decoding_spoc.cpp

Go to the documentation of this file.

//=============================================================================================================
 
//=============================================================================================================
// INCLUDES
//=============================================================================================================
 
#include "decoding_spoc.h"
 
//=============================================================================================================
// EIGEN INCLUDES
//=============================================================================================================
 
#include <Eigen/Eigenvalues>
 
//=============================================================================================================
// STL INCLUDES
//=============================================================================================================
 
#include <algorithm>
#include <cmath>
#include <numeric>
#include <stdexcept>
 
//=============================================================================================================
// USED NAMESPACES
//=============================================================================================================
 
using namespace DECODINGLIB;
using namespace Eigen;
 
//=============================================================================================================
 
DecodingSpoc::DecodingSpoc(int nComponents,
                           TransformMode transformInto,
                           bool useLog)
: m_nComponents(nComponents)
, m_transformInto(transformInto)
, m_useLog(useLog)
{
}
 
//=============================================================================================================
 
void DecodingSpoc::fit(const std::vector<MatrixXd>& epochs,
                       const VectorXd& y)
{
    if (epochs.empty()) {
        throw std::invalid_argument("DecodingSpoc::fit: epochs must be non-empty");
    }
    if (static_cast<int>(epochs.size()) != y.size()) {
        throw std::invalid_argument(
            "DecodingSpoc::fit: epochs and y must have the same length");
    }
 
    // Delegate core GED (SPoC algorithm)
    const auto n_epochs = static_cast<Index>(epochs.size());
    const auto n_ch = epochs[0].rows();
 
    // 1. Normalize target
    double z_mean = y.mean();
    double z_std = std::sqrt(
        (y.array() - z_mean).square().sum() / static_cast<double>(n_epochs - 1));
    VectorXd z = (y.array() - z_mean).matrix();
    if (z_std > 1e-15)
        z /= z_std;
 
    // 2. Per-epoch covariance → C and Cz
    MatrixXd C = MatrixXd::Zero(n_ch, n_ch);
    MatrixXd Cz = MatrixXd::Zero(n_ch, n_ch);
 
    for (Index e = 0; e < n_epochs; ++e) {
        const auto& X = epochs[static_cast<size_t>(e)];
        MatrixXd Xc = X.colwise() - X.rowwise().mean();
        MatrixXd cov_e = (Xc * Xc.transpose()) / static_cast<double>(Xc.cols());
        C += cov_e;
        Cz += z(e) * cov_e;
    }
    C /= static_cast<double>(n_epochs);
    Cz /= static_cast<double>(n_epochs);
 
    // 3. Generalized eigenvalue problem: Cz w = λ C w
    GeneralizedSelfAdjointEigenSolver<MatrixXd> solver(Cz, C);
    if (solver.info() != Eigen::Success) {
        throw std::runtime_error(
            "DecodingSpoc::fit: eigenvalue decomposition failed");
    }
 
    const VectorXd& all_evals = solver.eigenvalues();
    const MatrixXd& all_evecs = solver.eigenvectors();
 
    // 4. Sort by |eigenvalue| descending
    std::vector<int> idx(static_cast<size_t>(n_ch));
    std::iota(idx.begin(), idx.end(), 0);
    std::sort(idx.begin(), idx.end(), [&](int a, int b) {
        return std::abs(all_evals(a)) > std::abs(all_evals(b));
    });
 
    int n_comp = std::min(m_nComponents, static_cast<int>(n_ch));
    m_filters.resize(n_comp, n_ch);
 
    for (int i = 0; i < n_comp; ++i) {
        int j = idx[static_cast<size_t>(i)];
        VectorXd w = all_evecs.col(j);
        double norm = w.norm();
        if (norm > 0.0)
            w /= norm;
        m_filters.row(i) = w.transpose();
    }
 
    // 5. Patterns: A = C W inv(W^T C W)
    MatrixXd Wt = m_filters.transpose();
    MatrixXd CW = C * Wt;
    MatrixXd WtCW = Wt.transpose() * CW;
    m_patterns = (CW * WtCW.inverse()); // (n_ch × n_comp) — already correct shape
 
    // Compute mean band power for z-score normalisation
    MatrixXd powerFeatures = computePowerFeatures(epochs);
    m_mean = powerFeatures.colwise().mean();
 
    m_std = VectorXd(powerFeatures.cols());
    for (int c = 0; c < powerFeatures.cols(); ++c) {
        VectorXd centered = powerFeatures.col(c).array() - m_mean(c);
        m_std(c) = std::sqrt(centered.squaredNorm() / static_cast<double>(centered.size()));
    }
 
    m_fitted = true;
}
 
//=============================================================================================================
 
MatrixXd DecodingSpoc::transform(const std::vector<MatrixXd>& epochs) const
{
    if (!m_fitted) {
        throw std::runtime_error("DecodingSpoc::transform: not fitted");
    }
 
    if (m_transformInto == TransformMode::CspSpace) {
        const int nEpochs = static_cast<int>(epochs.size());
        const int nComp = static_cast<int>(m_filters.rows());
        const int nTimes = static_cast<int>(epochs[0].cols());
 
        MatrixXd result(nEpochs * nComp, nTimes);
        for (int e = 0; e < nEpochs; ++e) {
            result.middleRows(static_cast<Eigen::Index>(e) * nComp, nComp) = m_filters * epochs[static_cast<size_t>(e)];
        }
        return result;
    }
 
    // AveragePower mode
    MatrixXd X = computePowerFeatures(epochs);
 
    if (m_useLog) {
        X = X.array().max(1e-30).log().matrix();
    } else {
        for (int c = 0; c < X.cols(); ++c) {
            double s = m_std(c);
            if (s < 1e-15)
                s = 1.0;
            X.col(c) = (X.col(c).array() - m_mean(c)) / s;
        }
    }
 
    return X;
}
 
//=============================================================================================================
 
MatrixXd DecodingSpoc::fitTransform(const std::vector<MatrixXd>& epochs,
                                    const VectorXd& y)
{
    fit(epochs, y);
    return transform(epochs);
}
 
//=============================================================================================================
 
const MatrixXd& DecodingSpoc::filters() const
{
    if (!m_fitted) {
        throw std::runtime_error("DecodingSpoc::filters: not fitted");
    }
    return m_filters;
}
 
//=============================================================================================================
 
const MatrixXd& DecodingSpoc::patterns() const
{
    if (!m_fitted) {
        throw std::runtime_error("DecodingSpoc::patterns: not fitted");
    }
    return m_patterns;
}
 
//=============================================================================================================
 
const VectorXd& DecodingSpoc::mean() const
{
    if (!m_fitted) {
        throw std::runtime_error("DecodingSpoc::mean: not fitted");
    }
    return m_mean;
}
 
//=============================================================================================================
 
const VectorXd& DecodingSpoc::stddev() const
{
    if (!m_fitted) {
        throw std::runtime_error("DecodingSpoc::stddev: not fitted");
    }
    return m_std;
}
 
//=============================================================================================================
 
bool DecodingSpoc::isFitted() const
{
    return m_fitted;
}
 
//=============================================================================================================
 
MatrixXd DecodingSpoc::computePowerFeatures(
    const std::vector<MatrixXd>& epochs) const
{
    const int nEpochs = static_cast<int>(epochs.size());
    const int nComp = static_cast<int>(m_filters.rows());
 
    MatrixXd features(nEpochs, nComp);
    for (int e = 0; e < nEpochs; ++e) {
        MatrixXd filtered = m_filters * epochs[static_cast<size_t>(e)];
        for (int c = 0; c < nComp; ++c) {
            features(e, c) = filtered.row(c).squaredNorm() / static_cast<double>(filtered.cols());
        }
    }
 
    return features;
}