MomentumTests.cpp

Go to the documentation of this file.

#include <gtest/gtest.h>
#include <cmath>

#include <optimization/ArtificialLandscapesTests.hpp>

// Redefine word "public" so every class field/method will be accessible for tests.
#define private public
#include <optimization/Momentum.hpp>




// Momentum


TEST_F(Sphere1DLandscape, Momentum_Convergence) {
    // Optimization function - 1d momentum with learning rate = 0.1 (and default 0.9 momentum).
    mic::neural_nets::optimization::Momentum<double> opt(x->rows(), x->cols());

    // Simulate a simple gradient descent.
    size_t iteration = 0;
    double abs_diff = 1.0;
    while (abs_diff > eps) {
        mic::types::MatrixPtr<double> dx = fun.calculateGradient(x);
        opt.update(x, dx, 0.1);

        // Check whether value of the function is finite.
        double value = fun.calculateValue(x);
        ASSERT_EQ(true, std::isfinite(value)) << " at iteration i=" << iteration;

        // Calculate diff - std:abs!!
        abs_diff = std::abs(value - fun.minValue());
        iteration++;
    }//: while
    ASSERT_GE(eps, std::abs(fun.calculateValue(x) - fun.minValue()));
    std::cout << "              -> Converged after " << iteration << " iterations\n";
}


TEST_F(Sphere20DLandscape, Momentum_Convergence) {
    // Optimization function - 20d momentum with learning rate = 0.1 (and default 0.9 momentum).
    mic::neural_nets::optimization::Momentum<double> opt(x->rows(), x->cols());

    // Simulate a simple gradient descent.
    size_t iteration = 0;
    double abs_diff = 1.0;
    while (abs_diff > eps) {
        mic::types::MatrixPtr<double> dx = fun.calculateGradient(x);
        opt.update(x, dx, 0.01);

        // Check whether value of the function is finite.
        double value = fun.calculateValue(x);
        ASSERT_EQ(true, std::isfinite(value)) << " at iteration i=" << iteration;

        // Calculate diff - std:abs!!
        abs_diff = std::abs(value - fun.minValue());
        iteration++;
    }//: while
    ASSERT_GE(eps, std::abs(fun.calculateValue(x) - fun.minValue()));
    std::cout << "              -> Converged after " << iteration << " iterations\n";
}


TEST_F(Beale2DLandscape, Momentum_Convergence) {
    // Optimization function - 2d momentum with learning rate = 0.1 (and default 0.9 momentum).
    mic::neural_nets::optimization::Momentum<double> opt(x->rows(), x->cols());

    // Simulate a simple gradient descent.
    size_t iteration = 0;
    double abs_diff = 1.0;
    while (abs_diff > eps) {
        mic::types::MatrixPtr<double> dx = fun.calculateGradient(x);
        opt.update(x, dx, 0.01);

        // Check whether value of the function is finite.
        double value = fun.calculateValue(x);
        ASSERT_EQ(true, std::isfinite(value)) << " at iteration i=" << iteration;

        // Calculate diff - std:abs!!
        abs_diff = std::abs(value - fun.minValue());
        iteration++;
    }//: while
    ASSERT_GE(eps, std::abs(fun.calculateValue(x) - fun.minValue()));
    std::cout << "              -> Converged after " << iteration << " iterations\n";
}


TEST_F(Rosenbrock2DLandscape, Momentum_Convergence) {
    // Optimization function - 2d momentum with learning rate = 0.00001 (and default 0.9 momentum).
    mic::neural_nets::optimization::Momentum<double> opt(x->rows(), x->cols());

    // Simulate a simple gradient descent.
    size_t iteration = 0;
    double abs_diff = 1.0;
    while (abs_diff > eps) {
        mic::types::MatrixPtr<double> dx = fun.calculateGradient(x);
        opt.update(x, dx, 0.00001);

        // Check whether value of the function is finite.
        double value = fun.calculateValue(x);
        ASSERT_EQ(true, std::isfinite(value)) << " at iteration i=" << iteration;

        // Calculate diff - std:abs!!
        abs_diff = std::abs(value - fun.minValue());
        iteration++;
    }//: while
    ASSERT_GE(eps, std::abs(fun.calculateValue(x) - fun.minValue()));
    std::cout << "              -> Converged after " << iteration << " iterations\n";
}