumCols() != numVisibleUnits ){
errorLog << "train_(MatrixDouble &data) - Weights matrix row size does not match the number of visible units!" << endl;
return false;
}
if( visibleLayerBias.size() != numVisibleUnits ){
errorLog << "train_(MatrixDouble &data) - Visible layer bias size does not match the number of visible units!" << endl;
return false;
}
if( hiddenLayerBias.size() != numHiddenUnits ){
errorLog << "train_(MatrixDouble &data) - Hidden layer bias size does not match the number of hidden units!" << endl;
return false;
}
}
//Flag the model has been trained encase the user wants to save the model during a training iteration using an observer
trained = true;
//Make sure the data is scaled between [0 1]
ranges = data.getRanges();
if( useScaling ){
for(UINT i=0; i<numTrainingSamples; i++){
for(UINT j=0; j<numInputDimensions; j++){
data[i][j] = scale(data[i][j], ranges[j].minValue, ranges[j].maxValue, 0, 1);
}
}
}
const UINT numBatches = (UINT)ceil( numTrainingSamples/batchSize );
//Setup the batch indexs
vector< BatchIndexs > batchIndexs( numBatches );
UINT startIndex = 0;
for(UINT i=0; i<numBatches; i++){
batchIndexs[i].startIndex = startIndex;
batchIndexs[i].endIndex = startIndex + batchSize;
//Make sure the last batch end index is not larger than the number of training examples
if( batchIndexs[i].endIndex >= numTrainingSamples ){
batchIndexs[i].endIndex = numTrainingSamples;
}
//Get the batch size
batchIndexs[i].batchSize = batchIndexs[i].endIndex - batchIndexs[i].startIndex;
//Set the start index for the next batch
startIndex = batchIndexs[i].endIndex;
}
Timer timer;
UINT i,j,n,epoch,noChangeCounter = 0;
double startTime = 0;
double alpha = learningRate;
double error = 0;
double err = 0;
double delta = 0;
double lastError = 0;
vector< UINT > indexList(numTrainingSamples);
TrainingResult trainingResult;
MatrixDouble wT( numVisibleUnits, numHiddenUnits ); //Stores a transposed copy of the weights vector
MatrixDouble vW( numHiddenUnits, numVisibleUnits ); //Stores the weight velocity updates
MatrixDouble tmpW( numHiddenUnits, numVisibleUnits ); //Stores the weight values that will be used to update the main weights matrix at each batch update
MatrixDouble v1( batchSize, numVisibleUnits ); //Stores the real batch data during a batch update
MatrixDouble v2( batchSize, numVisibleUnits ); //Stores the sampled batch data during a batch update
MatrixDouble h1( batchSize, numHiddenUnits ); //Stores the hidden states given v1 and the current weightsMatrix
MatrixDouble h2( batchSize, numHiddenUnits ); //Stores the sampled hidden states given v2 and the current weightsMatrix
MatrixDouble c1( numHiddenUnits, numVisibleUnits ); //Stores h1' * v1
MatrixDouble c2( numHiddenUnits, numVisibleUnits ); //Stores h2' * v2
MatrixDouble vDiff( batchSize, numVisibleUnits ); //Stores the difference between v1-v2
MatrixDouble hDiff( batchSize, numVisibleUnits ); //Stores the difference between h1-h2
MatrixDouble cDiff( numHiddenUnits, numVisibleUnits ); //Stores the difference between c1-c2
VectorDouble vDiffSum( numVisibleUnits ); //Stores the column sum of vDiff
VectorDouble hDiffSum( numHiddenUnits ); //Stores the column sum of hDiff
VectorDouble visibleLayerBiasVelocity( numVisibleUnits ); //Stores the velocity update of the visibleLayerBias
VectorDouble hiddenLayerBiasVelocity( numHiddenUnits ); //Stores the velocity update of the hiddenLayerBias
//Set all the velocity weights to zero
vW.setAllValues( 0 );
std::fill(visibleLayerBiasVelocity.begin(),visibleLayerBiasVelocity.end(),0);
std::fill(hiddenLayerBiasVelocity.begin(),hiddenLayerBiasVelocity.end(),0);
//Randomize the order that the training samples will be used in
for(UINT i=0; i<numTrainingSamples; i++) indexList[i] = i;
if( randomiseTrainingOrder ){
std::random_shuffle(indexList.begin(), indexList.end());
}
//Start the main training loop
timer.start();
for(epoch=0; epoch<maxNumEpochs; epoch++) {
startTime = timer.getMilliSeconds();
error = 0;
//Randomize the batch order
std::random_shuffle(batchIndexs.begin(),batchIndexs.end());
//Run each of the batch updates
for(UINT k=0; k<numBatches; k+=batchStepSize){
//Resize the data matrices, the matrices will only be resized if the rows cols are different
v1.resize( batchIndexs[k].batchSize, numVisibleUnits );
h1.resize( batchIndexs[k].batchSize, numHiddenUnits );
v2.resize( batchIndexs[k].batchSize, numVisibleUnits );
h2.resize( batchIndexs[k].batchSize, numHiddenUnits );
//Setup the data pointers, using data pointers saves a few ms on large matrix updates
double **w_p = weightsMatrix.getDataPointer();
double **wT_p = wT.getDataPointer();
double **vW_p = vW.getDataPointer();
double **data_p = data.getDataPointer();
double **v1_p = v1.getDataPointer();
double **v2_p = v2.getDataPointer();
double **h1_p = h1.getDataPointer();
double **h2_p = h2.getDataPointer();
double *vlb_p = &visibleLayerBias[0];
double *hlb_p = &hiddenLayerBias[0];
//Get the batch data
UINT index = 0;
for(i=batchIndexs[k].startIndex; i<batchIndexs[k].endIndex; i++){
for(j=0; j<numVisibleUnits; j++){
v1_p[index][j] = data_p[ indexList[i] ][j];
}
index++;
}
//Copy a transposed version of the weights matrix, this is used to compute h1 and h2
for(i=0; i<numHiddenUnits; i++)
for(j=0; j<numVisibleUnits; j++)
wT_p[j][i] = w_p[i][j];
//Compute h1
h1.multiple(v1, wT);
for(n=0; n<batchIndexs[k].batchSize; n++){
for(i=0; i<numHiddenUnits; i++){
h1_p[n][i] = sigmoidRandom( h1_p[n][i] + hlb_p[i] );
}
}
//Compute v2
v2.multiple(h1, weightsMatrix);
for(n=0; n<batchIndexs[k].batchSize; n++){
for(i=0; i<numVisibleUnits; i++){
v2_p[n][i] = sigmoidRandom( v2_p[n][i] + vlb_p[i] );
}
}
//Compute h2
h2.multiple(v2,wT);
for(n=0; n<batchIndexs[k].batchSize; n++){
for(i=0; i<numHiddenUnits; i++){
h2_p[n][i] = sigmoid( h2_p[n][i] + hlb_p[i] );
}
}
//Compute c1, c2 and the difference between v1-v2
c1.multiple(h1,v1,true);
c2.multiple(h2,v2,true);
vDiff.subtract(v1, v2);
//Compute the sum of vdiff
for(j=0; j<numVisibleUnits; j++){
vDiffSum[j] = 0;
for(i=0; i<batchIndexs[k].batchSize; i++){
vDiffSum[j] += vDiff[i][j];
}
}
//Compute the difference between h1 and h2
hDiff.subtract(h1, h2);
for(j=0; j<numHiddenUnits; j++){
hDiffSum[j] = 0;
for(i=0; i<batchIndexs[k].batchSize; i++){
hDiffSum[j] += hDiff[i][j];
}
}
//Compute the difference between c1 and c2
cDiff.subtract(c1,c2);
//Update the weight velocities
for(i=0; i<numHiddenUnits; i++){
for(j=0; j<numVisibleUnits; j++){
vW_p[i][j] = ((momentum * vW_p[i][j]) + (alpha * cDiff[i][j])) / batchIndexs[k].batchSize;
}
}
for(i=0; i<numVisibleUnits; i++){
visibleLayerBiasVelocity[i] = ((momentum * visibleLayerBiasVelocity[i]) + (alpha * vDiffSum[i])) / batchIndexs[k].batchSize;
}
for(i=0; i<numHiddenUnits; i++){
hiddenLayerBiasVelocity[i] = ((momentum * hiddenLayerBiasVelocity[i]) + (alpha * hDiffSum[i])) / batchIndexs[k].batchSize;
}
//Update the weights
weightsMatrix.add( vW );
//Update the bias for the visible layer
for(i=0; i<numVisibleUnits; i++){
visibleLayerBias[i] += visibleLayerBiasVelocity[i];
}
//Update the bias for the visible layer
for(i=0; i<numHiddenUnits; i++){
hiddenLayerBias[i] += hiddenLayerBiasVelocity[i];
}
//Compute the reconstruction error
err = 0;
for(i=0; i<batchIndexs[k].batchSize; i++){
for(j=0; j<numVisibleUnits; j++){
err += SQR( v1[i][j] - v2[i][j] );
}
}
error += err / batchIndexs[k].batchSize;
}
error /= numBatches;
delta = lastError - error;
lastError = error;
trainingLog << "Epoch: " << epoch+1 << "/" << maxNumEpochs;
trainingLog << " Epoch time: " << (timer.getMilliSeconds()-startTime)/1000.0 << " seconds";
trainingLog << " Learning rate: " << alpha;
trainingLog << " Momentum: " << momentum;
trainingLog << " Average reconstruction error: " << error;
trainingLog << " Delta: " << delta << endl;
//Update the learning rate
alpha *= learningRateUpdate;
trainingResult.setClassificationResult(epoch, error, this);
trainingResults.push_back(trainingResult);
trainingResultsObserverManager.notifyObservers( trainingResult );
//Check for convergance
if( fabs(delta) < minChange ){
if( ++noChangeCounter >= minNumEpochs ){
trainingLog << "Stopping training. MinChange limit reached!" << endl;
break;
}
}else noChangeCounter = 0;
}
trainingLog << "Training complete after " << epoch << " epochs. Total training time: " << timer.getMilliSeconds()/1000.0 << " seconds" << endl;
trained = true;
return true;
}
