adaboost_impl.hpp

/*
 * @file methods/adaboost/adaboost_impl.hpp
 * @author Udit Saxena
 *
 * Implementation of the AdaBoost class.
 *
 * @code
 * @article{schapire1999improved,
 *   author = {Schapire, Robert E. and Singer, Yoram},
 *   title = {Improved Boosting Algorithms Using Confidence-rated Predictions},
 *   journal = {Machine Learning},
 *   volume = {37},
 *   number = {3},
 *   month = dec,
 *   year = {1999},
 *   issn = {0885-6125},
 *   pages = {297--336},
 * }
 * @endcode
 *
 * mlpack is free software; you may redistribute it and/or modify it under the
 * terms of the 3-clause BSD license.  You should have received a copy of the
 * 3-clause BSD license along with mlpack.  If not, see
 * http://www.opensource.org/licenses/BSD-3-Clause for more information.
 */
#ifndef MLPACK_METHODS_ADABOOST_ADABOOST_IMPL_HPP
#define MLPACK_METHODS_ADABOOST_ADABOOST_IMPL_HPP

#include "adaboost.hpp"

namespace mlpack {
namespace adaboost {

template<typename WeakLearnerType, typename MatType>
AdaBoost<WeakLearnerType, MatType>::AdaBoost(
    const MatType& data,
    const arma::Row<size_t>& labels,
    const size_t numClasses,
    const WeakLearnerType& other,
    const size_t iterations,
    const double tol)
{
  Train(data, labels, numClasses, other, iterations, tol);
}

// Empty constructor.
template<typename WeakLearnerType, typename MatType>
AdaBoost<WeakLearnerType, MatType>::AdaBoost(const double tolerance) :
    numClasses(0),
    tolerance(tolerance)
{
  // Nothing to do.
}

// Train AdaBoost.
template<typename WeakLearnerType, typename MatType>
double AdaBoost<WeakLearnerType, MatType>::Train(
    const MatType& data,
    const arma::Row<size_t>& labels,
    const size_t numClasses,
    const WeakLearnerType& other,
    const size_t iterations,
    const double tolerance)
{
  // Clear information from previous runs.
  wl.clear();
  alpha.clear();

  this->tolerance = tolerance;
  this->numClasses = numClasses;

  // crt is the cumulative rt value for terminating the optimization when rt is
  // changing by less than the tolerance.
  double rt, crt = 0.0, alphat = 0.0, zt;

  double ztProduct = 1.0;

  // To be used for prediction by the weak learner.
  arma::Row<size_t> predictedLabels(labels.n_cols);

  // Use tempData to modify input data for incorporating weights.
  MatType tempData(data);

  // This matrix is a helper matrix used to calculate the final hypothesis.
  arma::mat sumFinalH = arma::zeros<arma::mat>(numClasses,
      predictedLabels.n_cols);

  // Load the initial weights into a 2-D matrix.
  const double initWeight = 1.0 / double(data.n_cols * numClasses);
  arma::mat D(numClasses, data.n_cols);
  D.fill(initWeight);

  // Weights are stored in this row vector.
  arma::rowvec weights(predictedLabels.n_cols);

  // This is the final hypothesis.
  arma::Row<size_t> finalH(predictedLabels.n_cols);

  // Now, start the boosting rounds.
  for (size_t i = 0; i < iterations; ++i)
  {
    // Initialized to zero in every round.  rt is used for calculation of
    // alphat; it is the weighted error.
    // rt = (sum) D(i) y(i) ht(xi)
    rt = 0.0;

    // zt is used for weight normalization.
    zt = 0.0;

    // Build the weight vectors.
    weights = arma::sum(D);

    // Use the existing weak learner to train a new one with new weights.
    WeakLearnerType w(other, tempData, labels, numClasses, weights);
    w.Classify(tempData, predictedLabels);

    // Now from predictedLabels, build ht, the weak hypothesis
    // buildClassificationMatrix(ht, predictedLabels);

    // Now, calculate alpha(t) using ht.
    for (size_t j = 0; j < D.n_cols; ++j) // instead of D, ht
    {
      if (predictedLabels(j) == labels(j))
        rt += arma::accu(D.col(j));
      else
        rt -= arma::accu(D.col(j));
    }

    if ((i > 0) && (std::abs(rt - crt) < tolerance))
      break;

    // Check if model has converged.
    if (rt >= 1.0)
    {
      // Save the weak learner and terminate.
      alpha.push_back(1.0);
      wl.push_back(w);
      break;
    }

    crt = rt;

    // Our goal is to find alphat which mizimizes or approximately minimizes the
    // value of Z as a function of alpha.
    alphat = 0.5 * log((1 + rt) / (1 - rt));

    alpha.push_back(alphat);
    wl.push_back(w);

    // Now start modifying the weights.
    for (size_t j = 0; j < D.n_cols; ++j)
    {
      const double expo = exp(alphat);
      if (predictedLabels(j) == labels(j))
      {
        for (size_t k = 0; k < D.n_rows; ++k)
        {
          // We calculate zt, the normalization constant.
          D(k, j) /= expo;
          zt += D(k, j); // * exp(-1 * alphat * yt(j,k) * ht(j,k));

          // Add to the final hypothesis matrix.
          // sumFinalH(k, j) += (alphat * ht(k, j));
          if (k == labels(j))
            sumFinalH(k, j) += (alphat); // * ht(k, j));
          else
            sumFinalH(k, j) -= (alphat);
        }
      }
      else
      {
        for (size_t k = 0; k < D.n_rows; ++k)
        {
          // We calculate zt, the normalization constant.
          D(k, j) *= expo;
          zt += D(k, j);

          // Add to the final hypothesis matrix.
          if (k == labels(j))
            sumFinalH(k, j) += alphat; // * ht(k, j));
          else
            sumFinalH(k, j) -= alphat;
        }
      }
    }

    // Normalize D.
    D /= zt;

    // Accumulate the value of zt for the Hamming loss bound.
    ztProduct *= zt;
  }
  return ztProduct;
}

template<typename WeakLearnerType, typename MatType>
void AdaBoost<WeakLearnerType, MatType>::Classify(
    const MatType& test,
    arma::Row<size_t>& predictedLabels)
{
  arma::Row<size_t> tempPredictedLabels(test.n_cols);
  arma::mat probabilities;

  Classify(test, predictedLabels, probabilities);
}

template<typename WeakLearnerType, typename MatType>
void AdaBoost<WeakLearnerType, MatType>::Classify(
    const MatType& test,
    arma::Row<size_t>& predictedLabels,
    arma::mat& probabilities)
{
  arma::Row<size_t> tempPredictedLabels(test.n_cols);

  probabilities.zeros(numClasses, test.n_cols);
  predictedLabels.set_size(test.n_cols);

  for (size_t i = 0; i < wl.size(); ++i)
  {
    wl[i].Classify(test, tempPredictedLabels);

    for (size_t j = 0; j < tempPredictedLabels.n_cols; ++j)
      probabilities(tempPredictedLabels(j), j) += alpha[i];
  }

  arma::colvec pRow;
  arma::uword maxIndex = 0;

  for (size_t i = 0; i < predictedLabels.n_cols; ++i)
  {
    probabilities.col(i) /= arma::accu(probabilities.col(i));
    pRow = probabilities.unsafe_col(i);
    pRow.max(maxIndex);
    predictedLabels(i) = maxIndex;
  }
}

template<typename WeakLearnerType, typename MatType>
template<typename Archive>
void AdaBoost<WeakLearnerType, MatType>::serialize(Archive& ar,
                                                   const uint32_t /* version */)
{
  ar(CEREAL_NVP(numClasses));
  ar(CEREAL_NVP(tolerance));
  ar(CEREAL_NVP(alpha));

  // Now serialize each weak learner.
  if (cereal::is_loading<Archive>())
  {
    wl.clear();
    wl.resize(alpha.size());
  }
  ar(CEREAL_NVP(wl));
}

} // namespace adaboost
} // namespace mlpack

#endif