void random_initialize_model(vblda_model * model, vblda_corpus* corpus, vblda_ss* ss){
int n, j, d;
double* beta = malloc(sizeof(double)*model->n);
double* nu = malloc(sizeof(double)*model->n);
for (n = 0; n < model->n; n++){
beta[n] = 0.0;
nu[n] = 0.0;
}
for (j = 0; j < model->m; j++){
for (n = 0; n < model->n; n++){
nu[n] = model->nu;
}
gsl_ran_dirichlet (r, model->n, nu, beta);
for (n = 0; n < model->n; n++){
//model->mu[j][n] = beta[n];
//model->mu[j][n] = gsl_ran_gamma(r, 100.0, 1.0/100.);
model->mu[j][n] = model->nu + 1.0 + gsl_rng_uniform(r);
}
}
free(beta);
free(nu);
}
/*
Sample-based estimate of Normalization constant
Assume the normalized distribution P(x), and a sample-generating
distribution p(x), such that:
P(x) = zp(x)
We would like to find z.
The domain of x here is the 4-simplex with volume V. Each sample
point is centered around a sub-volume, where all N sub-volumes
constitute V (a 'tesselation' if you like). One notional definition
of "fair sample points" is that one can estimate the total mass of the
density in the sub-volume as volume times density, where the estimate
of the density over that subvolume is taken to be constant and having
the value of the density at that point.
Taking total mass to be 1, each sub-volume contributes 1/N mass. And,
volume = mass/density, so by summing over sub-volumes, we have:
sum_i{(1/N) / P(x_i)} = V
sum_i{(1/N) / (zp(x_i)} = V
(1/Nz) sum_i{1/p(x_i)} = V
z = (1/NV) sum_i{1/p(x_i)}
~ (1/N) sum_i{1/p(x_i)}
*/
void populate_points(struct wpoint *points, unsigned npoints,
double *alpha_dist, double *alpha_weights,
gsl_rng *rand,
const char *name,
double *dist_norm)
{
struct wpoint *s, *se = points + npoints;
/* double ln_dist_norm = NAN; */
double dn = 0;
for (s = points; s != se; ++s)
{
gsl_ran_dirichlet(rand, NUCS, alpha_dist, s->x);
s->d = gsl_ran_dirichlet_pdf(NUCS, alpha_dist, s->x);
s->o = gsl_ran_dirichlet_pdf(NUCS, alpha_weights, s->x);
s->ln_d = gsl_ran_dirichlet_lnpdf(NUCS, alpha_dist, s->x);
s->ln_o = gsl_ran_dirichlet_lnpdf(NUCS, alpha_weights, s->x);
strcpy(s->dist, name);
s->num_samples = npoints;
dn += 1.0 / s->d;
/* ln_dist_norm = safe_sum(ln_dist_norm, -s->ln_d); */
}
/* ln_dist_norm -= gsl_sf_log(npoints); */
/* *dist_norm = gsl_sf_exp(ln_dist_norm); */
dn /= (double)npoints;
*dist_norm = dn;
}
void test_dirichlet(void){
double alpha[SIZE] = { .4, .9, .4, .2 };
double theta[SIZE] = { 0.0, 0.0, 0.0, 0.0 };
double prob;
int i;
gsl_ran_dirichlet (rng, SIZE, alpha, theta);
printf("gsl_ran_dirichlet\t");
print_double_array(alpha, SIZE);
printf("\t");
print_double_array(theta, SIZE);
printf("\n");
theta[0] = 0.107873072217;
theta[1] = 0.518033738502;
theta[2] = 0.220000000209;
theta[3] = 0.154093189072;
/* theta and alpha flipped in perl interface */
prob = gsl_ran_dirichlet_pdf (SIZE, alpha, theta);
printf("gsl_ran_dirichlet_pdf\t");
print_double_array(theta, SIZE);
printf("\t");
print_double_array(alpha, SIZE);
printf("\t%.12f\n", prob);
prob = gsl_ran_dirichlet_lnpdf (SIZE, alpha, theta);
printf("gsl_ran_dirichlet_lnpdf\t");
print_double_array(theta, SIZE);
printf("\t");
print_double_array(alpha, SIZE);
printf("\t%.12f\n", prob);
}
void rdirichlet(double alpha, int* n, int K, double* theta) {
//alpha - symmetric hyperparameter
//n - integer hyperparameter offsets (can be posterior counts)
int k;
double* a = (double*) malloc(sizeof(double)*K);
for (k=0; k < K; k++) a[k] = alpha + n[k];
gsl_ran_dirichlet (RANDOM_NUMBER, (size_t)K, a, theta);
free(a);
}
/* DIRICHLET */
CAMLprim value ml_gsl_ran_dirichlet(value rng, value alpha, value theta)
{
const size_t K = Double_array_length(alpha);
if(Double_array_length(theta) != K)
GSL_ERROR("alpha and theta must have same size", GSL_EBADLEN);
gsl_ran_dirichlet(Rng_val(rng), K, Double_array_val(alpha),
Double_array_val(theta));
return Val_unit;
}
/* linear interpolation: */
if (i >= PDFLEN - 1) {
i = PDFLEN - 1;
pdf_x = y * pdf_stdnormal[i];
}
else
pdf_x = y * (pdf_stdnormal[i] +
(z - i * X_STEP_PDF) *
(pdf_stdnormal[i + 1] - pdf_stdnormal[i]) / X_STEP_PDF);
return (pdf_x);
STOP:
return (-1.0);
# undef CUR_PROC
} /* double ighmm_rand_normal_density_approx */
double ighmm_rand_dirichlet(int seed, int len, double *alpha, double *theta){
if (seed != 0) {
GHMM_RNG_SET(RNG, seed);
}
#ifdef DO_WITH_GSL
gsl_ran_dirichlet(RNG, len, alpha, theta);
#else
printf("not implemted without gsl. Compile with gsl to use dirichlet");
#endif
}
double
test_dirichlet_small (void)
{
size_t K = 2;
double alpha[2] = { 2.5e-3, 5.0e-3};
double theta[2] = { 0.0, 0.0 };
gsl_ran_dirichlet (r_global, K, alpha, theta);
return theta[0];
}
double
test_dirichlet (void)
{
/* This is a bit of a lame test, since when K=2, the Dirichlet distribution
becomes a beta distribution */
size_t K = 2;
double alpha[2] = { 2.5, 5.0 };
double theta[2] = { 0.0, 0.0 };
gsl_ran_dirichlet (r_global, K, alpha, theta);
return theta[0];
}
void librdist_dirichlet(gsl_rng *rng, int argc, void *argv, int bufc, float *buf){
t_atom *av = (t_atom *)argv;
int i, j;
size_t k = argc;
double alpha[argc];
double theta[argc];
for(i = 0; i < k; i++){
alpha[i] = librdist_atom_getfloat(av + i);
}
for(j = 0; j < floor(bufc / k); j++){
gsl_ran_dirichlet(rng, k, alpha, theta);
for(i = 0; i < k; i++){
buf[i] = theta[i];
}
}
}
//#define SEQ_DEPTH 100//测序深度
int
create_sampling(double *alpha,double *theta,int nclass)
{
//int n;
//n=SAMPLE_CNT;//dirchlet sampling count
/* set up GSL RNG */
const gsl_rng_type *T;
gsl_rng *r ;
gsl_rng_env_setup();
T=gsl_rng_default;
r=gsl_rng_alloc(T);
/* end of GSL setup */
gsl_ran_dirichlet(r,nclass,alpha,theta);
gsl_rng_free(r);
return(0);
}
double mc_integrate(gsl_matrix *saxs_pre, gsl_vector *saxs_exp,
gsl_vector *err_saxs, int k, int N) {
double energy_final;
double *alphas;
double *samples;
double *alpha_ens;
size_t Ntrials = 10000;
alphas = (double * ) malloc( k * sizeof( double ));
samples = (double * ) malloc( k * sizeof( double ));
alpha_ens = (double * ) malloc( k * sizeof( double ));
const gsl_rng_type *T;
gsl_rng *r;
gsl_vector *weights = gsl_vector_alloc(k);
gsl_vector *saxs_ens = gsl_vector_alloc(N);
for (int i = 0; i<k; i++) alphas[i] = 0.5;
double energy_trial=0.0;
double saxs_scale = 0.0;
double alpha_zero;
for (int i=0; i<Ntrials; i++) {
gsl_ran_dirichlet(r, k, alphas, samples);
alpha_zero = 0.0;
for (int j = 0; j < k; j++) {
alpha_zero += samples[j];
}
for( int j = 0; j< N; j++) {
alpha_ens[j] = 0.0;
for (int l = 0; l < k; l++) {
alpha_ens[j]+=gsl_matrix_get(saxs_pre,j,l)*samples[l];
}
gsl_vector_set(weights, j, alpha_ens[j]/alpha_zero);
}
saxs_scale = SaxsScaleMean(weights,saxs_exp,err_saxs,N);
gsl_blas_dgemv(CblasNoTrans, 1.0, saxs_pre, weights, 0.0, saxs_ens);
energy_trial+=ModelEvidenceEnergy(saxs_ens,saxs_exp,err_saxs,saxs_scale,N);
}
energy_final/=Ntrials;
gsl_rng_free (r);
return energy_final;
}
void
test_dirichlet_moments (void)
{
double alpha[DIRICHLET_K];
double theta[DIRICHLET_K];
double theta_sum[DIRICHLET_K];
double alpha_sum = 0.0;
double mean, obs_mean, sd, sigma;
int status, k, n;
for (k = 0; k < DIRICHLET_K; k++)
{
alpha[k] = gsl_ran_exponential (r_global, 0.1);
alpha_sum += alpha[k];
theta_sum[k] = 0.0;
}
for (n = 0; n < N; n++)
{
gsl_ran_dirichlet (r_global, DIRICHLET_K, alpha, theta);
for (k = 0; k < DIRICHLET_K; k++)
theta_sum[k] += theta[k];
}
for (k = 0; k < DIRICHLET_K; k++)
{
mean = alpha[k] / alpha_sum;
sd =
sqrt ((alpha[k] * (1. - alpha[k] / alpha_sum)) /
(alpha_sum * (alpha_sum + 1.)));
obs_mean = theta_sum[k] / N;
sigma = sqrt ((double) N) * fabs (mean - obs_mean) / sd;
status = (sigma > 3.0);
gsl_test (status,
"test gsl_ran_dirichlet: mean (%g observed vs %g expected)",
obs_mean, mean);
}
}
// Sample topic parameters
// Function written from perspective of sampling user topics parameters
// Switch roles of user-item inputs to sample item topics parameters
void sampleTopicParams(const mxArray* exampsByUser, int KU, int numUsers,
double alpha, double* logthetaU, uint32_t* zU){
// Array of static random number generators
gsl_rng** rngs = getRngArray();
// Prior term for Dirichlet
const double ratio = alpha/KU;
// Allocate memory for storing topic counts
double* counts[MAX_NUM_THREADS];
for(int thread = 0; thread < MAX_NUM_THREADS; thread++)
counts[thread] = mxMalloc(KU*sizeof(**counts));
#pragma omp parallel for
for(int u = 0; u < numUsers; u++){
int thread = omp_get_thread_num();
// Initialize to prior term
for(int i = 0; i < KU; i++)
counts[thread][i] = ratio;
// Iterate over user's examples computing sufficient stats
mxArray* exampsArray = mxGetCell(exampsByUser, u);
mwSize len = mxGetN(exampsArray);
uint32_t* examps = (uint32_t*) mxGetData(exampsArray);
for(int j = 0; j < len; j++)
counts[thread][zU[examps[j]-1]-1]++;
// Sample new topic parameters
double* logthetaPtr = logthetaU + u*KU;
gsl_ran_dirichlet(rngs[omp_get_thread_num()], KU, counts[thread],
logthetaPtr);
// Take logs
for(int i = 0; i < KU; i++)
logthetaPtr[i] = log(logthetaPtr[i]);
}
// Clean up
for(int thread = 0; thread < MAX_NUM_THREADS; thread++)
mxFree(counts[thread]);
}
struct pair_comp gen_pair_comp(double *alpha, double dist, gsl_rng *rg)
{
struct pair_comp pc;
double s1[3], min_v, max_dist = -1.0;
unsigned i, min_i;
double alpha_low[] = { 0.1, 0.1, 0.1, 0.1 };
if (dist == 1)
{
unsigned long c1 = gsl_rng_uniform_int(rg, 4);
unsigned long c2 = (c1 + 1 + gsl_rng_uniform_int(rg, 3)) % 4;
memcpy(pc.c1, bary_corners[c1], sizeof(pc.c1));
memcpy(pc.c2, bary_corners[c2], sizeof(pc.c2));
}
else if (dist == 0)
{
gsl_ran_dirichlet(rg, 4, (dist > 0.9 ? alpha_low : alpha), pc.c1);
memcpy(pc.c2, pc.c1, sizeof(pc.c2));
}
else if (dist > 0.95)
{
fprintf(stderr, "Cannot handle distances not equal to 1 but > 0.95\n");
exit(1);
}
else
{
/* generate a valid c1 */
double s_corner[3];
while (max_dist < dist * (1.0 / 0.95))
{
gsl_ran_dirichlet(rg, 4, (dist > 0.9 ? alpha_low : alpha), pc.c1);
min_v = pc.c1[0], min_i = 0;
for (i = 1; i != 4; ++i)
if (min_v > pc.c1[i])
min_v = pc.c1[i], min_i = i;
barycentric_to_simplex(pc.c1, s1);
barycentric_to_simplex(bary_corners[min_i], s_corner);
max_dist = dist3(s_corner, s1);
}
/* sample from the unit sphere until we find a point in the
simplex */
double sph[3], s1p[3];
while (1)
{
unit_sphere3(sph, rg);
for (i = 0; i != 3; ++i) s1p[i] = s1[i] + dist * sph[i];
if (inside_simplex(s1p)) break;
}
simplex_to_barycentric(s1p, pc.c2);
}
/* fprintf(stderr, */
/* "%5.3f,%5.3f,%5.3f,%5.3f\t" */
/* "%5.3f,%5.3f,%5.3f,%5.3f\t" */
/* "%f\t%f\n", */
/* pc.c1[0], pc.c1[1], pc.c1[2], pc.c1[3], */
/* pc.c2[0], pc.c2[1], pc.c2[2], pc.c2[3], */
/* dist, */
/* dist4_scaled(pc.c1, pc.c2) */
/* ); */
return pc;
}
int main()
{
gsl_rng * r;
const gsl_rng_type * T;
gsl_rng_env_setup();
T = gsl_rng_default;
r = gsl_rng_alloc (T);
// printf ("generator type: %s\n", gsl_rng_name (r));
// printf ("seed = %lu\n", gsl_rng_default_seed);
// printf ("first value = %lu\n", gsl_rng_get (r));
// Step 1
int k = 3; // k is the # of communities
const double alpha = double (1.0) / k;
// printf ("k = %d and alpha = %f\n", k, alpha);
double eta = 1; // eta is the hidden variable of the Beta distribution
double comm_str [k];
for (uint32_t i = 0; i < k; ++i)
comm_str[i] = gsl_ran_beta(r, eta, eta);
// Step 2
int n = 10; // Number of nodes
double theta = 1; // Used for gsl_ran_dirichlet
double alpha_array [k]; // Array for gsl_ran_dirichlet
for (uint32_t i = 0; i < k; ++i)
alpha_array[i] = alpha;
double pi [n][k]; // Array to store pi
// Populate pi
for (uint32_t i = 0; i < n; ++i)
gsl_ran_dirichlet(r, k, alpha_array, pi[i]);
// // Convert to binary
// for (uint32_t i = 0; i < n; ++i){
// for (uint32_t j = 0; j < k ; ++j){
// if (pi[i][j] > 0.5){
// pi[i][j] = 1;
// }
// else{
// pi[i][j] = 0;
// }
// }
// }
// for (uint32_t i = 0; i < n; ++i){
// for (uint32_t j = 0; j < k ; ++j){
// printf("%d, %d, %f\n", i, j ,pi[i][j]);
// }
// }
// Step 4
double mean = 0.0; // Set mean
double var = 1; // Set var
double prob = 0.5; // Set probability for Bernoulli
// Save input values
std::ofstream inputs ("inputs.txt");
if (inputs.is_open()){
inputs << "k = " << k << "\n";
inputs << "alpha = " << alpha << "\n";
inputs << "eta = " << eta << "\n";
inputs << "mean = " << mean << "\n";
inputs << "var = " << var << "\n";
inputs << "prob = " << prob << "\n";
inputs << "comm_str = ";
for (uint32_t i = 0; i < k; ++i)
inputs << comm_str[i] << " ";
inputs << "\n" << "n = " << n << "\n" << "pi = " << "\n";
for (uint32_t i = 0; i < n; ++i){
for (uint32_t j = 0; j < k ; ++j){
inputs << pi[i][j] << "\t";
}
inputs << "\n";
}
inputs.close();
}
// Save attribute matrix
std::ofstream attributes ("attributes.txt");
if (attributes.is_open()){
for (uint32_t i = 0; i < n; ++i){
attributes << gsl_ran_gaussian(r, var) << "\t";
attributes << gsl_ran_gaussian(r, var) << "\t";
attributes << gsl_ran_bernoulli(r, prob) << "\t";
attributes << gsl_ran_bernoulli(r, prob) << "\t";
attributes << "\n";
}
attributes.close();
}
// Step 3
int num_edge = 20; // Define number of edges
double epsilon = 1e-30;
int adj_matrix [n][n];
// Populate adjancency matrix with 0s
for (uint32_t i = 0; i < n; ++i)
for (uint32_t j = 0; j < n ; ++j)
adj_matrix[i][j] = 0;
int count = 0;
while (count < num_edge){
// printf("count %d\n", count);
int a = gsl_rng_uniform_int(r, n);
int b = gsl_rng_uniform_int(r, n);
// printf("%d, %d\n", a, b);
if (a == b){
continue;
}
double a_probs [k];
double b_probs [k];
for (uint32_t i = 0; i < k; ++i){
a_probs[i] = pi[a][i];
b_probs[i] = pi[b][i];
}
int a_val = gsl_ran_discrete(r, gsl_ran_discrete_preproc(k, a_probs));
int b_val = gsl_ran_discrete(r, gsl_ran_discrete_preproc(k, b_probs));
if (a_val == b_val){
// printf("%d, %d\n", a_val, b_val);
int x = gsl_ran_bernoulli(r, comm_str[a_val]);
if (x == 1){
// printf("x = 1\n");
if (adj_matrix[a][b] == 0){
adj_matrix[a][b] = 1;
++count;
}
}
}
else{
int x = gsl_ran_bernoulli(r, epsilon);
if (x == 1){
// printf("x = 1\n");
if (adj_matrix[a][b] == 0){
adj_matrix[a][b] = 1;
++count;
}
}
}
// for (uint32_t j = 0; j < k ; ++j){
// if (std::pi[a][j] == pi[b][j]){
// int x = gsl_ran_bernoulli(r, comm_str[j]);
// // printf("Match 1 %d\n", x);
// if (x == 1){
// if (adj_matrix[a][b])
// continue;
// else
// adj_matrix[a][b] = 1;
// run_eps = 0;
// ++count;
// continue;
// }
// }
// if (run_eps == 1){
// int x = gsl_ran_bernoulli(r, epsilon);
// // printf("0 %d\n", x);
// if (x == 1){
// if (adj_matrix[a][b])
// continue;
// else
// adj_matrix[a][b] = 1;
// ++count;
// continue;
// }
// }
// }
}
std::ofstream matrix ("matrix.txt");
if (matrix.is_open()){
for (uint32_t i = 0; i < n; ++i){
for (uint32_t j = 0; j < n ; ++j){
if (adj_matrix [i][j])
matrix << i << "\t" << j << '\n';
}
}
matrix.close();
}
}
