void EM::read(const FileNode& fn)
{
Algorithm::read(fn);
decomposeCovs();
computeLogWeightDivDet();
}
void clusterTrainSamples()
{
int nclusters = params.nclusters;
int nsamples = trainSamples.rows;
// Cluster samples, compute/update means
// Convert samples and means to 32F, because kmeans requires this type.
Mat trainSamplesFlt, meansFlt;
if(trainSamples.type() != CV_32FC1)
trainSamples.convertTo(trainSamplesFlt, CV_32FC1);
else
trainSamplesFlt = trainSamples;
if(!means.empty())
{
if(means.type() != CV_32FC1)
means.convertTo(meansFlt, CV_32FC1);
else
meansFlt = means;
}
Mat labels;
kmeans(trainSamplesFlt, nclusters, labels,
TermCriteria(TermCriteria::COUNT, means.empty() ? 10 : 1, 0.5),
10, KMEANS_PP_CENTERS, meansFlt);
// Convert samples and means back to 64F.
CV_Assert(meansFlt.type() == CV_32FC1);
if(trainSamples.type() != CV_64FC1)
{
Mat trainSamplesBuffer;
trainSamplesFlt.convertTo(trainSamplesBuffer, CV_64FC1);
trainSamples = trainSamplesBuffer;
}
meansFlt.convertTo(means, CV_64FC1);
// Compute weights and covs
weights = Mat(1, nclusters, CV_64FC1, Scalar(0));
covs.resize(nclusters);
for(int clusterIndex = 0; clusterIndex < nclusters; clusterIndex++)
{
Mat clusterSamples;
for(int sampleIndex = 0; sampleIndex < nsamples; sampleIndex++)
{
if(labels.at<int>(sampleIndex) == clusterIndex)
{
const Mat sample = trainSamples.row(sampleIndex);
clusterSamples.push_back(sample);
}
}
CV_Assert(!clusterSamples.empty());
calcCovarMatrix(clusterSamples, covs[clusterIndex], means.row(clusterIndex),
CV_COVAR_NORMAL + CV_COVAR_ROWS + CV_COVAR_USE_AVG + CV_COVAR_SCALE, CV_64FC1);
weights.at<double>(clusterIndex) = static_cast<double>(clusterSamples.rows)/static_cast<double>(nsamples);
}
decomposeCovs();
}
void read(const FileNode& fn)
{
clear();
read_params(fn["training_params"]);
fn["weights"] >> weights;
fn["means"] >> means;
FileNode cfn = fn["covs"];
FileNodeIterator cfn_it = cfn.begin();
int i, n = (int)cfn.size();
covs.resize(n);
for( i = 0; i < n; i++, ++cfn_it )
(*cfn_it) >> covs[i];
decomposeCovs();
computeLogWeightDivDet();
}
bool EM::doTrain(int startStep, OutputArray logLikelihoods, OutputArray labels, OutputArray probs)
{
int dim = trainSamples.cols;
// Precompute the empty initial train data in the cases of EM::START_E_STEP and START_AUTO_STEP
if(startStep != EM::START_M_STEP)
{
if(covs.empty())
{
CV_Assert(weights.empty());
clusterTrainSamples();
}
}
if(!covs.empty() && covsEigenValues.empty() )
{
CV_Assert(invCovsEigenValues.empty());
decomposeCovs();
}
if(startStep == EM::START_M_STEP)
mStep();
double trainLogLikelihood, prevTrainLogLikelihood = 0.;
for(int iter = 0; ; iter++)
{
eStep();
trainLogLikelihood = sum(trainLogLikelihoods)[0];
if(iter >= maxIters - 1)
break;
double trainLogLikelihoodDelta = trainLogLikelihood - prevTrainLogLikelihood;
if( iter != 0 &&
(trainLogLikelihoodDelta < -DBL_EPSILON ||
trainLogLikelihoodDelta < epsilon * std::fabs(trainLogLikelihood)))
break;
mStep();
prevTrainLogLikelihood = trainLogLikelihood;
}
if( trainLogLikelihood <= -DBL_MAX/10000. )
{
clear();
return false;
}
// postprocess covs
covs.resize(nclusters);
for(int clusterIndex = 0; clusterIndex < nclusters; clusterIndex++)
{
if(covMatType == EM::COV_MAT_SPHERICAL)
{
covs[clusterIndex].create(dim, dim, CV_64FC1);
setIdentity(covs[clusterIndex], Scalar(covsEigenValues[clusterIndex].at<double>(0)));
}
else if(covMatType == EM::COV_MAT_DIAGONAL)
{
covs[clusterIndex] = Mat::diag(covsEigenValues[clusterIndex]);
}
}
if(labels.needed())
trainLabels.copyTo(labels);
if(probs.needed())
trainProbs.copyTo(probs);
if(logLikelihoods.needed())
trainLogLikelihoods.copyTo(logLikelihoods);
trainSamples.release();
trainProbs.release();
trainLabels.release();
trainLogLikelihoods.release();
return true;
}
