MultiLayerNeuralNetworkTests.hpp

Go to the documentation of this file.

#ifndef MULTILAYERNEURALNETSTESTS_HPP_
#define MULTILAYERNEURALNETSTESTS_HPP_


#include <gtest/gtest.h>

// Redefine "private" and "protected" so every class field/method will be accessible for tests.
#define private public
#define protected public
#include <mlnn/BackpropagationNeuralNetwork.hpp>


namespace mic { namespace neural_nets { namespace unit_tests {


class Simple2LayerRegressionNN : public ::testing::Test {
public:
    // Constructor. Sets layer size.
    Simple2LayerRegressionNN () :
        nn("simple_linear_network")
    {
        // Add
        nn.pushLayer(new mic::mlnn::fully_connected::Linear<double>(10, 20, "First Linear"));
        nn.pushLayer(new mic::mlnn::activation_function::ReLU<double>(20, "First ReLU"));
        nn.pushLayer(new mic::mlnn::fully_connected::Linear<double>(20, 4, "Second Linear"));
        nn.pushLayer(new mic::mlnn::activation_function::ReLU<double>(4, "Second ReLU"));
        nn.setLoss< mic::neural_nets::loss::SquaredErrorLoss<double> >();
    }

protected:
    virtual void SetUp() {
        // Reset the gradients.
        nn.resetGrads();
        // Set batch size to 1.
        nn.resizeBatch(1);
    }

private:
    // Neural network.
    mic::mlnn::BackpropagationNeuralNetwork<double> nn;
};



class Tutorial2LayerNN : public ::testing::Test {
public:
    // Constructor. Sets layer size.
    Tutorial2LayerNN () :
        nn("simple_linear_network")
    {
        // Add
        nn.pushLayer(new mic::mlnn::fully_connected::Linear<double>(2, 2, "Linear1"));
        nn.pushLayer(new mic::mlnn::activation_function::Sigmoid<double>(2, "Sigmoid1"));
        nn.pushLayer(new mic::mlnn::fully_connected::Linear<double>(2, 2, "Linear2"));
        nn.pushLayer(new mic::mlnn::activation_function::Sigmoid<double>(2, "Sigmoid2"));
        nn.setLoss< mic::neural_nets::loss::SquaredErrorLoss<double> >();

        input_x = MAKE_MATRIX_PTR(double, 2, 1);
        target_y = MAKE_MATRIX_PTR(double, 2, 1);
        ffpass1_lin1_y = MAKE_MATRIX_PTR(double, 2, 1);
        ffpass1_sig1_y = MAKE_MATRIX_PTR(double, 2, 1);
        ffpass1_lin2_y = MAKE_MATRIX_PTR(double, 2, 1);
        ffpass1_sig2_y = MAKE_MATRIX_PTR(double, 2, 1);
        ffpass1_dy = MAKE_MATRIX_PTR(double, 2, 1);
        bwpass1_lin2_dW = MAKE_MATRIX_PTR(double, 2, 2);
        bwpass1_lin2_pW_updated = MAKE_MATRIX_PTR(double, 2, 2);
        bwpass1_lin1_dW = MAKE_MATRIX_PTR(double, 2, 2);
        bwpass1_lin1_pW_updated = MAKE_MATRIX_PTR(double, 2, 2);
}

protected:
    virtual void SetUp() {
        // Reset the gradients.
        nn.resetGrads();
        // Set batch size to 1.
        nn.resizeBatch(1);

        // Set weights and biases.
        (*nn.layers[0]->p["W"]) << 0.15, .20, .25, .30;
        (*nn.layers[0]->p["b"]) << .35, 0.35;

        (*nn.layers[2]->p["W"]) << 0.4, .45, .5, .55;
        (*nn.layers[2]->p["b"]) << .6, 0.6;

        (*input_x) << 0.05, 0.1;
        (*target_y) << 0.01, 0.99;

        // Values after the first forward pass.
        // Linear1
        (*ffpass1_lin1_y) << 0.3775, 0.3925;
        // Sigmoid1
        (*ffpass1_sig1_y) << 0.59327, 0.596884;
        // Linear2
        (*ffpass1_lin2_y) << 1.10591, 1.22492;
        // Sigmoid2
        (*ffpass1_sig2_y) << 0.751365, 0.772928;

        // Loss and gradient.
        ffpass1_loss = 0.298371;
        (*ffpass1_dy) <<  0.741365, -0.217072;

        // backward pass - weights gradients.
        (*bwpass1_lin2_dW) << 0.082167, 0.0826676, -0.0226025, -0.0227402;
        (*bwpass1_lin1_dW) << 0.000438568, 0.000877135, 0.000497713, 0.000995425;

        // Weights after update with  gradient descent learning rule with learning rate = 0.5 (in this case: delta rule).
        (*bwpass1_lin2_pW_updated) << 0.358916, 0.408666, 0.511301, 0.56137;
        (*bwpass1_lin1_pW_updated) << 0.149781, 0.199561, 0.249751, 0.299502;
    }


private:
    // Neural network.
    mic::mlnn::BackpropagationNeuralNetwork<double> nn;

    // Test input x - used in forward pass.
    mic::types::MatrixPtr<double> input_x;

    // Target y values.
    mic::types::MatrixPtr<double> target_y;

    //  Results of the first forward pass.
    mic::types::MatrixPtr<double> ffpass1_lin1_y, ffpass1_sig1_y, ffpass1_lin2_y, ffpass1_sig2_y;

    // Loss and gradient.
    double ffpass1_loss;
    mic::types::MatrixPtr<double> ffpass1_dy;

    // Backward pass.
    mic::types::MatrixPtr<double> bwpass1_lin2_dW;
    mic::types::MatrixPtr<double> bwpass1_lin2_pW_updated;
    mic::types::MatrixPtr<double> bwpass1_lin1_dW;
    mic::types::MatrixPtr<double> bwpass1_lin1_pW_updated;
};

} } }//: namespaces

#endif /* MULTILAYERNEURALNETSTESTS_HPP_ */