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);
}
double optimizeMultiVariatePerfMeasure(SAMPLE sample, int datasetStartIdx, int chunkSz, STRUCTMODEL *sm, STRUCT_LEARN_PARM *sparm,
double C, double Cdash, double epsilon, int MAX_ITER, LEARN_PARM *learn_parm, char *trainfile,
double ***w_iters, int eid, int chunkid, int numChunks, double *zeroes){
int i;
time_t time_start, time_end;
double decrement;
double primal_obj, last_primal_obj;
double cooling_eps;
double stop_crit;
LATENT_VAR *imputed_h = NULL;
int dataset_sz = sample.n;
SVECTOR **fycache, *diff, *fy;
EXAMPLE *ex = sample.examples;
/* some training information */
printf("C: %.8g\n", C);
printf("Cdash: %.8g\n", Cdash);
printf("epsilon: %.8g\n", epsilon);
printf("sample.n: %ld\n", dataset_sz);
printf("sm->sizePsi: %ld\n", sm->sizePsi); fflush(stdout);
/* prepare feature vector cache for correct labels with imputed latent variables */
fycache = (SVECTOR**)malloc(dataset_sz*sizeof(SVECTOR*));
for (i=0;i<dataset_sz;i++) {
fy = psi(ex[i].x, ex[i].y, ex[i].h, sm, sparm);
diff = add_list_ss(fy);
free_svector(fy);
fy = diff;
fycache[i] = fy;
}
/* time taken stats */
time(&time_start);
/* outer loop: latent variable imputation */
int outer_iter = 0;
last_primal_obj = 0;
decrement = 0;
cooling_eps = 0.5*MAX(C,Cdash)*epsilon;
while ((outer_iter<2)||((!stop_crit)&&(outer_iter<MAX_OUTER_ITER))) {
printf("OUTER ITER %d\n", outer_iter); fflush(stdout);
/* cutting plane algorithm */
time_t cp_start, cp_end;
time(&cp_start);
/// NOTE : Change of variables (Create 'u' by subtracting w_prev from w)
create_u_variables(w_iters, eid, chunkid, numChunks, sm, zeroes);
if(chunkid == 0 && eid == 0){ // First Chunk of First Epoch
primal_obj = cutting_plane_algorithm(w_iters[eid][chunkid], dataset_sz, MAX_ITER, C, cooling_eps,
fycache, ex, sm, sparm, learn_parm->tmpdir, trainfile, learn_parm->frac_sim,
learn_parm->Fweight, learn_parm->dataset_stats_file, learn_parm->rho_admm,
learn_parm->isExhaustive, learn_parm->isLPrelaxation, Cdash, datasetStartIdx, chunkSz,
eid, chunkid, zeroes, numChunks); // pass the zeroes vector
}
else if(chunkid == 0){ // First chunk of the new Epoch
primal_obj = cutting_plane_algorithm(w_iters[eid][chunkid], dataset_sz, MAX_ITER, C, cooling_eps,
fycache, ex, sm, sparm, learn_parm->tmpdir, trainfile, learn_parm->frac_sim,
learn_parm->Fweight, learn_parm->dataset_stats_file, learn_parm->rho_admm,
learn_parm->isExhaustive, learn_parm->isLPrelaxation, Cdash, datasetStartIdx, chunkSz,
eid, chunkid, w_iters[eid-1][numChunks-1], numChunks); // Last chunk of previous epoch
}
else {
primal_obj = cutting_plane_algorithm(w_iters[eid][chunkid], dataset_sz, MAX_ITER, C, cooling_eps,
fycache, ex, sm, sparm, learn_parm->tmpdir, trainfile, learn_parm->frac_sim,
learn_parm->Fweight, learn_parm->dataset_stats_file, learn_parm->rho_admm,
learn_parm->isExhaustive, learn_parm->isLPrelaxation, Cdash, datasetStartIdx, chunkSz,
eid, chunkid, w_iters[eid][chunkid-1], numChunks); // previous chunk id of current epoch
}
time(&cp_end);
#if(DEBUG_LEVEL==1)
char msg[20];
sprintf(msg,"OUTER ITER %d", outer_iter);
print_time(cp_start, cp_end, msg);
#endif
/* compute decrement in objective in this outer iteration */
decrement = last_primal_obj - primal_obj;
last_primal_obj = primal_obj;
printf("primal objective: %.4f\n", primal_obj);
printf("decrement: %.4f\n", decrement); fflush(stdout);
stop_crit = (decrement<MAX(C, Cdash)*epsilon)&&(cooling_eps<0.5*MAX(C, Cdash)*epsilon+1E-8);
cooling_eps = -decrement*0.01;
cooling_eps = MAX(cooling_eps, 0.5*MAX(C,Cdash)*epsilon);
printf("cooling_eps: %.8g\n", cooling_eps);
/* impute latent variable using updated weight vector */
for(i = 0; i < dataset_sz; i ++)
free_latent_var(ex[i].h);
if(imputed_h != NULL)
free(imputed_h);
imputed_h = (LATENT_VAR*)malloc(sizeof(LATENT_VAR) * dataset_sz);
infer_latent_variables_all(imputed_h, sm, sparm, dataset_sz, learn_parm->tmpdir, trainfile, datasetStartIdx, chunkSz, eid, chunkid);
for (i=0;i<dataset_sz;i++) {
// free_latent_var(ex[i].h);
// ex[i].h = infer_latent_variables(ex[i].x, ex[i].y, &sm, &sparm); // ILP for Pr (Z | Y_i, X_i) in our case
ex[i].h = imputed_h[i];
}
/* re-compute feature vector cache */
for (i=0;i<dataset_sz;i++) {
free_svector(fycache[i]);
fy = psi(ex[i].x, ex[i].y, ex[i].h, &sm, &sparm);
diff = add_list_ss(fy);
free_svector(fy);
fy = diff;
fycache[i] = fy;
}
printf("(OnlineSVM) .. finished outer_iter %d\n",outer_iter);
outer_iter++;
/// NOTE: Restore the 'w' by adding the current 'u' to w_prev
restore_w_variables(w_iters, eid, chunkid, numChunks, sm, zeroes);
} // end outer loop
time(&time_end);
#if (DEBUG_LEVEL==1)
print_time(time_start, time_end, "Total time");
#endif
for(i=0;i<dataset_sz;i++) {
free_svector(fycache[i]);
}
free(fycache);
return primal_obj;
}
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);
}
int main(int argc, char* argv[]) {
double *w; /* weight vector */
long m, i;
double C, epsilon;
LEARN_PARM learn_parm;
KERNEL_PARM kernel_parm;
char trainfile[1024];
char modelfile[1024];
int MAX_ITER;
/* new struct variables */
SVECTOR **fycache, *diff, *fy;
EXAMPLE *ex;
SAMPLE alldata;
SAMPLE sample;
SAMPLE val;
STRUCT_LEARN_PARM sparm;
STRUCTMODEL sm;
double primal_obj;
double stop_crit;
char itermodelfile[2000];
/* self-paced learning variables */
double init_spl_weight;
double spl_weight;
double spl_factor;
int *valid_examples;
/* read input parameters */
my_read_input_parameters(argc, argv, trainfile, modelfile, &learn_parm, &kernel_parm, &sparm,
&init_spl_weight, &spl_factor);
epsilon = learn_parm.eps;
C = learn_parm.svm_c;
MAX_ITER = learn_parm.maxiter;
/* read in examples */
alldata = read_struct_examples(trainfile,&sparm);
int ntrain = (int) round(1.0*alldata.n); /* no validation set */
if(ntrain < alldata.n)
{
long *perm = randperm(alldata.n);
sample = generate_train_set(alldata, perm, ntrain);
val = generate_validation_set(alldata, perm, ntrain);
free(perm);
}
else
{
sample = alldata;
}
ex = sample.examples;
m = sample.n;
/* initialization */
init_struct_model(alldata,&sm,&sparm,&learn_parm,&kernel_parm);
w = create_nvector(sm.sizePsi);
clear_nvector(w, sm.sizePsi);
sm.w = w; /* establish link to w, as long as w does not change pointer */
/* some training information */
printf("C: %.8g\n", C);
printf("spl weight: %.8g\n",init_spl_weight);
printf("epsilon: %.8g\n", epsilon);
printf("sample.n: %d\n", sample.n);
printf("sm.sizePsi: %ld\n", sm.sizePsi); fflush(stdout);
/* prepare feature vector cache for correct labels with imputed latent variables */
fycache = (SVECTOR**)malloc(m*sizeof(SVECTOR*));
for (i=0;i<m;i++) {
fy = psi(ex[i].x, ex[i].y, &sm, &sparm);
diff = add_list_ss(fy);
free_svector(fy);
fy = diff;
fycache[i] = fy;
}
/* learn initial weight vector using all training examples */
valid_examples = (int *) malloc(m*sizeof(int));
/* errors for validation set */
double cur_loss, best_loss = DBL_MAX;
int loss_iter;
/* initializations */
spl_weight = init_spl_weight;
/* solve biconvex self-paced learning problem */
primal_obj = alternate_convex_search(w, m, MAX_ITER, C, epsilon, fycache, ex, &sm, &sparm, valid_examples, spl_weight);
printf("primal objective: %.4f\n", primal_obj);
fflush(stdout);
//alternate_convex_search(w, m, MAX_ITER, C, epsilon, fycache, ex, &sm, &sparm, valid_examples, spl_weight);
int nValid = 0;
for (i=0;i<m;i++) {
if(valid_examples[i]) {
nValid++;
}
}
if(ntrain < alldata.n) {
cur_loss = compute_current_loss(val,&sm,&sparm);
printf("CURRENT LOSS: %f\n",cur_loss);
}
/* write structural model */
write_struct_model(modelfile, &sm, &sparm);
// skip testing for the moment
/* free memory */
free_struct_sample(alldata);
if(ntrain < alldata.n)
{
free(sample.examples);
free(val.examples);
}
free_struct_model(sm, &sparm);
for(i=0;i<m;i++) {
free_svector(fycache[i]);
}
free(fycache);
free(valid_examples);
return(0);
}
int main(int argc, char* argv[]) {
double *w; /* weight vector */
int outer_iter;
long m, i;
double C, epsilon;
LEARN_PARM learn_parm;
KERNEL_PARM kernel_parm;
char trainfile[1024];
char modelfile[1024];
int MAX_ITER;
/* new struct variables */
SVECTOR **fycache, *diff, *fy;
EXAMPLE *ex;
SAMPLE sample;
STRUCT_LEARN_PARM sparm;
STRUCTMODEL sm;
//double decrement;
double primal_obj;//, last_primal_obj;
//double cooling_eps;
//double stop_crit;
DebugConfiguration::VerbosityLevel = VerbosityLevel::None;
/* read input parameters */
my_read_input_parameters(argc, argv, trainfile, modelfile, &learn_parm, &kernel_parm, &sparm);
epsilon = learn_parm.eps;
C = learn_parm.svm_c;
MAX_ITER = learn_parm.maxiter;
/* read in examples */
sample = read_struct_examples(trainfile,&sparm);
ex = sample.examples;
m = sample.n;
/* initialization */
init_struct_model(sample,&sm,&sparm,&learn_parm,&kernel_parm);
w = sm.w;
//w = create_nvector(sm.sizePsi);
//clear_nvector(w, sm.sizePsi);
//sm.w = w; /* establish link to w, as long as w does not change pointer */
/* some training information */
printf("C: %.8g\n", C);
printf("epsilon: %.8g\n", epsilon);
printf("sample.n: %ld\n", sample.n);
printf("sm.sizePsi: %ld\n", sm.sizePsi);
fflush(stdout);
/* impute latent variable for first iteration */
init_latent_variables(&sample,&learn_parm,&sm,&sparm);
/* prepare feature vector cache for correct labels with imputed latent variables */
fycache = (SVECTOR**)malloc(m*sizeof(SVECTOR*));
for (i=0; i<m; i++) {
fy = psi(ex[i].x, ex[i].y, ex[i].h, &sm, &sparm);
/* DEBUG */
printf("true_psi[%d]=", i);
for (int j = 0; j < sm.sizePsi; ++j)
printf("%.4lf ", fy->words[j].weight);
printf("\n");
diff = add_list_ss(fy);
free_svector(fy);
fy = diff;
fycache[i] = fy;
}
/* outer loop: latent variable imputation */
outer_iter = 1;
//last_primal_obj = 0;
//decrement = 0;
//cooling_eps = 0.5*C*epsilon;
//while ((outer_iter<=MIN_OUTER_ITER)||((!stop_crit)&&(outer_iter<MAX_OUTER_ITER))) {
while (outer_iter<MAX_OUTER_ITER) {
LearningTracker::NextOuterIteration();
printf("OUTER ITER %d\n", outer_iter);
/* cutting plane algorithm */
primal_obj = cutting_plane_algorithm(w, m, MAX_ITER, C, /*cooling_eps, */fycache, ex, &sm, &sparm);
/* compute decrement in objective in this outer iteration */
/*
decrement = last_primal_obj - primal_obj;
last_primal_obj = primal_obj;
printf("primal objective: %.4f\n", primal_obj);
printf("decrement: %.4f\n", decrement); fflush(stdout);
stop_crit = (decrement<C*epsilon)&&(cooling_eps<0.5*C*epsilon+1E-8);
cooling_eps = -decrement*0.01;
cooling_eps = MAX(cooling_eps, 0.5*C*epsilon);
printf("cooling_eps: %.8g\n", cooling_eps); */
/* print new weights */
printf("W=");
for (i = 1; i <= sm.sizePsi; ++i)
printf("%.3f ", sm.w[i]);
printf("\n");
/* Save model */
char modelfile_tmp[1024];
sprintf(modelfile_tmp, "%s.%d", modelfile, outer_iter);
write_struct_model(modelfile_tmp, &sm, &sparm);
/* impute latent variable using updated weight vector */
for (i=0; i<m; i++) {
free_latent_var(ex[i].h);
ex[i].h = infer_latent_variables(ex[i].x, ex[i].y, &sm, &sparm);
}
/* re-compute feature vector cache */
for (i=0; i<m; i++) {
free_svector(fycache[i]);
fy = psi(ex[i].x, ex[i].y, ex[i].h, &sm, &sparm);
/* DEBUG */
printf("true_psi[%d]=", i);
for (int j = 0; j < sm.sizePsi; ++j)
printf("%.4lf ", fy->words[j].weight);
printf("\n");
diff = add_list_ss(fy);
free_svector(fy);
fy = diff;
fycache[i] = fy;
}
outer_iter++;
} // end outer loop
/* write structural model */
write_struct_model(modelfile, &sm, &sparm);
// skip testing for the moment
/* free memory */
free_struct_sample(sample);
free_struct_model(sm, &sparm);
for(i=0; i<m; i++) {
free_svector(fycache[i]);
}
free(fycache);
return(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 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);
}
