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);
}
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);
}
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);
}
int _svm_learn (int argc, char* argv[])
{
char docfile[200]; /* file with training examples */
char modelfile[200]; /* file for resulting classifier */
char restartfile[200]; /* file with initial alphas */
DOC **docs; /* training examples */
long totwords,totdoc,i;
double *target;
double *alpha_in=NULL;
KERNEL_CACHE *kernel_cache;
LEARN_PARM learn_parm;
KERNEL_PARM kernel_parm;
MODEL *model=(MODEL *)my_malloc(sizeof(MODEL));
HIDEO_ENV *hideo_env=create_env();
model->td_pred=NULL;
model->n_td_pred=0;
_read_input_parameters(argc,argv,docfile,modelfile,restartfile,&verbosity,
&learn_parm,&kernel_parm);
read_documents(docfile,&docs,&target,&totwords,&totdoc);
if(restartfile[0]) alpha_in=read_alphas(restartfile,totdoc);
if(kernel_parm.kernel_type == LINEAR) { /* don't need the cache */
kernel_cache=NULL;
}
else {
/* Always get a new kernel cache. It is not possible to use the
same cache for two different training runs */
kernel_cache=kernel_cache_init(totdoc,learn_parm.kernel_cache_size);
}
if(learn_parm.type == CLASSIFICATION) {
svm_learn_classification(docs,target,totdoc,totwords,&learn_parm,
&kernel_parm,kernel_cache,model,alpha_in,hideo_env);
}
else if(learn_parm.type == REGRESSION) {
svm_learn_regression(docs,target,totdoc,totwords,&learn_parm,
&kernel_parm,&kernel_cache,model,hideo_env);
}
else if(learn_parm.type == RANKING) {
svm_learn_ranking(docs,target,totdoc,totwords,&learn_parm,
&kernel_parm,&kernel_cache,model,hideo_env);
}
else if(learn_parm.type == OPTIMIZATION) {
svm_learn_optimization(docs,target,totdoc,totwords,&learn_parm,
&kernel_parm,kernel_cache,model,alpha_in,hideo_env);
}
if(kernel_cache) {
/* Free the memory used for the cache. */
kernel_cache_cleanup(kernel_cache);
}
/* Warning: The model contains references to the original data 'docs'.
If you want to free the original data, and only keep the model, you
have to make a deep copy of 'model'. */
/* deep_copy_of_model=copy_model(model); */
write_model(modelfile,model);
free(alpha_in);
free_model(model,0);
for(i=0;i<totdoc;i++)
free_example(docs[i],1);
free(docs);
free(target);
free_env(hideo_env);
return(0);
}
int SVMLightRunner::librarySVMLearnMain(
int argc, char **argv, bool use_gmumr, SVMConfiguration &config
) {
LOG(
config.log,
LogLevel::DEBUG_LEVEL,
__debug_prefix__ + ".librarySVMLearnMain() Started."
);
DOC **docs; /* training examples */
long totwords,totdoc,i;
double *target;
double *alpha_in=NULL;
KERNEL_CACHE *kernel_cache;
LEARN_PARM learn_parm;
KERNEL_PARM kernel_parm;
MODEL *model=(MODEL *)my_malloc(sizeof(MODEL));
// GMUM.R changes {
librarySVMLearnReadInputParameters(
argc, argv, docfile, modelfile, restartfile, &verbosity, &learn_parm,
&kernel_parm, use_gmumr, config
);
kernel_parm.kernel_type = static_cast<long int>(config.kernel_type);
libraryReadDocuments(
docfile, &docs, &target, &totwords, &totdoc, use_gmumr, config
);
// GMUM.R changes }
if(restartfile[0]) alpha_in=read_alphas(restartfile,totdoc);
if(kernel_parm.kernel_type == LINEAR) { /* don't need the cache */
kernel_cache=NULL;
}
else {
/* Always get a new kernel cache. It is not possible to use the
* same cache for two different training runs */
kernel_cache=kernel_cache_init(totdoc,learn_parm.kernel_cache_size);
}
//gmum.r
init_global_params_QP();
if(learn_parm.type == CLASSIFICATION) {
svm_learn_classification(docs,target,totdoc,totwords,&learn_parm,
&kernel_parm,kernel_cache,model,alpha_in);
}
else if(learn_parm.type == REGRESSION) {
svm_learn_regression(docs,target,totdoc,totwords,&learn_parm,
&kernel_parm,&kernel_cache,model);
}
else if(learn_parm.type == RANKING) {
svm_learn_ranking(docs,target,totdoc,totwords,&learn_parm,
&kernel_parm,&kernel_cache,model);
}
else if(learn_parm.type == OPTIMIZATION) {
svm_learn_optimization(docs,target,totdoc,totwords,&learn_parm,
&kernel_parm,kernel_cache,model,alpha_in);
}
//gmum.r
config.iter = learn_parm.iterations;
if(kernel_cache) {
/* Free the memory used for the cache. */
kernel_cache_cleanup(kernel_cache);
}
/* Warning: The model contains references to the original data 'docs'.
If you want to free the original data, and only keep the model, you
have to make a deep copy of 'model'. */
/* deep_copy_of_model=copy_model(model); */
// GMUM.R changes {
if (!use_gmumr) {
write_model(modelfile,model);
} else {
SVMLightModelToSVMConfiguration(model, config);
}
// GMUM.R changes }
free(alpha_in);
free_model(model,0);
for(i=0;i<totdoc;i++)
free_example(docs[i],1);
free(docs);
free(target);
LOG(
config.log,
LogLevel::DEBUG_LEVEL,
__debug_prefix__ + ".librarySVMLearnMain() Done."
);
return(0);
}
double cutting_plane_algorithm(double *w, long m, int MAX_ITER, double C, double epsilon, SVECTOR **fycache,
EXAMPLE *ex, STRUCTMODEL *sm, STRUCT_LEARN_PARM *sparm) {
long i, j;
double xi;
double *alpha;
double **G; /* Gram matrix */
DOC **dXc; /* constraint matrix */
double *delta; /* rhs of constraints */
SVECTOR *new_constraint;
double dual_obj, alphasum;
int iter, size_active;
double value;
int r;
int *idle; /* for cleaning up */
double margin;
double primal_obj;
double *proximal_rhs;
double *gammaG0 = NULL;
double min_rho = 0.001;
double max_rho;
double serious_counter = 0;
double rho = 1.0; /* temporarily set it to 1 first */
double expected_descent, primal_obj_b = -1, reg_master_obj;
int null_step = 1;
double *w_b;
double kappa = 0.1;
double temp_var;
double proximal_term, primal_lower_bound;
double v_k;
double obj_difference;
double *cut_error; // cut_error[i] = alpha_{k,i} at current center x_k
double sigma_k;
double m2 = 0.2;
double m3 = 0.9;
double gTd;
double last_sigma_k = 0;
double initial_primal_obj;
int suff_decrease_cond = 0;
double decrease_proportion = 0.2; // start from 0.2 first
double z_k_norm;
double last_z_k_norm = 0;
/* set parameters for hideo solver */
LEARN_PARM lparm;
KERNEL_PARM kparm;
MODEL *svmModel = NULL;
lparm.biased_hyperplane = 0;
lparm.epsilon_crit = MIN(epsilon, 0.001);
lparm.svm_c = C;
lparm.sharedslack = 1;
kparm.kernel_type = LINEAR;
lparm.remove_inconsistent = 0;
lparm.skip_final_opt_check = 0;
lparm.svm_maxqpsize = 10;
lparm.svm_newvarsinqp = 0;
lparm.svm_iter_to_shrink = -9999;
lparm.maxiter = 100000;
lparm.kernel_cache_size = 40;
lparm.eps = epsilon;
lparm.transduction_posratio = -1.0;
lparm.svm_costratio = 1.0;
lparm.svm_costratio_unlab = 1.0;
lparm.svm_unlabbound = 1E-5;
lparm.epsilon_a = 1E-10; /* changed from 1e-15 */
lparm.compute_loo = 0;
lparm.rho = 1.0;
lparm.xa_depth = 0;
strcpy(lparm.alphafile, "");
kparm.poly_degree = 3;
kparm.rbf_gamma = 1.0;
kparm.coef_lin = 1;
kparm.coef_const = 1;
strcpy(kparm.custom, "empty");
w_b = create_nvector(sm->sizePsi);
clear_nvector(w_b, sm->sizePsi);
/* warm start */
for (i = 1; i < sm->sizePsi + 1; i++) {
w_b[i] = w[i];
}
iter = 0;
size_active = 0;
xi = 0.0;
alpha = NULL;
G = NULL;
dXc = NULL;
delta = NULL;
idle = NULL;
proximal_rhs = NULL;
cut_error = NULL;
new_constraint = find_cutting_plane(ex, fycache, &margin, m, sm, sparm);
value = margin - sprod_ns(w, new_constraint);
primal_obj_b = 0.5 * sprod_nn(w_b, w_b, sm->sizePsi) + C * value;
primal_obj = 0.5 * sprod_nn(w, w, sm->sizePsi) + C * value;
primal_lower_bound = 0;
expected_descent = -primal_obj_b;
initial_primal_obj = primal_obj_b;
//max_rho = C;
max_rho = 100 * C; // tree-edge loss not within 0-1
printf("Running CCCP inner loop solver: ");
fflush(stdout);
while ((!suff_decrease_cond) && (expected_descent < -epsilon) && (iter < MAX_ITER)) {
iter += 1;
size_active += 1;
#if (DEBUG_LEVEL > 0)
printf("ITER %d\n", iter);
#endif
printf(".");
fflush(stdout);
/* add constraint */
dXc = (DOC **) realloc(dXc, sizeof(DOC *) * size_active);
assert(dXc != NULL);
dXc[size_active - 1] = (DOC *) malloc(sizeof(DOC));
dXc[size_active - 1]->fvec = new_constraint;
dXc[size_active - 1]->slackid = 1; // only one common slackid (one-slack)
dXc[size_active - 1]->costfactor = 1.0;
delta = (double *) realloc(delta, sizeof(double) * size_active);
assert(delta != NULL);
delta[size_active - 1] = margin;
alpha = (double *) realloc(alpha, sizeof(double) * size_active);
assert(alpha != NULL);
alpha[size_active - 1] = 0.0;
idle = (int *) realloc(idle, sizeof(int) * size_active);
assert(idle != NULL);
idle[size_active - 1] = 0;
/* proximal point */
proximal_rhs = (double *) realloc(proximal_rhs, sizeof(double) * size_active);
assert(proximal_rhs != NULL);
cut_error = (double *) realloc(cut_error, sizeof(double) * size_active);
assert(cut_error != NULL);
// note g_i = - new_constraint
cut_error[size_active - 1] = C * (sprod_ns(w_b, new_constraint) - sprod_ns(w, new_constraint));
cut_error[size_active - 1] += (primal_obj_b - 0.5 * sprod_nn(w_b, w_b, sm->sizePsi));
cut_error[size_active - 1] -= (primal_obj - 0.5 * sprod_nn(w, w, sm->sizePsi));
gammaG0 = (double *) realloc(gammaG0, sizeof(double) * size_active);
assert(gammaG0 != NULL);
/* update Gram matrix */
G = (double **) realloc(G, sizeof(double *) * size_active);
assert(G != NULL);
G[size_active - 1] = NULL;
for (j = 0; j < size_active; j++) {
G[j] = (double *) realloc(G[j], sizeof(double) * size_active);
assert(G[j] != NULL);
}
for (j = 0; j < size_active - 1; j++) {
G[size_active - 1][j] = sprod_ss(dXc[size_active - 1]->fvec, dXc[j]->fvec);
G[j][size_active - 1] = G[size_active - 1][j];
}
G[size_active - 1][size_active - 1] = sprod_ss(dXc[size_active - 1]->fvec, dXc[size_active - 1]->fvec);
/* update gammaG0 */
if (null_step == 1) {
gammaG0[size_active - 1] = sprod_ns(w_b, dXc[size_active - 1]->fvec);
} else {
for (i = 0; i < size_active; i++) {
gammaG0[i] = sprod_ns(w_b, dXc[i]->fvec);
}
}
/* update proximal_rhs */
for (i = 0; i < size_active; i++) {
proximal_rhs[i] = (1 + rho) * delta[i] - rho * gammaG0[i];
}
/* solve QP to update alpha */
//dual_obj = 0;
//r = mosek_qp_optimize(G, proximal_rhs, alpha, (long) size_active, C, &dual_obj,rho);
if (size_active > 1) {
if (svmModel != NULL) free_model(svmModel, 0);
svmModel = (MODEL *) my_malloc(sizeof(MODEL));
svm_learn_optimization(dXc, proximal_rhs, size_active, sm->sizePsi, &lparm, &kparm, NULL, svmModel, alpha);
} else {
assert(size_active == 1);
alpha[0] = C;
}
/* DEBUG */
//printf("r: %d\n", r); fflush(stdout);
/* END DEBUG */
clear_nvector(w, sm->sizePsi);
for (j = 0; j < size_active; j++) {
if (alpha[j] > C * ALPHA_THRESHOLD) {
add_vector_ns(w, dXc[j]->fvec, alpha[j] / (1 + rho));
}
}
/* compute dual obj */
dual_obj = +0.5 * (1 + rho) * sprod_nn(w, w, sm->sizePsi);
for (j = 0; j < size_active; j++) {
dual_obj -= proximal_rhs[j] / (1 + rho) * alpha[j];
}
z_k_norm = sqrt(sprod_nn(w, w, sm->sizePsi));
add_vector_nn(w, w_b, sm->sizePsi, rho / (1 + rho));
/* detect if step size too small */
sigma_k = 0;
alphasum = 0;
for (j = 0; j < size_active; j++) {
sigma_k += alpha[j] * cut_error[j];
alphasum += alpha[j];
}
sigma_k /= C;
gTd = -C * (sprod_ns(w, new_constraint) - sprod_ns(w_b, new_constraint));
#if (DEBUG_LEVEL > 0)
for (j=0;j<size_active;j++) {
printf("alpha[%d]: %.8g, cut_error[%d]: %.8g\n", j, alpha[j], j, cut_error[j]);
}
printf("sigma_k: %.8g\n", sigma_k);
printf("alphasum: %.8g\n", alphasum);
printf("g^T d: %.8g\n", gTd);
fflush(stdout);
#endif
/* update cleanup information */
for (j = 0; j < size_active; j++) {
if (alpha[j] < ALPHA_THRESHOLD * C) {
idle[j]++;
} else {
idle[j] = 0;
}
}
new_constraint = find_cutting_plane(ex, fycache, &margin, m, sm, sparm);
value = margin - sprod_ns(w, new_constraint);
/* print primal objective */
primal_obj = 0.5 * sprod_nn(w, w, sm->sizePsi) + C * value;
#if (DEBUG_LEVEL > 0)
printf("ITER PRIMAL_OBJ %.4f\n", primal_obj); fflush(stdout);
#endif
temp_var = sprod_nn(w_b, w_b, sm->sizePsi);
proximal_term = 0.0;
for (i = 1; i < sm->sizePsi + 1; i++) {
proximal_term += (w[i] - w_b[i]) * (w[i] - w_b[i]);
}
reg_master_obj = -dual_obj + 0.5 * rho * temp_var / (1 + rho);
expected_descent = reg_master_obj - primal_obj_b;
v_k = (reg_master_obj - proximal_term * rho / 2) - primal_obj_b;
primal_lower_bound = MAX(primal_lower_bound, reg_master_obj - 0.5 * rho * (1 + rho) * proximal_term);
#if (DEBUG_LEVEL > 0)
printf("ITER REG_MASTER_OBJ: %.4f\n", reg_master_obj);
printf("ITER EXPECTED_DESCENT: %.4f\n", expected_descent);
printf("ITER PRIMLA_OBJ_B: %.4f\n", primal_obj_b);
printf("ITER RHO: %.4f\n", rho);
printf("ITER ||w-w_b||^2: %.4f\n", proximal_term);
printf("ITER PRIMAL_LOWER_BOUND: %.4f\n", primal_lower_bound);
printf("ITER V_K: %.4f\n", v_k);
#endif
obj_difference = primal_obj - primal_obj_b;
if (primal_obj < primal_obj_b + kappa * expected_descent) {
/* extra condition to be met */
if ((gTd > m2 * v_k) || (rho < min_rho + 1E-8)) {
#if (DEBUG_LEVEL > 0)
printf("SERIOUS STEP\n");
#endif
/* update cut_error */
for (i = 0; i < size_active; i++) {
cut_error[i] -= (primal_obj_b - 0.5 * sprod_nn(w_b, w_b, sm->sizePsi));
cut_error[i] -= C * sprod_ns(w_b, dXc[i]->fvec);
cut_error[i] += (primal_obj - 0.5 * sprod_nn(w, w, sm->sizePsi));
cut_error[i] += C * sprod_ns(w, dXc[i]->fvec);
}
primal_obj_b = primal_obj;
for (i = 1; i < sm->sizePsi + 1; i++) {
w_b[i] = w[i];
}
null_step = 0;
serious_counter++;
} else {
/* increase step size */
#if (DEBUG_LEVEL > 0)
printf("NULL STEP: SS(ii) FAILS.\n");
#endif
serious_counter--;
rho = MAX(rho / 10, min_rho);
}
} else { /* no sufficient decrease */
serious_counter--;
if ((cut_error[size_active - 1] > m3 * last_sigma_k) &&
(fabs(obj_difference) > last_z_k_norm + last_sigma_k)) {
#if (DEBUG_LEVEL > 0)
printf("NULL STEP: NS(ii) FAILS.\n");
#endif
rho = MIN(10 * rho, max_rho);
}
#if (DEBUG_LEVEL > 0)
else printf("NULL STEP\n");
#endif
}
/* update last_sigma_k */
last_sigma_k = sigma_k;
last_z_k_norm = z_k_norm;
/* break away from while loop if more than certain proportioal decrease in primal objective */
if (primal_obj_b / initial_primal_obj < 1 - decrease_proportion) {
suff_decrease_cond = 1;
}
/* clean up */
if (iter % CLEANUP_CHECK == 0) {
//size_active = resize_cleanup(size_active, idle, alpha, delta, gammaG0, proximal_rhs, G, dXc, cut_error);
size_active = resize_cleanup(size_active, &idle, &alpha, &delta, &gammaG0, &proximal_rhs, &G, &dXc,
&cut_error);
}
} // end cutting plane while loop
printf(" Inner loop optimization finished.\n");
fflush(stdout);
/* free memory */
for (j = 0; j < size_active; j++) {
free(G[j]);
free_example(dXc[j], 0);
}
free(G);
free(dXc);
free(alpha);
free(delta);
free_svector(new_constraint);
free(idle);
free(gammaG0);
free(proximal_rhs);
free(cut_error);
/* copy and free */
for (i = 1; i < sm->sizePsi + 1; i++) {
w[i] = w_b[i];
}
free(w_b);
return (primal_obj_b);
}
/* call as model = mexsvmlearn(data,labels,options) */
void mexFunction(int nlhs, mxArray *plhs[],
int nrhs, const mxArray *prhs[])
{
char **argv;
int argc;
DOC **docs; /* training examples */
long totwords,totdoc,i;
double *target;
double *alpha_in=NULL;
KERNEL_CACHE *kernel_cache;
LEARN_PARM learn_parm;
KERNEL_PARM kernel_parm;
MODEL model;
/* check for valid calling format */
if ((nrhs != 3) || (nlhs != 1))
mexErrMsgTxt(ERR001);
if (mxGetM(prhs[0]) != mxGetM(prhs[1]))
mexErrMsgTxt(ERR002);
if (mxGetN(prhs[1]) != 1)
mexErrMsgTxt(ERR003);
/* reset static variables -- as a .DLL, static things are sticky */
global_init( );
/* convert the parameters (given in prhs[2]) into an argv/argc combination */
argv = make_argv((mxArray *)prhs[2],&argc); /* send the options */
/* this was originally supposed to be argc, argv, re-written for MATLAB ...
its cheesy - but it workss :: convert the options array into an argc,
argv pair and let svm_lite handle it from there. */
read_input_parameters(argc,argv,docfile,modelfile,restartfile,&verbosity,
&learn_parm,&kernel_parm);
extract_user_opts((mxArray *)prhs[2], &kernel_parm);
totdoc = mxGetM(prhs[0]);
totwords = mxGetN(prhs[0]);
/* prhs[0] = samples (mxn) array
prhs[1] = labels (mx1) array */
mexToDOC((mxArray *)prhs[0], (mxArray *)prhs[1], &docs, &target, NULL, NULL);
/* TODO modify to accept this array
if(restartfile[0]) alpha_in=read_alphas(restartfile,totdoc); */
if(kernel_parm.kernel_type == LINEAR) { /* don't need the cache */
kernel_cache=NULL;
}
else {
/* Always get a new kernel cache. It is not possible to use the
same cache for two different training runs */
kernel_cache=kernel_cache_init(totdoc,learn_parm.kernel_cache_size);
}
if(learn_parm.type == CLASSIFICATION) {
svm_learn_classification(docs,target,totdoc,totwords,&learn_parm,
&kernel_parm,kernel_cache,&model,alpha_in);
}
else if(learn_parm.type == REGRESSION) {
svm_learn_regression(docs,target,totdoc,totwords,&learn_parm,
&kernel_parm,&kernel_cache,&model);
}
else if(learn_parm.type == RANKING) {
svm_learn_ranking(docs,target,totdoc,totwords,&learn_parm,
&kernel_parm,&kernel_cache,&model);
}
else if(learn_parm.type == OPTIMIZATION) {
svm_learn_optimization(docs,target,totdoc,totwords,&learn_parm,
&kernel_parm,kernel_cache,&model,alpha_in);
}
else {
mexErrMsgTxt(ERR004);
}
if(kernel_cache) {
/* Free the memory used for the cache. */
kernel_cache_cleanup(kernel_cache);
}
/* **********************************
* After the training/learning portion has finished,
* copy the model back to the output arrays for MATLAB
* ********************************** */
store_model(&model, plhs);
free_kernel();
global_destroy( );
}
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);
}