static void fullfill_param(Lda *lda) {
for (int i = 0; i < lda->t; i++) {
int uid = lda->tokens[i][0];
int vid = lda->tokens[i][1];
int top = lda->tokens[i][2];
lda->nd[uid * lda->p.k + top] += 1;
lda->nw[vid * lda->p.k + top] += 1;
lda->nkw[top] += 1;
}
save_lda(lda, 0);
}
void est_lda(Lda *lda) {
fullfill_param(lda);
int *p = (int *) malloc(sizeof(int) * lda->t);
int st = 0;
double *prob = (double *) malloc(sizeof(double) * lda->p.k);
double vb = lda->p.b * lda->v;
for (int i = 0; i < lda->t; i++) p[i] = i;
for (int i = 1; i <= lda->p.niters; i++) {
fprintf(stderr, "iteration %d estimate begin ... ", i);
shuffle(p, lda->t);
for (int j = 0; j < lda->t; j++) {
int id = p[j];
int uid = lda->tokens[id][0];
int vid = lda->tokens[id][1];
int top = lda->tokens[id][2];
lda->nd[uid * lda->p.k + top] -= 1;
lda->nw[vid * lda->p.k + top] -= 1;
lda->nkw[top] -= 1;
for (int l = 0; l < lda->p.k; l++) {
prob[l] = 1.0 * (lda->nd[uid * lda->p.k + l] + lda->p.a) *
(lda->nw[vid * lda->p.k + l] + lda->p.b) /
(lda->nkw[l] + vb);
if (l > 0) prob[l] += prob[l - 1];
}
double rnd = prob[lda->p.k - 1] * (0.1 + rand()) / (0.1 + RAND_MAX);
for (st = 0; st < lda->p.k; st++) {
if (prob[st] > rnd) break;
}
lda->nd[uid * lda->p.k + st] += 1;
lda->nw[vid * lda->p.k + st] += 1;
lda->nkw[st] += 1;
lda->tokens[id][2] = st;
}
fprintf(stderr, " done\n");
if (i % lda->p.savestep == 0) {
save_lda(lda, i);
}
}
free(p); p = NULL;
free(prob); prob = NULL;
}
void est_lda(Lda *lda) {
// first full fill theta & phi matrix
// and link the nonzero elements
fullfill_param(lda);
// est iteration for lda
for (int n = 1; n <= lda->p.niters; n++) {
long sec1 = time(NULL);
gibbs_sample(lda);
long sec2 = time(NULL);
fprintf(stderr, "iter %d done, using %ld seconds\n", n, sec2 - sec1);
if (n % lda->p.savestep == 0) {
save_lda(lda, n);
}
}
}
static void fullfill_param(Lda *lda) {
for (int i = 0; i < lda->t; i++) {
int uid = lda->tokens[i][0];
int vid = lda->tokens[i][1];
int tid = lda->tokens[i][2];
lda->nd[uid * (lda->p.k + 1) + tid].count += 1;
lda->nw[vid * (lda->p.k + 1) + tid].count += 1;
lda->nkw[tid] += 1;
}
for (int d = 0; d < lda->d; d++) {
int offs = d * (lda->p.k + 1);
int p = 0;
for (int k = 1; k <= lda->p.k; k++) {
if (lda->nd[offs + k].count > 0) {
lda->nd[offs + p].next = k;
lda->nd[offs + k].prev = p;
p = k;
}
}
lda->nd[offs + p].next = 0;
lda->nd[offs].prev = p;
}
for (int v = 0; v < lda->v; v++) {
int offs = v * (lda->p.k + 1);
int p = 0;
for (int k = 1; k <= lda->p.k; k++) {
if (lda->nw[offs + k].count > 0) {
lda->nw[offs + p].next = k;
lda->nw[offs + k].prev = p;
p = k;
}
}
lda->nw[offs + p].next = 0;
lda->nw[offs].prev = p;
}
save_lda(lda, 0);
}