Beispiel #1
0
LABEL       find_most_violated_constraint_slackrescaling(PATTERNX x, LABEL y, 
						     STRUCTMODEL *sm, 
						     STRUCT_LEARN_PARM *sparm)
{
  /* Finds the label ybar for pattern x that that is responsible for
     the most violated constraint for the slack rescaling
     formulation. It has to take into account the scoring function in
     sm, especially the weights sm.w, as well as the loss
     function. The weights in sm.w correspond to the features defined
     by psi() and range from index 1 to index sm->sizePsi. Most simple
     is the case of the zero/one loss function. For the zero/one loss,
     this function should return the highest scoring label ybar, if
     ybar is unequal y; if it is equal to the correct label y, then
     the function shall return the second highest scoring label. If
     the function cannot find a label, it shall return an empty label
     as recognized by the function empty_label(y). */
  LABEL ybar;
  DOC doc;
  long classlabel, bestclass=-1, first=1;
  double score, score_y, score_ybar, bestscore=-1;

  /* NOTE: This function could be made much more efficient by not
     always computing a new PSI vector. */
  doc = *(x.doc);
  doc.fvec = psi(x,y,sm,sparm);
  score_y = classify_example(sm->svm_model,&doc);
  free_svector(doc.fvec);

  ybar.scores = NULL;
  ybar.num_classes = sparm->num_classes;
  for(classlabel=1; classlabel<=sparm->num_classes; classlabel++) {
    ybar.classlabel = classlabel;
    doc.fvec=psi(x,ybar,sm,sparm);
    score_ybar=classify_example(sm->svm_model,&doc);
    free_svector(doc.fvec);
    score=loss(y,ybar,sparm,x.doc->fvec)*(1.0-score_y+score_ybar);
    if((bestscore<score)  || (first)) {
      bestscore=score;
      bestclass = classlabel;
      first=0;
    }
  }
  if(bestclass == -1) 
    printf("ERROR: Only one class\n");
  ybar.classlabel = bestclass;
  if(struct_verbosity>=3)
    printf("[%ld:%.2f] ",bestclass,bestscore);
  return(ybar);
}
double SVMModule::classify(std::vector<float>& instance)
{
	if(m_pTrainModel == NULL)
		return -1;

	WORD *words;
	int vecLen = instance.size();
	words = (WORD*)my_malloc(sizeof(WORD)*(vecLen + 1));

	for(int i = 0; i < vecLen; i++) {
		words[i].wnum = (i+1);
		words[i].weight = instance[i];
	}
	words[vecLen].wnum = 0;

	// make svector
	SVECTOR* svec = create_svector(words, "", 1.0); 
	DOC* doc = create_example(-1, 0, 0, 0.0, svec);

	double prob = classify_example(m_pTrainModel, doc);

	free_example(doc, 1);
	
	return prob;
}
Beispiel #3
0
/* call as  model = mexsvmlearn(data,labels,options) */
void mexFunction(int nlhs, mxArray *plhs[],
		  int nrhs, const mxArray *prhs[])
{

  
  DOC **docs; /* hold a test example */
  double *target; /* hold labels */
  WORD *words;  /* the words read from the example */
  long rows, cols; /* the number of rows and cols in the test data */
  double dist, doc_label, costfactor;
  double *err,*pred;
  long int correct=0, incorrect=0, none=0,i;
  MODEL *model;
  checkParameters(nlhs, plhs, nrhs, prhs);

  global_init( );

  /* load model parameters from the "model" parameter */
  model = restore_model((mxArray *)prhs[2]);

  rows = mxGetM(prhs[0]);
  cols = mxGetN(prhs[0]);

  /* load the testing arrays into docs */

  mexToDOC((mxArray *)prhs[0], (mxArray *)prhs[1], &docs, &target, NULL, NULL);
  
  /* setup output environment */
  plhs[0] = mxCreateDoubleMatrix(1,1,mxREAL);
  plhs[1] = mxCreateDoubleMatrix(rows,1,mxREAL);

  err = mxGetPr(plhs[0]);
  pred = mxGetPr(plhs[1]);

  /* classify examples */
  for (i = 0; i < rows; i++) {

    dist = classify_example(model, docs[i]);
    pred[i] = dist;

    if (dist > 0) {
      if (target[i] > 0) correct++;
      else incorrect++;
    } else {
      if (target[i] < 0) correct++;
      else incorrect++;
    }

    if ((int)(0.1+(target[i] * target[i]))!=1)
      none++;

  }

  err[0] = incorrect / (double) rows;

  
  global_destroy( );

}
double find_most_violated_joint_constraint_in_cache_old(int n, int cache_size, SVECTOR ***fydelta_cache, double **loss_cache, MODEL *svmModel, SVECTOR **lhs, double *margin)
{
  int     i,j;
  double  progress,progress_old;
  double  maxviol=0,sumviol,viol,lossval;
  double  dist_ydelta;
  SVECTOR *fydelta;
  DOC     *doc_fydelta;

  (*lhs)=NULL;
  (*margin)=0;
  sumviol=0;

  progress=0;
  progress_old=progress;

  for(i=0; i<n; i++) { /*** example loop ***/
	  
    progress+=10.0/n;
    if((struct_verbosity==1) && (((int)progress_old) != ((int)progress)))
      {printf("+");fflush(stdout); progress_old=progress;}
    if(struct_verbosity>=2)
      {printf("+"); fflush(stdout);}
    
    fydelta=NULL;
    lossval=0;
    for(j=0;j<cache_size;j++) {
      doc_fydelta=create_example(1,0,1,1,fydelta_cache[j][i]);
      dist_ydelta=classify_example(svmModel,doc_fydelta);
      free_example(doc_fydelta,0);

      viol=loss_cache[j][i]-dist_ydelta;
      if((viol > maxviol) || (!fydelta)) {
	fydelta=fydelta_cache[j][i];
	lossval=loss_cache[j][i];
	maxviol=viol;
      }
    }

    /**** add current fydelta to joint constraint ****/
    fydelta=copy_svector(fydelta);
    append_svector_list(fydelta,(*lhs));     /* add fydelta to lhs */
    (*lhs)=fydelta;
    (*margin)+=lossval;                      /* add loss to rhs */
    sumviol+=maxviol;
  }

  return(sumviol);
}
Beispiel #5
0
void update_constraint_cache_for_model(CCACHE *ccache, MODEL *svmModel)
     /* update the violation scores according to svmModel and find the
	most violated constraints for each example */
{ 
  int     i;
  long    progress=0;
  double  maxviol=0;
  double  dist_ydelta;
  DOC     *doc_fydelta;
  CCACHEELEM *celem,*prev,*maxviol_celem,*maxviol_prev;

  doc_fydelta=create_example(1,0,1,1,NULL);
  for(i=0; i<ccache->n; i++) { /*** example loop ***/
	  
    if(struct_verbosity>=3) 
      print_percent_progress(&progress,ccache->n,10,"+");

    maxviol=0;
    prev=NULL;
    maxviol_celem=NULL;
    maxviol_prev=NULL;
    for(celem=ccache->constlist[i];celem;celem=celem->next) {
      doc_fydelta->fvec=celem->fydelta;
      dist_ydelta=classify_example(svmModel,doc_fydelta);
      celem->viol=celem->rhs-dist_ydelta;
      if((celem->viol > maxviol) || (!maxviol_celem)) {
	maxviol=celem->viol;
	maxviol_celem=celem;
	maxviol_prev=prev;
      }
      prev=celem;
    }
    ccache->changed[i]=0;
    if(maxviol_prev) { /* move max violated constraint to the top of list */
      maxviol_prev->next=maxviol_celem->next;
      maxviol_celem->next=ccache->constlist[i];
      ccache->constlist[i]=maxviol_celem;
      ccache->changed[i]=1;
    }
  }
  free_example(doc_fydelta,0);
}
void update_constraint_cache_for_model(CCACHE *ccache, MODEL *svmModel)
     /* update the violation scores according to svmModel and find the
	most violated constraints for each example */
{ 
  int     i;
  double  progress=0,progress_old=0;
  double  maxviol=0;
  double  dist_ydelta;
  DOC     *doc_fydelta;
  CCACHEELEM *celem,*prev,*maxviol_celem,*maxviol_prev;

  for(i=0; i<ccache->n; i++) { /*** example loop ***/
	  
    progress+=10.0/ccache->n;
    if((struct_verbosity==1) && (((int)progress_old) != ((int)progress)))
      {printf("+");fflush(stdout); progress_old=progress;}
    if(struct_verbosity>=2)
      {printf("+"); fflush(stdout);}
    
    maxviol=0;
    prev=NULL;
    maxviol_celem=NULL;
    maxviol_prev=NULL;
    for(celem=ccache->constlist[i];celem;celem=celem->next) {
      doc_fydelta=create_example(1,0,1,1,celem->fydelta);
      dist_ydelta=classify_example(svmModel,doc_fydelta);
      free_example(doc_fydelta,0);
      celem->viol=celem->rhs-dist_ydelta;
      if((celem->viol > maxviol) || (!maxviol_celem)) {
	maxviol=celem->viol;
	maxviol_celem=celem;
	maxviol_prev=prev;
      }
      prev=celem;
    }
    if(maxviol_prev) { /* move max violated constraint to the top of list */
      maxviol_prev->next=maxviol_celem->next;
      maxviol_celem->next=ccache->constlist[i];
      ccache->constlist[i]=maxviol_celem;
    }
  }
}
void add_constraint_to_constraint_cache(CCACHE *ccache, MODEL *svmModel, int exnum, SVECTOR *fydelta, double rhs, int maxconst)
     /* add new constraint fydelta*w>rhs for example exnum to cache,
	if it is more violated than the currently most violated
	constraint in cache. if this grows the number of constraint
	for this example beyond maxconst, then the most unused
	constraint is deleted. the funciton assumes that
	update_constraint_cache_for_model has been run. */
{
  double  viol;
  double  dist_ydelta;
  DOC     *doc_fydelta;
  CCACHEELEM *celem;
  int     cnum;

  doc_fydelta=create_example(1,0,1,1,fydelta);
  dist_ydelta=classify_example(svmModel,doc_fydelta);
  free_example(doc_fydelta,0);  
  viol=rhs-dist_ydelta;

  if((viol-0.000000000001) > ccache->constlist[exnum]->viol) {
    celem=ccache->constlist[exnum];
    ccache->constlist[exnum]=(CCACHEELEM *)malloc(sizeof(CCACHEELEM));
    ccache->constlist[exnum]->next=celem;
    ccache->constlist[exnum]->fydelta=fydelta;
    ccache->constlist[exnum]->rhs=rhs;
    ccache->constlist[exnum]->viol=viol;

    /* remove last constraint in list, if list is longer than maxconst */
    cnum=2;
    for(celem=ccache->constlist[exnum];celem && celem->next && celem->next->next;celem=celem->next)
      cnum++;
    if(cnum>maxconst) {
      free_svector(celem->next->fydelta);
      free(celem->next);
      celem->next=NULL;
    }
  }
  else {
    free_svector(fydelta);
  }
}
Beispiel #8
0
LABEL       classify_struct_example(PATTERNX x, STRUCTMODEL *sm, 
				    STRUCT_LEARN_PARM *sparm)
{
  /* Finds the label yhat for pattern x that scores the highest
     according to the linear evaluation function in sm, especially the
     weights sm.w. The returned label is taken as the prediction of sm
     for the pattern x. The weights correspond to the features defined
     by psi() and range from index 1 to index sm->sizePsi. If the
     function cannot find a label, it shall return an empty label as
     recognized by the function empty_label(y). */
  LABEL y;
  DOC doc;
  long classlabel, bestclass=-1, first=1, j;
  double score, bestscore=-1;
  TOKEN *words;

  doc = *(x.doc);
  y.scores = (double *)my_malloc(sizeof(double)*(sparm->num_classes+1));
  y.num_classes = sparm->num_classes;
  words = doc.fvec->words;
  for(j=0; (words[j]).wnum != 0; j++) {       /* Check if feature numbers   */
    if((words[j]).wnum>sparm->num_features) /* are not larger than in     */
      (words[j]).wnum=0;                    /* model. Remove feature if   */
  }                                         /* necessary.                 */
  for(classlabel=1; classlabel<=sparm->num_classes; classlabel++) 
  {
	  y.classlabel = classlabel;
	  doc.fvec = psi(x,y,sm,sparm);
	  score = classify_example(sm->svm_model,&doc);
	  free_svector(doc.fvec);
	  y.scores[classlabel] = score;
	  if((bestscore<score)  || (first)) {
		  bestscore = score;
		  bestclass = classlabel;
		  first = 0;
	  }
  }
  y.classlabel = bestclass;
  return(y);
}
Beispiel #9
0
void svm_learn_struct_joint(SAMPLE sample, STRUCT_LEARN_PARM *sparm,
			    LEARN_PARM *lparm, KERNEL_PARM *kparm, 
			    STRUCTMODEL *sm, int alg_type)
{
  int         i,j;
  int         numIt=0;
  long        argmax_count=0;
  long        totconstraints=0;
  long        kernel_type_org;
  double      epsilon,epsilon_cached;
  double      lhsXw,rhs_i;
  double      rhs=0;
  double      slack,ceps;
  double      dualitygap,modellength,alphasum;
  long        sizePsi;
  double      *alpha=NULL;
  long        *alphahist=NULL,optcount=0;
  CONSTSET    cset;
  SVECTOR     *diff=NULL;
  double      *lhs_n=NULL;
  SVECTOR     *fy, *fydelta, **fycache, *lhs;
  MODEL       *svmModel=NULL;
  DOC         *doc;

  long        n=sample.n;
  EXAMPLE     *ex=sample.examples;
  double      rt_total=0,rt_opt=0,rt_init=0,rt_psi=0,rt_viol=0,rt_kernel=0;
  double      rt_cacheupdate=0,rt_cacheconst=0,rt_cacheadd=0,rt_cachesum=0;
  double      rt1=0,rt2=0;
  long        progress;

  /*
  SVECTOR     ***fydelta_cache=NULL;
  double      **loss_cache=NULL;
  int         cache_size=0;
  */
  CCACHE      *ccache=NULL;
  int         cached_constraint;
  double      viol,viol_est,epsilon_est=0;
  long        uptr=0;
  long        *randmapping=NULL;
  long        batch_size=n;

  rt1=get_runtime();

  if(sparm->batch_size<100)
    batch_size=sparm->batch_size*n/100.0;

  init_struct_model(sample,sm,sparm,lparm,kparm); 
  sizePsi=sm->sizePsi+1;          /* sm must contain size of psi on return */

  if(sparm->slack_norm == 1) {
    lparm->svm_c=sparm->C;          /* set upper bound C */
    lparm->sharedslack=1;
  }
  else if(sparm->slack_norm == 2) {
    printf("ERROR: The joint algorithm does not apply to L2 slack norm!"); 
    fflush(stdout);
    exit(0); 
  }
  else {
    printf("ERROR: Slack norm must be L1 or L2!"); fflush(stdout);
    exit(0);
  }


  lparm->biased_hyperplane=0;     /* set threshold to zero */
  epsilon=100.0;                  /* start with low precision and
				     increase later */
  epsilon_cached=epsilon;         /* epsilon to use for iterations
				     using constraints constructed
				     from the constraint cache */

  cset=init_struct_constraints(sample, sm, sparm);
  if(cset.m > 0) {
    alpha=(double *)realloc(alpha,sizeof(double)*cset.m);
    alphahist=(long *)realloc(alphahist,sizeof(long)*cset.m);
    for(i=0; i<cset.m; i++) {
      alpha[i]=0;
      alphahist[i]=-1; /* -1 makes sure these constraints are never removed */
    }
  }
  kparm->gram_matrix=NULL;
  if((alg_type == ONESLACK_DUAL_ALG) || (alg_type == ONESLACK_DUAL_CACHE_ALG))
    kparm->gram_matrix=init_kernel_matrix(&cset,kparm);

  /* set initial model and slack variables */
  svmModel=(MODEL *)my_malloc(sizeof(MODEL));
  lparm->epsilon_crit=epsilon;
  svm_learn_optimization(cset.lhs,cset.rhs,cset.m,sizePsi,
			 lparm,kparm,NULL,svmModel,alpha);
  add_weight_vector_to_linear_model(svmModel);
  sm->svm_model=svmModel;
  sm->w=svmModel->lin_weights; /* short cut to weight vector */

  /* create a cache of the feature vectors for the correct labels */
  fycache=(SVECTOR **)my_malloc(n*sizeof(SVECTOR *));
  for(i=0;i<n;i++) {
    if(USE_FYCACHE) {
      fy=psi(ex[i].x,ex[i].y,sm,sparm);
      if(kparm->kernel_type == LINEAR_KERNEL) { /* store difference vector directly */
	diff=add_list_sort_ss_r(fy,COMPACT_ROUNDING_THRESH); 
	free_svector(fy);
	fy=diff;
      }
    }
    else
      fy=NULL;
    fycache[i]=fy;
  }

  /* initialize the constraint cache */
  if(alg_type == ONESLACK_DUAL_CACHE_ALG) {
    ccache=create_constraint_cache(sample,sparm,sm);
    /* NOTE:  */
    for(i=0;i<n;i++) 
      if(loss(ex[i].y,ex[i].y,sparm) != 0) {
	printf("ERROR: Loss function returns non-zero value loss(y_%d,y_%d)\n",i,i);
	printf("       W4 algorithm assumes that loss(y_i,y_i)=0 for all i.\n");
	exit(1);
      }
  }
  
  if(kparm->kernel_type == LINEAR_KERNEL)
    lhs_n=create_nvector(sm->sizePsi);

  /* randomize order or training examples */
  if(batch_size<n)
    randmapping=random_order(n);

  rt_init+=MAX(get_runtime()-rt1,0);
  rt_total+=rt_init;

    /*****************/
   /*** main loop ***/
  /*****************/
  do { /* iteratively find and add constraints to working set */

      if(struct_verbosity>=1) { 
	printf("Iter %i: ",++numIt); 
	fflush(stdout);
      }
      
      rt1=get_runtime();

      /**** compute current slack ****/
      alphasum=0;
      for(j=0;(j<cset.m);j++) 
	  alphasum+=alpha[j];
      for(j=0,slack=-1;(j<cset.m) && (slack==-1);j++)  
	if(alpha[j] > alphasum/cset.m)
	  slack=MAX(0,cset.rhs[j]-classify_example(svmModel,cset.lhs[j]));
      slack=MAX(0,slack);

      rt_total+=MAX(get_runtime()-rt1,0);

      /**** find a violated joint constraint ****/
      lhs=NULL;
      rhs=0;
      if(alg_type == ONESLACK_DUAL_CACHE_ALG) {
	rt1=get_runtime();
	/* Compute violation of constraints in cache for current w */
	if(struct_verbosity>=2) rt2=get_runtime();
	update_constraint_cache_for_model(ccache, svmModel);
	if(struct_verbosity>=2) rt_cacheupdate+=MAX(get_runtime()-rt2,0);
	/* Is there is a sufficiently violated constraint in cache? */
	viol=compute_violation_of_constraint_in_cache(ccache,epsilon_est/2);
	if(viol-slack > MAX(epsilon_est/10,sparm->epsilon)) { 
	  /* There is a sufficiently violated constraint in cache, so
	     use this constraint in this iteration. */
	  if(struct_verbosity>=2) rt2=get_runtime();
	  viol=find_most_violated_joint_constraint_in_cache(ccache,
					       epsilon_est/2,lhs_n,&lhs,&rhs);
	  if(struct_verbosity>=2) rt_cacheconst+=MAX(get_runtime()-rt2,0);
	  cached_constraint=1;
	}
	else {
	  /* There is no sufficiently violated constraint in cache, so
	     update cache by computing most violated constraint
	     explicitly for batch_size examples. */
	  viol_est=0;
	  progress=0;
	  viol=compute_violation_of_constraint_in_cache(ccache,0);
	  for(j=0;(j<batch_size) || ((j<n)&&(viol-slack<sparm->epsilon));j++) {
	    if(struct_verbosity>=1) 
	      print_percent_progress(&progress,n,10,".");
	    uptr=uptr % n;
	    if(randmapping) 
	      i=randmapping[uptr];
	    else
	      i=uptr;
	    /* find most violating fydelta=fy-fybar and rhs for example i */
	    find_most_violated_constraint(&fydelta,&rhs_i,&ex[i],
					  fycache[i],n,sm,sparm,
					  &rt_viol,&rt_psi,&argmax_count);
	    /* add current fy-fybar and loss to cache */
	    if(struct_verbosity>=2) rt2=get_runtime();
	    viol+=add_constraint_to_constraint_cache(ccache,sm->svm_model,
			     i,fydelta,rhs_i,0.0001*sparm->epsilon/n,
			     sparm->ccache_size,&rt_cachesum);
	    if(struct_verbosity>=2) rt_cacheadd+=MAX(get_runtime()-rt2,0);
	    viol_est+=ccache->constlist[i]->viol;
	    uptr++;
	  }
	  cached_constraint=(j<n);
	  if(struct_verbosity>=2) rt2=get_runtime();
	  if(cached_constraint)
	    viol=find_most_violated_joint_constraint_in_cache(ccache,
					       epsilon_est/2,lhs_n,&lhs,&rhs);
	  else
	    viol=find_most_violated_joint_constraint_in_cache(ccache,0,lhs_n,
							 &lhs,&rhs);
	  if(struct_verbosity>=2) rt_cacheconst+=MAX(get_runtime()-rt2,0);
	  viol_est*=((double)n/j);
	  epsilon_est=(1-(double)j/n)*epsilon_est+(double)j/n*(viol_est-slack);
	  if((struct_verbosity >= 1) && (j!=n))
	    printf("(upd=%5.1f%%,eps^=%.4f,eps*=%.4f)",
		   100.0*j/n,viol_est-slack,epsilon_est);
	}
	lhsXw=rhs-viol;

	rt_total+=MAX(get_runtime()-rt1,0);
      }
      else { 
	/* do not use constraint from cache */
	rt1=get_runtime();
	cached_constraint=0;
	if(kparm->kernel_type == LINEAR_KERNEL)
	  clear_nvector(lhs_n,sm->sizePsi);
	progress=0;
	rt_total+=MAX(get_runtime()-rt1,0);

	for(i=0; i<n; i++) {
	  rt1=get_runtime();

	  if(struct_verbosity>=1) 
	    print_percent_progress(&progress,n,10,".");

	  /* compute most violating fydelta=fy-fybar and rhs for example i */
	  find_most_violated_constraint(&fydelta,&rhs_i,&ex[i],fycache[i],n,
				      sm,sparm,&rt_viol,&rt_psi,&argmax_count);
	  /* add current fy-fybar to lhs of constraint */
	  if(kparm->kernel_type == LINEAR_KERNEL) {
	    add_list_n_ns(lhs_n,fydelta,1.0); /* add fy-fybar to sum */
	    free_svector(fydelta);
	  }
	  else {
	    append_svector_list(fydelta,lhs); /* add fy-fybar to vector list */
	    lhs=fydelta;
	  }
	  rhs+=rhs_i;                         /* add loss to rhs */
	  
	  rt_total+=MAX(get_runtime()-rt1,0);

	} /* end of example loop */

	rt1=get_runtime();

	/* create sparse vector from dense sum */
	if(kparm->kernel_type == LINEAR_KERNEL)
	  lhs=create_svector_n_r(lhs_n,sm->sizePsi,NULL,1.0,
				 COMPACT_ROUNDING_THRESH);
	doc=create_example(cset.m,0,1,1,lhs);
	lhsXw=classify_example(svmModel,doc);
	free_example(doc,0);
	viol=rhs-lhsXw;

	rt_total+=MAX(get_runtime()-rt1,0);

      } /* end of finding most violated joint constraint */

      rt1=get_runtime();

      /**** if `error', then add constraint and recompute QP ****/
      if(slack > (rhs-lhsXw+0.000001)) {
	printf("\nWARNING: Slack of most violated constraint is smaller than slack of working\n");
	printf("         set! There is probably a bug in 'find_most_violated_constraint_*'.\n");
	printf("slack=%f, newslack=%f\n",slack,rhs-lhsXw);
	/* exit(1); */
      }
      ceps=MAX(0,rhs-lhsXw-slack);
      if((ceps > sparm->epsilon) || cached_constraint) { 
	/**** resize constraint matrix and add new constraint ****/
	cset.lhs=(DOC **)realloc(cset.lhs,sizeof(DOC *)*(cset.m+1));
	cset.lhs[cset.m]=create_example(cset.m,0,1,1,lhs);
	cset.rhs=(double *)realloc(cset.rhs,sizeof(double)*(cset.m+1));
	cset.rhs[cset.m]=rhs;
	alpha=(double *)realloc(alpha,sizeof(double)*(cset.m+1));
	alpha[cset.m]=0;
	alphahist=(long *)realloc(alphahist,sizeof(long)*(cset.m+1));
	alphahist[cset.m]=optcount;
	cset.m++;
	totconstraints++;
	if((alg_type == ONESLACK_DUAL_ALG) 
	   || (alg_type == ONESLACK_DUAL_CACHE_ALG)) {
	  if(struct_verbosity>=2) rt2=get_runtime();
	  kparm->gram_matrix=update_kernel_matrix(kparm->gram_matrix,cset.m-1,
						  &cset,kparm);
	  if(struct_verbosity>=2) rt_kernel+=MAX(get_runtime()-rt2,0);
	}
	
	/**** get new QP solution ****/
	if(struct_verbosity>=1) {
	  printf("*");fflush(stdout);
	}
	if(struct_verbosity>=2) rt2=get_runtime();
	/* set svm precision so that higher than eps of most violated constr */
	if(cached_constraint) {
	  epsilon_cached=MIN(epsilon_cached,ceps); 
	  lparm->epsilon_crit=epsilon_cached/2; 
	}
	else {
	  epsilon=MIN(epsilon,ceps); /* best eps so far */
	  lparm->epsilon_crit=epsilon/2; 
	  epsilon_cached=epsilon;
	}
	free_model(svmModel,0);
	svmModel=(MODEL *)my_malloc(sizeof(MODEL));
	/* Run the QP solver on cset. */
	kernel_type_org=kparm->kernel_type;
	if((alg_type == ONESLACK_DUAL_ALG) 
	   || (alg_type == ONESLACK_DUAL_CACHE_ALG))
	  kparm->kernel_type=GRAM; /* use kernel stored in kparm */
	svm_learn_optimization(cset.lhs,cset.rhs,cset.m,sizePsi,
			       lparm,kparm,NULL,svmModel,alpha);
	kparm->kernel_type=kernel_type_org; 
	svmModel->kernel_parm.kernel_type=kernel_type_org;
	/* Always add weight vector, in case part of the kernel is
	   linear. If not, ignore the weight vector since its
	   content is bogus. */
	add_weight_vector_to_linear_model(svmModel);
	sm->svm_model=svmModel;
	sm->w=svmModel->lin_weights; /* short cut to weight vector */
	optcount++;
	/* keep track of when each constraint was last
	   active. constraints marked with -1 are not updated */
	for(j=0;j<cset.m;j++) 
	  if((alphahist[j]>-1) && (alpha[j] != 0))  
	    alphahist[j]=optcount;
	if(struct_verbosity>=2) rt_opt+=MAX(get_runtime()-rt2,0);
	
	/* Check if some of the linear constraints have not been
	   active in a while. Those constraints are then removed to
	   avoid bloating the working set beyond necessity. */
	if(struct_verbosity>=3)
	  printf("Reducing working set...");fflush(stdout);
	remove_inactive_constraints(&cset,alpha,optcount,alphahist,50);
	if(struct_verbosity>=3)
	  printf("done. ");
      }
      else {
	free_svector(lhs);
      }

      if(struct_verbosity>=1)
	printf("(NumConst=%d, SV=%ld, CEps=%.4f, QPEps=%.4f)\n",cset.m,
	       svmModel->sv_num-1,ceps,svmModel->maxdiff);

      rt_total+=MAX(get_runtime()-rt1,0);

  } while(finalize_iteration(ceps,cached_constraint,sample,sm,cset,alpha,sparm)|| cached_constraint || (ceps > sparm->epsilon) );

  // originally like below ... finalize_iteration was not called because of short-circuit evaluation
//  } while(cached_constraint || (ceps > sparm->epsilon) || 
//	  finalize_iteration(ceps,cached_constraint,sample,sm,cset,alpha,sparm)
//	 );
  
  if(struct_verbosity>=1) {
    printf("Final epsilon on KKT-Conditions: %.5f\n",
	   MAX(svmModel->maxdiff,ceps));

    slack=0;
    for(j=0;j<cset.m;j++) 
      slack=MAX(slack,
		cset.rhs[j]-classify_example(svmModel,cset.lhs[j]));
    alphasum=0;
    for(i=0; i<cset.m; i++)  
      alphasum+=alpha[i]*cset.rhs[i];
    if(kparm->kernel_type == LINEAR_KERNEL)
      modellength=model_length_n(svmModel);
    else
      modellength=model_length_s(svmModel);
    dualitygap=(0.5*modellength*modellength+sparm->C*viol)
               -(alphasum-0.5*modellength*modellength);
    
    printf("Upper bound on duality gap: %.5f\n", dualitygap);
    printf("Dual objective value: dval=%.5f\n",
	    alphasum-0.5*modellength*modellength);
    printf("Primal objective value: pval=%.5f\n",
	    0.5*modellength*modellength+sparm->C*viol);
    printf("Total number of constraints in final working set: %i (of %i)\n",(int)cset.m,(int)totconstraints);
    printf("Number of iterations: %d\n",numIt);
    printf("Number of calls to 'find_most_violated_constraint': %ld\n",argmax_count);
    printf("Number of SV: %ld \n",svmModel->sv_num-1);
    printf("Norm of weight vector: |w|=%.5f\n",modellength);
    printf("Value of slack variable (on working set): xi=%.5f\n",slack);
    printf("Value of slack variable (global): xi=%.5f\n",viol);
    printf("Norm of longest difference vector: ||Psi(x,y)-Psi(x,ybar)||=%.5f\n",
	   length_of_longest_document_vector(cset.lhs,cset.m,kparm));
    if(struct_verbosity>=2) 
      printf("Runtime in cpu-seconds: %.2f (%.2f%% for QP, %.2f%% for kernel, %.2f%% for Argmax, %.2f%% for Psi, %.2f%% for init, %.2f%% for cache update, %.2f%% for cache const, %.2f%% for cache add (incl. %.2f%% for sum))\n",
	   rt_total/100.0, (100.0*rt_opt)/rt_total, (100.0*rt_kernel)/rt_total,
	   (100.0*rt_viol)/rt_total, (100.0*rt_psi)/rt_total, 
	   (100.0*rt_init)/rt_total,(100.0*rt_cacheupdate)/rt_total,
	   (100.0*rt_cacheconst)/rt_total,(100.0*rt_cacheadd)/rt_total,
	   (100.0*rt_cachesum)/rt_total);
    else if(struct_verbosity==1) 
      printf("Runtime in cpu-seconds: %.2f\n",rt_total/100.0);
  }
  if(ccache) {
    long cnum=0;
    CCACHEELEM *celem;
    for(i=0;i<n;i++) 
      for(celem=ccache->constlist[i];celem;celem=celem->next) 
	cnum++;
    printf("Final number of constraints in cache: %ld\n",cnum);
  }
  if(struct_verbosity>=4)
    printW(sm->w,sizePsi,n,lparm->svm_c);

  if(svmModel) {
    sm->svm_model=copy_model(svmModel);
    sm->w=sm->svm_model->lin_weights; /* short cut to weight vector */
    free_model(svmModel,0);
  }

  print_struct_learning_stats(sample,sm,cset,alpha,sparm);

  if(lhs_n)
    free_nvector(lhs_n);
  if(ccache)    
    free_constraint_cache(ccache);
  for(i=0;i<n;i++)
    if(fycache[i])
      free_svector(fycache[i]);
  free(fycache);
  free(alpha); 
  free(alphahist); 
  free(cset.rhs); 
  for(i=0;i<cset.m;i++) 
    free_example(cset.lhs[i],1);
  free(cset.lhs);
  if(kparm->gram_matrix)
    free_matrix(kparm->gram_matrix);
}
Beispiel #10
0
void svm_learn_struct(SAMPLE sample, STRUCT_LEARN_PARM *sparm,
		      LEARN_PARM *lparm, KERNEL_PARM *kparm, 
		      STRUCTMODEL *sm, int alg_type)
{
  int         i,j;
  int         numIt=0;
  long        argmax_count=0;
  long        newconstraints=0, totconstraints=0, activenum=0; 
  int         opti_round, *opti, fullround, use_shrinking;
  long        old_totconstraints=0;
  double      epsilon,svmCnorm;
  long        tolerance,new_precision=1,dont_stop=0;
  double      lossval,factor,dist;
  double      margin=0;
  double      slack, *slacks, slacksum, ceps;
  double      dualitygap,modellength,alphasum;
  long        sizePsi;
  double      *alpha=NULL;
  long        *alphahist=NULL,optcount=0,lastoptcount=0;
  CONSTSET    cset;
  SVECTOR     *diff=NULL;
  SVECTOR     *fy, *fybar, *f, **fycache=NULL;
  SVECTOR     *slackvec;
  WORD        slackv[2];
  MODEL       *svmModel=NULL;
  KERNEL_CACHE *kcache=NULL;
  LABEL       ybar;
  DOC         *doc;

  long        n=sample.n;
  EXAMPLE     *ex=sample.examples;
  double      rt_total=0, rt_opt=0, rt_init=0, rt_psi=0, rt_viol=0;
  double      rt1,rt2;

  rt1=get_runtime();

  init_struct_model(sample,sm,sparm,lparm,kparm); 
  sizePsi=sm->sizePsi+1;          /* sm must contain size of psi on return */

  /* initialize shrinking-style example selection heuristic */ 
  if(alg_type == NSLACK_SHRINK_ALG)
    use_shrinking=1;
  else
    use_shrinking=0;
  opti=(int*)my_malloc(n*sizeof(int));
  for(i=0;i<n;i++) {
    opti[i]=0;
  }
  opti_round=0;

  /* normalize regularization parameter C by the number of training examples */
  svmCnorm=sparm->C/n;

  if(sparm->slack_norm == 1) {
    lparm->svm_c=svmCnorm;          /* set upper bound C */
    lparm->sharedslack=1;
  }
  else if(sparm->slack_norm == 2) {
    lparm->svm_c=999999999999999.0; /* upper bound C must never be reached */
    lparm->sharedslack=0;
    if(kparm->kernel_type != LINEAR_KERNEL) {
      printf("ERROR: Kernels are not implemented for L2 slack norm!"); 
      fflush(stdout);
      exit(0); 
    }
  }
  else {
    printf("ERROR: Slack norm must be L1 or L2!"); fflush(stdout);
    exit(0);
  }


  epsilon=100.0;                  /* start with low precision and
				     increase later */
  tolerance=MIN(n/3,MAX(n/100,5));/* increase precision, whenever less
                                     than that number of constraints
                                     is not fulfilled */
  lparm->biased_hyperplane=0;     /* set threshold to zero */

  cset=init_struct_constraints(sample, sm, sparm);
  if(cset.m > 0) {
    alpha=(double *)realloc(alpha,sizeof(double)*cset.m);
    alphahist=(long *)realloc(alphahist,sizeof(long)*cset.m);
    for(i=0; i<cset.m; i++) {
      alpha[i]=0;
      alphahist[i]=-1; /* -1 makes sure these constraints are never removed */
    }
  }

  /* set initial model and slack variables*/
  svmModel=(MODEL *)my_malloc(sizeof(MODEL));
  lparm->epsilon_crit=epsilon;
  if(kparm->kernel_type != LINEAR_KERNEL)
    kcache=kernel_cache_init(MAX(cset.m,1),lparm->kernel_cache_size);
  svm_learn_optimization(cset.lhs,cset.rhs,cset.m,sizePsi+n,
			 lparm,kparm,kcache,svmModel,alpha);
  if(kcache)
    kernel_cache_cleanup(kcache);
  add_weight_vector_to_linear_model(svmModel);
  sm->svm_model=svmModel;
  sm->w=svmModel->lin_weights; /* short cut to weight vector */

  /* create a cache of the feature vectors for the correct labels */
  if(USE_FYCACHE) {
    fycache=(SVECTOR **)my_malloc(n*sizeof(SVECTOR *));
    for(i=0;i<n;i++) {
      fy=psi(ex[i].x,ex[i].y,sm,sparm);
      if(kparm->kernel_type == LINEAR_KERNEL) {
	diff=add_list_ss(fy); /* store difference vector directly */
	free_svector(fy);
	fy=diff;
      }
      fycache[i]=fy;
    }
  }

  rt_init+=MAX(get_runtime()-rt1,0);
  rt_total+=MAX(get_runtime()-rt1,0);

    /*****************/
   /*** main loop ***/
  /*****************/
  do { /* iteratively increase precision */

    epsilon=MAX(epsilon*0.49999999999,sparm->epsilon);
    new_precision=1;
    if(epsilon == sparm->epsilon)   /* for final precision, find all SV */
      tolerance=0; 
    lparm->epsilon_crit=epsilon/2;  /* svm precision must be higher than eps */
    if(struct_verbosity>=1)
      printf("Setting current working precision to %g.\n",epsilon);

    do { /* iteration until (approx) all SV are found for current
            precision and tolerance */
      
      opti_round++;
      activenum=n;
      dont_stop=0;
      old_totconstraints=totconstraints;

      do { /* with shrinking turned on, go through examples that keep
	      producing new constraints */

	if(struct_verbosity>=1) { 
	  printf("Iter %i (%ld active): ",++numIt,activenum); 
	  fflush(stdout);
	}
	
	ceps=0;
	fullround=(activenum == n);

	for(i=0; i<n; i++) { /*** example loop ***/
	  
	  rt1=get_runtime();
	    
	  if((!use_shrinking) || (opti[i] != opti_round)) {
	                                /* if the example is not shrunk
	                                away, then see if it is necessary to 
					add a new constraint */
	    rt2=get_runtime();
	    argmax_count++;
	    if(sparm->loss_type == SLACK_RESCALING) 
	      ybar=find_most_violated_constraint_slackrescaling(ex[i].x,
								ex[i].y,sm,
								sparm);
	    else
	      ybar=find_most_violated_constraint_marginrescaling(ex[i].x,
								 ex[i].y,sm,
								 sparm);
	    rt_viol+=MAX(get_runtime()-rt2,0);
	    
	    if(empty_label(ybar)) {
	      if(opti[i] != opti_round) {
		activenum--;
		opti[i]=opti_round; 
	      }
	      if(struct_verbosity>=2)
		printf("no-incorrect-found(%i) ",i);
	      continue;
	    }
	  
	    /**** get psi(y)-psi(ybar) ****/
	    rt2=get_runtime();
	    if(fycache) 
	      fy=copy_svector(fycache[i]);
	    else
	      fy=psi(ex[i].x,ex[i].y,sm,sparm);
	    fybar=psi(ex[i].x,ybar,sm,sparm);
	    rt_psi+=MAX(get_runtime()-rt2,0);
	    
	    /**** scale feature vector and margin by loss ****/
	    lossval=loss(ex[i].y,ybar,sparm);
	    if(sparm->slack_norm == 2)
	      lossval=sqrt(lossval);
	    if(sparm->loss_type == SLACK_RESCALING)
	      factor=lossval;
	    else               /* do not rescale vector for */
	      factor=1.0;      /* margin rescaling loss type */
	    for(f=fy;f;f=f->next)
	      f->factor*=factor;
	    for(f=fybar;f;f=f->next)
	      f->factor*=-factor;
	    margin=lossval;

	    /**** create constraint for current ybar ****/
	    append_svector_list(fy,fybar);/* append the two vector lists */
	    doc=create_example(cset.m,0,i+1,1,fy);

	    /**** compute slack for this example ****/
	    slack=0;
	    for(j=0;j<cset.m;j++) 
	      if(cset.lhs[j]->slackid == i+1) {
		if(sparm->slack_norm == 2) /* works only for linear kernel */
		  slack=MAX(slack,cset.rhs[j]
			          -(classify_example(svmModel,cset.lhs[j])
				    -sm->w[sizePsi+i]/(sqrt(2*svmCnorm))));
		else
		  slack=MAX(slack,
			   cset.rhs[j]-classify_example(svmModel,cset.lhs[j]));
	      }
	    
	    /**** if `error' add constraint and recompute ****/
	    dist=classify_example(svmModel,doc);
	    ceps=MAX(ceps,margin-dist-slack);
	    if(slack > (margin-dist+0.0001)) {
	      printf("\nWARNING: Slack of most violated constraint is smaller than slack of working\n");
	      printf("         set! There is probably a bug in 'find_most_violated_constraint_*'.\n");
	      printf("Ex %d: slack=%f, newslack=%f\n",i,slack,margin-dist);
	      /* exit(1); */
	    }
	    if((dist+slack)<(margin-epsilon)) { 
	      if(struct_verbosity>=2)
		{printf("(%i,eps=%.2f) ",i,margin-dist-slack); fflush(stdout);}
	      if(struct_verbosity==1)
		{printf("."); fflush(stdout);}
	      
	      /**** resize constraint matrix and add new constraint ****/
	      cset.m++;
	      cset.lhs=(DOC **)realloc(cset.lhs,sizeof(DOC *)*cset.m);
	      if(kparm->kernel_type == LINEAR_KERNEL) {
		diff=add_list_ss(fy); /* store difference vector directly */
		if(sparm->slack_norm == 1) 
		  cset.lhs[cset.m-1]=create_example(cset.m-1,0,i+1,1,
						    copy_svector(diff));
		else if(sparm->slack_norm == 2) {
		  /**** add squared slack variable to feature vector ****/
		  slackv[0].wnum=sizePsi+i;
		  slackv[0].weight=1/(sqrt(2*svmCnorm));
		  slackv[1].wnum=0; /*terminator*/
		  slackvec=create_svector(slackv,NULL,1.0);
		  cset.lhs[cset.m-1]=create_example(cset.m-1,0,i+1,1,
						    add_ss(diff,slackvec));
		  free_svector(slackvec);
		}
		free_svector(diff);
	      }
	      else { /* kernel is used */
		if(sparm->slack_norm == 1) 
		  cset.lhs[cset.m-1]=create_example(cset.m-1,0,i+1,1,
						    copy_svector(fy));
		else if(sparm->slack_norm == 2)
		  exit(1);
	      }
	      cset.rhs=(double *)realloc(cset.rhs,sizeof(double)*cset.m);
	      cset.rhs[cset.m-1]=margin;
	      alpha=(double *)realloc(alpha,sizeof(double)*cset.m);
	      alpha[cset.m-1]=0;
	      alphahist=(long *)realloc(alphahist,sizeof(long)*cset.m);
	      alphahist[cset.m-1]=optcount;
	      newconstraints++;
	      totconstraints++;
	    }
	    else {
	      printf("+"); fflush(stdout); 
	      if(opti[i] != opti_round) {
		activenum--;
		opti[i]=opti_round; 
	      }
	    }

	    free_example(doc,0);
	    free_svector(fy); /* this also free's fybar */
	    free_label(ybar);
	  }

	  /**** get new QP solution ****/
	  if((newconstraints >= sparm->newconstretrain) 
	     || ((newconstraints > 0) && (i == n-1))
	     || (new_precision && (i == n-1))) {
	    if(struct_verbosity>=1) {
	      printf("*");fflush(stdout);
	    }
	    rt2=get_runtime();
	    free_model(svmModel,0);
	    svmModel=(MODEL *)my_malloc(sizeof(MODEL));
	    /* Always get a new kernel cache. It is not possible to use the
	       same cache for two different training runs */
	    if(kparm->kernel_type != LINEAR_KERNEL)
	      kcache=kernel_cache_init(MAX(cset.m,1),lparm->kernel_cache_size);
	    /* Run the QP solver on cset. */
	    svm_learn_optimization(cset.lhs,cset.rhs,cset.m,sizePsi+n,
				   lparm,kparm,kcache,svmModel,alpha);
	    if(kcache)
	      kernel_cache_cleanup(kcache);
	    /* Always add weight vector, in case part of the kernel is
	       linear. If not, ignore the weight vector since its
	       content is bogus. */
	    add_weight_vector_to_linear_model(svmModel);
	    sm->svm_model=svmModel;
	    sm->w=svmModel->lin_weights; /* short cut to weight vector */
	    optcount++;
	    /* keep track of when each constraint was last
	       active. constraints marked with -1 are not updated */
	    for(j=0;j<cset.m;j++) 
	      if((alphahist[j]>-1) && (alpha[j] != 0))  
		alphahist[j]=optcount;
	    rt_opt+=MAX(get_runtime()-rt2,0);
	    
	    if(new_precision && (epsilon <= sparm->epsilon))  
	      dont_stop=1; /* make sure we take one final pass */
	    new_precision=0;
	    newconstraints=0;
	  }	

	  rt_total+=MAX(get_runtime()-rt1,0);

	} /* end of example loop */

	rt1=get_runtime();
	
	if(struct_verbosity>=1)
	  printf("(NumConst=%d, SV=%ld, CEps=%.4f, QPEps=%.4f)\n",cset.m,
		 svmModel->sv_num-1,ceps,svmModel->maxdiff);
	
	/* Check if some of the linear constraints have not been
	   active in a while. Those constraints are then removed to
	   avoid bloating the working set beyond necessity. */
	if(struct_verbosity>=2)
	  printf("Reducing working set...");fflush(stdout);
	remove_inactive_constraints(&cset,alpha,optcount,alphahist,
				    MAX(50,optcount-lastoptcount));
	lastoptcount=optcount;
	if(struct_verbosity>=2)
	  printf("done. (NumConst=%d)\n",cset.m);
	
	rt_total+=MAX(get_runtime()-rt1,0);
	
      } while(use_shrinking && (activenum > 0)); /* when using shrinking, 
						    repeat until all examples 
						    produced no constraint at
						    least once */

    } while(((totconstraints - old_totconstraints) > tolerance) || dont_stop);

  } while((epsilon > sparm->epsilon) 
	  || finalize_iteration(ceps,0,sample,sm,cset,alpha,sparm));  

  if(struct_verbosity>=1) {
    /**** compute sum of slacks ****/
    /**** WARNING: If positivity constraints are used, then the
	  maximum slack id is larger than what is allocated
	  below ****/
    slacks=(double *)my_malloc(sizeof(double)*(n+1));
    for(i=0; i<=n; i++) { 
      slacks[i]=0;
    }
    if(sparm->slack_norm == 1) {
      for(j=0;j<cset.m;j++) 
	slacks[cset.lhs[j]->slackid]=MAX(slacks[cset.lhs[j]->slackid],
			   cset.rhs[j]-classify_example(svmModel,cset.lhs[j]));
      }
    else if(sparm->slack_norm == 2) {
      for(j=0;j<cset.m;j++) 
	slacks[cset.lhs[j]->slackid]=MAX(slacks[cset.lhs[j]->slackid],
		cset.rhs[j]
	         -(classify_example(svmModel,cset.lhs[j])
		   -sm->w[sizePsi+cset.lhs[j]->slackid-1]/(sqrt(2*svmCnorm))));
    }
    slacksum=0;
    for(i=1; i<=n; i++)  
      slacksum+=slacks[i];
    free(slacks);
    alphasum=0;
    for(i=0; i<cset.m; i++)  
      alphasum+=alpha[i]*cset.rhs[i];
    modellength=model_length_s(svmModel);
    dualitygap=(0.5*modellength*modellength+svmCnorm*(slacksum+n*ceps))
               -(alphasum-0.5*modellength*modellength);
    
    printf("Final epsilon on KKT-Conditions: %.5f\n",
	   MAX(svmModel->maxdiff,epsilon));
    printf("Upper bound on duality gap: %.5f\n", dualitygap);
    printf("Dual objective value: dval=%.5f\n",
	    alphasum-0.5*modellength*modellength);
    printf("Total number of constraints in final working set: %i (of %i)\n",(int)cset.m,(int)totconstraints);
    printf("Number of iterations: %d\n",numIt);
    printf("Number of calls to 'find_most_violated_constraint': %ld\n",argmax_count);
    if(sparm->slack_norm == 1) {
      printf("Number of SV: %ld \n",svmModel->sv_num-1);
      printf("Number of non-zero slack variables: %ld (out of %ld)\n",
	     svmModel->at_upper_bound,n);
      printf("Norm of weight vector: |w|=%.5f\n",modellength);
    }
    else if(sparm->slack_norm == 2){ 
      printf("Number of SV: %ld (including %ld at upper bound)\n",
	     svmModel->sv_num-1,svmModel->at_upper_bound);
      printf("Norm of weight vector (including L2-loss): |w|=%.5f\n",
	     modellength);
    }
    printf("Norm. sum of slack variables (on working set): sum(xi_i)/n=%.5f\n",slacksum/n);
    printf("Norm of longest difference vector: ||Psi(x,y)-Psi(x,ybar)||=%.5f\n",
	   length_of_longest_document_vector(cset.lhs,cset.m,kparm));
    printf("Runtime in cpu-seconds: %.2f (%.2f%% for QP, %.2f%% for Argmax, %.2f%% for Psi, %.2f%% for init)\n",
	   rt_total/100.0, (100.0*rt_opt)/rt_total, (100.0*rt_viol)/rt_total, 
	   (100.0*rt_psi)/rt_total, (100.0*rt_init)/rt_total);
  }
  if(struct_verbosity>=4)
    printW(sm->w,sizePsi,n,lparm->svm_c);

  if(svmModel) {
    sm->svm_model=copy_model(svmModel);
    sm->w=sm->svm_model->lin_weights; /* short cut to weight vector */
  }

  print_struct_learning_stats(sample,sm,cset,alpha,sparm);

  if(fycache) {
    for(i=0;i<n;i++)
      free_svector(fycache[i]);
    free(fycache);
  }
  if(svmModel)
    free_model(svmModel,0);
  free(alpha); 
  free(alphahist); 
  free(opti); 
  free(cset.rhs); 
  for(i=0;i<cset.m;i++) 
    free_example(cset.lhs[i],1);
  free(cset.lhs);
}
Beispiel #11
0
double add_constraint_to_constraint_cache(CCACHE *ccache, MODEL *svmModel, int exnum, SVECTOR *fydelta, double rhs, double gainthresh, int maxconst, double *rt_cachesum)
     /* add new constraint fydelta*w>rhs for example exnum to cache,
	if it is more violated (by gainthresh) than the currently most
	violated constraint in cache. if this grows the number of
	cached constraints for this example beyond maxconst, then the
	least recently used constraint is deleted. the function
	assumes that update_constraint_cache_for_model has been
	run. */
{
  double  viol,viol_gain,viol_gain_trunc;
  double  dist_ydelta;
  DOC     *doc_fydelta;
  SVECTOR *fydelta_new;
  CCACHEELEM *celem;
  int     cnum;
  double  rt2=0;

  /* compute violation of new constraint */
  doc_fydelta=create_example(1,0,1,1,fydelta);
  dist_ydelta=classify_example(svmModel,doc_fydelta);
  free_example(doc_fydelta,0);  
  viol=rhs-dist_ydelta;
  viol_gain=viol-ccache->constlist[exnum]->viol;
  viol_gain_trunc=viol-MAX(ccache->constlist[exnum]->viol,0);
  ccache->avg_viol_gain[exnum]=viol_gain;

  /* check if violation of new constraint is larger than that of the
     best cache element */
  if(viol_gain > gainthresh) {
    fydelta_new=fydelta;
    if(struct_verbosity>=2) rt2=get_runtime();
    if(svmModel->kernel_parm.kernel_type == LINEAR_KERNEL) {
      if(COMPACT_CACHED_VECTORS == 1) { /* eval sum for linear */
	fydelta_new=add_list_sort_ss_r(fydelta,COMPACT_ROUNDING_THRESH);  
	free_svector(fydelta);
      }
      else if(COMPACT_CACHED_VECTORS == 2) {
	fydelta_new=add_list_ss_r(fydelta,COMPACT_ROUNDING_THRESH); 
	free_svector(fydelta);
      }
      else if(COMPACT_CACHED_VECTORS == 3) {
	fydelta_new=add_list_ns_r(fydelta,COMPACT_ROUNDING_THRESH); 
	free_svector(fydelta);
      }
    }
    if(struct_verbosity>=2) (*rt_cachesum)+=MAX(get_runtime()-rt2,0);
    celem=ccache->constlist[exnum];
    ccache->constlist[exnum]=(CCACHEELEM *)my_malloc(sizeof(CCACHEELEM));
    ccache->constlist[exnum]->next=celem;
    ccache->constlist[exnum]->fydelta=fydelta_new;
    ccache->constlist[exnum]->rhs=rhs;
    ccache->constlist[exnum]->viol=viol;
    ccache->changed[exnum]+=2;

    /* remove last constraint in list, if list is longer than maxconst */
    cnum=2;
    for(celem=ccache->constlist[exnum];celem && celem->next && celem->next->next;celem=celem->next)
      cnum++;
    if(cnum>maxconst) {
      free_svector(celem->next->fydelta);
      free(celem->next);
      celem->next=NULL;
    }
  }
  else {
    free_svector(fydelta);
  }
  return(viol_gain_trunc);
}
Beispiel #12
0
int main_classify (int argc, char* argv[])
{
  DOC *doc;   /* test example */
  WORDSVM *words;
  long max_docs,max_words_doc,lld;
  long totdoc=0,queryid,slackid;
  long correct=0,incorrect=0,no_accuracy=0;
  long res_a=0,res_b=0,res_c=0,res_d=0,wnum,pred_format;
  long j;
  double t1,runtime=0;
  double dist,doc_label,costfactor;
  char *line,*comment; 
  FILE *predfl,*docfl;
  MODEL *model; 

  read_input_parameters(argc,argv,docfile,modelfile,predictionsfile,
			&verbosity,&pred_format);

  nol_ll(docfile,&max_docs,&max_words_doc,&lld); /* scan size of input file */
  max_words_doc+=2;
  lld+=2;

  line = (char *)my_malloc(sizeof(char)*lld);
  words = (WORDSVM *)my_malloc(sizeof(WORDSVM)*(max_words_doc+10));

  model=read_model(modelfile);

  if(model->kernel_parm.kernel_type == 0) { /* linear kernel */
    /* compute weight vector */
    add_weight_vector_to_linear_model(model);
  }
  
  if(verbosity>=2) {
    printf("Classifying test examples.."); fflush(stdout);
  }

  if ((docfl = fopen (docfile, "r")) == NULL)
  { perror (docfile); exit (1); }
  if ((predfl = fopen (predictionsfile, "w")) == NULL)
  { perror (predictionsfile); exit (1); }

  while((!feof(docfl)) && fgets(line,(int)lld,docfl)) {
    if(line[0] == '#') continue;  /* line contains comments */
    parse_document(line,words,&doc_label,&queryid,&slackid,&costfactor,&wnum,
		   max_words_doc,&comment);
    totdoc++;
    if(model->kernel_parm.kernel_type == 0) {   /* linear kernel */
      for(j=0;(words[j]).wnum != 0;j++) {  /* Check if feature numbers   */
	if((words[j]).wnum>model->totwords) /* are not larger than in     */
	  (words[j]).wnum=0;               /* model. Remove feature if   */
      }                                        /* necessary.                 */
      doc = create_example(-1,0,0,0.0,create_svector(words,comment,1.0));
      t1=get_runtime();
      dist=classify_example_linear(model,doc);
      runtime+=(get_runtime()-t1);
      free_example(doc,1);
    }
    else {                             /* non-linear kernel */
      doc = create_example(-1,0,0,0.0,create_svector(words,comment,1.0));
      t1=get_runtime();
      dist=classify_example(model,doc);
      runtime+=(get_runtime()-t1);
      free_example(doc,1);
    }
    if(dist>0) {
      if(pred_format==0) { /* old weired output format */
	fprintf(predfl,"%.8g:+1 %.8g:-1\n",dist,-dist);
      }
      if(doc_label>0) correct++; else incorrect++;
      if(doc_label>0) res_a++; else res_b++;
    }
    else {
      if(pred_format==0) { /* old weired output format */
	fprintf(predfl,"%.8g:-1 %.8g:+1\n",-dist,dist);
      }
      if(doc_label<0) correct++; else incorrect++;
      if(doc_label>0) res_c++; else res_d++;
    }
    if(pred_format==1) { /* output the value of decision function */
      fprintf(predfl,"%.8g\n",dist);
    }
    if((int)(0.01+(doc_label*doc_label)) != 1) 
      { no_accuracy=1; } /* test data is not binary labeled */
    if(verbosity>=2) {
      if(totdoc % 100 == 0) {
	printf("%ld..",totdoc); fflush(stdout);
      }
    }
  }  
  free(line);
  free(words);
  free_model(model,1);

  if(verbosity>=2) {
    printf("done\n");

/*   Note by Gary Boone                     Date: 29 April 2000        */
/*      o Timing is inaccurate. The timer has 0.01 second resolution.  */
/*        Because classification of a single vector takes less than    */
/*        0.01 secs, the timer was underflowing.                       */
    printf("Runtime (without IO) in cpu-seconds: %.2f\n",
	   (float)(runtime/100.0));
    
  }
  if((!no_accuracy) && (verbosity>=1)) {
    printf("Accuracy on test set: %.2f%% (%ld correct, %ld incorrect, %ld total)\n",(float)(correct)*100.0/totdoc,correct,incorrect,totdoc);
    printf("Precision/recall on test set: %.2f%%/%.2f%%\n",(float)(res_a)*100.0/(res_a+res_b),(float)(res_a)*100.0/(res_a+res_c));
  }

  return(0);
}
Beispiel #13
0
int Classifier::getZone(IplImage* frame, double& confidence, FrameAnnotation& fa) {
    if (!leftEye || !rightEye || !nose) {
        string err = "Classifier::getZone. Location extractors malformed.";
        throw (err);
    }

    // the roi offset
    CvPoint offset;

    // LOIs
    CvPoint leftEyeLocation;
    CvPoint rightEyeLocation;
    CvPoint noseLocation;

    // computing the confidence of the location identification
    double leftPSR;
    double rightPSR;
    double nosePSR;

    CvPoint center = fa.getLOI(Annotations::Face);
    if (!center.x || !center.y) {
        center.x = Globals::imgWidth / 2;
        center.y = Globals::imgHeight / 2;
        fa.setFace(center);
    }

    offset.x = offset.y = 0;
    IplImage* roi = (roiFunction)? roiFunction(frame, fa, offset, Annotations::Face) : 0;

    // all location extractors do identical preprocessing. Therefore, preprocess
    // once using say the left eye extractor and re-use it for all three extractors
    fftw_complex* preprocessedImage = leftEye->getPreprocessedImage((roi)? roi : frame);

    #pragma omp parallel sections num_threads(2)
    {
        #pragma omp section
        {
            leftEye->setImage(preprocessedImage);
            leftEye->apply();
            leftEye->getMaxLocation(leftEyeLocation, leftPSR);
            leftEyeLocation.x += offset.x;
            leftEyeLocation.y += offset.y;
        }

        #pragma omp section
        {
            // get the location of the right eye
            rightEye->setImage(preprocessedImage);
            rightEye->apply();
            rightEye->getMaxLocation(rightEyeLocation, rightPSR);
            rightEyeLocation.x += offset.x;
            rightEyeLocation.y += offset.y;
        }
    }

    if (roi)
        cvReleaseImage(&roi);

    center.x = (leftEyeLocation.x + rightEyeLocation.x) / 2;
    center.y = leftEyeLocation.y + Globals::noseDrop;

    fa.setNose(center);

    offset.x = offset.y = 0;
    roi = (roiFunction)? roiFunction(frame, fa, offset, Annotations::Nose) : 0;

    // free the preprocessed image
    fftw_free(preprocessedImage);

    // all location extractors do identical preprocessing. Therefore, preprocess
    // once using say the left eye extractor and re-use it for all three extractors
    preprocessedImage = nose->getPreprocessedImage((roi)? roi : frame);

    // get the location of the nose
    nose->setImage(preprocessedImage);
    nose->apply();
    nose->getMaxLocation(noseLocation, nosePSR);
    noseLocation.x += offset.x;
    noseLocation.y += offset.y;

    // free the preprocessed image
    fftw_free(preprocessedImage);

    fa.setLeftIris(leftEyeLocation);
    fa.setRightIris(rightEyeLocation);
    fa.setNose(noseLocation);

    // we are done with the images at this point. Free roi if not zero
    if (roi)
        cvReleaseImage(&roi);

    //  cout << "Confidence (L, R, N) = (" << leftPSR << ", " <<
    //    rightPSR << ")" << endl;

    // extract features vector
    vector<double> data;
    for (int i = 0; i < nFeatures; i++) {
        double value = featureExtractors[i]->extract(&fa);
        data.push_back(value);
    }

    // normalize
    normalize(data);

    // create SVM Light objects to classify
    DOC* doc;
    WORD* words = (WORD*)malloc(sizeof(WORD) * (nFeatures + 1));

    for (int i = 0; i < nFeatures; i++) {
        words[i].wnum = featureExtractors[i]->getId();
        words[i].weight = data[i];
    }

    // SVM Light expects that the features vector has a zero element
    // to indicate termination and hence
    words[nFeatures].wnum = 0;
    words[nFeatures].weight = 0.0;

    // create doc
    string comment = "Gaze SVM";
    doc = create_example(-1, 0, 0, 0.0, create_svector(words, (char*)comment.c_str(), 1.0));

    int maxIndex = 0;
    confidence = -FLT_MAX;

    double dists[Globals::numZones];

    // classify using each zone model
    #pragma omp parallel for num_threads(Globals::numZones)
    for (unsigned int i = 0; i < Globals::numZones; i++) {
        if (kernelType == Trainer::Linear)
            dists[i] = classify_example_linear(models[i], doc);
        else
            dists[i] = classify_example(models[i], doc);
    }

    for (unsigned int i = 0; i < Globals::numZones; i++) {
        if (confidence < dists[i]) {
            confidence = dists[i];
            maxIndex = i + 1;
        }
    }

    free_example(doc, 1);
    free(words);

    return maxIndex;
}
void svm_learn_struct_joint(SAMPLE sample, STRUCT_LEARN_PARM *sparm,
			    LEARN_PARM *lparm, KERNEL_PARM *kparm, 
			    STRUCTMODEL *sm, int alg_type)
{
  int         i,j;
  int         numIt=0;
  long        argmax_count=0;
  long        totconstraints=0;
  long        kernel_type_org;
  double      epsilon,epsilon_cached;
  double      lossval,factor,dist;
  double      margin=0;
  double      slack, slacksum, ceps;
  double      dualitygap,modellength,alphasum;
  long        sizePsi;
  double      *alpha=NULL;
  long        *alphahist=NULL,optcount=0;
  CONSTSET    cset;
  SVECTOR     *diff=NULL;
  double      *diff_n=NULL;
  SVECTOR     *fy, *fybar, *f, **fycache, *lhs;
  MODEL       *svmModel=NULL;
  LABEL       ybar;
  DOC         *doc;

  long        n=sample.n;
  EXAMPLE     *ex=sample.examples;
  double      rt_total=0,rt_opt=0,rt_init=0,rt_psi=0,rt_viol=0,rt_kernel=0;
  double      rt1,rt2;
  double      progress,progress_old;

  /*
  SVECTOR     ***fydelta_cache=NULL;
  double      **loss_cache=NULL;
  int         cache_size=0;
  */
  CCACHE      *ccache=NULL;
  int         cached_constraint;

  rt1=get_runtime();

  init_struct_model(sample,sm,sparm,lparm,kparm); 
  sizePsi=sm->sizePsi+1;          /* sm must contain size of psi on return */

  if(sparm->slack_norm == 1) {
    lparm->svm_c=sparm->C;          /* set upper bound C */
    lparm->sharedslack=1;
  }
  else if(sparm->slack_norm == 2) {
    printf("ERROR: The joint algorithm does not apply to L2 slack norm!"); 
    fflush(stdout);
    exit(0); 
  }
  else {
    printf("ERROR: Slack norm must be L1 or L2!"); fflush(stdout);
    exit(0);
  }


  lparm->biased_hyperplane=0;     /* set threshold to zero */
  epsilon=100.0;                  /* start with low precision and
				     increase later */
  epsilon_cached=epsilon;         /* epsilon to use for iterations
				     using constraints constructed
				     from the constraint cache */

  cset=init_struct_constraints(sample, sm, sparm);
  if(cset.m > 0) {
    alpha=(double *)realloc(alpha,sizeof(double)*cset.m);
    alphahist=(long *)realloc(alphahist,sizeof(long)*cset.m);
    for(i=0; i<cset.m; i++) {
      alpha[i]=0;
      alphahist[i]=-1; /* -1 makes sure these constraints are never removed */
    }
  }
  kparm->gram_matrix=NULL;
  if((alg_type == DUAL_ALG) || (alg_type == DUAL_CACHE_ALG))
    kparm->gram_matrix=init_kernel_matrix(&cset,kparm);

  /* set initial model and slack variables */
  svmModel=(MODEL *)my_malloc(sizeof(MODEL));
  lparm->epsilon_crit=epsilon;
  svm_learn_optimization(cset.lhs,cset.rhs,cset.m,sizePsi+n,
			 lparm,kparm,NULL,svmModel,alpha);
  add_weight_vector_to_linear_model(svmModel);
  sm->svm_model=svmModel;
  sm->w=svmModel->lin_weights; /* short cut to weight vector */

  /* create a cache of the feature vectors for the correct labels */
  fycache=(SVECTOR **)malloc(n*sizeof(SVECTOR *));
  for(i=0;i<n;i++) {
    fy=psi(ex[i].x,ex[i].y,sm,sparm);
    if(kparm->kernel_type == LINEAR) {
      diff=add_list_ss(fy); /* store difference vector directly */
      free_svector(fy);
      fy=diff;
    }
    fycache[i]=fy;
  }

  /* initialize the constraint cache */
  if(alg_type == DUAL_CACHE_ALG) {
    ccache=create_constraint_cache(sample,sparm);
  }

  rt_init+=MAX(get_runtime()-rt1,0);
  rt_total+=MAX(get_runtime()-rt1,0);

    /*****************/
   /*** main loop ***/
  /*****************/
  do { /* iteratively find and add constraints to working set */

      if(struct_verbosity>=1) { 
	printf("Iter %i: ",++numIt); 
	fflush(stdout);
      }
      
      rt1=get_runtime();

      /**** compute current slack ****/
      slack=0;
      for(j=0;j<cset.m;j++) 
	slack=MAX(slack,cset.rhs[j]-classify_example(svmModel,cset.lhs[j]));
      
      /**** find a violated joint constraint ****/
      lhs=NULL;
      dist=0;
      if(alg_type == DUAL_CACHE_ALG) {
	/* see if it is possible to construct violated constraint from cache */
	update_constraint_cache_for_model(ccache, svmModel);
	dist=find_most_violated_joint_constraint_in_cache(ccache,&lhs,&margin);
      }

      rt_total+=MAX(get_runtime()-rt1,0);

      /* Is there a sufficiently violated constraint in cache? */
      if(dist-slack > MAX(epsilon/10,sparm->epsilon)) { 
	/* use constraint from cache */
	rt1=get_runtime();
	cached_constraint=1;
	if(kparm->kernel_type == LINEAR) {
	  diff=add_list_ns(lhs); /* Linear case: compute weighted sum */
	  free_svector_shallow(lhs);
	}
	else { /* Non-linear case: make sure we have deep copy for cset */
	  diff=copy_svector(lhs); 
	  free_svector_shallow(lhs);
	}
	rt_total+=MAX(get_runtime()-rt1,0);
      }
      else { 
	/* do not use constraint from cache */
	rt1=get_runtime();
	cached_constraint=0;
	if(lhs)
	  free_svector_shallow(lhs);
	lhs=NULL;
	if(kparm->kernel_type == LINEAR) {
	  diff_n=create_nvector(sm->sizePsi);
	  clear_nvector(diff_n,sm->sizePsi);
	}
	margin=0;
	progress=0;
	progress_old=progress;
	rt_total+=MAX(get_runtime()-rt1,0);

	/**** find most violated joint constraint ***/
	for(i=0; i<n; i++) {
	  
	  rt1=get_runtime();
      
	  progress+=10.0/n;
	  if((struct_verbosity==1) && (((int)progress_old) != ((int)progress)))
	    {printf(".");fflush(stdout); progress_old=progress;}
	  if(struct_verbosity>=2)
	    {printf("."); fflush(stdout);}

	  rt2=get_runtime();
	  argmax_count++;
	  if(sparm->loss_type == SLACK_RESCALING) 
	    ybar=find_most_violated_constraint_slackrescaling(ex[i].x,
							      ex[i].y,sm,
							      sparm);
	  else
	    ybar=find_most_violated_constraint_marginrescaling(ex[i].x,
							       ex[i].y,sm,
							       sparm);
	  rt_viol+=MAX(get_runtime()-rt2,0);
	  
	  if(empty_label(ybar)) {
	    printf("ERROR: empty label was returned for example (%i)\n",i);
	    /* exit(1); */
	    continue;
	  }
	  
	  /**** get psi(x,y) and psi(x,ybar) ****/
	  rt2=get_runtime();
	  fy=copy_svector(fycache[i]); /*<= fy=psi(ex[i].x,ex[i].y,sm,sparm);*/
	  fybar=psi(ex[i].x,ybar,sm,sparm);
	  rt_psi+=MAX(get_runtime()-rt2,0);
	  lossval=loss(ex[i].y,ybar,sparm);
	  free_label(ybar);
	  
	  /**** scale feature vector and margin by loss ****/
	  if(sparm->loss_type == SLACK_RESCALING)
	    factor=lossval/n;
	  else                 /* do not rescale vector for */
	    factor=1.0/n;      /* margin rescaling loss type */
	  for(f=fy;f;f=f->next)
	    f->factor*=factor;
	  for(f=fybar;f;f=f->next)
	    f->factor*=-factor;
	  append_svector_list(fybar,fy);   /* compute fy-fybar */
	  
	  /**** add current fy-fybar and loss to cache ****/
	  if(alg_type == DUAL_CACHE_ALG) {
	    if(kparm->kernel_type == LINEAR) 
	      add_constraint_to_constraint_cache(ccache,svmModel,i,
						 add_list_ss(fybar),
						 lossval/n,sparm->ccache_size);
	    else
	      add_constraint_to_constraint_cache(ccache,svmModel,i,
						 copy_svector(fybar),
						 lossval/n,sparm->ccache_size);
	  }

	  /**** add current fy-fybar to constraint and margin ****/
	  if(kparm->kernel_type == LINEAR) {
	    add_list_n_ns(diff_n,fybar,1.0); /* add fy-fybar to sum */
	    free_svector(fybar);
	  }
	  else {
	    append_svector_list(fybar,lhs);  /* add fy-fybar to vector list */
	    lhs=fybar;
	  }
	  margin+=lossval/n;                 /* add loss to rhs */
	  
	  rt_total+=MAX(get_runtime()-rt1,0);

	} /* end of example loop */

	rt1=get_runtime();

	/* create sparse vector from dense sum */
	if(kparm->kernel_type == LINEAR) {
	  diff=create_svector_n(diff_n,sm->sizePsi,"",1.0);
	  free_nvector(diff_n);
	}
	else {
	  diff=lhs;
	}

	rt_total+=MAX(get_runtime()-rt1,0);

      } /* end of finding most violated joint constraint */

      rt1=get_runtime();

      /**** if `error', then add constraint and recompute QP ****/
      doc=create_example(cset.m,0,1,1,diff);
      dist=classify_example(svmModel,doc);
      ceps=MAX(0,margin-dist-slack);
      if(slack > (margin-dist+0.000001)) {
	printf("\nWARNING: Slack of most violated constraint is smaller than slack of working\n");
	printf("         set! There is probably a bug in 'find_most_violated_constraint_*'.\n");
	printf("slack=%f, newslack=%f\n",slack,margin-dist);
	/* exit(1); */
      }
      if(ceps > sparm->epsilon) { 
	/**** resize constraint matrix and add new constraint ****/
	cset.lhs=(DOC **)realloc(cset.lhs,sizeof(DOC *)*(cset.m+1));
	if(sparm->slack_norm == 1) 
	  cset.lhs[cset.m]=create_example(cset.m,0,1,1,diff);
	else if(sparm->slack_norm == 2)
	  exit(1);
	cset.rhs=(double *)realloc(cset.rhs,sizeof(double)*(cset.m+1));
	cset.rhs[cset.m]=margin;
	alpha=(double *)realloc(alpha,sizeof(double)*(cset.m+1));
	alpha[cset.m]=0;
	alphahist=(long *)realloc(alphahist,sizeof(long)*(cset.m+1));
	alphahist[cset.m]=optcount;
	cset.m++;
	totconstraints++;
	if((alg_type == DUAL_ALG) || (alg_type == DUAL_CACHE_ALG)) {
	  if(struct_verbosity>=1) {
	    printf(":");fflush(stdout);
	  }
	  rt2=get_runtime();
	  kparm->gram_matrix=update_kernel_matrix(kparm->gram_matrix,cset.m-1,
						  &cset,kparm);
	  rt_kernel+=MAX(get_runtime()-rt2,0);
	}
	
	/**** get new QP solution ****/
	if(struct_verbosity>=1) {
	  printf("*");fflush(stdout);
	}
	rt2=get_runtime();
	/* set svm precision so that higher than eps of most violated constr */
	if(cached_constraint) {
	  epsilon_cached=MIN(epsilon_cached,MAX(ceps,sparm->epsilon)); 
	  lparm->epsilon_crit=epsilon_cached/2; 
	}
	else {
	  epsilon=MIN(epsilon,MAX(ceps,sparm->epsilon)); /* best eps so far */
	  lparm->epsilon_crit=epsilon/2; 
	  epsilon_cached=epsilon;
	}
	free_model(svmModel,0);
	svmModel=(MODEL *)my_malloc(sizeof(MODEL));
	/* Run the QP solver on cset. */
	kernel_type_org=kparm->kernel_type;
	if((alg_type == DUAL_ALG) || (alg_type == DUAL_CACHE_ALG))
	  kparm->kernel_type=GRAM; /* use kernel stored in kparm */
	svm_learn_optimization(cset.lhs,cset.rhs,cset.m,sizePsi+n,
			       lparm,kparm,NULL,svmModel,alpha);
	kparm->kernel_type=kernel_type_org; 
	svmModel->kernel_parm.kernel_type=kernel_type_org;
	/* Always add weight vector, in case part of the kernel is
	   linear. If not, ignore the weight vector since its
	   content is bogus. */
	add_weight_vector_to_linear_model(svmModel);
	sm->svm_model=svmModel;
	sm->w=svmModel->lin_weights; /* short cut to weight vector */
	optcount++;
	/* keep track of when each constraint was last
	   active. constraints marked with -1 are not updated */
	for(j=0;j<cset.m;j++) 
	  if((alphahist[j]>-1) && (alpha[j] != 0))  
	    alphahist[j]=optcount;
	rt_opt+=MAX(get_runtime()-rt2,0);
	
	/* Check if some of the linear constraints have not been
	   active in a while. Those constraints are then removed to
	   avoid bloating the working set beyond necessity. */
	if(struct_verbosity>=2)
	  printf("Reducing working set...");fflush(stdout);
	remove_inactive_constraints(&cset,alpha,optcount,alphahist,50);
	if(struct_verbosity>=2)
	  printf("done. (NumConst=%d) ",cset.m);
      }
      else {
	free_svector(diff);
      }

      if(struct_verbosity>=1)
	printf("(NumConst=%d, SV=%ld, CEps=%.4f, QPEps=%.4f)\n",cset.m,
	       svmModel->sv_num-1,ceps,svmModel->maxdiff);

      free_example(doc,0);
	
      rt_total+=MAX(get_runtime()-rt1,0);

  } while((ceps > sparm->epsilon) || 
	  finalize_iteration(ceps,cached_constraint,sample,sm,cset,alpha,sparm)
	 );
  

  if(struct_verbosity>=1) {
    /**** compute sum of slacks ****/
    /**** WARNING: If positivity constraints are used, then the
	  maximum slack id is larger than what is allocated
	  below ****/
    slacksum=0;
    if(sparm->slack_norm == 1) {
      for(j=0;j<cset.m;j++) 
	slacksum=MAX(slacksum,
		     cset.rhs[j]-classify_example(svmModel,cset.lhs[j]));
      }
    else if(sparm->slack_norm == 2) {
      exit(1);
    }
    alphasum=0;
    for(i=0; i<cset.m; i++)  
      alphasum+=alpha[i]*cset.rhs[i];
    modellength=model_length_s(svmModel,kparm);
    dualitygap=(0.5*modellength*modellength+sparm->C*(slacksum+ceps))
               -(alphasum-0.5*modellength*modellength);
    
    printf("Final epsilon on KKT-Conditions: %.5f\n",
	   MAX(svmModel->maxdiff,ceps));
    printf("Upper bound on duality gap: %.5f\n", dualitygap);
    printf("Dual objective value: dval=%.5f\n",
	    alphasum-0.5*modellength*modellength);
    printf("Total number of constraints in final working set: %i (of %i)\n",(int)cset.m,(int)totconstraints);
    printf("Number of iterations: %d\n",numIt);
    printf("Number of calls to 'find_most_violated_constraint': %ld\n",argmax_count);
    if(sparm->slack_norm == 1) {
      printf("Number of SV: %ld \n",svmModel->sv_num-1);
      printf("Norm of weight vector: |w|=%.5f\n",
	     model_length_s(svmModel,kparm));
    }
    else if(sparm->slack_norm == 2){ 
      printf("Number of SV: %ld (including %ld at upper bound)\n",
	     svmModel->sv_num-1,svmModel->at_upper_bound);
      printf("Norm of weight vector (including L2-loss): |w|=%.5f\n",
	     model_length_s(svmModel,kparm));
    }
    printf("Value of slack variable (on working set): xi=%.5f\n",slacksum);
    printf("Norm of longest difference vector: ||Psi(x,y)-Psi(x,ybar)||=%.5f\n",
	   length_of_longest_document_vector(cset.lhs,cset.m,kparm));
    printf("Runtime in cpu-seconds: %.2f (%.2f%% for QP, %.2f%% for kernel, %.2f%% for Argmax, %.2f%% for Psi, %.2f%% for init)\n",
	   rt_total/100.0, (100.0*rt_opt)/rt_total, (100.0*rt_kernel)/rt_total,
	   (100.0*rt_viol)/rt_total, (100.0*rt_psi)/rt_total, 
	   (100.0*rt_init)/rt_total);
  }
  if(ccache) {
    long cnum=0;
    CCACHEELEM *celem;
    for(i=0;i<n;i++) 
      for(celem=ccache->constlist[i];celem;celem=celem->next) 
	cnum++;
    printf("Final number of constraints in cache: %ld\n",cnum);
  }
  if(struct_verbosity>=4)
    printW(sm->w,sizePsi,n,lparm->svm_c);

  if(svmModel) {
    sm->svm_model=copy_model(svmModel);
    sm->w=sm->svm_model->lin_weights; /* short cut to weight vector */
  }

  print_struct_learning_stats(sample,sm,cset,alpha,sparm);

  if(ccache)    
    free_constraint_cache(ccache);
  for(i=0;i<n;i++)
    free_svector(fycache[i]);
  free(fycache);
  if(svmModel)
    free_model(svmModel,0);
  free(alpha); 
  free(alphahist); 
  free(cset.rhs); 
  for(i=0;i<cset.m;i++) 
    free_example(cset.lhs[i],1);
  free(cset.lhs);
  if(kparm->gram_matrix)
    free_matrix(kparm->gram_matrix);
}
Beispiel #15
0
IMP_S32 ipProcessTargetClassifierInternal( IpTargetClassifier *pstClassifier )
{
	IpTrackedTarget *pstTarget;
	IpTargetPosition *pstPos0;
	IMP_U8 *value1,*value2;
	IMP_S32 i,j,m,n,cnt,c,targetType;
	IpClassFeature *astClassFeature;

	WORDSVM words[15];

	svm_node x[15];
	DOC *doc;

 	IMP_RECT_S *rc;
	IMP_RECT_S pstRectImg;
	IMP_S32 j13;//=blobObjCur.pt.y+blobObjCur.blobHeight/2-blobObjCur.blobHeight/3;
	IMP_S32 j23;//=blobObjCur.pt.y+blobObjCur.blobHeight/2-2*blobObjCur.blobHeight/3;

	IMP_S32 is_possible_human,is_possible_vehicle;

	IMP_FLOAT type;
	IMP_S32 m13,m23,mBG,mFG,mPerimeter;
	PEA_RESULT_S *pstResult = pstClassifier->pstResult;

	PEA_DETECTED_REGIONSET_S *pstRegions = &pstResult->stDRegionSet;
	GRAY_IMAGE_S *pstImgFgOrg = pstRegions->pstImgFgOrg;
	GRAY_IMAGE_S *pstImgMediate = pstRegions->pstImgMediate;

	IMP_S32 s32Width = pstImgFgOrg->s32W;
	IMP_S32 s32Height = pstImgFgOrg->s32H;

	IMP_S32 blobWidth,blobHeight;
	IpClassifierPara *pstParams=&pstClassifier->stPara;
	IMP_S32 s32UseBorderConstrain=pstParams->s32UseBorderConstrain;
	IMP_S32 s32SceneMode = 0;

	pstTarget = pstResult->stTrackedTargetSet.astTargets;
	cnt = pstResult->stTrackedTargetSet.s32UsedTotal;
	astClassFeature=pstResult->stTrackedTargetSet.astClassFeature;

	if (pstParams->pstRule->stEvrmentTgtPara.stScenePara.s32SceneLmt)
	{
		s32SceneMode = pstParams->pstRule->stEvrmentTgtPara.stScenePara.s32SceneMode;
	}

	if (pstParams->s8SvmFuncType==DISABLE)
	{
		return -1;
	}
	IMP_GrayImageClone(pstImgFgOrg,pstImgMediate);
	for( i=0, j=0; i<IMP_MAX_TGT_CNT; i++ )
	{
		if (ipTrackedTargetIsActive(pstTarget))
		{
#if 0		
			{//added by mzhang, just for debug.
				pstTarget->stTargetInfo.s32HumanLikehood = 100;
				pstTarget->stTargetInfo.s32VehicleLikehood = -1;
				pstTarget->stTargetInfo.u32Type=IMP_TGT_TYPE_HUMAN; 
				continue; 
			}
#endif
			{
				targetType = 0;

				if(!targetType)
				{
					pstPos0 = ipTargetTrajectoryGetPosition( &pstTarget->stTrajectory, 0 );

					rc=&pstPos0->stRg;

					if( s32UseBorderConstrain )
					{
						pstRectImg.s16X1 = 0;
						pstRectImg.s16Y1 = 0;
						pstRectImg.s16X2 = s32Width - 1;
						pstRectImg.s16Y2 = s32Height - 1;

					}
					if (!s32UseBorderConstrain ||
						s32UseBorderConstrain && !ipBoundaryConditionsJudgment( rc, &pstRectImg, s32UseBorderConstrain )//加入了限定要求并且满足边界条件
						)
					{
						m13=0;
						m23=0;
						mBG=0;
						mFG=0;
						mPerimeter=0;

						j13=rc->s16Y2-(rc->s16Y2-rc->s16Y1)/3;
						j23=rc->s16Y2-2*(rc->s16Y2-rc->s16Y1)/3;

						for (m=rc->s16X1;m<rc->s16X2;m++)
						{
							value1=pstImgFgOrg->pu8Data+ s32Width*j13 + m;
							value2=pstImgFgOrg->pu8Data+ s32Width*j23 + m;
							if(*value1>0)
							{
								m13++;
							}
							if(*value2>0)
							{
								m23++;
							}
						}

						for (m=rc->s16Y1;m<rc->s16Y2;m++)
						{
							for (n=rc->s16X1;n<rc->s16X2;n++)
							{
								value1=pstImgFgOrg->pu8Data+ s32Width*m + n;
								if (*value1==0)
								{
									mBG++;
								}
							}
						}



						for (m=rc->s16Y1;m<rc->s16Y2;m++)
						{

							for (n=rc->s16X1;n<rc->s16X2;n++)
							{
								value1=pstImgMediate->pu8Data+ s32Width*m + n;
								if (*value1>0)
								{
									mPerimeter++;
									*value1=100;
									break;
								}
							}

							for (n=rc->s16X2;n<rc->s16X1;n--)
							{
								value1=pstImgMediate->pu8Data+ s32Width*m + n;
								if (*value1==100)
								{
									break;
								}
								else if (*value1>0)
								{
									mPerimeter++;
									*value1=100;
									break;
								}
							}
						}

						for (n=rc->s16X1;n<rc->s16X2;n++)
						{

							for (m=rc->s16Y1;m<rc->s16Y2;m++)
							{
								value1=pstImgMediate->pu8Data+ s32Width*m + n;
								if (*value1==100)
									break;
								if (*value1>0)
								{
									mPerimeter++;
									*value1=100;
									break;
								}
							}

							for (m=rc->s16Y2;m<rc->s16Y1;m--)
							{
								value1=pstImgMediate->pu8Data+ s32Width*m + n;
								if (*value1==100)
									break;
								if (*value1>0)
								{
									mPerimeter++;
									*value1=100;
									break;
								}
							}
						}
						blobWidth=rc->s16X2-rc->s16X1+1;
						blobHeight=rc->s16Y2-rc->s16Y1+1;

						mFG=blobWidth*blobHeight-mBG;
						astClassFeature->_P=(IMP_DOUBLE)(blobWidth)/blobHeight;
						astClassFeature->_P13=(IMP_DOUBLE)(m13)/blobHeight;
						astClassFeature->_P23=(IMP_DOUBLE)(m23)/blobHeight;

						astClassFeature->_I=(IMP_DOUBLE)(mBG)/(blobWidth*blobHeight);
						astClassFeature->_D=(IMP_DOUBLE)mPerimeter/mFG;


						ipComputeHuMomentInvariants(pstImgFgOrg,rc,s32Width,s32Height,astClassFeature->_Hu,&astClassFeature->_Axis);

						astClassFeature->_Delta=0;

						astClassFeature->label=pstParams->s32ClassType;
						if (pstParams->s8SvmFuncType==CLASSIFY)
						{

							for (i=0;i<=13;i++)
							{
								x[i].index = i+1;
							}

							x[14].index = -1;
							c=-1;
							x[0].value = astClassFeature->_Axis;
							x[1].value = astClassFeature->_D;
							x[2].value = astClassFeature->_Delta;
							x[3].value = astClassFeature->_Hu[0];
							x[4].value = astClassFeature->_Hu[1];
							x[5].value = astClassFeature->_Hu[2];
							x[6].value = astClassFeature->_Hu[3];
							x[7].value = astClassFeature->_Hu[4];
							x[8].value = astClassFeature->_Hu[5];
							x[9].value = astClassFeature->_Hu[6];
							x[10].value = astClassFeature->_I;
							x[11].value = astClassFeature->_P;
							x[12].value = astClassFeature->_P13;
							x[13].value = astClassFeature->_P23;


							// 				(*label)[i]=classFeatureTemp.label;
							// 				(queryid)=0;
							// 				(slackid)=0;
							// 				(costfactor)=1;

							words[0].wnum = 1;
							words[0].weight = (float)astClassFeature->_Axis;
							words[1].wnum = 2;
							words[1].weight = (float)astClassFeature->_D;
							words[2].wnum = 3;
							words[2].weight = (float)astClassFeature->_Delta;

							words[3].wnum = 4;
							words[3].weight = (float)astClassFeature->_Hu[0];
							words[4].wnum = 5;
							words[4].weight = (float)astClassFeature->_Hu[1];
							words[5].wnum = 6;
							words[5].weight = (float)astClassFeature->_Hu[2];
							words[6].wnum = 7;
							words[6].weight = (float)astClassFeature->_Hu[3];
							words[7].wnum = 8;
							words[7].weight = (float)astClassFeature->_Hu[4];
							words[8].wnum = 9;
							words[8].weight = (float)astClassFeature->_Hu[5];
							words[9].wnum = 10;
							words[9].weight = (float)astClassFeature->_Hu[6];

							words[10].wnum = 11;
							words[10].weight = (float)astClassFeature->_I;
							words[11].wnum = 12;
							words[11].weight = (float)astClassFeature->_P;
							words[12].wnum = 13;
							words[12].weight = (float)astClassFeature->_P13;
							words[13].wnum = 14;
							words[13].weight = (float)astClassFeature->_P23;

							words[14].wnum=0;

							//c= (IMP_S32)svm_predict(pstParams->m_model, x);

							doc = create_example(-1,0,0,0.0,create_svector(words,"",1.0));

							type=(float)classify_example(pstParams->pstModel,doc);
							if (type>0 && type <2)
							{
								c=IMP_TGT_TYPE_HUMAN;
								pstTarget->stTargetInfo.s32HumanLikehood++;
								pstTarget->stTargetInfo.s32VehicleLikehood--;

							}
							else if (type>2 && type <4)
							{
								c=IMP_TGT_TYPE_VEHICLE;

								pstTarget->stTargetInfo.s32VehicleLikehood++;
								pstTarget->stTargetInfo.s32HumanLikehood--;
							}

							free_example(doc,1);
						}
				}
			}
			}

			if (pstTarget->stTargetInfo.s32VehicleLikehood>100)
			{
				pstTarget->stTargetInfo.s32VehicleLikehood=100;
			}
			else if(pstTarget->stTargetInfo.s32VehicleLikehood<-1)
			{
				pstTarget->stTargetInfo.s32VehicleLikehood=-1;
			}

			if (pstTarget->stTargetInfo.s32VehicleLikehood>100)
			{
				pstTarget->stTargetInfo.s32VehicleLikehood=100;
			}
			else if(pstTarget->stTargetInfo.s32VehicleLikehood<-1)
			{
				pstTarget->stTargetInfo.s32VehicleLikehood=-1;
			}

			if (pstTarget->stTargetInfo.s32HumanLikehood &&
				pstTarget->stTargetInfo.s32HumanLikehood >= pstTarget->stTargetInfo.s32VehicleLikehood
				)
			{
				pstTarget->stTargetInfo.u32Type=IMP_TGT_TYPE_HUMAN;
				//printf("target Type is HUMAN\n");
			}
			else if (pstTarget->stTargetInfo.s32VehicleLikehood &&
				pstTarget->stTargetInfo.s32VehicleLikehood >= pstTarget->stTargetInfo.s32HumanLikehood
				)
			{
				pstTarget->stTargetInfo.u32Type=IMP_TGT_TYPE_VEHICLE;
				//printf("target Type is VEHICLE\n");
			}
			else
			{
				pstTarget->stTargetInfo.u32Type=IMP_TGT_TYPE_UNKNOWN;
				//printf("target Type is UNKNOWN\n");
			}
		}

		j += pstTarget->s32Used ? 1 : 0;
		if( j>=cnt ) break;
		astClassFeature++;
		pstTarget++;
	}


	return 1;
}
Beispiel #16
0
void svm_learn_struct(SAMPLE sample, STRUCT_LEARN_PARM *sparm,
		      LEARN_PARM *lparm, KERNEL_PARM *kparm, 
		      STRUCTMODEL *sm)
{
  int         i,j;
  int         numIt=0;
  long        newconstraints=0, activenum=0; 
  int         opti_round, *opti;
  long        old_numConst=0;
  double      epsilon;
  long        tolerance;
  double      lossval,factor;
  double      margin=0;
  double      slack, *slacks, slacksum;
  long        sizePsi;
  double      *alpha=NULL;
  CONSTSET    cset;
  SVECTOR     *diff=NULL;
  SVECTOR     *fy, *fybar, *f;
  SVECTOR     *slackvec;
  WORD        slackv[2];
  MODEL       *svmModel=NULL;
  KERNEL_CACHE *kcache=NULL;
  LABEL       ybar;
  DOC         *doc;

  long        n=sample.n;
  EXAMPLE     *ex=sample.examples;
  double      rt_total=0.0, rt_opt=0.0;
  long        rt1,rt2;

  init_struct_model(sample,sm,sparm); 
  sizePsi=sm->sizePsi+1;          /* sm must contain size of psi on return */

  /* initialize example selection heuristic */ 
  opti=(int*)my_malloc(n*sizeof(int));
  for(i=0;i<n;i++) {
    opti[i]=0;
  }
  opti_round=0;

  if(sparm->slack_norm == 1) {
    lparm->svm_c=sparm->C;          /* set upper bound C */
    lparm->sharedslack=1;
  }
  else if(sparm->slack_norm == 2) {
    lparm->svm_c=999999999999999.0; /* upper bound C must never be reached */
    lparm->sharedslack=0;
    if(kparm->kernel_type != LINEAR) {
      printf("ERROR: Kernels are not implemented for L2 slack norm!"); 
      fflush(stdout);
      exit(0);
    }
  }
  else {
    printf("ERROR: Slack norm must be L1 or L2!"); fflush(stdout);
    exit(0);
  }


  epsilon=1.0;                    /* start with low precision and
				     increase later */
  tolerance=n/100;                /* increase precision, whenever less
                                     than that number of constraints
                                     is not fulfilled */
  lparm->biased_hyperplane=0;     /* set threshold to zero */

  cset=init_struct_constraints(sample, sm, sparm);
  if(cset.m > 0) {
    alpha=realloc(alpha,sizeof(double)*cset.m);
    for(i=0; i<cset.m; i++) 
      alpha[i]=0;
  }

  /* set initial model and slack variables*/
  svmModel=(MODEL *)my_malloc(sizeof(MODEL));
  svm_learn_optimization(cset.lhs,cset.rhs,cset.m,sizePsi+n,
			 lparm,kparm,NULL,svmModel,alpha);
  add_weight_vector_to_linear_model(svmModel);
  sm->svm_model=svmModel;
  sm->w=svmModel->lin_weights; /* short cut to weight vector */

  printf("Starting Iterations\n");

    /*****************/
   /*** main loop ***/
  /*****************/
  do { /* iteratively increase precision */

    epsilon=MAX(epsilon*0.09999999999,sparm->epsilon);
    if(epsilon == sparm->epsilon)   /* for final precision, find all SV */
      tolerance=0;
    lparm->epsilon_crit=epsilon/2;  /* svm precision must be higher than eps */
    if(struct_verbosity>=1)
      printf("Setting current working precision to %g.\n",epsilon);

    do { /* iteration until (approx) all SV are found for current
            precision and tolerance */
      
      old_numConst=cset.m;
      opti_round++;
      activenum=n;

      do { /* go through examples that keep producing new constraints */

	if(struct_verbosity>=1) { 
	  printf("--Iteration %i (%ld active): ",++numIt,activenum); 
	  fflush(stdout);
	}
	
	for(i=0; i<n; i++) { /*** example loop ***/
	  
	  rt1=get_runtime();
	    
	  if(opti[i] != opti_round) {/* if the example is not shrunk
	                                away, then see if it is necessary to 
					add a new constraint */
	    if(sparm->loss_type == SLACK_RESCALING) 
	      ybar=find_most_violated_constraint_slackrescaling(ex[i].x,
								ex[i].y,sm,
								sparm);
	    else
	      ybar=find_most_violated_constraint_marginrescaling(ex[i].x,
								 ex[i].y,sm,
								 sparm);
	    
	    if(empty_label(ybar)) {
	      if(opti[i] != opti_round) {
		activenum--;
		opti[i]=opti_round; 
	      }
	      if(struct_verbosity>=2)
		printf("no-incorrect-found(%i) ",i);
	      continue;
	    }
	  
	    /**** get psi(y)-psi(ybar) ****/
	    fy=psi(ex[i].x,ex[i].y,sm,sparm);
	    fybar=psi(ex[i].x,ybar,sm,sparm);
	    
	    /**** scale feature vector and margin by loss ****/
	    lossval=loss(ex[i].y,ybar,sparm);
	    if(sparm->slack_norm == 2)
	      lossval=sqrt(lossval);
	    if(sparm->loss_type == SLACK_RESCALING)
	      factor=lossval;
	    else               /* do not rescale vector for */
	      factor=1.0;      /* margin rescaling loss type */
	    for(f=fy;f;f=f->next)
	      f->factor*=factor;
	    for(f=fybar;f;f=f->next)
	      f->factor*=-factor;
	    margin=lossval;

	    /**** create constraint for current ybar ****/
	    append_svector_list(fy,fybar);/* append the two vector lists */
	    doc=create_example(cset.m,0,i+1,1,fy);

	    /**** compute slack for this example ****/
	    slack=0;
	    for(j=0;j<cset.m;j++) 
	      if(cset.lhs[j]->slackid == i+1) {
		if(sparm->slack_norm == 2) /* works only for linear kernel */
		  slack=MAX(slack,cset.rhs[j]
			          -(classify_example(svmModel,cset.lhs[j])
				    -sm->w[sizePsi+i]/(sqrt(2*sparm->C))));
		else
		  slack=MAX(slack,
			   cset.rhs[j]-classify_example(svmModel,cset.lhs[j]));
	      }
	    
	    /**** if `error' add constraint and recompute ****/
	    if((classify_example(svmModel,doc)+slack)<(margin-epsilon)) { 
	      if(struct_verbosity>=2)
		{printf("(%i) ",i); fflush(stdout);}
	      if(struct_verbosity==1)
		{printf("."); fflush(stdout);}
	      
	      /**** resize constraint matrix and add new constraint ****/
	      cset.m++;
	      cset.lhs=realloc(cset.lhs,sizeof(DOC *)*cset.m);
	      if(kparm->kernel_type == LINEAR) {
		diff=add_list_ss(fy); /* store difference vector directly */
		if(sparm->slack_norm == 1) 
		  cset.lhs[cset.m-1]=create_example(cset.m-1,0,i+1,1,
						    copy_svector(diff));
		else if(sparm->slack_norm == 2) {
		  /**** add squared slack variable to feature vector ****/
		  slackv[0].wnum=sizePsi+i;
		  slackv[0].weight=1/(sqrt(2*sparm->C));
		  slackv[1].wnum=0; /*terminator*/
		  slackvec=create_svector(slackv,"",1.0);
		  cset.lhs[cset.m-1]=create_example(cset.m-1,0,i+1,1,
						    add_ss(diff,slackvec));
		  free_svector(slackvec);
		}
		free_svector(diff);
	      }
	      else { /* kernel is used */
		if(sparm->slack_norm == 1) 
		  cset.lhs[cset.m-1]=create_example(cset.m-1,0,i+1,1,
						    copy_svector(fy));
		else if(sparm->slack_norm == 2)
		  exit(1);
	      }
	      cset.rhs=realloc(cset.rhs,sizeof(double)*cset.m);
	      cset.rhs[cset.m-1]=margin;
	      alpha=realloc(alpha,sizeof(double)*cset.m);
	      alpha[cset.m-1]=0;
	      newconstraints++;
	    }
	    else {
	      printf("+"); fflush(stdout); 
	      if(opti[i] != opti_round) {
		activenum--;
		opti[i]=opti_round; 
	      }
	    }

	    free_example(doc,0);
	    free_svector(fy); /* this also free's fybar */
	    free_label(ybar);
	  }

	  /**** get new QP solution ****/
	  if((newconstraints >= sparm->newconstretrain) 
	     || ((newconstraints > 0) && (i == n-1))) {
	    if(struct_verbosity>=1) {
	      printf("*");fflush(stdout);
	    }
	    rt2=get_runtime();
	    free_model(svmModel,0);
	    svmModel=(MODEL *)my_malloc(sizeof(MODEL));
	    /* Always get a new kernel cache. It is not possible to use the
	       same cache for two different training runs */
	    if(kparm->kernel_type != LINEAR)
	      kcache=kernel_cache_init(cset.m,lparm->kernel_cache_size);
	    /* Run the QP solver on cset. */
	    svm_learn_optimization(cset.lhs,cset.rhs,cset.m,sizePsi+n,
				   lparm,kparm,kcache,svmModel,alpha);
	    if(kcache)
	      kernel_cache_cleanup(kcache);
	    /* Always add weight vector, in case part of the kernel is
	       linear. If not, ignore the weight vector since its
	       content is bogus. */
	    add_weight_vector_to_linear_model(svmModel);
	    sm->svm_model=svmModel;
	    sm->w=svmModel->lin_weights; /* short cut to weight vector */
	    rt_opt+=MAX(get_runtime()-rt2,0);
	    
	    newconstraints=0;
	  }	

	  rt_total+=MAX(get_runtime()-rt1,0);
	} /* end of example loop */

	if(struct_verbosity>=1)
	  printf("(NumConst=%d, SV=%ld, Eps=%.4f)\n",cset.m,svmModel->sv_num-1,
		 svmModel->maxdiff);

      } while(activenum > 0);   /* repeat until all examples produced no
				   constraint at least once */

    } while((cset.m - old_numConst) > tolerance) ;

  } while(epsilon > sparm->epsilon);  

  if(struct_verbosity>=1) {
    /**** compute sum of slacks ****/
    slacks=(double *)my_malloc(sizeof(double)*(n+1));
    for(i=0; i<=n; i++) { 
      slacks[i]=0;
    }
    if(sparm->slack_norm == 1) {
      for(j=0;j<cset.m;j++) 
	slacks[cset.lhs[j]->slackid]=MAX(slacks[cset.lhs[j]->slackid],
			   cset.rhs[j]-classify_example(svmModel,cset.lhs[j]));
      }
    else if(sparm->slack_norm == 2) {
      for(j=0;j<cset.m;j++) 
	slacks[cset.lhs[j]->slackid]=MAX(slacks[cset.lhs[j]->slackid],
		cset.rhs[j]
	         -(classify_example(svmModel,cset.lhs[j])
		   -sm->w[sizePsi+cset.lhs[j]->slackid-1]/(sqrt(2*sparm->C))));
    }
    slacksum=0;
    for(i=0; i<=n; i++)  
      slacksum+=slacks[i];
    free(slacks);

    printf("Final epsilon on KKT-Conditions: %.5f\n",
	   MAX(svmModel->maxdiff,epsilon));
    printf("Total number of constraints added: %i\n",(int)cset.m);
    if(sparm->slack_norm == 1) {
      printf("Number of SV: %ld \n",svmModel->sv_num-1);
      printf("Number of non-zero slack variables: %ld (out of %ld)\n",
	     svmModel->at_upper_bound,n);
      printf("Norm of weight vector: |w|=%.5f\n",
	     model_length_s(svmModel,kparm));
    }
    else if(sparm->slack_norm == 2){ 
      printf("Number of SV: %ld (including %ld at upper bound)\n",
	     svmModel->sv_num-1,svmModel->at_upper_bound);
      printf("Norm of weight vector (including L2-loss): |w|=%.5f\n",
	     model_length_s(svmModel,kparm));
    }
    printf("Sum of slack variables: sum(xi_i)=%.5f\n",slacksum);
    printf("Norm of longest difference vector: ||Psi(x,y)-Psi(x,ybar)||=%.5f\n",
	   length_of_longest_document_vector(cset.lhs,cset.m,kparm));
    printf("Runtime in cpu-seconds: %.2f (%.2f%% for SVM optimization)\n",
	   rt_total/100.0, 100.0*rt_opt/rt_total);
  }
  if(struct_verbosity>=4)
    printW(sm->w,sizePsi,n,lparm->svm_c);

  if(svmModel) {
    sm->svm_model=copy_model(svmModel);
    sm->w=sm->svm_model->lin_weights; /* short cut to weight vector */
  }

  print_struct_learning_stats(sample,sm,cset,alpha,sparm);

  if(svmModel)
    free_model(svmModel,0);
  free(alpha); 
  free(opti); 
  free(cset.rhs); 
  for(i=0;i<cset.m;i++) 
    free_example(cset.lhs[i],1);
  free(cset.lhs);
}
Beispiel #17
0
int SVMLightRunner::librarySVMClassifyMain(
    int argc, char **argv, bool use_gmumr, SVMConfiguration &config
) {
    LOG(
        config.log,
        LogLevel::DEBUG_LEVEL,
        __debug_prefix__ + ".librarySVMClassifyMain() Started."
    );
    DOC *doc;   /* test example */
    WORD *words;
    long max_docs,max_words_doc,lld;
    long totdoc=0,queryid,slackid;
    long correct=0,incorrect=0,no_accuracy=0;
    long res_a=0,res_b=0,res_c=0,res_d=0,wnum,pred_format;
    long j;
    double t1,runtime=0;
    double dist,doc_label,costfactor;
    char *line,*comment; 
    FILE *predfl,*docfl;
    MODEL *model; 

    // GMUM.R changes {
    librarySVMClassifyReadInputParameters(
        argc, argv, docfile, modelfile, predictionsfile, &verbosity,
        &pred_format, use_gmumr, config);

    if (!use_gmumr) {
        nol_ll(docfile,&max_docs,&max_words_doc,&lld); /* scan size of input file */
        lld+=2;

        line = (char *)my_malloc(sizeof(char)*lld);
    } else {
        max_docs = config.target.n_rows;
        max_words_doc = config.getDataDim();
        config.result = arma::zeros<arma::vec>(max_docs);
        // Prevent writing to the file
        pred_format = -1;
        // lld used only for file reading
    }
    max_words_doc+=2;
    words = (WORD *)my_malloc(sizeof(WORD)*(max_words_doc+10));
    // GMUM.R changes }

    model=libraryReadModel(modelfile, use_gmumr, config);
    // GMUM.R changes }

    if(model->kernel_parm.kernel_type == 0) { /* linear kernel */
      /* compute weight vector */
      add_weight_vector_to_linear_model(model);
    }
    
    if(verbosity>=2) {
      C_PRINTF("Classifying test examples.."); C_FFLUSH(stdout);
    }

    // GMUM.R changes {
    bool newline;
    if (!use_gmumr) {
        if ((predfl = fopen (predictionsfile, "w")) == NULL)
        { perror (predictionsfile); EXIT (1); }
        if ((docfl = fopen (docfile, "r")) == NULL)
        { perror (docfile); EXIT (1); }

        newline = (!feof(docfl)) && fgets(line,(int)lld,docfl);
    } else {
        newline = false;
        if (totdoc < config.getDataExamplesNumber()) {
            newline = true;
            std::string str = SVMConfigurationToSVMLightLearnInputLine(config, totdoc);
            line = new char[str.size() + 1];
            std::copy(str.begin(), str.end(), line);
            line[str.size()] = '\0';
        }
    }
    while(newline) {
      if (use_gmumr) {
            std::string stringline = "";
      }
      // GMUM.R changes }
      if(line[0] == '#') continue;  /* line contains comments */
      parse_document(line,words,&doc_label,&queryid,&slackid,&costfactor,&wnum,
             max_words_doc,&comment);
      totdoc++;
      if(model->kernel_parm.kernel_type == 0) {   /* linear kernel */
        for(j=0;(words[j]).wnum != 0;j++) {  /* Check if feature numbers   */
      if((words[j]).wnum>model->totwords) /* are not larger than in     */
        (words[j]).wnum=0;               /* model. Remove feature if   */
        }                                        /* necessary.                 */
        doc = create_example(-1,0,0,0.0,create_svector(words,comment,1.0));
        t1=get_runtime();
        dist=classify_example_linear(model,doc);
        runtime+=(get_runtime()-t1);
        free_example(doc,1);
      }
      else {                             /* non-linear kernel */
        doc = create_example(-1,0,0,0.0,create_svector(words,comment,1.0));
        t1=get_runtime();
        dist=classify_example(model,doc);
        runtime+=(get_runtime()-t1);
        free_example(doc,1);
      }
      if(dist>0) {
        if(pred_format==0) { /* old weired output format */
      C_FPRINTF(predfl,"%.8g:+1 %.8g:-1\n",dist,-dist);
        }
        if(doc_label>0) correct++; else incorrect++;
        if(doc_label>0) res_a++; else res_b++;
      }
      else {
        if(pred_format==0) { /* old weired output format */
      C_FPRINTF(predfl,"%.8g:-1 %.8g:+1\n",-dist,dist);
        }
        if(doc_label<0) correct++; else incorrect++;
        if(doc_label>0) res_c++; else res_d++;
      }
      if(pred_format==1) { /* output the value of decision function */
        C_FPRINTF(predfl,"%.8g\n",dist);
      }
      if((int)(0.01+(doc_label*doc_label)) != 1)
        { no_accuracy=1; } /* test data is not binary labeled */
      if(verbosity>=2) {
        if(totdoc % 100 == 0) {
      C_PRINTF("%ld..",totdoc); C_FFLUSH(stdout);
        }
      }
      // GMUM.R changes {
      if (!use_gmumr) {
          newline = (!feof(docfl)) && fgets(line,(int)lld,docfl);
      } else {
          newline = false;
          // Store prediction result in config
          config.result[totdoc-1] = dist;
          // Read next line
          if (totdoc < config.getDataExamplesNumber()) {
              newline = true;
              std::string str = SVMConfigurationToSVMLightLearnInputLine(config, totdoc);
              line = new char[str.size() + 1];
              std::copy(str.begin(), str.end(), line);
              line[str.size()] = '\0';
          }
      }
    }
    if (!use_gmumr) {
        fclose(predfl);
        fclose(docfl);
        free(line);
    }
    // GMUM.R changes }
    free(words);
    free_model(model,1);

    if(verbosity>=2) {
      C_PRINTF("done\n");

  /*   Note by Gary Boone                     Date: 29 April 2000        */
  /*      o Timing is inaccurate. The timer has 0.01 second resolution.  */
  /*        Because classification of a single vector takes less than    */
  /*        0.01 secs, the timer was underflowing.                       */
      C_PRINTF("Runtime (without IO) in cpu-seconds: %.2f\n",
         (float)(runtime/100.0));

    }
    if((!no_accuracy) && (verbosity>=1)) {
      C_PRINTF("Accuracy on test set: %.2f%% (%ld correct, %ld incorrect, %ld total)\n",(float)(correct)*100.0/totdoc,correct,incorrect,totdoc);
      C_PRINTF("Precision/recall on test set: %.2f%%/%.2f%%\n",(float)(res_a)*100.0/(res_a+res_b),(float)(res_a)*100.0/(res_a+res_c));
    }

    return(0);
}