static void
dot_cleanup_node(node_t * n)
{
free_list(ND_in(n));
free_list(ND_out(n));
free_list(ND_flat_out(n));
free_list(ND_flat_in(n));
free_list(ND_other(n));
free_label(ND_label(n));
free_label(ND_xlabel(n));
if (ND_shape(n))
ND_shape(n)->fns->freefn(n);
agdelrec(n, "Agnodeinfo_t");
}
void graph_cleanup(graph_t *g)
{
free(GD_drawing(g));
GD_drawing(g) = NULL;
free_label(GD_label(g));
memset(&(g->u), 0, sizeof(Agraphinfo_t));
}
static void
dot_cleanup_edge(edge_t * e)
{
dot_free_splines(e);
free_label(ED_label(e));
memset(&(e->u), 0, sizeof(Agedgeinfo_t));
}
static void free_chip(sensors_chip *chip)
{
int i;
for (i = 0; i < chip->chips.fits_count; i++)
free_chip_name(&chip->chips.fits[i]);
free(chip->chips.fits);
chip->chips.fits_count = chip->chips.fits_max = 0;
for (i = 0; i < chip->labels_count; i++)
free_label(&chip->labels[i]);
free(chip->labels);
chip->labels_count = chip->labels_max = 0;
for (i = 0; i < chip->sets_count; i++)
free_set(&chip->sets[i]);
free(chip->sets);
chip->sets_count = chip->sets_max = 0;
for (i = 0; i < chip->computes_count; i++)
free_compute(&chip->computes[i]);
free(chip->computes);
chip->computes_count = chip->computes_max = 0;
for (i = 0; i < chip->ignores_count; i++)
free_ignore(&chip->ignores[i]);
free(chip->ignores);
chip->ignores_count = chip->ignores_max = 0;
}
void gtk_scrollbox_set_label(GtkScrollbox *self, gint n, gchar *value)
{
gboolean append = TRUE;
GdkPixmap *newpixmap;
struct label *newlbl;
if (n != -1)
append = FALSE;
if (!append)
{
struct label *lbl = (struct label*)g_ptr_array_index(self->labels, n);
if (lbl)
free_label(lbl);
newlbl = lbl;
}
else
{
newlbl = g_new0(struct label, 1);
g_ptr_array_add(self->labels, newlbl);
}
newpixmap = make_pixmap(self, value);
newlbl->pixmap = newpixmap;
newlbl->msg = g_strdup(value);
}
static void twopi_cleanup_node(node_t * n)
{
free(ND_alg(n));
if (ND_shape(n))
ND_shape(n)->fns->freefn(n);
free_label(ND_label(n));
memset(&(n->u), 0, sizeof(Agnodeinfo_t));
}
void neato_cleanup_node(node_t * n)
{
if (ND_shape(n)) {
ND_shape(n)->fns->freefn(n);
}
free(ND_pos(n));
free_label(ND_label(n));
memset(&(n->u), 0, sizeof(Agnodeinfo_t));
}
void free_struct_sample(SAMPLE s) {
int i;
for (i=0;i<s.n;i++) {
free_pattern(s.examples[i].x);
free_label(s.examples[i].y);
free_latent_var(s.examples[i].h);
}
free(s.examples);
}
void free_struct_sample(SAMPLE s)
{
/* Frees the memory of sample s. */
int i;
for(i=0;i<s.n;i++) {
free_pattern(s.examples[i].x);
free_label(s.examples[i].y);
}
free(s.examples);
}
void
clean_labels (GList *labels)
{
GList *current = g_list_first (labels);
while (current != NULL)
{
Label *label;
label = (Label *) current->data;
free_label (label);
current = g_list_next (current);
}
}
void free_struct_sample(SAMPLE s) {
/*
Free the whole training sample.
*/
int i;
for (i=0;i<s.n;i++) {
free_pattern(s.examples[i].x);
free_label(s.examples[i].y);
}
free(s.examples);
}
int main(int argc, char* argv[]) {
double avgloss,l;
long i, correct;
char testfile[1024];
char modelfile[1024];
STRUCTMODEL model;
STRUCT_LEARN_PARM sparm;
LEARN_PARM lparm;
KERNEL_PARM kparm;
SAMPLE testsample;
LABEL y;
LATENT_VAR h;
/* read input parameters */
read_input_parameters(argc,argv,testfile,modelfile,&sparm);
/* read model file */
printf("Reading model..."); fflush(stdout);
// model = read_struct_model(modelfile, &sparm);
printf("done.\n");
/* read test examples */
printf("Reading test examples..."); fflush(stdout);
testsample = read_struct_examples(testfile,&sparm);
printf("done.\n");
init_struct_model(testsample,&model,&sparm,&lparm,&kparm);
avgloss = 0.0;
correct = 0;
for (i=0;i<testsample.n;i++) {
classify_struct_example(testsample.examples[i].x,&y,&h,&model,&sparm);
l = loss(testsample.examples[i].y,y,h,&sparm);
avgloss += l;
if (l==0) correct++;
free_label(y);
free_latent_var(h);
}
printf("Average loss on test set: %.4f\n", avgloss/testsample.n);
printf("Zero/one error on test set: %.4f\n", 1.0 - ((float) correct)/testsample.n);
free_struct_sample(testsample);
free_struct_model(model,&sparm);
return(0);
}
static void
dot_cleanup_node(node_t * n)
{
free_list(ND_in(n));
free_list(ND_out(n));
free_list(ND_flat_out(n));
free_list(ND_flat_in(n));
free_list(ND_other(n));
free_label(ND_label(n));
if (ND_shape(n))
ND_shape(n)->fns->freefn(n);
memset(&(n->u), 0, sizeof(Agnodeinfo_t));
}
void graph_cleanup(graph_t *g)
{
if (GD_drawing(g) && GD_drawing(g)->xdots)
freeXDot ((xdot*)GD_drawing(g)->xdots);
if (GD_drawing(g) && GD_drawing(g)->id)
free (GD_drawing(g)->id);
free(GD_drawing(g));
GD_drawing(g) = NULL;
free_label(GD_label(g));
//FIX HERE , STILL SHALLOW
//memset(&(g->u), 0, sizeof(Agraphinfo_t));
agclean(g, AGRAPH,"Agraphinfo_t");
}
void find_most_violated_constraint(SVECTOR **fydelta, double *rhs,
EXAMPLE *ex, SVECTOR *fycached, long n,
STRUCTMODEL *sm, STRUCT_LEARN_PARM *sparm,
double *rt_viol, double *rt_psi,
long *argmax_count)
/* returns fydelta=fy-fybar and rhs scalar value that correspond
to the most violated constraint for example ex */
{
double rt2=0;
LABEL ybar;
SVECTOR *fybar, *fy;
double factor,lossval;
if(struct_verbosity>=2) rt2=get_runtime();
(*argmax_count)++;
if(sparm->loss_type == SLACK_RESCALING)
ybar=find_most_violated_constraint_slackrescaling(ex->x,ex->y,sm,sparm);
else
ybar=find_most_violated_constraint_marginrescaling(ex->x,ex->y,sm,sparm);
if(struct_verbosity>=2) (*rt_viol)+=MAX(get_runtime()-rt2,0);
if(empty_label(ybar)) {
printf("ERROR: empty label was returned for example\n");
/* exit(1); */
/* continue; */
}
/**** get psi(x,y) and psi(x,ybar) ****/
if(struct_verbosity>=2) rt2=get_runtime();
if(fycached)
fy=copy_svector(fycached);
else
fy=psi(ex->x,ex->y,sm,sparm);
fybar=psi(ex->x,ybar,sm,sparm);
if(struct_verbosity>=2) (*rt_psi)+=MAX(get_runtime()-rt2,0);
lossval=loss(ex->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 */
mult_svector_list(fy,factor);
mult_svector_list(fybar,-factor);
append_svector_list(fybar,fy); /* compute fy-fybar */
(*fydelta)=fybar;
(*rhs)=lossval/n;
}
static void
dot_cleanup_node(node_t * n)
{
free_list(ND_in(n));
free_list(ND_out(n));
free_list(ND_flat_out(n));
free_list(ND_flat_in(n));
free_list(ND_other(n));
free_label(ND_label(n));
if (ND_shape(n))
ND_shape(n)->fns->freefn(n);
#ifndef WITH_CGRAPH
memset(&(n->u), 0, sizeof(Agnodeinfo_t));
#else /* WITH_CGRAPH */
agclean(agraphof(n), AGNODE,"Agnodeinfo_t");
#endif /* WITH_CGRAPH */
}
static void cleanup_subgs(graph_t * g)
{
graph_t *mg;
edge_t *me;
node_t *mn;
graph_t *subg;
mg = g->meta_node->graph;
for (me = agfstout(mg, g->meta_node); me; me = agnxtout(mg, me)) {
mn = me->head;
subg = agusergraph(mn);
free_label(GD_label(subg));
if (GD_alg(subg)) {
free(PORTS(subg));
free(GD_alg(subg));
}
cleanup_subgs(subg);
}
}
void gtk_scrollbox_clear (GtkScrollbox *self)
{
stop_callback(self);
DEBUG_PRINT("coming in %d\n", self->labels->len);
while(self->labels->len > 0)
{
struct label *lbl = (struct label*)g_ptr_array_index(self->labels, 0);
free_label(lbl);
g_ptr_array_remove_index(self->labels, 0);
}
DEBUG_PRINT("going out %d\n", self->labels->len);
self->pixmap = NULL;
gtk_widget_set_size_request(GTK_WIDGET(self), 0, 0);
self->draw_middle = 0;
self->draw_maxmiddle = 0;
}
gboolean
mcus_compiler_parse (MCUSCompiler *self, const gchar *code, GError **error)
{
/* In EBNF:
* comment ::= ";" , ? any characters ? , "\n"
* terminating-whitespace ::= comment | whitespace | "\n"
* assembly ::= { terminating-whitespace , { terminating-whitespace } , ( instruction | label | lookup_table ) } , { terminating-whitespace } , "\0" */
/* Set up parser variables */
reset_state (self);
self->priv->code = code;
self->priv->i = code;
self->priv->dirty = TRUE;
skip_whitespace (self, TRUE, FALSE);
while (*(self->priv->i) != '\0') {
MCUSInstruction instruction;
MCUSLabel label;
MCUSLookupTable lookup_table;
GError *child_error = NULL;
/* Are we finished? */
if (*(self->priv->i) == '\0')
break;
if (lex_lookup_table (self, &lookup_table, &child_error) == TRUE) {
/* Lookup table */
if (store_lookup_table (self, &lookup_table, &child_error) == FALSE) {
free_lookup_table (&lookup_table);
goto throw_error;
}
skip_whitespace (self, TRUE, FALSE);
continue;
} else if (g_error_matches (child_error, MCUS_COMPILER_ERROR, MCUS_COMPILER_ERROR_INVALID_LOOKUP_TABLE) == FALSE) {
goto throw_error;
} else {
g_clear_error (&child_error);
}
if (lex_label (self, &label, &child_error) == TRUE) {
/* Label */
if (store_label (self, &label, &child_error) == FALSE) {
free_label (&label);
goto throw_error;
}
skip_whitespace (self, TRUE, FALSE);
continue;
} else if (g_error_matches (child_error, MCUS_COMPILER_ERROR, MCUS_COMPILER_ERROR_INVALID_LABEL_DELIMITATION) == FALSE) {
goto throw_error;
} else {
g_clear_error (&child_error);
}
if (lex_instruction (self, &instruction, &child_error) == TRUE) {
/* Instruction */
store_instruction (self, &instruction);
skip_whitespace (self, TRUE, FALSE);
continue;
}
throw_error:
/* Throw the error */
g_propagate_error (error, child_error);
return FALSE;
}
return TRUE;
}
void neato_cleanup_edge(edge_t * e)
{
neato_free_splines(e);
free_label(ED_label(e));
memset(&(e->u), 0, sizeof(Agedgeinfo_t));
}
/* select training samples for aspectlet */
void select_examples_aspectlet(char *trainfile, CAD **cads, int cad_index)
{
int i, j, n, count, count_pos, part_num, view_label;
int *flag;
char *pname;
float azimuth, elevation, distance;
FILE *fp;
LABEL y;
EXAMPLE example, *examples_aspectlet;
/* find corresponding parts between two cads */
part_num = cads[cad_index]->part_num;
flag = (int*)my_malloc(sizeof(int)*part_num);
for(i = 0; i < part_num; i++)
{
pname = cads[cad_index]->part_names[i];
for(j = 0; j < cads[0]->part_num; j++)
{
if(strcmp(pname, cads[0]->part_names[j]) == 0)
{
flag[i] = j;
break;
}
}
}
examples_aspectlet = NULL;
/* select positive samples */
/* open data file */
if((fp = fopen(trainfile, "r")) == NULL)
{
printf("Can not open data file %s\n", trainfile);
exit(1);
}
fscanf(fp, "%d", &n);
count = 0;
for(i = 0; i < n; i++)
{
/* read example */
example = read_one_example(fp, cads);
y = example.y;
if(y.object_label == 1)
{
azimuth = cads[0]->objects2d[y.view_label]->azimuth;
elevation = cads[0]->objects2d[y.view_label]->elevation;
distance = cads[0]->objects2d[y.view_label]->distance;
/* find the view label */
view_label = -1;
for(j = 0; j < cads[cad_index]->view_num; j++)
{
if(cads[cad_index]->objects2d[j]->azimuth == azimuth && cads[cad_index]->objects2d[j]->elevation == elevation && cads[cad_index]->objects2d[j]->distance == distance)
{
view_label = j;
break;
}
}
/* select the sample */
if(view_label != -1)
{
examples_aspectlet = (EXAMPLE*)realloc(examples_aspectlet, sizeof(EXAMPLE)*(count+1));
if(examples_aspectlet == NULL)
{
printf("out of memory\n");
exit(1);
}
/* construct the sample */
examples_aspectlet[count].y.object_label = y.object_label;
examples_aspectlet[count].y.cad_label = y.cad_label;
examples_aspectlet[count].y.view_label = view_label;
examples_aspectlet[count].y.part_num = part_num;
examples_aspectlet[count].y.part_label = (float*)my_malloc(sizeof(float)*2*part_num);
/* get part labels */
for(j = 0; j < part_num; j++)
{
if(cads[cad_index]->roots[j] != -1) /* not root template */
{
examples_aspectlet[count].y.part_label[j] = y.part_label[flag[j]];
examples_aspectlet[count].y.part_label[j+part_num] = y.part_label[flag[j]+cads[0]->part_num];
}
else /* is root template */
{
if(y.part_label[flag[j]] == 0 && y.part_label[flag[j]+cads[0]->part_num] == 0) /* root not visible */
{
examples_aspectlet[count].y.part_label[j] = 0;
examples_aspectlet[count].y.part_label[j+part_num] = 0;
}
else /* compute root location and bounding box */
compute_bbox_root(&(examples_aspectlet[count].y), j, cads[cad_index]);
}
}
/* get occlusion label */
examples_aspectlet[count].y.occlusion = (int*)my_malloc(sizeof(int)*part_num);
for(j = 0; j < part_num; j++)
examples_aspectlet[count].y.occlusion[j] = y.occlusion[flag[j]];
examples_aspectlet[count].y.energy = 0;
/* copy the image */
examples_aspectlet[count].x.image = copy_cumatrix(example.x.image);
count++;
} /* end if view_label != -1 */
free_pattern(example.x);
free_label(example.y);
} /* end if y.object_label == 1 */
else
break;
}
/* select negative samples */
count_pos = count;
for(; i < n; i++)
{
if(y.object_label == -1)
{
examples_aspectlet = (EXAMPLE*)realloc(examples_aspectlet, sizeof(EXAMPLE)*(count+1));
if(examples_aspectlet == NULL)
{
printf("out of memory\n");
exit(1);
}
examples_aspectlet[count].y.object_label = -1;
examples_aspectlet[count].y.cad_label = -1;
examples_aspectlet[count].y.view_label = -1;
examples_aspectlet[count].y.part_num = -1;
examples_aspectlet[count].y.part_label = NULL;
examples_aspectlet[count].y.occlusion = NULL;
for(j = 0; j < 4; j++)
examples_aspectlet[count].y.bbox[j] = 0;
examples_aspectlet[count].y.energy = 0;
/* copy the image */
examples_aspectlet[count].x.image = copy_cumatrix(example.x.image);
count++;
if(count >= 2*count_pos)
break;
}
free_pattern(example.x);
free_label(example.y);
example = read_one_example(fp, cads);
y = example.y;
}
free_pattern(example.x);
free_label(example.y);
free(flag);
fclose(fp);
/* write examples */
printf("Writing data to temp_part.dat: %d positive samples, %d negative samples...\n", count_pos, count-count_pos);
fp = fopen("temp_part.dat", "w");
fprintf(fp, "%d\n", count);
for(i = 0; i < count; i++)
{
/* write object label */
fprintf(fp, "%d ", examples_aspectlet[i].y.object_label);
if(examples_aspectlet[i].y.object_label == 1)
{
/* write cad label */
fprintf(fp, "%d ", examples_aspectlet[i].y.cad_label);
/* write view label */
fprintf(fp, "%d ", examples_aspectlet[i].y.view_label);
/* write part label */
for(j = 0; j < 2*part_num; j++)
fprintf(fp, "%f ", examples_aspectlet[i].y.part_label[j]);
/* write occlusion label */
for(j = 0; j < part_num; j++)
fprintf(fp, "%d ", examples_aspectlet[i].y.occlusion[j]);
/* write bounding box */
for(j = 0; j < 4; j++)
fprintf(fp, "%f ", examples_aspectlet[i].y.bbox[j]);
}
/* write image size */
fprintf(fp, "%d ", examples_aspectlet[i].x.image.dims_num);
for(j = 0; j < examples_aspectlet[i].x.image.dims_num; j++)
fprintf(fp, "%d ", examples_aspectlet[i].x.image.dims[j]);
/* write image pixel */
for(j = 0; j < examples_aspectlet[i].x.image.length; j++)
fprintf(fp, "%u ", (unsigned int)examples_aspectlet[i].x.image.data[j]);
fprintf(fp, "\n");
}
fclose(fp);
printf("Done\n");
/* clean up */
for(i = 0; i < count; i++)
{
free_pattern(examples_aspectlet[i].x);
free_label(examples_aspectlet[i].y);
}
free(examples_aspectlet);
}
int main (int argc, char* argv[])
{
long correct=0,incorrect=0,no_accuracy=0;
long i;
double t1,runtime=0;
double avgloss=0,l;
FILE *predfl;
STRUCTMODEL model;
STRUCT_LEARN_PARM sparm;
STRUCT_TEST_STATS teststats;
SAMPLE testsample;
LABEL y;
svm_struct_classify_api_init(argc,argv);
read_input_parameters(argc,argv,testfile,modelfile,predictionsfile,&sparm,
&verbosity,&struct_verbosity);
if(struct_verbosity>=1) {
printf("Reading model..."); fflush(stdout);
}
model=read_struct_model(modelfile,&sparm);
if(struct_verbosity>=1) {
fprintf(stdout, "done.\n");
}
if(model.svm_model->kernel_parm.kernel_type == LINEAR) { /* linear kernel */
/* compute weight vector */
//add_weight_vector_to_linear_model(model.svm_model);
//model.w=model.svm_model->lin_weights;
}
if(struct_verbosity>=1) {
printf("Reading test examples..."); fflush(stdout);
}
testsample=read_struct_examples(testfile,&sparm);
if(struct_verbosity>=1) {
printf("done.\n"); fflush(stdout);
}
if(struct_verbosity>=1) {
printf("Classifying test examples..."); fflush(stdout);
}
if ((predfl = fopen (predictionsfile, "w")) == NULL)
{ perror (predictionsfile); exit (1); }
for(i=0;i<testsample.n;i++) {
t1=get_runtime();
y=classify_struct_example(testsample.examples[i].x,&model,&sparm);
runtime+=(get_runtime()-t1);
write_label(predfl,y);
l=loss(testsample.examples[i].y,y,&sparm);
avgloss+=l;
if(l == 0)
correct++;
else
incorrect++;
eval_prediction(i,testsample.examples[i],y,&model,&sparm,&teststats);
if(empty_label(testsample.examples[i].y))
{ no_accuracy=1; } /* test data is not labeled */
if(struct_verbosity>=2) {
if((i+1) % 100 == 0) {
printf("%ld..",i+1); fflush(stdout);
}
}
free_label(y);
}
avgloss/=testsample.n;
fclose(predfl);
if(struct_verbosity>=1) {
printf("done\n");
printf("Runtime (without IO) in cpu-seconds: %.2f\n",
(float)(runtime/100.0));
}
if((!no_accuracy) && (struct_verbosity>=1)) {
printf("Average loss on test set: %.4f\n",(float)avgloss);
printf("Zero/one-error on test set: %.2f%% (%ld correct, %ld incorrect, %d total)\n",(float)100.0*incorrect/testsample.n,correct,incorrect,testsample.n);
}
print_struct_testing_stats(testsample,&model,&sparm,&teststats);
free_struct_sample(testsample);
free_struct_model(model);
svm_struct_classify_api_exit();
return(0);
}
/* data mining hard negative samples */
void data_mining_hard_examples(char *trainfile, char *testfile, STRUCT_LEARN_PARM *sparm, STRUCTMODEL *sm)
{
int i, j, n, ntrain, ntest, num, object_label, count, count_pos, count_neg;
char line[BUFFLE_SIZE];
double *energy, *energies;
SAMPLE testsample;
LABEL *y = NULL;
ENERGYINDEX *energy_index;
FILE *fp, *ftrain, *ftest;
/* MPI process */
int rank;
int procs_num;
int start, end, block_size;
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &procs_num);
/* read negative examples */
printf("Reading negative samples for data mining...");
testsample = read_struct_examples(testfile, sparm, sm);
printf("Done\n");
n = testsample.n;
block_size = (n+procs_num-1) / procs_num;
start = rank*block_size;
end = start+block_size-1 > n-1 ? n-1 : start+block_size-1;
energy = (double*)my_malloc(sizeof(double)*n);
energies = (double*)my_malloc(sizeof(double)*n);
memset(energy, 0, sizeof(double)*n);
memset(energies, 0, sizeof(double)*n);
for(i = start; i <= end; i++)
{
y = classify_struct_example(testsample.examples[i].x, &num, 0, sm, sparm);
count = 0;
for(j = 0; j < num; j++)
{
if(y[j].object_label)
count++;
}
printf("Data mining hard negative example %d/%d: %d objects detected\n", i+1, n, count);
if(count == 0)
energy[i] = -sparm->loss_value;
else
{
for(j = 0; j < num; j++)
{
if(y[j].object_label)
{
energy[i] = y[j].energy;
break;
}
}
}
/* free labels */
for(j = 0; j < num; j++)
free_label(y[j]);
free(y);
}
MPI_Allreduce(energy, energies, n, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
if(rank == 0)
{
energy_index = (ENERGYINDEX*)my_malloc(sizeof(ENERGYINDEX)*n);
for(i = 0; i < n; i++)
{
energy_index[i].index = i;
energy_index[i].energy = energies[i];
}
/* sort energies */
qsort(energy_index, n, sizeof(ENERGYINDEX), compare_energy);
/* construct new training data and write to file */
printf("Writing data to temp.dat...\n");
fp = fopen("temp.dat", "w");
ftrain = fopen(trainfile, "r");
if(ftrain == NULL)
{
printf("Cannot open file %s to read\n", trainfile);
exit(1);
}
ftest = fopen(testfile, "r");
if(ftest == NULL)
{
printf("Cannot open file %s to read\n", testfile);
exit(1);
}
/* positive samples from training data file */
fscanf(ftrain, "%d\n", &ntrain);
count_pos = ntrain/2;
fscanf(ftest, "%d\n", &ntest);
if(ntest < ntrain/2)
n = ntest + ntrain/2;
else
n = ntrain;
count_neg = n - count_pos;
fprintf(fp, "%d\n", n);
for(i = 0; i < ntrain; i++)
{
fgets(line, BUFFLE_SIZE, ftrain);
sscanf(line, "%d", &object_label);
if(object_label == 1)
fputs(line, fp);
else
break;
}
fclose(ftrain);
/* negative samples from hard negative examples */
count = 0;
for(i = 0; i < ntest; i++)
{
fgets(line, BUFFLE_SIZE, ftest);
for(j = 0; j < count_neg; j++)
{
if(i == energy_index[j].index)
{
count++;
fputs(line, fp);
break;
}
}
if(count >= count_neg)
break;
}
fclose(ftest);
fclose(fp);
free(energy_index);
printf("Done\n");
}
MPI_Barrier(MPI_COMM_WORLD);
free(energy);
free(energies);
free_struct_sample(testsample);
}
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);
}
SVECTOR* find_cutting_plane(EXAMPLE *ex, SVECTOR **fycache, double *margin, long m, STRUCTMODEL *sm, STRUCT_LEARN_PARM *sparm,
int *valid_examples) {
long i, j;
SVECTOR *f, *fy, *fybar, *lhs;
LABEL ybar;
double lossval;
double *new_constraint;
long valid_count = 0;
long l,k;
SVECTOR *fvec;
WORD *words;
/* find cutting plane */
lhs = NULL;
*margin = 0;
for (i=0;i<m;i++) {
if (valid_examples[i]) {
valid_count++;
}
}
for (i=0;i<m;i++) {
if (!valid_examples[i]) {
continue;
}
find_most_violated_constraint_marginrescaling(ex[i].x, ex[i].y, &ybar, sm, sparm);
/* get difference vector */
fy = copy_svector(fycache[i]);
fybar = psi(ex[i].x,ybar,sm,sparm);
lossval = loss(ex[i].y,ybar,sparm);
free_label(ybar);
/* scale difference vector */
for (f=fy;f;f=f->next) {
//f->factor*=1.0/m;
//f->factor*=ex[i].x.example_cost/m;
f->factor*=ex[i].x.example_cost/valid_count;
}
for (f=fybar;f;f=f->next) {
//f->factor*=-1.0/m;
//f->factor*=-ex[i].x.example_cost/m;
f->factor*=-ex[i].x.example_cost/valid_count;
}
/* add ybar to constraint */
append_svector_list(fy,lhs);
append_svector_list(fybar,fy);
lhs = fybar;
//*margin+=lossval/m;
//*margin+=lossval*ex[i].x.example_cost/m;
*margin+=lossval*ex[i].x.example_cost/valid_count;
}
/* compact the linear representation */
new_constraint = add_list_nn(lhs, sm->sizePsi);
free_svector(lhs);
l=0;
for (i=1;i<sm->sizePsi+1;i++) {
if (fabs(new_constraint[i])>1E-10) l++; // non-zero
}
words = (WORD*)my_malloc(sizeof(WORD)*(l+1));
assert(words!=NULL);
k=0;
for (i=1;i<sm->sizePsi+1;i++) {
if (fabs(new_constraint[i])>1E-10) {
words[k].wnum = i;
words[k].weight = new_constraint[i];
k++;
}
}
words[k].wnum = 0;
words[k].weight = 0.0;
fvec = create_svector(words,"",1);
free(words);
free(new_constraint);
return(fvec);
}
SVECTOR* find_cutting_plane(EXAMPLE *ex, SVECTOR **fycache, double *margin, long m, STRUCTMODEL *sm,
STRUCT_LEARN_PARM *sparm, char* tmpdir, char *trainfile, double frac_sim, double Fweight,
char *dataset_stats_file, double rho_admm, long isExhaustive, long isLPrelaxation,
double *margin2, int datasetStartIdx, int chunkSz, int eid, int chunkid) {
long i;
SVECTOR *f, *fy, *fybar, *lhs;
LABEL ybar;
LATENT_VAR hbar;
double lossval;
double *new_constraint;
long l,k;
SVECTOR *fvec;
WORD *words;
LABEL *ybar_all = (LABEL*) malloc(sizeof(LABEL) * m);
LATENT_VAR *hbar_all = (LATENT_VAR*) malloc (sizeof(LATENT_VAR) * m);
time_t mv_start, mv_end;
time(&mv_start);
find_most_violated_constraint_marginrescaling_all_online(ybar_all, hbar_all, sm, sparm, m,
tmpdir, trainfile, frac_sim, dataset_stats_file, rho_admm, isExhaustive, isLPrelaxation,
Fweight, datasetStartIdx, chunkSz, eid, chunkid);
time(&mv_end);
#if (DEBUG_LEVEL==1)
print_time(mv_start, mv_end, "Max violators");
#endif
/* find cutting plane */
lhs = NULL;
lossval = lossF1(ex, m, ybar_all, sparm, Fweight);
*margin = lossval;
*margin2 = 0;
for (i=0;i<m;i++) {
//find_most_violated_constraint_marginrescaling(ex[i].x, ex[i].y, &ybar, &hbar, sm, sparm);
ybar = ybar_all[i];
hbar = hbar_all[i];
/* get difference vector */
fy = copy_svector(fycache[i]);
fybar = psi(ex[i].x,ybar,hbar,sm,sparm);
lossval = loss(ex[i].y,ybar,hbar,sparm);
free_label(ybar);
free_latent_var(hbar);
/* scale difference vector */
for (f=fy;f;f=f->next) {
f->factor*=1.0/m;
//f->factor*=ex[i].x.example_cost/m;
}
for (f=fybar;f;f=f->next) {
f->factor*=-1.0/m;
//f->factor*=-ex[i].x.example_cost/m;
}
/* add ybar to constraint */
append_svector_list(fy,lhs);
append_svector_list(fybar,fy);
lhs = fybar;
*margin2+=lossval/m;
//*margin+=lossval*ex[i].x.example_cost/m;
}
free(ybar_all);
free(hbar_all);
/* compact the linear representation */
new_constraint = add_list_nn(lhs, sm->sizePsi);
// printf("After this segfault ? \n");fflush(stdout);
// printf("%x\n",new_constraint);
free_svector(lhs);
l=0;
for (i=1;i<sm->sizePsi+1;i++) {
if (fabs(new_constraint[i])>1E-10) l++; // non-zero
}
words = (WORD*)my_malloc(sizeof(WORD)*(l+1));
assert(words!=NULL);
k=0;
for (i=1;i<sm->sizePsi+1;i++) {
if (fabs(new_constraint[i])>1E-10) {
words[k].wnum = i;
words[k].weight = new_constraint[i];
k++;
}
}
words[k].wnum = 0;
words[k].weight = 0.0;
fvec = create_svector(words,"",1);
free(words);
free(new_constraint);
return(fvec);
}
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);
}
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 mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
#endif
{
long correct=0,incorrect=0,no_accuracy=0;
long i;
double t1,runtime=0;
double avgloss=0,l;
#ifndef COMPILE_MEX_INTERFACE
FILE *predfl;
#endif
STRUCTMODEL model;
STRUCT_LEARN_PARM sparm;
STRUCT_TEST_STATS teststats;
SAMPLE testsample;
LABEL y;
#ifdef COMPILE_MEX_INTERFACE
int argc;
char **argv;
if (nrhs < 3) {
print_help();
return;
}
else if (nrhs==3) {
argc=1;
argv=(char **)my_malloc(MAX_ARGVS*sizeof(char *));
argv[0]="OLR";
}
else
create_argc_argv(prhs[3],&argc,&argv);
#endif
svm_struct_classify_api_init(argc,argv);
#ifndef COMPILE_MEX_INTERFACE
read_input_parameters(argc,argv,testfile,modelfile,predictionsfile,&sparm,
&verbosity,&struct_verbosity);
#else
read_input_parameters(argc,argv,&sparm,&verbosity,&struct_verbosity);
#endif
if(struct_verbosity>=1) {
printf("Reading model..."); fflush(stdout);
}
#ifndef COMPILE_MEX_INTERFACE
model=read_struct_model(modelfile,&sparm);
#else
model=read_struct_model(prhs[2],&sparm);
#endif
if(struct_verbosity>=1) {
fprintf(stdout, "done.\n");
}
if(model.svm_model->kernel_parm.kernel_type == LINEAR) { /* linear kernel */
/* compute weight vector */
add_weight_vector_to_linear_model(model.svm_model);
model.w=model.svm_model->lin_weights;
}
if(struct_verbosity>=1) {
printf("Reading test examples..."); fflush(stdout);
}
#ifndef COMPILE_MEX_INTERFACE
testsample=read_struct_examples(testfile,&sparm);
#else
testsample=read_struct_examples(prhs,&sparm);
#endif
if(struct_verbosity>=1) {
printf("done.\n"); fflush(stdout);
}
if(struct_verbosity>=1) {
printf("Classifying test examples..."); fflush(stdout);
}
#ifndef COMPILE_MEX_INTERFACE
if ((predfl = fopen (predictionsfile, "w")) == NULL)
{ perror (predictionsfile); exit (1); }
#else
mwSize rows=mxGetM(prhs[0]);
mxArray *predictions=mxCreateDoubleMatrix(rows,1,mxREAL);
double *pred_ptr=mxGetPr(predictions);
#endif
for(i=0;i<testsample.n;i++) {
t1=get_runtime();
y=classify_struct_example(testsample.examples[i].x,&model,&sparm);
runtime+=(get_runtime()-t1);
#ifndef COMPILE_MEX_INTERFACE
write_label(predfl,y);
#else
write_label(&pred_ptr,y);
#endif
l=loss(testsample.examples[i].y,y,&sparm);
avgloss+=l;
if(l == 0)
correct++;
else
incorrect++;
eval_prediction(i,testsample.examples[i],y,&model,&sparm,&teststats);
if(empty_label(testsample.examples[i].y))
{ no_accuracy=1; } /* test data is not labeled */
if(struct_verbosity>=2) {
if((i+1) % 100 == 0) {
printf("%ld..",i+1); fflush(stdout);
}
}
free_label(y);
}
avgloss/=testsample.n;
#ifndef COMPILE_MEX_INTERFACE
fclose(predfl);
#endif
if(struct_verbosity>=1) {
printf("done\n");
printf("Runtime (without IO) in cpu-seconds: %.2f\n",
(float)(runtime/100.0));
}
if((!no_accuracy) && (struct_verbosity>=1)) {
printf("Average loss on test set: %.4f\n",(float)avgloss);
printf("Zero/one-error on test set: %.2f%% (%ld correct, %ld incorrect, %d total)\n",(float)100.0*incorrect/testsample.n,correct,incorrect,testsample.n);
}
print_struct_testing_stats(testsample,&model,&sparm,&teststats);
free_struct_sample(testsample);
free_struct_model(model);
svm_struct_classify_api_exit();
#ifndef COMPILE_MEX_INTERFACE
return(0);
#else
plhs[0]=predictions;
#endif
}
