Class ParameterUpdaterMomentum

Inheritance Relationships

Base Type

Class Documentation

class ParameterUpdaterMomentum : public parpe::ParameterUpdater

Mini-batch optimizer: Momentum Updater A classical gradient based optimizer using a vanilla momentum formula.

Public Functions

ParameterUpdaterMomentum() = default
virtual void updateParameters(double learningRate, int iteration, gsl::span<const double> gradient, gsl::span<double> parameters, gsl::span<const double> lowerBounds = gsl::span<const double>(), gsl::span<const double> upperBounds = gsl::span<const double>()) override

Update parameter vector.

Parameters:

  • learningRate – Current learning rate, i.e., step-size

  • iteration – Current iteration, i.e., epoch

  • gradient – Cost function gradient at parameters

  • parameters – In: Current parameters, Out: Updated parameters

virtual void undoLastStep() override

If ODE becomes non-integrable, the last step must be undone using this method

virtual void clearCache() override

If the ODE is repeatedly non-integrable, a cold restart is performed using this method

virtual void initialize(unsigned int numParameters) override

Initialize the parameter updater