static int
do_search(struct saucy *s)
{
int min;
unprepare_permutation(s);
/* Backtrack to the ancestor with zeta */
if (s->lev > s->anc) s->lev = s->anc + 1;
/* Perform additional backtracking */
min = backtrack(s);
/* Keep going while there are tree nodes to expand */
while (s->lev) {
/* Descend to a new leaf node */
if (descend(s, &s->right, s->start[s->lev], min)
&& descend_left(s)) {
/* Prepare permutation */
prepare_permutation(s);
/* If we found an automorphism, return it */
if (s->is_automorphism(s)) {
++s->stats->gens;
s->stats->support += s->ndiffs;
update_theta(s);
return s->consumer(s->n, s->gamma,
s->ndiffs, s->unsupp, s->arg);
}
else {
unprepare_permutation(s);
}
}
/* If we get here, something went wrong; backtrack */
++s->stats->bads;
min = backtrack_bad(s);
}
/* Normalize group size */
while (s->stats->grpsize_base >= 10.0) {
s->stats->grpsize_base /= 10;
++s->stats->grpsize_exp;
}
return 0;
}
void SPF::learn() {
double old_likelihood, delta_likelihood, likelihood = -1e10;
int likelihood_decreasing_count = 0;
time_t start_time, end_time;
int iteration = 0;
char iter_as_str[4];
bool converged = false;
bool on_final_pass = false;
while (!converged) {
time(&start_time);
iteration++;
printf("iteration %d\n", iteration);
reset_helper_params();
// update rate for user preferences
b_theta.each_col() += sum(beta, 1);
set<int> items;
int user = -1, item, rating;
for (int i = 0; i < settings->sample_size; i++) {
if (on_final_pass && settings->final_pass_test) {
user++;
while (data->test_users.count(user)==0) {
user++;
}
} else if (settings->svi) {
user = gsl_rng_uniform_int(rand_gen, data->user_count());
} else {
user = i;
}
bool user_converged = false;
int user_iters = 0;
while (!user_converged) {
user_iters++;
a_beta_user.zeros();
a_delta_user.zeros();
// look at all the user's items
for (int j = 0; j < data->item_count(user); j++) {
item = data->get_item(user, j);
items.insert(item);
rating = 1;
//TODO: rating = data->get_train_rating(i);
update_shape(user, item, rating);
}
// update per-user parameters
double user_change = 0;
if (!settings->factor_only && !settings->fix_influence)
user_change += update_tau(user);
if (!settings->social_only)
user_change += update_theta(user);
if (!settings->social_only && !settings->factor_only && !settings->fix_influence) {
user_change /= 2;
// if the updates are less than 1% change, the local params have converged
if (user_change < 0.01)
user_converged = true;
} else {
// if we're only looking at social or factor (not combined)
// then the user parameters will always have converged with
// a single pass (since there's nothing to balance against)
user_converged = true;
}
}
if (settings->verbose)
printf("%d\tuser %d took %d iters to converge\n", iteration, user, user_iters);
a_beta += a_beta_user;
a_delta += a_delta_user;
}
if (!settings->social_only) {
// update rate for item attributes
b_beta.each_col() += sum(theta, 1);
// update per-item parameters
set<int>::iterator it;
for (it = items.begin(); it != items.end(); it++) {
item = *it;
if (iter_count[item] == 0)
iter_count[item] = 0;
iter_count[item]++;
update_beta(item);
if (settings->item_bias)
update_delta(item);
}
} else if (settings->item_bias) {
set<int>::iterator it;
for (it = items.begin(); it != items.end(); it++) {
item = *it;
if (iter_count[item] == 0)
iter_count[item] = 0;
iter_count[item]++;
if (settings->item_bias)
update_delta(item);
}
}
// check for convergence
if (on_final_pass) {
printf("Final pass complete\n");
converged = true;
old_likelihood = likelihood;
likelihood = get_ave_log_likelihood();
delta_likelihood = abs((old_likelihood - likelihood) /
old_likelihood);
log_convergence(iteration, likelihood, delta_likelihood);
} else if (iteration >= settings->max_iter) {
printf("Reached maximum number of iterations.\n");
converged = true;
old_likelihood = likelihood;
likelihood = get_ave_log_likelihood();
delta_likelihood = abs((old_likelihood - likelihood) /
old_likelihood);
log_convergence(iteration, likelihood, delta_likelihood);
} else if (iteration % settings->conv_freq == 0) {
old_likelihood = likelihood;
likelihood = get_ave_log_likelihood();
if (likelihood < old_likelihood)
likelihood_decreasing_count += 1;
else
likelihood_decreasing_count = 0;
delta_likelihood = abs((old_likelihood - likelihood) /
old_likelihood);
log_convergence(iteration, likelihood, delta_likelihood);
if (settings->verbose) {
printf("delta: %f\n", delta_likelihood);
printf("old: %f\n", old_likelihood);
printf("new: %f\n", likelihood);
}
if (iteration >= settings->min_iter &&
delta_likelihood < settings->likelihood_delta) {
printf("Model converged.\n");
converged = true;
} else if (iteration >= settings->min_iter &&
likelihood_decreasing_count >= 2) {
printf("Likelihood decreasing.\n");
converged = true;
}
}
// save intermediate results
if (!converged && settings->save_freq > 0 &&
iteration % settings->save_freq == 0) {
printf(" saving\n");
sprintf(iter_as_str, "%04d", iteration);
save_parameters(iter_as_str);
}
// intermediate evaluation
if (!converged && settings->eval_freq > 0 &&
iteration % settings->eval_freq == 0) {
sprintf(iter_as_str, "%04d", iteration);
evaluate(iter_as_str);
}
time(&end_time);
log_time(iteration, difftime(end_time, start_time));
if (converged && !on_final_pass &&
(settings->final_pass || settings->final_pass_test)) {
printf("final pass on all users.\n");
on_final_pass = true;
converged = false;
// we need to modify some settings for the final pass
// things should look exactly like batch for all users
if (settings->final_pass) {
settings->set_stochastic_inference(false);
settings->set_sample_size(data->user_count());
scale = 1;
} else {
settings->set_sample_size(data->test_users.size());
scale = data->user_count() / settings->sample_size;
}
}
}
save_parameters("final");
}
int *
saucy_search(void)
{
int i;
/* Do initialization steps if necessary */
if (!init) {
init = 1;
/* Initialize stats */
stats->grpsize_base = 1.0;
stats->grpsize_exp = stats->nodes = stats->bads = stats->gens = 0;
/* Preprocessing after initial coloring */
refine();
}
/* Keep going while there are tree nodes to expand */
while (lev) {
if (desc) { desc = 0; descend(); }
/* Find a target cell */
++stats->nodes; target_cell();
if (target != n+1) { descend_left(); continue; }
/* We don't have a target, so we're discrete */
if (!have_zeta) {
have_zeta = 1;
for (i = 0; i < n; ++i) zeta[lab[i]] = i;
if (canon != NULL) memcpy(canon, lab, n * sizeof(int));
}
else {
/* Prepare permutation and check */
for (i = 0; i < n; ++i) gamma[i] = lab[zeta[i]];
is_automorphism();
/* Do some stuff if we actually got an automorphism */
if (flag) {
++stats->gens;
update_theta();
}
/* Keep track of leaf nodes that are not automorphisms */
else {
++stats->bads;
}
}
/* Since we're discrete, we need to backtrack */
backtrack(stats); desc = 1;
if (flag) return gamma;
}
/* Normalize group size */
while (stats->grpsize_base >= 10.0) {
stats->grpsize_base /= 10;
++stats->grpsize_exp;
}
/* All done */
return NULL;
}
