int find_winner_euc(struct entries *codes, struct data_entry *sample,
struct winner_info *win, int knn)
{
struct data_entry *codetmp;
int dim, i, masked;
float diffsf, diff, difference;
eptr p;
dim = codes->dimension;
win->index = -1;
win->winner = NULL;
win->diff = -1.0;
/* Go through all code vectors */
codetmp = rewind_entries(codes, &p);
diffsf = FLT_MAX;
while (codetmp != NULL) {
difference = 0.0;
masked = 0;
/* Compute the distance between codebook and input entry */
for (i = 0; i < dim; i++)
{
if ((sample->mask != NULL) && (sample->mask[i] != 0))
{
masked++;
continue; /* ignore vector components that have 1 in mask */
}
diff = codetmp->points[i] - sample->points[i];
difference += diff * diff;
if (difference > diffsf) break;
}
if (masked == dim)
return 0; /* can't calculate winner, empty data vector */
/* If distance is smaller than previous distances */
if (difference < diffsf) {
win->winner = codetmp;
win->index = p.index;
win->diff = difference;
diffsf = difference;
}
codetmp = next_entry(&p);
}
if (win->index < 0)
ifverbose(3)
fprintf(stderr, "find_winner_euc: can't find winner\n");
return 1; /* number of neighbours */
}
int find_winner_euc2(struct entries *codes, struct data_entry *sample,
struct winner_info *win, int knn)
{
struct data_entry *codetmp;
int dim, i;
float diffsf, diff, difference, *s, *c;
eptr p;
if (sample->mask != NULL)
return find_winner_euc(codes, sample, win, knn);
dim = codes->dimension;
win->index = -1;
win->winner = NULL;
win->diff = -1.0;
/* Go through all code vectors */
codetmp = rewind_entries(codes, &p);
diffsf = FLT_MAX;
while (codetmp != NULL) {
difference = 0.0;
/* Compute the distance between codebook and input entry */
c = codetmp->points; s = sample->points;
for (i = 0; i < dim; i++)
{
diff = *c++ - *s++;
difference += diff * diff;
}
/* If distance is smaller than previous distances */
if (difference < diffsf) {
win->winner = codetmp;
win->index = p.index;
win->diff = difference;
diffsf = difference;
}
codetmp = next_entry(&p);
}
if (win->index < 0)
ifverbose(3)
fprintf(stderr, "find_winner_euc: can't find winner\n");
return 1; /* number of neighbours */
}
int main(int argc, char **argv)
{
long nod, buffer;
float qerror;
char *in_data_file;
char *in_code_file;
struct teach_params teach;
int qmode;
struct entries *data, *codes;
global_options(argc, argv);
if (extract_parameter(argc, argv, "-help", OPTION2))
{
printhelp();
exit(0);
}
in_data_file = extract_parameter(argc, argv, IN_DATA_FILE, ALWAYS);
in_code_file = extract_parameter(argc, argv, IN_CODE_FILE, ALWAYS);
buffer = oatoi(extract_parameter(argc, argv, "-buffer", OPTION), 0);
teach.radius = oatof(extract_parameter(argc, argv, TRAINING_RADIUS, OPTION), 1.0);
qmode = oatoi(extract_parameter(argc, argv, "-qetype", OPTION), 0);
label_not_needed(1);
ifverbose(2)
fprintf(stdout, "Input entries are read from file %s\n", in_data_file);
data = open_entries(in_data_file);
if (data == NULL)
{
fprintf(stderr, "Can't open datafile: %s\n", in_data_file);
exit(1);
}
ifverbose(2)
fprintf(stdout, "Codebook entries are read from file %s\n", in_code_file);
codes = open_entries(in_code_file);
if (codes == NULL)
{
fprintf(stderr, "Can't open code file '%s'\n", in_code_file);
close_entries(data);
exit(1);
}
if (codes->topol < TOPOL_HEXA) {
fprintf(stderr, "File %s is not a map file\n", in_code_file);
close_entries(data);
close_entries(codes);
exit(1);
}
if (data->dimension != codes->dimension) {
fprintf(stderr, "Data and codebook vectors have different dimensions (%d != %d)",
data->dimension, codes->dimension);
close_entries(data);
close_entries(codes);
exit(1);
}
set_teach_params(&teach, codes, data, buffer);
set_som_params(&teach);
if (qmode > 0)
qerror = find_qerror2(&teach);
else
qerror = find_qerror(&teach);
nod = teach.data->num_entries;
ifverbose(1)
fprintf(stdout, "Quantization error of %s with map %s is %f per sample (%d samples)\n",
in_data_file, in_code_file, qerror / (float) nod, nod);
else
int main(int argc, char **argv)
{
int not, bnot, error;
int xdim, ydim;
int topol, neigh;
float alpha1, radius1;
int fixed, weights;
float alpha2, radius2;
float qerror, qerrorb;
char *in_data_file, *in_test_file, *out_code_file, *alpha_s;
struct entries *data = NULL;
struct entries *testdata = NULL;
struct entries *codes = NULL;
struct entries *codess = NULL;
struct entries *tmp;
struct teach_params params;
long buffer, length1, length2, noc, nod;
int qmode;
struct typelist *type_tmp;
error = 0;
global_options(argc, argv);
print_description();
not = get_int("Give the number of trials", 0);
in_data_file = get_str("Give the input data file name");
in_test_file = get_str("Give the input test file name");
out_code_file = get_str("Give the output map file name");
topol = topol_type(get_str("Give the topology type"));
if (topol == TOPOL_UNKNOWN) {
ifverbose(2)
fprintf(stderr, "Unknown topology type, using hexagonal\n");
topol = TOPOL_HEXA;
}
neigh = neigh_type(get_str("Give the neighborhood type"));
if (neigh == NEIGH_UNKNOWN) {
ifverbose(2)
fprintf(stderr, "Unknown neighborhood type, using bubble\n");
neigh = NEIGH_BUBBLE;
}
xdim = get_int("Give the x-dimension", 0);
ydim = get_int("Give the y-dimension", 0);
length1 = get_int("Give the training length of first part", 0);
alpha1 = get_float("Give the training rate of first part", 0.0);
radius1 = get_float("Give the radius in first part", 0.0);
length2 = get_int("Give the training length of second part", 0);
alpha2 = get_float("Give the training rate of second part", 0.0);
radius2 = get_float("Give the radius in second part", 0.0);
printf("\n");
fixed = (int) oatoi(extract_parameter(argc, argv, FIXPOINTS, OPTION), 0);
weights = (int) oatoi(extract_parameter(argc, argv, WEIGHTS, OPTION), 0);
buffer = oatoi(extract_parameter(argc, argv, "-buffer", OPTION), 0);
alpha_s = extract_parameter(argc, argv, "-alpha_type", OPTION);
qmode = oatoi(extract_parameter(argc, argv, "-qetype", OPTION), 0);
use_fixed(fixed);
use_weights(weights);
label_not_needed(1);
ifverbose(2)
fprintf(stderr, "Input entries are read from file %s\n", in_data_file);
data = open_entries(in_data_file);
if (data == NULL)
{
fprintf(stderr, "Can't open data file '%s'\n", in_data_file);
error = 1;
goto end;
}
set_buffer(data, buffer);
ifverbose(2)
fprintf(stderr, "Test entries are read from file %s\n", in_test_file);
testdata = open_entries(in_test_file);
if (testdata == NULL)
{
fprintf(stderr, "Can't open test data file '%s'\n", in_test_file);
error = 1;
goto end;
}
set_buffer(testdata, buffer);
noc = xdim * ydim;
if (noc <= 0)
{
fprintf(stderr, "Dimensions of map (%d %d) are incorrect\n", xdim, ydim);
error = 1;
goto end;
}
if (xdim < 0)
{
fprintf(stderr, "Dimensions of map (%d %d) are incorrect\n", xdim, ydim);
error = 1;
goto end;
}
if (alpha_s)
{
type_tmp = get_type_by_str(alpha_list, alpha_s);
if (type_tmp->data == NULL)
{
fprintf(stderr, "Unknown alpha type %s\n", alpha_s);
error = 1;
goto end;
}
}
else
type_tmp = get_type_by_id(alpha_list, ALPHA_LINEAR);
params.alpha_type = type_tmp->id;
params.alpha_func = type_tmp->data;
codess = NULL;
qerrorb = FLT_MAX;
bnot = 0;
while (not) {
init_random(not);
ifverbose(2)
fprintf(stderr, "Initializing codebook\n");
codes = randinit_codes(data, topol, neigh, xdim, ydim);
if (codes == NULL)
{
fprintf(stderr, "Error initializing random codebook, aborting\n");
error = 1;
goto end;
}
set_teach_params(¶ms, codes, NULL, 0);
set_som_params(¶ms);
params.data = data;
params.length = length1;
params.alpha = alpha1;
params.radius = radius1;
ifverbose(2)
fprintf(stderr, "Training map, first part, rlen: %ld alpha: %f\n",
params.length, params.alpha);
codes = som_training(¶ms);
params.length = length2;
params.alpha = alpha2;
params.radius = radius2;
ifverbose(2)
fprintf(stderr, "Training map, second part, rlen: %ld alpha: %f\n",
params.length, params.alpha);
codes = som_training(¶ms);
params.data = testdata;
ifverbose(2)
fprintf(stderr, "Calculating quantization error\n");
if (qmode > 0)
qerror = find_qerror2(¶ms);
else
qerror = find_qerror(¶ms);
nod = testdata->num_entries;
if (qerror < qerrorb) {
qerrorb = qerror;
bnot = not;
tmp = codess;
codess = codes;
codes = tmp;
}
close_entries(codes);
codes = NULL;
ifverbose(1)
fprintf(stderr, "%3d: %f\n", not, qerror/(float) nod);
not--;
}
if (codess != NULL)
{
ifverbose(2)
fprintf(stdout, "Codebook entries are saved to file %s\n", out_code_file);
save_entries(codess, out_code_file);
ifverbose(1)
fprintf(stdout, "Smallest error with random seed %3d: %f\n",
bnot, qerrorb/(float) nod);
}
end:
if (codess)
close_entries(codess);
if (data)
close_entries(data);
if (testdata)
close_entries(testdata);
return(error);
}
struct snapshot_info *get_snapshot(char *filename, long interval, int type)
{
struct snapshot_info *shot;
int len = 0;
shot = malloc(sizeof(struct snapshot_info));
if (shot == NULL)
{
fprintf(stderr, "get_snapshot: Can't allocate structure\n");
perror("get_snapshot");
return NULL;
}
shot->flags = 0;
shot->fi = NULL;
#ifndef NO_BACKGROUND_SNAP
shot->pid = -1;
#endif
shot->counter = 0;
shot->data = NULL;
/* allocate room for string */
shot->filename = NULL;
if (filename)
{
len = strlen(filename);
if ((shot->filename = malloc(len + 1)) == NULL)
{
fprintf(stderr, "get_snapshot: Can't allocate mem for string\n");
perror("get_snapshot");
free(shot);
return NULL;
}
else
strcpy(shot->filename, filename);
}
if (len > 0)
if (shot->filename[len - 1] == '&')
{
shot->filename[len - 1] = '\0';
shot->flags |= SNAPFLAG_BACKGROUND;
}
shot->interval = interval;
shot->type = type;
switch (type)
{
case SNAPSHOT_KEEPOPEN:
shot->flags |= SNAPFLAG_KEEPOPEN;
shot->flags &= (~SNAPFLAG_NOWAIT);
break;
#ifndef NO_BACKGROUND_SNAP
case SNAPSHOT_ASYNC:
shot->flags |= SNAPFLAG_BACKGROUND;
break;
case SNAPSHOT_ASYNC_NOWAIT:
shot->flags |= SNAPFLAG_BACKGROUND|SNAPFLAG_NOWAIT;
break;
#endif
default:
break;
}
ifverbose(2)
fprintf(stderr, "snapshot: filename: '%s', interval: %ld, type: %s\n",
shot->filename, shot->interval, get_str_by_id(snapshot_list, type));
return shot;
}
int save_snapshot(struct teach_params *teach, long iter)
{
struct entries *codes = teach->codes;
struct snapshot_info *shot = teach->snapshot;
struct data_entry *entry;
eptr p;
char filename[1024]; /* hope this is enough */
struct file_info *fi = NULL;
int retcode = 0;
int bg = shot->flags & SNAPFLAG_BACKGROUND;
int ko = shot->flags & SNAPFLAG_KEEPOPEN;
shot->counter++;
if (ko)
if ((fi = shot->fi) == NULL)
{
if ((shot->fi = open_file(shot->filename, "w")) == NULL)
return 1;
fi = shot->fi;
}
else
fi = shot->fi;
#ifndef NO_BACKGROUND_SNAP
if (bg)
{
if (shot->pid > 0)
{
int statptr;
if (!(shot->flags & SNAPFLAG_NOWAIT))
{
/* fprintf(stderr, "Snap: waiting for pid %d\n", (int)shot->pid); */
/* wait for previous child before forking a new one */
waitpid(shot->pid, &statptr, 0);
shot->pid = -1;
}
}
if ((shot->pid = fork()) < 0)
{
perror("save_snapshot(background):");
return 1;
}
if (shot->pid == 0)
{
/* fprintf(stderr, "saving snapshot\n"); */
}
else
{
/* fprintf(stderr, "forked pid %d\n", (int)shot->pid); */
return 0;
}
}
#endif /* NO_BACKGROUND_SNAP */
if (fi == NULL)
{
/* make filename */
sprintf(filename, shot->filename, iter);
if ((fi = open_file(filename, "w")) == NULL)
return 1;
}
if (ko)
fprintf(fi2fp(fi), "#start %d\n", shot->counter);
if (write_header(fi, codes))
{
fprintf(stderr, "save_snapshot: Error writing headers\n");
close_file(shot->fi);
shot->fi = NULL;
return 1;
}
/* open file for writing */
ifverbose(3)
fprintf(stderr, "saving snapshot: file '%s', type '%s'\n", filename,
get_str_by_id(snapshot_list, shot->type));
fprintf(fi2fp(fi), "#SNAPSHOT FILE\n#iterations: %ld/%ld\n",
iter, teach->length);
for (entry = rewind_entries(codes, &p); entry != NULL; entry = next_entry(&p))
{
if (write_entry(fi, codes, entry))
{
fprintf(stderr, "save_entries: Error writing entry, aborting\n");
retcode = 1;
break;
}
}
if (ko)
{
fprintf(fi2fp(fi), "#end\n");
fflush(fi2fp(fi));
}
else
close_file(fi);
#ifndef NO_BACKGROUND_SNAP
if (bg)
exit(retcode);
else
#endif
return(retcode);
}
int find_winner_knn2(struct entries *codes, struct data_entry *sample,
struct winner_info *win, int knn)
{
struct data_entry *codetmp;
int dim, i, j;
float difference, diff, *s, *c;
eptr p;
if (sample->mask != NULL)
return find_winner_knn(codes, sample, win, knn);
if (knn == 1) /* might be a little faster */
return find_winner_euc2(codes, sample, win, 1);
dim = codes->dimension;
for (i = 0; i < knn; i++)
{
win[i].index = -1;
win[i].winner = NULL;
win[i].diff = FLT_MAX;
}
/* Go through all code vectors */
codetmp = rewind_entries(codes, &p);
while (codetmp != NULL) {
difference = 0.0;
/* Compute the distance between codebook and input entry */
c = codetmp->points; s = sample->points;
for (i = 0; i < dim; i++)
{
diff = *c++ - *s++;
difference += diff * diff;
}
/* If distance is smaller than previous distances */
for (i = 0; (i < knn) && (difference > win[i].diff); i++);
if (i < knn)
{
for (j = knn - 1; j > i; j--)
{
win[j].diff = win[j - 1].diff;
win[j].index = win[j - 1].index;
win[j].winner = win[j - 1].winner;
}
win[i].diff = difference;
win[i].index = p.index;
win[i].winner = codetmp;
}
codetmp = next_entry(&p);
}
if (win->index < 0)
ifverbose(3)
fprintf(stderr, "find_winner_knn: can't find winner\n");
return knn; /* number of neighbours */
}
int find_winner_knn(struct entries *codes, struct data_entry *sample,
struct winner_info *win, int knn)
{
struct data_entry *codetmp;
int dim, i, j, masked;
float difference, diff;
eptr p;
if (knn == 1) /* might be a little faster */
return find_winner_euc(codes, sample, win, 1);
dim = codes->dimension;
for (i = 0; i < knn; i++)
{
win[i].index = -1;
win[i].winner = NULL;
win[i].diff = FLT_MAX;
}
/* Go through all code vectors */
codetmp = rewind_entries(codes, &p);
while (codetmp != NULL) {
difference = 0.0;
masked = 0;
/* Compute the distance between codebook and input entry */
for (i = 0; i < dim; i++)
{
/* pitaisiko ottaa huomioon myos codebookissa olevat?? */
if ((sample->mask != NULL) && (sample->mask[i] != 0))
{
masked++;
continue; /* ignore vector components that have 1 in mask */
}
diff = codetmp->points[i] - sample->points[i];
difference += diff * diff;
if (difference > win[knn-1].diff) break;
}
if (masked == dim)
return 0;
/* If distance is smaller than previous distances */
for (i = 0; (i < knn) && (difference > win[i].diff); i++);
if (i < knn)
{
for (j = knn - 1; j > i; j--)
{
win[j].diff = win[j - 1].diff;
win[j].index = win[j - 1].index;
win[j].winner = win[j - 1].winner;
}
win[i].diff = difference;
win[i].index = p.index;
win[i].winner = codetmp;
}
codetmp = next_entry(&p);
}
if (win->index < 0)
ifverbose(3)
fprintf(stderr, "find_winner_knn: can't find winner\n");
return knn; /* number of neighbours */
}