WDSVMOcas.cpp

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) 2007-2008 Vojtech Franc
 * Written (W) 2007-2009 Soeren Sonnenburg
 * Copyright (C) 2007-2009 Fraunhofer Institute FIRST and Max-Planck-Society
 */

#include <shogun/labels/Labels.h>
#include <shogun/mathematics/Math.h>
#include <shogun/lib/DynamicArray.h>
#include <shogun/lib/Time.h>
#include <shogun/base/Parallel.h>
#include <shogun/machine/Machine.h>
#include <shogun/lib/external/libocas.h>
#include <shogun/classifier/svm/WDSVMOcas.h>
#include <shogun/features/StringFeatures.h>
#include <shogun/features/Alphabet.h>
#include <shogun/labels/Labels.h>
#include <shogun/labels/BinaryLabels.h>

using namespace shogun;

#ifndef DOXYGEN_SHOULD_SKIP_THIS
struct wdocas_thread_params_output
{
    float32_t* out;
    int32_t* val;
    float64_t* output;
    CWDSVMOcas* wdocas;
    int32_t start;
    int32_t end;
};

struct wdocas_thread_params_add
{
    CWDSVMOcas* wdocas;
    float32_t* new_a;
    uint32_t* new_cut;
    int32_t start;
    int32_t end;
    uint32_t cut_length;
};
#endif // DOXYGEN_SHOULD_SKIP_THIS

CWDSVMOcas::CWDSVMOcas()
: CMachine(), use_bias(false), bufsize(3000), C1(1), C2(1),
    epsilon(1e-3), method(SVM_OCAS)
{
    SG_UNSTABLE("CWDSVMOcas::CWDSVMOcas()", "\n");

    w=NULL;
    old_w=NULL;
    features=NULL;
    degree=6;
    from_degree=40;
    wd_weights=NULL;
    w_offsets=NULL;
    normalization_const=1.0;
}

CWDSVMOcas::CWDSVMOcas(E_SVM_TYPE type)
: CMachine(), use_bias(false), bufsize(3000), C1(1), C2(1),
    epsilon(1e-3), method(type)
{
    w=NULL;
    old_w=NULL;
    features=NULL;
    degree=6;
    from_degree=40;
    wd_weights=NULL;
    w_offsets=NULL;
    normalization_const=1.0;
}

CWDSVMOcas::CWDSVMOcas(
    float64_t C, int32_t d, int32_t from_d, CStringFeatures<uint8_t>* traindat,
    CLabels* trainlab)
: CMachine(), use_bias(false), bufsize(3000), C1(C), C2(C), epsilon(1e-3),
    degree(d), from_degree(from_d)
{
    w=NULL;
    old_w=NULL;
    method=SVM_OCAS;
    features=traindat;
    set_labels(trainlab);
    wd_weights=NULL;
    w_offsets=NULL;
    normalization_const=1.0;
}


CWDSVMOcas::~CWDSVMOcas()
{
}

CBinaryLabels* CWDSVMOcas::apply_binary(CFeatures* data)
{
    SGVector<float64_t> outputs = apply_get_outputs(data);
    return new CBinaryLabels(outputs);
}

CRegressionLabels* CWDSVMOcas::apply_regression(CFeatures* data)
{
    SGVector<float64_t> outputs = apply_get_outputs(data);
    return new CRegressionLabels(outputs);
}

SGVector<float64_t> CWDSVMOcas::apply_get_outputs(CFeatures* data)
{
    if (data)
    {
        if (data->get_feature_class() != C_STRING ||
                data->get_feature_type() != F_BYTE)
        {
            SG_ERROR("Features not of class string type byte\n");
        }

        set_features((CStringFeatures<uint8_t>*) data);
    }
    ASSERT(features);

    set_wd_weights();
    set_normalization_const();

    SGVector<float64_t> outputs;
    if (features)
    {
        int32_t num=features->get_num_vectors();
        ASSERT(num>0);

        outputs = SGVector<float64_t>(num);

        for (int32_t i=0; i<num; i++)
            outputs[i] = apply_one(i);
    }

    return outputs;
}

int32_t CWDSVMOcas::set_wd_weights()
{
    ASSERT(degree>0 && degree<=8);
    SG_FREE(wd_weights);
    wd_weights=SG_MALLOC(float32_t, degree);
    SG_FREE(w_offsets);
    w_offsets=SG_MALLOC(int32_t, degree);
    int32_t w_dim_single_c=0;

    for (int32_t i=0; i<degree; i++)
    {
        w_offsets[i]=CMath::pow(alphabet_size, i+1);
        wd_weights[i]=sqrt(2.0*(from_degree-i)/(from_degree*(from_degree+1)));
        w_dim_single_c+=w_offsets[i];
    }
    return w_dim_single_c;
}

bool CWDSVMOcas::train_machine(CFeatures* data)
{
    SG_INFO("C=%f, epsilon=%f, bufsize=%d\n", get_C1(), get_epsilon(), bufsize);

    ASSERT(m_labels);
    ASSERT(m_labels->get_label_type() == LT_BINARY);
    if (data)
    {
        if (data->get_feature_class() != C_STRING ||
                data->get_feature_type() != F_BYTE)
        {
            SG_ERROR("Features not of class string type byte\n");
        }
        set_features((CStringFeatures<uint8_t>*) data);
    }

    ASSERT(get_features());
    CAlphabet* alphabet=get_features()->get_alphabet();
    ASSERT(alphabet && alphabet->get_alphabet()==RAWDNA);

    alphabet_size=alphabet->get_num_symbols();
    string_length=features->get_num_vectors();
    SGVector<float64_t> labvec=((CBinaryLabels*) m_labels)->get_labels();
    lab=labvec.vector;

    w_dim_single_char=set_wd_weights();
    //CMath::display_vector(wd_weights, degree, "wd_weights");
    SG_DEBUG("w_dim_single_char=%d\n", w_dim_single_char);
    w_dim=string_length*w_dim_single_char;
    SG_DEBUG("cutting plane has %d dims\n", w_dim);
    num_vec=get_features()->get_max_vector_length();

    set_normalization_const();
    SG_INFO("num_vec: %d num_lab: %d\n", num_vec, labvec.vlen);
    ASSERT(num_vec==labvec.vlen);
    ASSERT(num_vec>0);

    SG_FREE(w);
    w=SG_MALLOC(float32_t, w_dim);
    memset(w, 0, w_dim*sizeof(float32_t));

    SG_FREE(old_w);
    old_w=SG_MALLOC(float32_t, w_dim);
    memset(old_w, 0, w_dim*sizeof(float32_t));
    bias=0;
    old_bias=0;

    cuts=SG_MALLOC(float32_t*, bufsize);
    memset(cuts, 0, sizeof(*cuts)*bufsize);
    cp_bias=SG_MALLOC(float64_t, bufsize);
    memset(cp_bias, 0, sizeof(float64_t)*bufsize);

    /*float64_t* tmp = SG_MALLOC(float64_t, num_vec);
    float64_t start=CTime::get_curtime();
    CMath::random_vector(w, w_dim, (float32_t) 0, (float32_t) 1000);
    compute_output(tmp, this);
    start=CTime::get_curtime()-start;
    SG_PRINT("timing:%f\n", start);
    SG_FREE(tmp);
    exit(1);*/
    float64_t TolAbs=0;
    float64_t QPBound=0;
    uint8_t Method=0;
    if (method == SVM_OCAS)
        Method = 1;
    ocas_return_value_T result = svm_ocas_solver( get_C1(), num_vec, get_epsilon(),
            TolAbs, QPBound, get_max_train_time(), bufsize, Method,
            &CWDSVMOcas::compute_W,
            &CWDSVMOcas::update_W,
            &CWDSVMOcas::add_new_cut,
            &CWDSVMOcas::compute_output,
            &CWDSVMOcas::sort,
            &CWDSVMOcas::print,
            this);

    SG_INFO("Ocas Converged after %d iterations\n"
            "==================================\n"
            "timing statistics:\n"
            "output_time: %f s\n"
            "sort_time: %f s\n"
            "add_time: %f s\n"
            "w_time: %f s\n"
            "solver_time %f s\n"
            "ocas_time %f s\n\n", result.nIter, result.output_time, result.sort_time,
            result.add_time, result.w_time, result.qp_solver_time, result.ocas_time);

    for (int32_t i=bufsize-1; i>=0; i--)
        SG_FREE(cuts[i]);
    SG_FREE(cuts);

    lab=NULL;
    SG_UNREF(alphabet);

    return true;
}

/*----------------------------------------------------------------------------------
  sq_norm_W = sparse_update_W( t ) does the following:

  W = oldW*(1-t) + t*W;
  sq_norm_W = W'*W;

  ---------------------------------------------------------------------------------*/
float64_t CWDSVMOcas::update_W( float64_t t, void* ptr )
{
  float64_t sq_norm_W = 0;
  CWDSVMOcas* o = (CWDSVMOcas*) ptr;
  uint32_t nDim = (uint32_t) o->w_dim;
  float32_t* W=o->w;
  float32_t* oldW=o->old_w;
  float64_t bias=o->bias;
  float64_t old_bias=bias;

  for(uint32_t j=0; j <nDim; j++)
  {
      W[j] = oldW[j]*(1-t) + t*W[j];
      sq_norm_W += W[j]*W[j];
  }

  bias=old_bias*(1-t) + t*bias;
  sq_norm_W += CMath::sq(bias);

  o->bias=bias;
  o->old_bias=old_bias;

  return( sq_norm_W );
}

/*----------------------------------------------------------------------------------
  sparse_add_new_cut( new_col_H, new_cut, cut_length, nSel ) does the following:

    new_a = sum(data_X(:,find(new_cut ~=0 )),2);
    new_col_H = [sparse_A(:,1:nSel)'*new_a ; new_a'*new_a];
    sparse_A(:,nSel+1) = new_a;

  ---------------------------------------------------------------------------------*/
void* CWDSVMOcas::add_new_cut_helper( void* ptr)
{
    wdocas_thread_params_add* p = (wdocas_thread_params_add*) ptr;
    CWDSVMOcas* o = p->wdocas;
    int32_t start = p->start;
    int32_t end = p->end;
    int32_t string_length = o->string_length;
    //uint32_t nDim=(uint32_t) o->w_dim;
    uint32_t cut_length=p->cut_length;
    uint32_t* new_cut=p->new_cut;
    int32_t* w_offsets = o->w_offsets;
    float64_t* y = o->lab;
    int32_t alphabet_size = o->alphabet_size;
    float32_t* wd_weights = o->wd_weights;
    int32_t degree = o->degree;
    CStringFeatures<uint8_t>* f = o->features;
    float64_t normalization_const = o->normalization_const;

    // temporary vector
    float32_t* new_a = p->new_a;
    //float32_t* new_a = SG_MALLOC(float32_t, nDim);
    //memset(new_a, 0, sizeof(float32_t)*nDim);

    int32_t* val=SG_MALLOC(int32_t, cut_length);
    for (int32_t j=start; j<end; j++)
    {
        int32_t offs=o->w_dim_single_char*j;
        memset(val,0,sizeof(int32_t)*cut_length);
        int32_t lim=CMath::min(degree, string_length-j);
        int32_t len;

        for (int32_t k=0; k<lim; k++)
        {
            bool free_vec;
            uint8_t* vec = f->get_feature_vector(j+k, len, free_vec);
            float32_t wd = wd_weights[k]/normalization_const;

            for(uint32_t i=0; i < cut_length; i++)
            {
                val[i]=val[i]*alphabet_size + vec[new_cut[i]];
                new_a[offs+val[i]]+=wd * y[new_cut[i]];
            }
            offs+=w_offsets[k];
            f->free_feature_vector(vec, j+k, free_vec);
        }
    }

    //p->new_a=new_a;
    SG_FREE(val);
    return NULL;
}

int CWDSVMOcas::add_new_cut(
    float64_t *new_col_H, uint32_t *new_cut, uint32_t cut_length,
    uint32_t nSel, void* ptr)
{
    CWDSVMOcas* o = (CWDSVMOcas*) ptr;
    int32_t string_length = o->string_length;
    uint32_t nDim=(uint32_t) o->w_dim;
    float32_t** cuts=o->cuts;
    float64_t* c_bias = o->cp_bias;

    uint32_t i;
    wdocas_thread_params_add* params_add=SG_MALLOC(wdocas_thread_params_add, o->parallel->get_num_threads());
    pthread_t* threads=SG_MALLOC(pthread_t, o->parallel->get_num_threads());
    float32_t* new_a=SG_MALLOC(float32_t, nDim);
    memset(new_a, 0, sizeof(float32_t)*nDim);

    int32_t t;
    int32_t nthreads=o->parallel->get_num_threads()-1;
    int32_t step= string_length/o->parallel->get_num_threads();

    if (step<1)
    {
        nthreads=string_length-1;
        step=1;
    }

#ifdef HAVE_PTHREAD
    for (t=0; t<nthreads; t++)
    {
        params_add[t].wdocas=o;
        //params_add[t].new_a=NULL;
        params_add[t].new_a=new_a;
        params_add[t].new_cut=new_cut;
        params_add[t].start = step*t;
        params_add[t].end = step*(t+1);
        params_add[t].cut_length = cut_length;

        if (pthread_create(&threads[t], NULL, &CWDSVMOcas::add_new_cut_helper, (void*)&params_add[t]) != 0)
        {
            nthreads=t;
            SG_SWARNING("thread creation failed\n");
            break;
        }
    }

    params_add[t].wdocas=o;
    //params_add[t].new_a=NULL;
    params_add[t].new_a=new_a;
    params_add[t].new_cut=new_cut;
    params_add[t].start = step*t;
    params_add[t].end = string_length;
    params_add[t].cut_length = cut_length;
    add_new_cut_helper(&params_add[t]);
    //float32_t* new_a=params_add[t].new_a;

    for (t=0; t<nthreads; t++)
    {
        if (pthread_join(threads[t], NULL) != 0)
            SG_SWARNING( "pthread_join failed\n");

        //float32_t* a=params_add[t].new_a;
        //for (i=0; i<nDim; i++)
        //  new_a[i]+=a[i];
        //SG_FREE(a);
    }
#endif /* HAVE_PTHREAD */
    for(i=0; i < cut_length; i++)
    {
        if (o->use_bias)
            c_bias[nSel]+=o->lab[new_cut[i]];
    }

    // insert new_a into the last column of sparse_A
    for(i=0; i < nSel; i++)
        new_col_H[i] = SGVector<float32_t>::dot(new_a, cuts[i], nDim) + c_bias[nSel]*c_bias[i];
    new_col_H[nSel] = SGVector<float32_t>::dot(new_a, new_a, nDim) + CMath::sq(c_bias[nSel]);

    cuts[nSel]=new_a;
    //CMath::display_vector(new_col_H, nSel+1, "new_col_H");
    //CMath::display_vector(cuts[nSel], nDim, "cut[nSel]");
    //
    SG_FREE(threads);
    SG_FREE(params_add);

    return 0;
}

int CWDSVMOcas::sort( float64_t* vals, float64_t* data, uint32_t size)
{
    CMath::qsort_index(vals, data, size);
    return 0;
}

/*----------------------------------------------------------------------
  sparse_compute_output( output ) does the follwing:

  output = data_X'*W;
  ----------------------------------------------------------------------*/
void* CWDSVMOcas::compute_output_helper(void* ptr)
{
    wdocas_thread_params_output* p = (wdocas_thread_params_output*) ptr;
    CWDSVMOcas* o = p->wdocas;
    int32_t start = p->start;
    int32_t end = p->end;
    float32_t* out = p->out;
    float64_t* output = p->output;
    int32_t* val = p->val;

    CStringFeatures<uint8_t>* f=o->get_features();

    int32_t degree = o->degree;
    int32_t string_length = o->string_length;
    int32_t alphabet_size = o->alphabet_size;
    int32_t* w_offsets = o->w_offsets;
    float32_t* wd_weights = o->wd_weights;
    float32_t* w= o->w;

    float64_t* y = o->lab;
    float64_t normalization_const = o->normalization_const;


    for (int32_t j=0; j<string_length; j++)
    {
        int32_t offs=o->w_dim_single_char*j;
        for (int32_t i=start ; i<end; i++)
            val[i]=0;

        int32_t lim=CMath::min(degree, string_length-j);
        int32_t len;

        for (int32_t k=0; k<lim; k++)
        {
            bool free_vec;
            uint8_t* vec=f->get_feature_vector(j+k, len, free_vec);
            float32_t wd = wd_weights[k];

            for (int32_t i=start; i<end; i++) // quite fast 1.9s
            {
                val[i]=val[i]*alphabet_size + vec[i];
                out[i]+=wd*w[offs+val[i]];
            }

            /*for (int32_t i=0; i<nData/4; i++) // slowest 2s
            {
                uint32_t x=((uint32_t*) vec)[i];
                int32_t ii=4*i;
                val[ii]=val[ii]*alphabet_size + (x&255);
                val[ii+1]=val[ii+1]*alphabet_size + ((x>>8)&255);
                val[ii+2]=val[ii+2]*alphabet_size + ((x>>16)&255);
                val[ii+3]=val[ii+3]*alphabet_size + (x>>24);
                out[ii]+=wd*w[offs+val[ii]];
                out[ii+1]+=wd*w[offs+val[ii+1]];
                out[ii+2]+=wd*w[offs+val[ii+2]];
                out[ii+3]+=wd*w[offs+val[ii+3]];
            }*/

            /*for (int32_t i=0; i<nData>>3; i++) // fastest on 64bit: 1.5s
            {
                uint64_t x=((uint64_t*) vec)[i];
                int32_t ii=i<<3;
                val[ii]=val[ii]*alphabet_size + (x&255);
                val[ii+1]=val[ii+1]*alphabet_size + ((x>>8)&255);
                val[ii+2]=val[ii+2]*alphabet_size + ((x>>16)&255);
                val[ii+3]=val[ii+3]*alphabet_size + ((x>>24)&255);
                val[ii+4]=val[ii+4]*alphabet_size + ((x>>32)&255);
                val[ii+5]=val[ii+5]*alphabet_size + ((x>>40)&255);
                val[ii+6]=val[ii+6]*alphabet_size + ((x>>48)&255);
                val[ii+7]=val[ii+7]*alphabet_size + (x>>56);
                out[ii]+=wd*w[offs+val[ii]];
                out[ii+1]+=wd*w[offs+val[ii+1]];
                out[ii+2]+=wd*w[offs+val[ii+2]];
                out[ii+3]+=wd*w[offs+val[ii+3]];
                out[ii+4]+=wd*w[offs+val[ii+4]];
                out[ii+5]+=wd*w[offs+val[ii+5]];
                out[ii+6]+=wd*w[offs+val[ii+6]];
                out[ii+7]+=wd*w[offs+val[ii+7]];
            }*/
            offs+=w_offsets[k];
            f->free_feature_vector(vec, j+k, free_vec);
        }
    }

    for (int32_t i=start; i<end; i++)
        output[i]=y[i]*o->bias + out[i]*y[i]/normalization_const;

    //CMath::display_vector(o->w, o->w_dim, "w");
    //CMath::display_vector(output, nData, "out");
    return NULL;
}

int CWDSVMOcas::compute_output( float64_t *output, void* ptr )
{
    CWDSVMOcas* o = (CWDSVMOcas*) ptr;
    int32_t nData=o->num_vec;
    wdocas_thread_params_output* params_output=SG_MALLOC(wdocas_thread_params_output, o->parallel->get_num_threads());
    pthread_t* threads = SG_MALLOC(pthread_t, o->parallel->get_num_threads());

    float32_t* out=SG_MALLOC(float32_t, nData);
    int32_t* val=SG_MALLOC(int32_t, nData);
    memset(out, 0, sizeof(float32_t)*nData);

    int32_t t;
    int32_t nthreads=o->parallel->get_num_threads()-1;
    int32_t step= nData/o->parallel->get_num_threads();

    if (step<1)
    {
        nthreads=nData-1;
        step=1;
    }
#ifdef HAVE_PTHREAD
    for (t=0; t<nthreads; t++)
    {
        params_output[t].wdocas=o;
        params_output[t].output=output;
        params_output[t].out=out;
        params_output[t].val=val;
        params_output[t].start = step*t;
        params_output[t].end = step*(t+1);

        //SG_SPRINT("t=%d start=%d end=%d output=%p\n", t, params_output[t].start, params_output[t].end, params_output[t].output);
        if (pthread_create(&threads[t], NULL, &CWDSVMOcas::compute_output_helper, (void*)&params_output[t]) != 0)
        {
            nthreads=t;
            SG_SWARNING("thread creation failed\n");
            break;
        }
    }

    params_output[t].wdocas=o;
    params_output[t].output=output;
    params_output[t].out=out;
    params_output[t].val=val;
    params_output[t].start = step*t;
    params_output[t].end = nData;
    compute_output_helper(&params_output[t]);
    //SG_SPRINT("t=%d start=%d end=%d output=%p\n", t, params_output[t].start, params_output[t].end, params_output[t].output);

    for (t=0; t<nthreads; t++)
    {
        if (pthread_join(threads[t], NULL) != 0)
            SG_SWARNING( "pthread_join failed\n");
    }
    SG_FREE(threads);
    SG_FREE(params_output);
    SG_FREE(val);
    SG_FREE(out);
#endif /* HAVE_PTHREAD */
    return 0;
}
/*----------------------------------------------------------------------
  sq_norm_W = compute_W( alpha, nSel ) does the following:

  oldW = W;
  W = sparse_A(:,1:nSel)'*alpha;
  sq_norm_W = W'*W;
  dp_WoldW = W'*oldW';

  ----------------------------------------------------------------------*/
void CWDSVMOcas::compute_W(
    float64_t *sq_norm_W, float64_t *dp_WoldW, float64_t *alpha, uint32_t nSel,
    void* ptr)
{
    CWDSVMOcas* o = (CWDSVMOcas*) ptr;
    uint32_t nDim= (uint32_t) o->w_dim;
    CMath::swap(o->w, o->old_w);
    float32_t* W=o->w;
    float32_t* oldW=o->old_w;
    float32_t** cuts=o->cuts;
    memset(W, 0, sizeof(float32_t)*nDim);
    float64_t* c_bias = o->cp_bias;
    float64_t old_bias=o->bias;
    float64_t bias=0;

    for (uint32_t i=0; i<nSel; i++)
    {
        if (alpha[i] > 0)
            SGVector<float32_t>::vec1_plus_scalar_times_vec2(W, (float32_t) alpha[i], cuts[i], nDim);

        bias += c_bias[i]*alpha[i];
    }

    *sq_norm_W = SGVector<float32_t>::dot(W,W, nDim) +CMath::sq(bias);
    *dp_WoldW = SGVector<float32_t>::dot(W,oldW, nDim) + bias*old_bias;;
    //SG_PRINT("nSel=%d sq_norm_W=%f dp_WoldW=%f\n", nSel, *sq_norm_W, *dp_WoldW);

    o->bias = bias;
    o->old_bias = old_bias;
}