void nTM_each_document(WordIndex *words, unsigned int length)
{
// Increment the total number of documents seen
_static_nTM->documents++;
// Create a new SparseCount for N^d_z (count of topics in this document)
count_list_add(_static_nTM->ndzs);
for (int i=0; i<length; i++) {
// Sample a topic from P(z|w)
TopicIndex assigned_topic = nTM_sample_topic(_static_nTM, words[i], _static_nTM->documents-1);
// If we sampled the new topic, create it first
if (assigned_topic == _static_nTM->topics) {
nTM_create_topic(_static_nTM);
}
nTM_assign_topic(_static_nTM,words[i],assigned_topic,_static_nTM->documents);
}
progressbar_inc(_static_progressbar);
if (_static_nTM->documents % _static_nTM->interval == 0) {
progressbar_finish(_static_progressbar);
nTM_save_assignments(_static_nTM);
_static_progressbar = progressbar_new("Training", _static_nTM->interval);
}
}
/**
*Example for statusbar and progressbar usage
**/
int main(void)
{
// Status bar
statusbar *status = statusbar_new("Indeterminate");
for (int i=0; i<30; i++) {
usleep(SLEEP_MS);
statusbar_inc(status);
}
statusbar_finish(status);
status = statusbar_new("Status bar with a really long label");
for (int i=0; i<10; i++) {
usleep(SLEEP_MS);
statusbar_inc(status);
}
statusbar_finish(status);
status = statusbar_new_with_format("Custom","(|)|");
for (int i=0; i<30; i++) {
usleep(SLEEP_MS);
statusbar_inc(status);
}
statusbar_finish(status);
// Progress bar
int max = 240;
progressbar *progress = progressbar_new("Smooth",max);
for(int i=0;i<max;i++) {
usleep(SLEEP_MS);
progressbar_inc(progress);
}
progressbar_finish(progress);
progress = progressbar_new("Three Second Task with a long label",3);
for(int i=0;i<3;i++) {
progressbar_inc(progress);
sleep(1);
}
progressbar_finish(progress);
progress = progressbar_new("Fast",100);
for(int i=0;i<100;i++) {
usleep(SLEEP_MS);
progressbar_inc(progress);
}
progressbar_finish(progress);
}
void nTM_train(nTM *model)
{
_static_nTM = model;
_static_progressbar = progressbar_new("Training", model->interval);
count_list_add(model->ndzs);
line_corpus_each_document(model->corpus, &nTM_each_document);
progressbar_finish(_static_progressbar);
_static_nTM = NULL;
}
void CW_train(CW *model)
{
srand(time(NULL));
static_cw_model = model;
static_progress = progressbar_new("Building Representations", model->corpus->document_count);
target_corpus_each_document(model->corpus, &CW_each_document);
progressbar_finish(static_progress);
int iteration_count = 50;
progressbar *iteration_progress = progressbar_new("Iterating", iteration_count);
for (int i=0; i<iteration_count; i++) {
CW_iteration(model);
progressbar_inc(iteration_progress);
}
progressbar_finish(iteration_progress);
static_cw_model = NULL;
CW_save_target_wordmap(model);
}
ultranest_results ultranest(LikelihoodFunc,
const char * root, const int ndim, const int max_samples, const double logZtol,
const int nlive_points, unsigned int nsteps)
{
unsigned int i = 0;
double tolerance = logZtol;
int pbar_maxval = nlive_points;
char pbar_label[200];
pbar_label[0] = 0;
ultranest_draw_state * drawer = ultranest_draw_init(Like, ndim, nsteps, 1.);
#ifdef ENABLE_PROGRESSBAR
progressbar * pbar = progressbar_new("initialising...", pbar_maxval);
pbar->format[1] = '=';
#endif
ultranest_state * sampler = ultranest_sampler_init(Like, root, ndim, nlive_points, drawer);
#ifdef ENABLE_PROGRESSBAR
progressbar_update_label(pbar, "sampling...");
progressbar_update(pbar, i);
#endif
point * current = ultranest_sampler_next(sampler);
/* begin integration */
double logVolremaining = 0;
double logwidth = log(1 - exp(-1. / sampler->nlive_points));
weighted_point * weights = NULL;
ultranest_results res;
res.root = root;
double wi = logwidth + current->L;
double logZ = wi;
double H = current->L - logZ;
double logZerr;
while(1) {
logwidth = log(1 - exp(-1. / sampler->nlive_points)) + logVolremaining;
logVolremaining -= 1. / sampler->nlive_points;
weights = (weighted_point *) realloc(weights, (i+sampler->nlive_points+1) * sizeof(weighted_point));
weights[i].p = current;
weights[i].weight = logwidth;
i = i + 1;
logZerr = sqrt(H / sampler->nlive_points);
// double i_final = -sampler->nlive_points * (-sampler->Lmax + logsubexp(fmax(tolerance - logZerr, logZerr / 100.) + logZ, logZ));
// i_final = -sampler.nlive_points * (-sampler.Lmax + log(exp(max(tolerance - logZerr, logZerr / 100.) + logZ) - exp(logZ)))
assert(fmax(tolerance - logZerr, logZerr / 100.) > 0);
assert(fmax(tolerance - logZerr, logZerr / 100.) + logZ > logZ);
int i_final = -(sampler->nlive_points * (-sampler->Lmax + logsubexp(logZ, fmax(tolerance - logZerr, logZerr / 100.))));
pbar_maxval = (int) fmin(fmax(i+1, i_final), i+100000);
ultranest_sampler_integrate_remainder(sampler, logwidth, logVolremaining, logZ, NULL);
progressbar_settext(pbar_label, i, pbar_maxval, sampler->nlive_points, sampler->ndraws,
logZ, sampler->remainderZ, logZerr, sampler->remainderZerr, current, sampler->ndim);
#ifdef ENABLE_PROGRESSBAR
progressbar_update_label(pbar, pbar_label);
pbar->max = pbar_maxval;
progressbar_update(pbar, i);
#else
printf("%s\n", pbar_label);
#endif
if (i > sampler->nlive_points) {
// tolerance
double total_error = logZerr + sampler->remainderZerr;
if (max_samples > 0 && (unsigned) max_samples < sampler->ndraws) {
#ifdef ENABLE_PROGRESSBAR
progressbar_finish(pbar);
#endif
printf("maximum number of samples reached\n");
break;
}
if (total_error < tolerance) {
#ifdef ENABLE_PROGRESSBAR
progressbar_finish(pbar);
#endif
printf("tolerance reached\n");
break;
}
// we want to make maxContribution as small as possible
// but if it becomes 10% of logZerr, that is enough
if (sampler->remainderZerr < logZerr / 10.) {
#ifdef ENABLE_PROGRESSBAR
progressbar_finish(pbar);
#endif
printf("tolerance will not improve: remainder error (%.3f) is much smaller than systematic errors (%.3f)\n", sampler->remainderZerr, logZerr);
break;
}
}
current = ultranest_sampler_next(sampler);
wi = logwidth + current->L;
double logZnew = logaddexp(logZ, wi);
H = exp(wi - logZnew) * current->L + exp(logZ - logZnew) * (H + logZ) - logZnew;
logZ = logZnew;
}
// not needed for integral, but for posterior samples, otherwise there
// is a hole in the most likely parameter ranges.
i += ultranest_sampler_integrate_remainder(sampler, logwidth, logVolremaining, logZ, weights + i);
logZerr += sampler->remainderZerr;
logZ = logaddexp(logZ, sampler->remainderZ);
res.logZ = logZ;
res.logZerr = logZerr;
res.ndraws = sampler->ndraws;
res.niter = i;
res.H = H;
res.weighted_points = weights;
printf("ULTRANEST result: lnZ = %.2f +- %.2f\n", res.logZ, res.logZerr);
write_results(root, res, ndim);
return res;
}
