WeightedDegreePositionStringKernel.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 _WEIGHTEDDEGREEPOSITIONSTRINGKERNEL_H___
#define _WEIGHTEDDEGREEPOSITIONSTRINGKERNEL_H___

#include <shogun/lib/common.h>
#include <shogun/kernel/string/StringKernel.h>
#include <shogun/kernel/string/WeightedDegreeStringKernel.h>
#include <shogun/lib/Trie.h>

namespace shogun
{

class CSVM;

class CWeightedDegreePositionStringKernel: public CStringKernel<char>
{
    public:
        CWeightedDegreePositionStringKernel();

        CWeightedDegreePositionStringKernel(
            int32_t size, int32_t degree,
            int32_t max_mismatch=0, int32_t mkl_stepsize=1);

        CWeightedDegreePositionStringKernel(
            int32_t size, float64_t* weights, int32_t degree,
            int32_t max_mismatch, int32_t* shift, int32_t shift_len,
            int32_t mkl_stepsize=1);

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

        virtual ~CWeightedDegreePositionStringKernel();

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

        virtual void cleanup();

        virtual EKernelType get_kernel_type() { return K_WEIGHTEDDEGREEPOS; }

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

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

        virtual bool init_optimization(
            int32_t count, int32_t *IDX, float64_t * alphas, int32_t tree_num,
            int32_t upto_tree=-1);

        virtual bool delete_optimization();

        virtual float64_t compute_optimized(int32_t idx)
        {
            ASSERT(get_is_initialized());
            ASSERT(alphabet);
            ASSERT(alphabet->get_alphabet()==DNA || alphabet->get_alphabet()==RNA);
            return compute_by_tree(idx);
        }

        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 ((opt_type==FASTBUTMEMHUNGRY) && (tries.get_use_compact_terminal_nodes()))
            {
                tries.set_use_compact_terminal_nodes(false) ;
                SG_DEBUG( "disabling compact trie nodes with FASTBUTMEMHUNGRY\n") ;
            }

            if (get_is_initialized())
            {
                if (opt_type==SLOWBUTMEMEFFICIENT)
                    tries.delete_trees(true);
                else if (opt_type==FASTBUTMEMHUNGRY)
                    tries.delete_trees(false);  // still buggy
                else
                    SG_ERROR( "unknown optimization type\n");

                set_is_initialized(false);
            }
        }

        virtual void add_to_normal(int32_t idx, float64_t weight)
        {
            add_example_to_tree(idx, weight);
            set_is_initialized(true);
        }

        virtual int32_t get_num_subkernels()
        {
            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())
            {
                compute_by_tree(idx, subkernel_contrib);
                return ;
            }

            SG_ERROR( "CWeightedDegreePositionStringKernel 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 (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 (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] ;
            }
        }

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

        bool is_tree_initialized() { return tree_initialized; }

        inline int32_t get_max_mismatch() { return max_mismatch; }

        inline int32_t get_degree() { return degree; }

        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 (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;
        }

        void set_shifts(SGVector<int32_t> shifts);

        bool set_weights(SGMatrix<float64_t> new_weights);

        virtual bool set_wd_weights();

        virtual void set_position_weights(SGVector<float64_t> pws);

        bool set_position_weights_lhs(float64_t* pws, int32_t len, int32_t num);

        bool set_position_weights_rhs(float64_t* pws, int32_t len, int32_t num);

        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 delete_position_weights_lhs()
        {
            SG_FREE(position_weights_lhs);
            position_weights_lhs=NULL;
            return true;
        }

        bool delete_position_weights_rhs()
        {
            SG_FREE(position_weights_rhs);
            position_weights_rhs=NULL;
            return true;
        }

        virtual float64_t compute_by_tree(int32_t idx);

        virtual void compute_by_tree(int32_t idx, float64_t* LevelContrib);

        float64_t* compute_scoring(
            int32_t max_degree, int32_t& num_feat, int32_t& num_sym,
            float64_t* target, int32_t num_suppvec, int32_t* IDX,
            float64_t* weights);

        char* compute_consensus(
            int32_t &num_feat, int32_t num_suppvec, int32_t* IDX,
            float64_t* alphas);

        float64_t* extract_w(
            int32_t max_degree, int32_t& num_feat, int32_t& num_sym,
            float64_t* w_result, int32_t num_suppvec, int32_t* IDX,
            float64_t* alphas);

        float64_t* compute_POIM(
            int32_t max_degree, int32_t& num_feat, int32_t& num_sym,
            float64_t* poim_result, int32_t num_suppvec, int32_t* IDX,
            float64_t* alphas, float64_t* distrib);

        void prepare_POIM2(SGMatrix<float64_t> distrib);

        void compute_POIM2(int32_t max_degree, CSVM* svm);

        SGVector<float64_t> get_POIM2();

        void cleanup_POIM2();

    protected:
        void create_empty_tries();

        virtual 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);

        virtual 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_without_mismatch_position_weights(
            char* avec, float64_t *posweights_lhs, int32_t alen,
            char* bvec, float64_t *posweights_rhs, int32_t blen);

        virtual void remove_lhs();

        virtual void load_serializable_post() throw (ShogunException);

    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* position_weights_lhs;
        int32_t position_weights_lhs_len;
        float64_t* position_weights_rhs;
        int32_t position_weights_rhs_len;
        bool* position_mask;

        float64_t* weights_buffer;
        int32_t mkl_stepsize;

        int32_t degree;
        int32_t length;

        int32_t max_mismatch;
        int32_t seq_length;

        int32_t *shift;
        int32_t shift_len;
        int32_t max_shift;

        bool block_computation;

        float64_t* block_weights;
        EWDKernType type;
        int32_t which_degree;

        CTrie<DNATrie> tries;
        CTrie<POIMTrie> poim_tries;

        bool tree_initialized;
        bool use_poim_tries;

        float64_t* m_poim_distrib;
        float64_t* m_poim;

        int32_t m_poim_num_sym;
        int32_t m_poim_num_feat;
        int32_t m_poim_result_len;

        CAlphabet* alphabet;
};
}
#endif /* _WEIGHTEDDEGREEPOSITIONSTRINGKERNEL_H__ */