nadam_update.hpp

Go to the documentation of this file.

#ifndef MLPACK_CORE_OPTIMIZERS_ADAM_NADAM_UPDATE_HPP
#define MLPACK_CORE_OPTIMIZERS_ADAM_NADAM_UPDATE_HPP

#include <mlpack/prereqs.hpp>

namespace mlpack {
namespace optimization {

class NadamUpdate
{
 public:
  NadamUpdate(const double epsilon = 1e-8,
              const double beta1 = 0.9,
              const double beta2 = 0.99,
              const double scheduleDecay = 4e-3) :
      epsilon(epsilon),
      beta1(beta1),
      beta2(beta2),
      scheduleDecay(scheduleDecay),
      iteration(0),
      cumBeta1(1)
  {
    // Nothing to do.
  }

  void Initialize(const size_t rows, const size_t cols)
  {
    m = arma::zeros<arma::mat>(rows, cols);
    v = arma::zeros<arma::mat>(rows, cols);
  }

  void Update(arma::mat& iterate,
              const double stepSize,
              const arma::mat& gradient)
  {
    // Increment the iteration counter variable.
    ++iteration;

    // And update the iterate.
    m *= beta1;
    m += (1 - beta1) * gradient;

    v *= beta2;
    v += (1 - beta2) * gradient % gradient;

    double beta1T = beta1 * (1 - (0.5 *
        std::pow(0.96, iteration * scheduleDecay)));

    double beta1T1 = beta1 * (1 - (0.5 *
        std::pow(0.96, (iteration + 1) * scheduleDecay)));

    cumBeta1 *= beta1T;

    const double biasCorrection1 = 1.0 - cumBeta1;

    const double biasCorrection2 = 1.0 - std::pow(beta2, iteration);

    const double biasCorrection3 = 1.0 - (cumBeta1 * beta1T1);

    /* Note :- arma::sqrt(v) + epsilon * sqrt(biasCorrection2) is approximated
     * as arma::sqrt(v) + epsilon
     */
    iterate -= (stepSize * (((1 - beta1T) / biasCorrection1) * gradient
        + (beta1T1 / biasCorrection3) * m) * sqrt(biasCorrection2))
        / (arma::sqrt(v) + epsilon);
  }

  double Epsilon() const { return epsilon; }
  double& Epsilon() { return epsilon; }

  double CumBeta1() const { return cumBeta1; }
  double& CumBeta1() { return cumBeta1; }

  double Beta1() const { return beta1; }
  double& Beta1() { return beta1; }

  double Beta2() const { return beta2; }
  double& Beta2() { return beta2; }

  double ScheduleDecay() const { return scheduleDecay; }
  double& ScheduleDecay() { return scheduleDecay; }

 private:
  // The epsilon value used to initialise the squared gradient parameter.
  double epsilon;

  // The smoothing parameter.
  double beta1;

  // The second moment coefficient.
  double beta2;

  // The exponential moving average of gradient values.
  arma::mat m;

  // The exponential moving average of squared gradient values.
  arma::mat v;

  // The decay parameter for decay coefficients
  double scheduleDecay;

  // The number of iterations.
  double iteration;

  // The cumulative product of decay coefficients
  double cumBeta1;
};

} // namespace optimization
} // namespace mlpack

#endif