void drwnGaussian::train(const drwnSuffStats& stats, double lambda)
{
DRWN_ASSERT_MSG(stats.size() == _n, stats.size() << " != " << _n);
// normalize sufficient statistics
_mu = stats.firstMoments() / stats.count();
if (stats.isDiagonal()) {
_mSigma = stats.secondMoments().array() / stats.count();
_mSigma -= _mu.array().square().matrix().asDiagonal();
_mSigma += lambda * MatrixXd::Identity(_n, _n);
} else {
_mSigma = stats.secondMoments() / stats.count() -
_mu * _mu.transpose() + lambda * MatrixXd::Identity(_n, _n);
}
updateCachedParameters();
}
// training
double drwnFeatureWhitener::train(const drwnSuffStats& stats)
{
_nFeatures = stats.size();
_mu = VectorXd::Zero(_nFeatures);
_beta = VectorXd::Zero(_nFeatures);
for (int i = 0; i < _nFeatures; i++) {
_mu[i] = stats.sum(i) / (stats.count() + DRWN_DBL_MIN);
double sigma = stats.sum2(i, i) / (stats.count() + DRWN_DBL_MIN) -
_mu[i] * _mu[i];
if (sigma < DRWN_DBL_MIN) {
_beta[i] = (fabs(_mu[i]) > DRWN_DBL_MIN) ? (1.0 / _mu[i]) : 1.0;
_mu[i] = 0.0;
} else {
_beta[i] = 1.0 / sqrt(sigma);
}
}
DRWN_LOG_DEBUG("drwnFeatureWhitener::_mu = " << _mu.transpose());
DRWN_LOG_DEBUG("drwnFeatureWhitener::_beta = " << _beta.transpose());
_bValid = true;
return _beta.sum();
}
drwnGaussian::drwnGaussian(const drwnSuffStats& stats) :
_n(stats.size()), _invSigma(NULL), _logZ(0.0), _mL(NULL)
{
initialize(_n);
train(stats, DRWN_EPSILON);
}
double drwnGaussian::klDivergence(const drwnSuffStats& stats) const
{
DRWN_ASSERT(stats.size() == _n);
return klDivergence(drwnGaussian(stats));
}
