NeuralNetwork.h

Go to the documentation of this file.

/*
 * Copyright (c) 2014, Shogun Toolbox Foundation
 * All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions are met:

 * 1. Redistributions of source code must retain the above copyright notice,
 * this list of conditions and the following disclaimer.
 *
 * 2. Redistributions in binary form must reproduce the above copyright notice,
 * this list of conditions and the following disclaimer in the documentation
 * and/or other materials provided with the distribution.
 *
 * 3. Neither the name of the copyright holder nor the names of its
 * contributors may be used to endorse or promote products derived from this
 * software without specific prior written permission.

 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
 * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
 * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
 * POSSIBILITY OF SUCH DAMAGE.
 *
 * Written (W) 2014 Khaled Nasr
 */

#ifndef __NEURALNETWORK_H__
#define __NEURALNETWORK_H__

#include <shogun/lib/common.h>
#include <shogun/machine/Machine.h>
#include <shogun/lib/SGVector.h>
#include <shogun/lib/SGMatrix.h>

namespace shogun
{
template<class T> class CDenseFeatures;
class CDynamicObjectArray;
class CNeuralLayer;

enum ENNOptimizationMethod
{
    NNOM_GRADIENT_DESCENT=0,
    NNOM_LBFGS=1
};

class CNeuralNetwork : public CMachine
{
friend class CDeepBeliefNetwork;

public:
    CNeuralNetwork();

    CNeuralNetwork(CDynamicObjectArray* layers);

    virtual void set_layers(CDynamicObjectArray* layers);

    virtual void connect(int32_t i, int32_t j);

    virtual void quick_connect();

    virtual void disconnect(int32_t i, int32_t j);

    virtual void disconnect_all();

    virtual void initialize_neural_network(float64_t sigma = 0.01f);

    virtual ~CNeuralNetwork();

    virtual CBinaryLabels* apply_binary(CFeatures* data);
    virtual CRegressionLabels* apply_regression(CFeatures* data);
    virtual CMulticlassLabels* apply_multiclass(CFeatures* data);

    virtual CDenseFeatures<float64_t>* transform(
        CDenseFeatures<float64_t>* data);

    virtual void set_labels(CLabels* lab);

    virtual EMachineType get_classifier_type() { return CT_NEURALNETWORK; }

    virtual EProblemType get_machine_problem_type() const;

    virtual float64_t check_gradients(float64_t approx_epsilon=1.0e-3,
            float64_t s = 1.0e-9);

    SGVector<float64_t>* get_layer_parameters(int32_t i);

    int32_t get_num_parameters() { return m_total_num_parameters; }

    SGVector<float64_t> get_parameters() { return m_params; }

    int32_t get_num_inputs() { return m_num_inputs; }

    int32_t get_num_outputs();

    CDynamicObjectArray* get_layers();

    virtual const char* get_name() const { return "NeuralNetwork";}

    void set_optimization_method(ENNOptimizationMethod optimization_method)
    {
        m_optimization_method = optimization_method;
    }

    ENNOptimizationMethod get_optimization_method() const
    {
        return m_optimization_method;
    }
    void set_l2_coefficient(float64_t l2_coefficient)
    {
        m_l2_coefficient = l2_coefficient;
    }

    float64_t get_l2_coefficient() const
    {
        return m_l2_coefficient;
    }
    void set_l1_coefficient(float64_t l1_coefficient)
    {
        m_l1_coefficient = l1_coefficient;
    }

    float64_t get_l1_coefficient() const
    {
        return m_l1_coefficient;
    }

    void set_dropout_hidden(float64_t dropout_hidden)
    {
        m_dropout_hidden = dropout_hidden;
    }

    float64_t get_dropout_hidden() const
    {
        return m_dropout_hidden;
    }

    void set_dropout_input(float64_t dropout_input)
    {
        m_dropout_input = dropout_input;
    }

    float64_t get_dropout_input() const
    {
        return m_dropout_input;
    }

    void set_max_norm(float64_t max_norm)
    {
        m_max_norm = max_norm;
    }

    float64_t get_max_norm() const
    {
        return m_max_norm;
    }

    void set_epsilon(float64_t epsilon)
    {
        m_epsilon = epsilon;
    }

    float64_t get_epsilon() const
    {
        return m_epsilon;
    }

    void set_max_num_epochs(int32_t max_num_epochs)
    {
        m_max_num_epochs = max_num_epochs;
    }

    int32_t get_max_num_epochs() const
    {
        return m_max_num_epochs;
    }

    void set_gd_mini_batch_size(int32_t gd_mini_batch_size)
    {
        m_gd_mini_batch_size = gd_mini_batch_size;
    }

    int32_t get_gd_mini_batch_size() const
    {
        return m_gd_mini_batch_size;
    }

    void set_gd_learning_rate(float64_t gd_learning_rate)
    {
        m_gd_learning_rate = gd_learning_rate;
    }

    float64_t get_gd_learning_rate() const
    {
        return m_gd_learning_rate;
    }

    void set_gd_learning_rate_decay(float64_t gd_learning_rate_decay)
    {
        m_gd_learning_rate_decay = gd_learning_rate_decay;
    }

    float64_t get_gd_learning_rate_decay() const
    {
        return m_gd_learning_rate_decay;
    }

    void set_gd_momentum(float64_t gd_momentum)
    {
        m_gd_momentum = gd_momentum;
    }

    float64_t get_gd_momentum() const
    {
        return m_gd_momentum;
    }

    void set_gd_error_damping_coeff(float64_t gd_error_damping_coeff)
    {
        m_gd_error_damping_coeff = gd_error_damping_coeff;
    }

    float64_t get_gd_error_damping_coeff() const
    {
        return m_gd_error_damping_coeff;
    }

protected:
    virtual bool train_machine(CFeatures* data=NULL);

    virtual bool train_gradient_descent(SGMatrix<float64_t> inputs,
            SGMatrix<float64_t> targets);

    virtual bool train_lbfgs(SGMatrix<float64_t> inputs,
            SGMatrix<float64_t> targets);

    virtual SGMatrix<float64_t> forward_propagate(CFeatures* data, int32_t j=-1);

    virtual SGMatrix<float64_t> forward_propagate(SGMatrix<float64_t> inputs, int32_t j=-1);

    virtual void set_batch_size(int32_t batch_size);

    virtual float64_t compute_gradients(SGMatrix<float64_t> inputs,
            SGMatrix<float64_t> targets, SGVector<float64_t> gradients);

    virtual float64_t compute_error(SGMatrix<float64_t> inputs,
            SGMatrix<float64_t> targets);

    virtual float64_t compute_error(SGMatrix<float64_t> targets);

    virtual bool is_label_valid(CLabels *lab) const;

    CNeuralLayer* get_layer(int32_t i);

    SGMatrix<float64_t> features_to_matrix(CFeatures* features);

    SGMatrix<float64_t> labels_to_matrix(CLabels* labs);

private:
    void init();

    static float64_t lbfgs_evaluate(void *userdata,
            const float64_t *W,
            float64_t *grad,
            const int32_t n,
            const float64_t step);

    static int lbfgs_progress(void *instance,
            const float64_t *x,
            const float64_t *g,
            const float64_t fx,
            const float64_t xnorm,
            const float64_t gnorm,
            const float64_t step,
            int n,
            int k,
            int ls
            );

    template<class T>
    SGVector<T> get_section(SGVector<T> v, int32_t i);

protected:
    int32_t m_num_inputs;

    int32_t m_num_layers;

    CDynamicObjectArray* m_layers;

    SGMatrix<bool> m_adj_matrix;

    int32_t m_total_num_parameters;

    SGVector<float64_t> m_params;

    SGVector<bool> m_param_regularizable;

    SGVector<int32_t> m_index_offsets;

    int32_t m_batch_size;

    bool m_is_training;

    ENNOptimizationMethod m_optimization_method;

    float64_t m_l2_coefficient;

    float64_t m_l1_coefficient;

    float64_t m_dropout_hidden;

    float64_t m_dropout_input;

    float64_t m_max_norm;

    float64_t m_epsilon;

    int32_t m_max_num_epochs;

    int32_t m_gd_mini_batch_size;

    float64_t m_gd_learning_rate;

    float64_t m_gd_learning_rate_decay;

    float64_t m_gd_momentum;

    float64_t m_gd_error_damping_coeff;

private:
    const SGMatrix<float64_t>* m_lbfgs_temp_inputs;
    const SGMatrix<float64_t>* m_lbfgs_temp_targets;
};

}
#endif