bool rjMcMC1DSampler::propose_move(rjMcMC1DModel& mcur, rjMcMC1DModel& mpro)
{
np_move++;
size_t n = mcur.nlayers();
if (n <= 1)return false;
size_t index = irand((size_t)1, n - 1);
double pold = mcur.layers[index].ptop;
//double std = sd_move;
double std = DEFAULTMOVESTDFRACTION*pold;
double pnew = pold + std*nrand();
double qpdfforward = gaussian_pdf(pold, pold*DEFAULTMOVESTDFRACTION, pnew);
double qpdfreverse = gaussian_pdf(pnew, pnew*DEFAULTMOVESTDFRACTION, pold);
bool isvalid = mpro.move_interface(index, pnew);
if (isvalid == false)return false;
set_misfit(mpro);
double pqratio = qpdfreverse / qpdfforward;
double logpqratio = log(pqratio);
double loglr = -(mpro.misfit() - mcur.misfit()) / 2.0;
double logar = logpqratio + loglr;
double logu = log(urand());
if (logu < logar){
na_move++;
return true;
}
return false;
}
bool rjMcMC1DSampler::propose_birth(rjMcMC1DModel& mcur, rjMcMC1DModel& mpro)
{
np_birth++;
size_t n = mcur.nlayers();
if (n >= nl_max)return false;
double pos = urand(0.0, pmax);
size_t index = mcur.which_layer(pos);
double vold = mcur.layers[index].value;
double vnew, pqratio;
if (mBirthDeathFromPrior){
vnew = urand(vmin, vmax);
pqratio = 1.0;
}
else{
double vcpdf;
double logstd = DEFAULTLOGSTDDECADES;
if (param_value == LINEAR){
double m = (pow(10.0, logstd) - pow(10.0, -logstd)) / 2.0;
vnew = vold + m*vold*nrand();
vcpdf = gaussian_pdf(vold, m*vold, vnew);
}
else{
vnew = vold + logstd*nrand();
vcpdf = gaussian_pdf(vold, logstd, vnew);
}
pqratio = 1.0 / ((vmax - vmin)*vcpdf);
}
//pqratio *= (double)(n) / double(n + 1);
bool isvalid = mpro.insert_interface(pos, vnew);
if (isvalid == false)return false;
set_misfit(mpro);
double logpqratio = log(pqratio);
double loglikeratio = -(mpro.misfit() - mcur.misfit()) / 2.0;
double logar = logpqratio + loglikeratio;
if (log(urand()) < logar){
na_birth++;
return true;
}
return false;
}
/**
* for all data compute p(x|i) prob that state i generated data point x
* correspond to the E step of EM.
*
* @param *pix is an alloc'd float table of dimension nstates*data_len
* returns total log_likelihood
*/
_fgmm_real fgmm_e_step(struct gmm * GMM,
const _fgmm_real * data,
int data_len,
_fgmm_real * pix)
{
/* E step */
_fgmm_real log_lik=0;
_fgmm_real like;
_fgmm_real * pxi;
int data_i=0;
int state_i;
pxi = (_fgmm_real *) malloc(sizeof(_fgmm_real) * GMM->nstates);
for(;data_i<data_len;data_i++)
{
like=0;
for(state_i=0;state_i<GMM->nstates;state_i++)
{
pxi[state_i] = gaussian_pdf(&GMM->gauss[state_i], data + data_i*GMM->dim) ;
//printf("state %d -> lik : %f\n",state_i,pxi[state_i]);
like += pxi[state_i]* GMM->gauss[state_i].prior;
/* pdata++;
ppxi++; */
}
if(like<= FLT_MIN)
{
/* printf("too far from current distrib %d\n",data_i); */
// in order to avoid us some numerical instabilities,
// we are gooing to boost the likelihood for this point
//
// yep ..
for(state_i = 0;state_i<GMM->nstates;state_i++)
pxi[state_i] += FLT_MIN;
like = FLT_MIN * GMM->nstates;
}
log_lik += log(like);
/* if(isnan(log_lik) || isinf(log_lik))
exit(0); */
for(state_i=0;state_i<GMM->nstates;state_i++)
{
pix[data_i + state_i*data_len] = pxi[state_i] * GMM->gauss[state_i].prior / like;
if(pix[data_i + state_i*data_len] <= FLT_MIN)
pix[data_i + state_i*data_len] = FLT_MIN;
}
}
free(pxi);
return log_lik;
}
bool rjMcMC1DSampler::propose_death(rjMcMC1DModel& mcur, rjMcMC1DModel& mpro)
{
np_death++;
size_t n = mcur.nlayers();
if (n <= nl_min)return false;
size_t index = irand((size_t)1, n - 1);
bool isvalid = mpro.delete_interface(index);
if (isvalid == false)return false;
set_misfit(mpro);
double pqratio;
if (mBirthDeathFromPrior){
pqratio = 1.0;
}
else{
double logstd = DEFAULTLOGSTDDECADES;
double vnew = mcur.layers[index - 1].value;
double vold = mcur.layers[index].value;
double vcpdf;
if (param_value == LINEAR){
double m = (pow(10.0, logstd) - pow(10.0, -logstd)) / 2.0;
vcpdf = gaussian_pdf(vnew, m*vnew, vold);
}
else{
vcpdf = gaussian_pdf(vnew, logstd, vold);
}
pqratio = (vmax - vmin)*vcpdf;
}
//pqratio *= (double)(n) / double(n - 1);
double logpqratio = log(pqratio);
double loglikeratio = -(mpro.misfit() - mcur.misfit()) / 2.0;
double logar = logpqratio + loglikeratio;
if (log(urand()) < logar){
na_death++;
return true;
}
return false;
}
/**
* for all data compute p(x|i) prob that state i generated data point x
* correspond to the E step of EM.
*
* @param *pix is an alloc'd float table of dimension nstates*data_len
* returns total log_likelihood
*/
_fgmm_real fgmm_e_step(struct gmm * GMM,
const _fgmm_real * data,
int data_len,
_fgmm_real * pix)
{
/* E step */
_fgmm_real log_lik=0;
_fgmm_real like;
_fgmm_real * pxi;
int data_i=0;
int state_i;
pxi = (_fgmm_real *) malloc(sizeof(_fgmm_real) * GMM->nstates);
for(;data_i<data_len;data_i++)
{
like=0;
for(state_i=0;state_i<GMM->nstates;state_i++)
{
pxi[state_i] = gaussian_pdf(&GMM->gauss[state_i], data + data_i*GMM->dim) ;
//printf("state %d -> lik : %f\n",state_i,pxi[state_i]);
like += pxi[state_i]* GMM->gauss[state_i].prior;
/*
pdata++;
ppxi++;
*/
}
if(like<= FLT_MIN)
{
//printf("too far from current distrib %d\n",data_i);
// exit(0);
}
else log_lik += log(like);
/*
if(isnan(log_lik) || isinf(log_lik))
exit(0);
*/
for(state_i=0;state_i<GMM->nstates;state_i++)
{
pix[data_i + state_i*data_len] = pxi[state_i] * GMM->gauss[state_i].prior / like;
if(pix[data_i + state_i*data_len] <= FLT_MIN)
pix[data_i + state_i*data_len] = FLT_MIN;
}
}
free(pxi);
return log_lik;
}
bool rjMcMC1DSampler::propose_valuechange(rjMcMC1DModel& mcur, rjMcMC1DModel& mpro)
{
np_valuechange++;
size_t index = irand((size_t)0,mcur.nlayers() - 1);
double logstd = DEFAULTLOGSTDDECADES;
double vold = mcur.layers[index].value;
double vnew;
double pqratio;
if (param_value == LINEAR){
double m = (pow(10.0, logstd) - pow(10.0, -logstd)) / 2.0;
vnew = vold + m*vold*nrand();
double qpdfforward = gaussian_pdf(vold, m*vold, vnew);
double qpdfreverse = gaussian_pdf(vnew, m*vnew, vold);
pqratio = qpdfreverse / qpdfforward;
}
else{
vnew = vold + logstd*nrand();
pqratio = 1.0;
}
bool isvalid = isinbounds(vmin, vmax, vnew);
if (isvalid == false)return false;
mpro.layers[index].value = vnew;
set_misfit(mpro);
double logpqratio = log(pqratio);
double logliker = -(mpro.misfit() - mcur.misfit()) / 2.0;
double logar = logpqratio + logliker;
if (log(urand()) < logar){
na_valuechange++;
return true;
}
return false;
}
int main(int argc,char ** argv)
{
int i;
float vn=0;
int sig_perc=0, sig3_perc=0;
struct gaussian g;
float v[] = {1.,1.,1.};
float pdf=0;
_fgmm_real samp[3];
_fgmm_real * data;
_fgmm_real * weights;
int k=0;
struct smat * cv=NULL;
_fgmm_real mean[3];
_fgmm_real * pcv;
srand(time(NULL));
printf("gaussian test suite, each test uses %d samples \n",randn_samples);
printf("box_muller testing :\n");
for(i=0;i<randn_samples;i++)
{
vn = randn_boxmuller();
if( fabs(vn) < 1.)
sig_perc++;
else if( fabs(vn) >= 2.)
sig3_perc++;
}
//printf("%d %d\n",sig_perc,sig3_perc);
assert( abs(sig_perc - one_sigma*randn_samples) < randn_samples / 400);
assert( abs(sig3_perc - three_sigma*randn_samples) < randn_samples / 400);
printf("..pass \n");
/* ----------------------------------------------- */
printf("simple gaussian pdf test :\n");
gaussian_init(&g,3);
invert_covar(&g);
assert(g.dim == 3);
pdf = gaussian_pdf(&g,v);
assert(fabs(pdf - value) < 1e-5);
printf("..pass \n");
/* ----------------------------------------------- */
printf("drawing sample from gaussian test :\n");
data = (_fgmm_real *) malloc( sizeof(_fgmm_real) * randn_samples * 3);
weights = (_fgmm_real *) malloc(sizeof(_fgmm_real) * randn_samples);
for(i=0;i<randn_samples;i++)
{
weights[i] = 1.;
gaussian_draw(&g,samp);
//printf("%f %f %f\n",samp[0],samp[1],samp[2]);
for(k=0;k<3;k++)
{
data[3*i+k] = samp[k];
}
}
smat_zero(&cv,3);
smat_covariance(cv,randn_samples,weights,data,mean);
// smat_pmatrix(cv);
/* checking covariance is identity */
pcv = cv->_;
for(i=0;i<3;i++)
{
assert(fabs(mean[i]) < 1e-2);
assert( fabs(*(pcv++)- 1.) < 2e-2);
for(k=i+1;k<3;k++)
assert(fabs(*pcv++) < 1e-2);
}
printf("..pass\n");
return EXIT_SUCCESS;
}
