WeightedDegreeStringKernel.h

Go to the documentation of this file.

/*
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation; either version 3 of the License, or
 * (at your option) any later version.
 *
 * Written (W) 1999-2009 Soeren Sonnenburg
 * Written (W) 1999-2008 Gunnar Raetsch
 * Copyright (C) 1999-2009 Fraunhofer Institute FIRST and Max-Planck-Society
 */

#ifndef _WEIGHTEDDEGREESTRINGKERNEL_H___
#define _WEIGHTEDDEGREESTRINGKERNEL_H___

#include <shogun/lib/config.h>

#include <shogun/lib/common.h>
#include <shogun/lib/Trie.h>
#include <shogun/kernel/string/StringKernel.h>
#include <shogun/transfer/multitask/MultitaskKernelMklNormalizer.h>
#include <shogun/features/StringFeatures.h>

namespace shogun
{

enum EWDKernType
{
    E_WD=0,
    E_EXTERNAL=1,

    E_BLOCK_CONST=2,
    E_BLOCK_LINEAR=3,
    E_BLOCK_SQPOLY=4,
    E_BLOCK_CUBICPOLY=5,
    E_BLOCK_EXP=6,
    E_BLOCK_LOG=7,
};


class CWeightedDegreeStringKernel: public CStringKernel<char>
{
    public:

        CWeightedDegreeStringKernel();


        CWeightedDegreeStringKernel(int32_t degree, EWDKernType type=E_WD);

        CWeightedDegreeStringKernel(SGVector<float64_t> weights);

        CWeightedDegreeStringKernel(
            CStringFeatures<char>* l, CStringFeatures<char>* r, int32_t degree);

        virtual ~CWeightedDegreeStringKernel();

        virtual bool init(CFeatures* l, CFeatures* r);

        virtual void cleanup();

        EWDKernType get_type() const
        {
            return type;
        }

        virtual EKernelType get_kernel_type() { return K_WEIGHTEDDEGREE; }

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

        virtual bool init_optimization(
            int32_t count, int32_t *IDX, float64_t* alphas)
        {
            return init_optimization(count, IDX, alphas, -1);
        }

        virtual bool init_optimization(
            int32_t count, int32_t *IDX, float64_t* alphas, int32_t tree_num);

        virtual bool delete_optimization();

        virtual float64_t compute_optimized(int32_t idx)
        {
            if (get_is_initialized())
                return compute_by_tree(idx);

            SG_ERROR("CWeightedDegreeStringKernel optimization not initialized\n")
            return 0;
        }

        static void* compute_batch_helper(void* p);

        virtual void compute_batch(
            int32_t num_vec, int32_t* vec_idx, float64_t* target,
            int32_t num_suppvec, int32_t* IDX, float64_t* alphas,
            float64_t factor=1.0);

        virtual void clear_normal()
        {
            if (get_is_initialized())
            {

                if (normalizer && normalizer->get_normalizer_type()==N_MULTITASK)
                    SG_ERROR("not implemented")

                tries->delete_trees(max_mismatch==0);
                set_is_initialized(false);
            }
        }

        virtual void add_to_normal(int32_t idx, float64_t weight)
        {

            if (normalizer && normalizer->get_normalizer_type()==N_MULTITASK)
                SG_ERROR("not implemented")

            if (max_mismatch==0)
                add_example_to_tree(idx, weight);
            else
                add_example_to_tree_mismatch(idx, weight);

            set_is_initialized(true);
        }

        virtual int32_t get_num_subkernels()
        {
            if (normalizer && normalizer->get_normalizer_type()==N_MULTITASK)
                return ((CMultitaskKernelMklNormalizer*)normalizer)->get_num_betas();
            if (position_weights!=NULL)
                return (int32_t) ceil(1.0*seq_length/mkl_stepsize) ;
            if (length==0)
                return (int32_t) ceil(1.0*get_degree()/mkl_stepsize);
            return (int32_t) ceil(1.0*get_degree()*length/mkl_stepsize) ;
        }

        inline void compute_by_subkernel(
            int32_t idx, float64_t * subkernel_contrib)
        {

            if (get_is_initialized())
            {

                if (normalizer && normalizer->get_normalizer_type()==N_MULTITASK)
                    SG_ERROR("not implemented")

                compute_by_tree(idx, subkernel_contrib);
                return ;
            }

            SG_ERROR("CWeightedDegreeStringKernel optimization not initialized\n")
        }

        inline const float64_t* get_subkernel_weights(int32_t& num_weights)
        {

            num_weights = get_num_subkernels();

            SG_FREE(weights_buffer);
            weights_buffer = SG_MALLOC(float64_t, num_weights);

            if (normalizer && normalizer->get_normalizer_type()==N_MULTITASK)
                for (int32_t i=0; i<num_weights; i++)
                    weights_buffer[i] = ((CMultitaskKernelMklNormalizer*)normalizer)->get_beta(i);
            else if (position_weights!=NULL)
                for (int32_t i=0; i<num_weights; i++)
                    weights_buffer[i] = position_weights[i*mkl_stepsize];
            else
                for (int32_t i=0; i<num_weights; i++)
                    weights_buffer[i] = weights[i*mkl_stepsize];

            return weights_buffer;
        }

        virtual void set_subkernel_weights(SGVector<float64_t> w)
        {
            float64_t* weights2=w.vector;
            int32_t num_weights2=w.vlen;
            int32_t num_weights = get_num_subkernels();
            if (num_weights!=num_weights2)
                SG_ERROR("number of weights do not match\n")


            if (normalizer && normalizer->get_normalizer_type()==N_MULTITASK)
                for (int32_t i=0; i<num_weights; i++)
                    ((CMultitaskKernelMklNormalizer*)normalizer)->set_beta(i, weights2[i]);
            else if (position_weights!=NULL)
            {
                for (int32_t i=0; i<num_weights; i++)
                {
                    for (int32_t j=0; j<mkl_stepsize; j++)
                    {
                        if (i*mkl_stepsize+j<seq_length)
                            position_weights[i*mkl_stepsize+j] = weights2[i];
                    }
                }
            }
            else if (length==0)
            {
                for (int32_t i=0; i<num_weights; i++)
                {
                    for (int32_t j=0; j<mkl_stepsize; j++)
                    {
                        if (i*mkl_stepsize+j<get_degree())
                            weights[i*mkl_stepsize+j] = weights2[i];
                    }
                }
            }
            else
            {
                for (int32_t i=0; i<num_weights; i++)
                {
                    for (int32_t j=0; j<mkl_stepsize; j++)
                    {
                        if (i*mkl_stepsize+j<get_degree()*length)
                            weights[i*mkl_stepsize+j] = weights2[i];
                    }
                }
            }
        }

        virtual bool set_normalizer(CKernelNormalizer* normalizer_) {

            if (normalizer_ && strcmp(normalizer_->get_name(),"MultitaskKernelTreeNormalizer")==0) {
                unset_property(KP_LINADD);
                unset_property(KP_BATCHEVALUATION);
            }
            else
            {
                set_property(KP_LINADD);
                set_property(KP_BATCHEVALUATION);
            }


            return CStringKernel<char>::set_normalizer(normalizer_);

        }

        // other kernel tree operations
        float64_t *compute_abs_weights(int32_t & len);

        void compute_by_tree(int32_t idx, float64_t *LevelContrib);

        bool is_tree_initialized() { return tree_initialized; }

        inline float64_t *get_degree_weights(int32_t& d, int32_t& len)
        {
            d=degree;
            len=length;
            return weights;
        }

        inline float64_t *get_weights(int32_t& num_weights)
        {

            if (normalizer && normalizer->get_normalizer_type()==N_MULTITASK)
                SG_ERROR("not implemented")

            if (position_weights!=NULL)
            {
                num_weights = seq_length ;
                return position_weights ;
            }
            if (length==0)
                num_weights = degree ;
            else
                num_weights = degree*length ;
            return weights;
        }

        inline float64_t *get_position_weights(int32_t& len)
        {
            len=seq_length;
            return position_weights;
        }

        bool set_wd_weights_by_type(EWDKernType type);

        inline void set_wd_weights(SGVector<float64_t> new_weights)
        {
            SGMatrix<float64_t> matrix = SGMatrix<float64_t>(new_weights.vector,new_weights.vlen,0);
            set_weights(matrix);
            matrix.matrix = NULL;
        }

        bool set_weights(SGMatrix<float64_t> new_weights);

        bool set_position_weights(float64_t* pws, int32_t len);

        bool init_block_weights();

        bool init_block_weights_from_wd();

        bool init_block_weights_from_wd_external();

        bool init_block_weights_const();

        bool init_block_weights_linear();

        bool init_block_weights_sqpoly();

        bool init_block_weights_cubicpoly();

        bool init_block_weights_exp();

        bool init_block_weights_log();

        bool delete_position_weights()
        {
            SG_FREE(position_weights);
            position_weights=NULL;
            return true;
        }

        bool set_max_mismatch(int32_t max);

        inline int32_t get_max_mismatch() const { return max_mismatch; }

        inline bool set_degree(int32_t deg) { degree=deg; return true; }

        inline int32_t get_degree() const { return degree; }

        inline bool set_use_block_computation(bool block)
        {
            block_computation=block;
            return true;
        }

        inline bool get_use_block_computation() { return block_computation; }

        inline bool set_mkl_stepsize(int32_t step)
        {
            if (step<1)
                SG_ERROR("Stepsize must be a positive integer\n")
            mkl_stepsize=step;
            return true;
        }

        inline int32_t get_mkl_stepsize() { return mkl_stepsize; }

        inline bool set_which_degree(int32_t which)
        {
            which_degree=which;
            return true;
        }

        inline int32_t get_which_degree() { return which_degree; }

    protected:
        void create_empty_tries();

        void add_example_to_tree(int32_t idx, float64_t weight);

        void add_example_to_single_tree(
            int32_t idx, float64_t weight, int32_t tree_num);

        void add_example_to_tree_mismatch(int32_t idx, float64_t weight);

        void add_example_to_single_tree_mismatch(
            int32_t idx, float64_t weight, int32_t tree_num);

        float64_t compute_by_tree(int32_t idx);

        float64_t compute(int32_t idx_a, int32_t idx_b);

        float64_t compute_with_mismatch(
            char* avec, int32_t alen, char* bvec, int32_t blen);

        float64_t compute_without_mismatch(
            char* avec, int32_t alen, char* bvec, int32_t blen);

        float64_t compute_without_mismatch_matrix(
            char* avec, int32_t alen, char* bvec, int32_t blen);

        float64_t compute_using_block(char* avec, int32_t alen,
            char* bvec, int32_t blen);

        virtual void remove_lhs();

    private:
        void init();

    protected:
        float64_t* weights;
        int32_t weights_degree;
        int32_t weights_length;


        float64_t* position_weights;
        int32_t position_weights_len;
        float64_t* weights_buffer;
        int32_t mkl_stepsize;
        int32_t degree;
        int32_t length;

        int32_t max_mismatch;
        int32_t seq_length;

        bool initialized;

        bool block_computation;

        float64_t* block_weights;
        EWDKernType type;
        int32_t which_degree;

        CTrie<DNATrie>* tries;

        bool tree_initialized;

        CAlphabet* alphabet;
};

}

#endif /* _WEIGHTEDDEGREESTRINGKERNEL_H__ */