bool ContinuousHiddenMarkovModel::train_(TimeSeriesClassificationSample &trainingData){
//Clear any previous models
clear();
//The number of states is simply set as the number of samples in the training sample
timeseriesLength = trainingData.getLength();
numStates = (unsigned int)floor((double)(timeseriesLength/downsampleFactor));
numInputDimensions = trainingData.getNumDimensions();
classLabel = trainingData.getClassLabel();
//a is simply set as the number of 1/numStates
a.resize(numStates, numStates);
for(unsigned int i=0; i<numStates; i++){
for(unsigned int j=0; j<numStates; j++){
a[i][j] = 1.0/numStates;
}
}
//b is simply set as the downsampled training sample
b.resize(numStates, numInputDimensions);
unsigned int index = 0;
Float norm = 0;
for(unsigned int j=0; j<numInputDimensions; j++){
index = 0;
for(unsigned int i=0; i<numStates; i++){
norm = 0;
b[i][j] = 0;
for(unsigned int k=0; k<downsampleFactor; k++){
if( index < trainingData.getLength() ){
b[i][j] += trainingData[index++][j];
norm += 1;
}
}
if( norm > 1 )
b[i][j] /= norm;
}
}
//Estimate pi
pi.resize(numStates);
switch( modelType ){
case(HMM_ERGODIC):
for(UINT i=0; i<numStates; i++){
pi[i] = 1.0/numStates;
}
break;
case(HMM_LEFTRIGHT):
//Set the state transitions constraints
for(UINT i=0; i<numStates; i++){
norm = 0;
for(UINT j=0; j<numStates; j++){
if((j<i) || (j>i+delta)) a[i][j] = 0.0;
norm += a[i][j];
}
if( norm > 0 ){
for(UINT j=0; j<numStates; j++){
a[i][j] /= norm;
}
}
}
//Set pi to start in state 0
for(UINT i=0; i<numStates; i++){
pi[i] = i==0 ? 1 : 0;
}
break;
default:
throw("HMM_ERROR: Unkown model type!");
return false;
break;
}
//Setup sigma for each state
sigmaStates.resize( numStates, numInputDimensions );
if( autoEstimateSigma ){
//Estimate the standard dev for each dimension, for each state
MatrixFloat meanResults( numStates, numInputDimensions );
for(unsigned int j=0; j<numInputDimensions; j++){
//Estimate the mean for each state
index = 0;
for(unsigned int i=0; i<numStates; i++){
norm = 0;
meanResults[i][j] = 0;
for(unsigned int k=0; k<downsampleFactor; k++){
if( index < trainingData.getLength() ){
meanResults[i][j] += trainingData[index++][j];
norm += 1;
}
}
if( norm > 1 ){
meanResults[i][j] /= norm;
}
}
//Loop back over the data again and estimate the stddev for each state
index = 0;
for(unsigned int i=0; i<numStates; i++){
norm = 0;
sigmaStates[i][j] = 0;
for(unsigned int k=0; k<downsampleFactor; k++){
if( index < trainingData.getLength() ){
sigmaStates[i][j] += SQR( trainingData[index++][j]-meanResults[i][j] );
norm += 1;
}
}
if( norm > 1 ){
sigmaStates[i][j] = sqrt( 1.0/norm * sigmaStates[i][j] );
}
if( sigmaStates[i][j] < sigma ){
sigmaStates[i][j] = sigma;
}
}
}
}else{
sigmaStates.setAllValues(sigma);
}
//Setup the observation buffer for prediction
observationSequence.resize( timeseriesLength, VectorFloat(numInputDimensions,0) );
obsSequence.resize(timeseriesLength,numInputDimensions);
estimatedStates.resize( numStates );
//Finally, flag that the model was trained
trained = true;
return true;
}
bool TimeSeriesClassificationSampleTrimmer::trimTimeSeries(TimeSeriesClassificationSample &timeSeries) {
const UINT M = timeSeries.getLength();
const UINT N = timeSeries.getNumDimensions();
if( M == 0 ) {
warningLog << "trimTimeSeries(TimeSeriesClassificationSample &timeSeries) - can't trim data, the length of the input time series is 0!" << endl;
return false;
}
if( N == 0 ) {
warningLog << "trimTimeSeries(TimeSeriesClassificationSample &timeSeries) - can't trim data, the number of dimensions in the input time series is 0!" << endl;
return false;
}
//Compute the energy of the time series
double maxValue = 0;
VectorDouble x(M,0);
for(UINT i=1; i<M; i++) {
for(UINT j=0; j<N; j++) {
x[i] += fabs(timeSeries[i][j]-timeSeries[i-1][j]);
}
x[i] /= N;
if( x[i] > maxValue ) maxValue = x[i];
}
//Normalize x so that the maximum energy has a value of 1
//At the same time search for the first time x[i] passes the trim threshold
UINT firstIndex = 0;
for(UINT i=1; i<M; i++) {
x[i] /= maxValue;
if( x[i] > trimThreshold && firstIndex == 0 ) {
firstIndex = i;
}
}
//Search for the last time x[i] passes the trim threshold
UINT lastIndex = 0;
for(UINT i=M-1; i>firstIndex; i--) {
if( x[i] > trimThreshold && lastIndex == 0 ) {
lastIndex = i;
break;
}
}
if( firstIndex == 0 && lastIndex == 0 ) {
warningLog << "Failed to find either the first index or the last index!";
return false;
}
if( firstIndex == lastIndex ) {
warningLog << "The first index and last index are the same!";
return false;
}
if( firstIndex > lastIndex ) {
warningLog << "The first index is greater than the last index!";
return false;
}
if( lastIndex == 0 ) {
warningLog << "Failed to find the last index!";
lastIndex = M-1;
}
//Compute how long the new time series would be if we trimmed it
UINT newM = lastIndex-firstIndex;
double trimPercentage = (double(newM) / double(M)) * 100.0;
if( 100 - trimPercentage <= maximumTrimPercentage ) {
MatrixDouble newTimeSeries(newM,N);
UINT index = 0;
for(UINT i=firstIndex; i<lastIndex; i++) {
for(UINT j=0; j<N; j++) {
newTimeSeries[index][j] = timeSeries[i][j];
}
index++;
}
timeSeries.setTrainingSample(timeSeries.getClassLabel(), newTimeSeries);
return true;
}
warningLog << "Maximum Trim Percentage Excedded, Can't Trim Sample!";
warningLog << " Original Timeseries Length: " << M << " Trimmed Timeseries Length: " << newM;
warningLog << " Percentage: " << (100-trimPercentage) << " MaximumTrimPercentage: " << maximumTrimPercentage << endl;
return false;
}