unsigned int som::train(bool autostop) {
previous_weights = NULL;
for (unsigned int epouch = 1; epouch < (epouchs + 1); epouch++) {
/* Depression term of coupling */
local_radius = std::pow( ( params.init_radius * std::exp(-( (double) epouch / (double) epouchs)) ), 2);
learn_rate = params.init_learn_rate * std::exp(-( (double) epouch / (double) epouchs));
/* Feature SOM 0003: Clear statistics */
if (autostop == true) {
for (unsigned int i = 0; i < size; i++) {
(*awards)[i] = 0;
(*capture_objects)[i]->clear();
}
}
for (unsigned int i = 0; i < data->size(); i++) {
/* Step 1: Competition */
unsigned int index_winner = competition(&(*data)[i]);
/* Step 2: Adaptation */
adaptation(index_winner, &(*data)[i]);
/* Update statistics */
if ( (autostop == true) || (epouch == (epouchs - 1)) ) {
(*awards)[index_winner]++;
(*capture_objects)[index_winner]->push_back(i);
}
}
/* Feature SOM 0003: Check requirement of stopping */
if (autostop == true) {
if (previous_weights == NULL) {
previous_weights = new std::vector<std::vector<double> * >(size, NULL);
unsigned int dimensions = (*weights)[0]->size();
for (unsigned int i = 0; i < weights->size(); i++) {
(*previous_weights)[i] = new std::vector<double>(dimensions, 0.0);
std::copy((*weights)[i]->begin(), (*weights)[i]->end(), (*previous_weights)[i]->begin());
}
}
else {
double maximal_adaptation = calculate_maximal_adaptation();
if (maximal_adaptation < params.adaptation_threshold) {
return epouch;
}
for (unsigned int i = 0; i < weights->size(); i++) {
std::copy((*weights)[i]->begin(), (*weights)[i]->end(), (*previous_weights)[i]->begin());
}
}
}
}
return epouchs;
}
static void adaptation_of_medias(axl_network *mysys)
{
int i, j;
axl_mass_media *mm;
for(i = 0; i < mysys->nmm; i++)
{
for(j = 0; j < mysys->mass_media[i].f; j++)
{
mm = &(mysys->mass_media[i]);
adaptation(mm, *mysys, j);
}
}
return;
}
void function_minimizer::shmc_mcmc_routine(int nmcmc,int iseed0,double dscale,
int restart_flag) {
if (nmcmc<=0)
{
cerr << endl << "Error: Negative iterations for MCMC not meaningful" << endl;
ad_exit(1);
}
uostream * pofs_psave=NULL;
if (mcmc2_flag==1)
{
initial_params::restore_start_phase();
}
initial_params::set_inactive_random_effects();
initial_params::set_active_random_effects();
int nvar_re=initial_params::nvarcalc();
int nvar=initial_params::nvarcalc(); // get the number of active parameters
if (mcmc2_flag==0)
{
initial_params::set_inactive_random_effects();
nvar=initial_params::nvarcalc(); // get the number of active parameters
}
initial_params::restore_start_phase();
independent_variables parsave(1,nvar_re);
// dvector x(1,nvar);
// initial_params::xinit(x);
// dvector pen_vector(1,nvar);
// {
// initial_params::reset(dvar_vector(x),pen_vector);
// }
initial_params::mc_phase=1;
int old_Hybrid_bounded_flag=-1;
int on,nopt = 0;
//// ------------------------------ Parse input options
// Step size. If not specified, will be adapted. If specified must be >0
// and will not be adapted.
double eps=0.1;
double _eps=-1.0;
int useDA=1; // whether to adapt step size
if ( (on=option_match(ad_comm::argc,ad_comm::argv,"-hyeps",nopt))>-1)
{
if (!nopt) // not specified means to adapt, using function below to find reasonable one
{
cerr << "Warning: No step size given after -hyeps, ignoring" << endl;
useDA=1;
}
else // read in specified value and do not adapt
{
istringstream ist(ad_comm::argv[on+1]);
ist >> _eps;
if (_eps<=0)
{
cerr << "Error: step size (-hyeps argument) needs positive number";
ad_exit(1);
}
else
{
eps=_eps;
useDA=0;
}
}
}
// Chain number -- for console display purposes only
int chain=1;
if ( (on=option_match(ad_comm::argc,ad_comm::argv,"-chain",nopt))>-1) {
if (nopt) {
int iii=atoi(ad_comm::argv[on+1]);
if (iii <1) {
cerr << "Error: chain must be >= 1" << endl;
ad_exit(1);
} else {
chain=iii;
}
}
}
// Number of leapfrog steps. Defaults to 10.
int L=10;
if ( (on=option_match(ad_comm::argc,ad_comm::argv,"-hynstep",nopt))>-1)
{
if (nopt)
{
int _L=atoi(ad_comm::argv[on+1]);
if (_L < 1 )
{
cerr << "Error: hynstep argument must be integer > 0 " << endl;
ad_exit(1);
}
else
{
L=_L;
}
}
}
// Number of warmup samples if using adaptation of step size. Defaults to
// half of iterations.
int nwarmup= (int)nmcmc/2;
if ( (on=option_match(ad_comm::argc,ad_comm::argv,"-nwarmup",nopt))>-1)
{
if (nopt)
{
int iii=atoi(ad_comm::argv[on+1]);
if (iii <=0 || iii > nmcmc)
{
cerr << "Error: nwarmup must be 0 < nwarmup < nmcmc" << endl;
ad_exit(1);
}
else
{
nwarmup=iii;
}
}
}
// Target acceptance rate for step size adaptation. Must be
// 0<adapt_delta<1. Defaults to 0.8.
double adapt_delta=0.8; // target acceptance rate specified by the user
if ( (on=option_match(ad_comm::argc,ad_comm::argv,"-adapt_delta",nopt))>-1)
{
if (nopt)
{
istringstream ist(ad_comm::argv[on+1]);
double _adapt_delta;
ist >> _adapt_delta;
if (_adapt_delta < 0 || _adapt_delta > 1 )
{
cerr << "Error: adapt_delta must be between 0 and 1"
" using default of 0.8" << endl;
}
else
{
adapt_delta=_adapt_delta;
}
}
}
// Use diagnoal covariance (identity mass matrix)
int diag_option=0;
if ( (on=option_match(ad_comm::argc,ad_comm::argv,"-mcdiag"))>-1)
{
diag_option=1;
cout << " Setting covariance matrix to diagonal with entries " << dscale
<< endl;
}
// Restart chain from previous run?
int mcrestart_flag=option_match(ad_comm::argc,ad_comm::argv,"-mcr");
if(mcrestart_flag > -1){
cerr << endl << "Error: -mcr option not implemented for HMC" << endl;
ad_exit(1);
}
if ( (on=option_match(ad_comm::argc,ad_comm::argv,"-mcec"))>-1)
{
cerr << endl << "Error: -mcec option not yet implemented with HMC" << endl;
ad_exit(1);
// use_empirical_flag=1;
// read_empirical_covariance_matrix(nvar,S,ad_comm::adprogram_name);
}
// Prepare the mass matrix for use. Depends on many factors below.
dmatrix S(1,nvar,1,nvar);
dvector old_scale(1,nvar);
int old_nvar;
// Need to grab old_scale values still, since it is scaled below
read_covariance_matrix(S,nvar,old_Hybrid_bounded_flag,old_scale);
if (diag_option) // set covariance to be diagonal
{
S.initialize();
for (int i=1;i<=nvar;i++)
{
S(i,i)=dscale;
}
}
// How much to thin, for now fixed at 1.
if ( (on=option_match(ad_comm::argc,ad_comm::argv,"-mcsave"))>-1)
{
cerr << "Option -mcsave does not currently work with HMC -- every iteration is saved" << endl;
ad_exit(1);
}
//// ------------------------------ End of input processing
//// Setup more inputs and outputs
pofs_psave=
new uostream((char*)(ad_comm::adprogram_name + adstring(".psv")));
if (!pofs_psave|| !(*pofs_psave))
{
cerr << "Error trying to open file" <<
ad_comm::adprogram_name + adstring(".psv") << endl;
ad_exit(1);
}
if (mcrestart_flag == -1 )
{
(*pofs_psave) << nvar;
}
// need to rescale the hessian
// get the current scale
dvector x0(1,nvar);
dvector current_scale(1,nvar);
initial_params::xinit(x0);
int mctmp=initial_params::mc_phase;
initial_params::mc_phase=0;
initial_params::stddev_scale(current_scale,x0);
initial_params::mc_phase=mctmp;
// cout << "old scale=" << old_scale << endl;
// cout << "current scale=" << current_scale << endl;
// cout << "S before=" << S << endl;
// I think this is only needed if mcmc2 is used??
// for (int i=1;i<=nvar;i++)
// {
// for (int j=1;j<=nvar;j++)
// {
// S(i,j)*=old_scale(i)*old_scale(j);
// }
// }
if(diag_option){
for (int i=1;i<=nvar;i++)
{
for (int j=1;j<=nvar;j++)
{
S(i,j)*=current_scale(i)*current_scale(j);
}
}
}
// cout << "S after=" << S << endl;
gradient_structure::set_NO_DERIVATIVES();
if (mcmc2_flag==0)
{
initial_params::set_inactive_random_effects();
}
// Setup random number generator, based on seed passed
int iseed=2197;
if (iseed0) iseed=iseed0;
random_number_generator rng(iseed);
gradient_structure::set_YES_DERIVATIVES();
initial_params::xinit(x0);
// Dual averaging components
dvector epsvec(1,nmcmc+1), epsbar(1,nmcmc+1), Hbar(1,nmcmc+1);
epsvec.initialize(); epsbar.initialize(); Hbar.initialize();
double time_warmup=0;
double time_total=0;
std::clock_t start = clock();
time_t now = time(0);
tm* localtm = localtime(&now);
cout << endl << "Starting static HMC for model '" << ad_comm::adprogram_name <<
"' at " << asctime(localtm);
// write sampler parameters
ofstream adaptation("adaptation.csv", ios::trunc);
adaptation << "accept_stat__,stepsize__,int_time__,energy__,lp__" << endl;
// Declare and initialize the variables needed for the algorithm
dmatrix chd = choleski_decomp(S); // cholesky decomp of mass matrix
dvector y(1,nvar); // unbounded parameters
y.initialize();
// transformed params
independent_variables z(1,nvar); z=chd*y;
dvector gr(1,nvar); // gradients in unbounded space
// Need to run this to fill gr with current gradients and initial NLL.
double nllbegin=get_hybrid_monte_carlo_value(nvar,z,gr);
if(std::isnan(nllbegin)){
cerr << "Starting MCMC trajectory at NaN -- something is wrong!" << endl;
ad_exit(1);
}
// initial rotated gradient
dvector gr2(1,nvar); gr2=gr*chd;
dvector p(1,nvar); // momentum vector
p.fill_randn(rng);
// Copy initial value to parsave in case first trajectory rejected
initial_params::copy_all_values(parsave,1.0);
double iaccept=0.0;
// The gradient and params at beginning of trajectory, in case rejected.
dvector gr2begin(1,nvar); gr2begin=gr2;
dvector ybegin(1,nvar); ybegin=y;
double nll=nllbegin;
// if(useDA){
// eps=find_reasonable_stepsize(nvar,y,p,chd);
// epsvec(1)=eps; epsbar(1)=eps; Hbar(1)=0;
// }
double mu=log(10*eps);
// Start of MCMC chain
for (int is=1;is<=nmcmc;is++) {
// Random momentum for next iteration, only affects Ham values
p.fill_randn(rng);
double H0=nll+0.5*norm2(p);
// Generate trajectory
int divergence=0;
for (int i=1;i<=L;i++) {
// leapfrog updates gr, p, y, and gr2 by reference
nll=leapfrog(nvar, gr, chd, eps, p, y, gr2);
// Break trajectory early if a divergence occurs to save computation
if(std::isnan(nll)){
divergence=1; break;
}
} // end of trajectory
// Test whether to accept the proposed state
double Ham=nll+0.5*norm2(p); // Update Hamiltonian for proposed set
double alpha=min(1.0, exp(H0-Ham)); // acceptance ratio
double rr=randu(rng); // Runif(1)
if (rr<alpha && !divergence){ // accept
iaccept++;
// Update for next iteration: params, Hamiltonian and gr2
ybegin=y;
gr2begin=gr2;
nllbegin=nll;
initial_params::copy_all_values(parsave,1.0);
} else {
// Reject and don't update anything to reuse initials for next trajectory
y=ybegin;
gr2=gr2begin;
nll=nllbegin;
}
// Save parameters to psv file, duplicated if rejected
(*pofs_psave) << parsave;
// Adaptation of step size (eps).
if(useDA && is <= nwarmup){
eps=adapt_eps(is, eps, alpha, adapt_delta, mu, epsvec, epsbar, Hbar);
}
adaptation << alpha << "," << eps << "," << eps*L << "," << H0 << "," << -nll << endl;
if(is ==nwarmup) time_warmup = ( std::clock()-start)/(double) CLOCKS_PER_SEC;
print_mcmc_progress(is, nmcmc, nwarmup, chain);
} // end of MCMC chain
// This final ratio should closely match adapt_delta
if(useDA){
cout << "Final acceptance ratio=" << iaccept/nmcmc << " and target is " << adapt_delta<<endl;
cout << "Final step size=" << eps << "; after " << nwarmup << " warmup iterations"<< endl;
} else {
cout << "Final acceptance ratio=" << iaccept/nmcmc << endl;
}
time_total = ( std::clock() - start ) / (double) CLOCKS_PER_SEC;
print_mcmc_timing(time_warmup, time_total);
// I assume this closes the connection to the file??
if (pofs_psave)
{
delete pofs_psave;
pofs_psave=NULL;
}
} // end of HMC function
int main(int argc, char *argv[])
{
int step, ie, iside, i, j, k;
double mflops, tmax, nelt_tot = 0.0;
char Class;
logical ifmortar = false, verified;
double t2, trecs[t_last+1];
char *t_names[t_last+1];
//--------------------------------------------------------------------
// Initialize NUMA control
//--------------------------------------------------------------------
numa_initialize_env(NUMA_MIGRATE_EXISTING);
//---------------------------------------------------------------------
// Read input file (if it exists), else take
// defaults from parameters
//---------------------------------------------------------------------
FILE *fp;
if ((fp = fopen("timer.flag", "r")) != NULL) {
timeron = true;
t_names[t_total] = "total";
t_names[t_init] = "init";
t_names[t_convect] = "convect";
t_names[t_transfb_c] = "transfb_c";
t_names[t_diffusion] = "diffusion";
t_names[t_transf] = "transf";
t_names[t_transfb] = "transfb";
t_names[t_adaptation] = "adaptation";
t_names[t_transf2] = "transf+b";
t_names[t_add2] = "add2";
fclose(fp);
} else {
timeron = false;
}
printf("\n\n NAS Parallel Benchmarks (NPB3.3-OMP-C) - UA Benchmark\n\n");
if ((fp = fopen("inputua.data", "r")) != NULL) {
int result;
printf(" Reading from input file inputua.data\n");
result = fscanf(fp, "%d", &fre);
while (fgetc(fp) != '\n');
result = fscanf(fp, "%d", &niter);
while (fgetc(fp) != '\n');
result = fscanf(fp, "%d", &nmxh);
while (fgetc(fp) != '\n');
result = fscanf(fp, "%lf", &alpha);
Class = 'U';
fclose(fp);
} else {
printf(" No input file inputua.data. Using compiled defaults\n");
fre = FRE_DEFAULT;
niter = NITER_DEFAULT;
nmxh = NMXH_DEFAULT;
alpha = ALPHA_DEFAULT;
Class = CLASS_DEFAULT;
}
dlmin = pow(0.5, REFINE_MAX);
dtime = 0.04*dlmin;
printf(" Levels of refinement: %8d\n", REFINE_MAX);
printf(" Adaptation frequency: %8d\n", fre);
printf(" Time steps: %8d dt: %15.6E\n", niter, dtime);
printf(" CG iterations: %8d\n", nmxh);
printf(" Heat source radius: %8.4f\n", alpha);
printf(" Number of available threads: %8d\n", omp_get_max_threads());
printf("\n");
top_constants();
for (i = 1; i <= t_last; i++) {
timer_clear(i);
}
if (timeron) timer_start(t_init);
// set up initial mesh (single element) and solution (all zero)
create_initial_grid();
r_init_omp((double *)ta1, ntot, 0.0);
nr_init_omp((int *)sje, 4*6*nelt, -1);
init_locks();
// compute tables of coefficients and weights
coef();
geom1();
// compute the discrete laplacian operators
setdef();
// prepare for the preconditioner
setpcmo_pre();
// refine initial mesh and do some preliminary work
time = 0.0;
mortar();
prepwork();
adaptation(&ifmortar, 0);
if (timeron) timer_stop(t_init);
timer_clear(1);
time = 0.0;
for (step = 0; step <= niter; step++) {
if (step == 1) {
// reset the solution and start the timer, keep track of total no elms
r_init((double *)ta1, ntot, 0.0);
time = 0.0;
nelt_tot = 0.0;
for (i = 1; i <= t_last; i++) {
if (i != t_init) timer_clear(i);
}
timer_start(1);
}
// advance the convection step
convect(ifmortar);
if (timeron) timer_start(t_transf2);
// prepare the intital guess for cg
transf(tmort, (double *)ta1);
// compute residual for diffusion term based on intital guess
// compute the left hand side of equation, lapacian t
#pragma omp parallel default(shared) private(ie,k,j,i)
{
#pragma omp for
for (ie = 0; ie < nelt; ie++) {
laplacian(ta2[ie], ta1[ie], size_e[ie]);
}
// compute the residual
#pragma omp for
for (ie = 0; ie < nelt; ie++) {
for (k = 0; k < LX1; k++) {
for (j = 0; j < LX1; j++) {
for (i = 0; i < LX1; i++) {
trhs[ie][k][j][i] = trhs[ie][k][j][i] - ta2[ie][k][j][i];
}
}
}
}
} //end parallel
// get the residual on mortar
transfb(rmor, (double *)trhs);
if (timeron) timer_stop(t_transf2);
// apply boundary condition: zero out the residual on domain boundaries
// apply boundary conidtion to trhs
#pragma omp parallel for default(shared) private(ie,iside)
for (ie = 0; ie < nelt; ie++) {
for (iside = 0; iside < NSIDES; iside++) {
if (cbc[ie][iside] == 0) {
facev(trhs[ie], iside, 0.0);
}
}
}
// apply boundary condition to rmor
col2(rmor, tmmor, nmor);
// call the conjugate gradient iterative solver
diffusion(ifmortar);
// add convection and diffusion
if (timeron) timer_start(t_add2);
add2((double *)ta1, (double *)t, ntot);
if (timeron) timer_stop(t_add2);
// perform mesh adaptation
time = time + dtime;
if ((step != 0) && (step/fre*fre == step)) {
if (step != niter) {
adaptation(&ifmortar, step);
}
} else {
ifmortar = false;
}
nelt_tot = nelt_tot + (double)(nelt);
}
timer_stop(1);
tmax = timer_read(1);
verify(&Class, &verified);
// compute millions of collocation points advanced per second.
// diffusion: nmxh advancements, convection: 1 advancement
mflops = nelt_tot*(double)(LX1*LX1*LX1*(nmxh+1))/(tmax*1.e6);
print_results("UA", Class, REFINE_MAX, 0, 0, niter,
tmax, mflops, " coll. point advanced",
verified, NPBVERSION, COMPILETIME, CS1, CS2, CS3, CS4, CS5,
CS6, "(none)");
//---------------------------------------------------------------------
// More timers
//---------------------------------------------------------------------
if (timeron) {
for (i = 1; i <= t_last; i++) {
trecs[i] = timer_read(i);
}
if (tmax == 0.0) tmax = 1.0;
printf(" SECTION Time (secs)\n");
for (i = 1; i <= t_last; i++) {
printf(" %-10s:%9.3f (%6.2f%%)\n",
t_names[i], trecs[i], trecs[i]*100./tmax);
if (i == t_transfb_c) {
t2 = trecs[t_convect] - trecs[t_transfb_c];
printf(" --> %11s:%9.3f (%6.2f%%)\n",
"sub-convect", t2, t2*100./tmax);
} else if (i == t_transfb) {
t2 = trecs[t_diffusion] - trecs[t_transf] - trecs[t_transfb];
printf(" --> %11s:%9.3f (%6.2f%%)\n",
"sub-diffuse", t2, t2*100./tmax);
}
}
}
//--------------------------------------------------------------------
// Teardown NUMA control
//--------------------------------------------------------------------
numa_shutdown();
return 0;
}
bool yee_compare(CompareArgs &args)
{
if ((args.image_a_->get_width() != args.image_b_->get_width()) or
(args.image_a_->get_height() != args.image_b_->get_height()))
{
args.error_string_ = "Image dimensions do not match\n";
return false;
}
const auto w = args.image_a_->get_width();
const auto h = args.image_a_->get_height();
const auto dim = w * h;
auto identical = true;
for (auto i = 0u; i < dim; i++)
{
if (args.image_a_->get(i) != args.image_b_->get(i))
{
identical = false;
break;
}
}
if (identical)
{
args.error_string_ = "Images are binary identical\n";
return true;
}
// Assuming colorspaces are in Adobe RGB (1998) convert to XYZ.
std::vector<float> a_lum(dim);
std::vector<float> b_lum(dim);
std::vector<float> a_a(dim);
std::vector<float> b_a(dim);
std::vector<float> a_b(dim);
std::vector<float> b_b(dim);
if (args.verbose_)
{
std::cout << "Converting RGB to XYZ\n";
}
const auto gamma = args.gamma_;
const auto luminance = args.luminance_;
#pragma omp parallel for shared(args, a_lum, b_lum, a_a, a_b, b_a, b_b)
for (auto y = 0; y < static_cast<ptrdiff_t>(h); y++)
{
for (auto x = 0u; x < w; x++)
{
const auto i = x + y * w;
const auto a_color_r = powf(args.image_a_->get_red(i) / 255.0f,
gamma);
const auto a_color_g = powf(args.image_a_->get_green(i) / 255.0f,
gamma);
const auto a_color_b = powf(args.image_a_->get_blue(i) / 255.0f,
gamma);
float a_x;
float a_y;
float a_z;
adobe_rgb_to_xyz(a_color_r, a_color_g, a_color_b, a_x, a_y, a_z);
float l;
xyz_to_lab(a_x, a_y, a_z, l, a_a[i], a_b[i]);
const auto b_color_r = powf(args.image_b_->get_red(i) / 255.0f,
gamma);
const auto b_color_g = powf(args.image_b_->get_green(i) / 255.0f,
gamma);
const auto b_color_b = powf(args.image_b_->get_blue(i) / 255.0f,
gamma);
float b_x;
float b_y;
float b_z;
adobe_rgb_to_xyz(b_color_r, b_color_g, b_color_b, b_x, b_y, b_z);
xyz_to_lab(b_x, b_y, b_z, l, b_a[i], b_b[i]);
a_lum[i] = a_y * luminance;
b_lum[i] = b_y * luminance;
}
}
if (args.verbose_)
{
std::cout << "Constructing Laplacian Pyramids\n";
}
const LPyramid la(a_lum, w, h);
const LPyramid lb(b_lum, w, h);
const auto num_one_degree_pixels =
to_degrees(2 *
std::tan(args.field_of_view_ * to_radians(.5f)));
const auto pixels_per_degree = w / num_one_degree_pixels;
if (args.verbose_)
{
std::cout << "Performing test\n";
}
const auto adaptation_level = adaptation(num_one_degree_pixels);
float cpd[MAX_PYR_LEVELS];
cpd[0] = 0.5f * pixels_per_degree;
for (auto i = 1u; i < MAX_PYR_LEVELS; i++)
{
cpd[i] = 0.5f * cpd[i - 1];
}
const auto csf_max = csf(3.248f, 100.0f);
static_assert(MAX_PYR_LEVELS > 2,
"MAX_PYR_LEVELS must be greater than 2");
float f_freq[MAX_PYR_LEVELS - 2];
for (auto i = 0u; i < MAX_PYR_LEVELS - 2; i++)
{
f_freq[i] = csf_max / csf(cpd[i], 100.0f);
}
auto pixels_failed = 0u;
auto error_sum = 0.;
#pragma omp parallel for reduction(+ : pixels_failed, error_sum) \
shared(args, a_a, a_b, b_a, b_b, cpd, f_freq)
for (auto y = 0; y < static_cast<ptrdiff_t>(h); y++)
{
for (auto x = 0u; x < w; x++)
{
const auto index = y * w + x;
const auto adapt = std::max((la.get_value(x, y, adaptation_level) +
lb.get_value(x, y, adaptation_level)) * 0.5f,
1e-5f);
auto sum_contrast = 0.f;
auto factor = 0.f;
for (auto i = 0u; i < MAX_PYR_LEVELS - 2; i++)
{
const auto n1 =
fabsf(la.get_value(x, y, i) - la.get_value(x, y, i + 1));
const auto n2 =
fabsf(lb.get_value(x, y, i) - lb.get_value(x, y, i + 1));
const auto numerator = std::max(n1, n2);
const auto d1 = fabsf(la.get_value(x, y, i + 2));
const auto d2 = fabsf(lb.get_value(x, y, i + 2));
const auto denominator = std::max(std::max(d1, d2), 1e-5f);
const auto contrast = numerator / denominator;
const auto f_mask = mask(contrast * csf(cpd[i], adapt));
factor += contrast * f_freq[i] * f_mask;
sum_contrast += contrast;
}
sum_contrast = std::max(sum_contrast, 1e-5f);
factor /= sum_contrast;
factor = std::min(std::max(factor, 1.f), 10.f);
const auto delta =
fabsf(la.get_value(x, y, 0) - lb.get_value(x, y, 0));
error_sum += delta;
auto pass = true;
// pure luminance test
if (delta > factor * tvi(adapt))
{
pass = false;
}
if (not args.luminance_only_)
{
// CIE delta E test with modifications
auto color_scale = args.color_factor_;
// ramp down the color test in scotopic regions
if (adapt < 10.0f)
{
// Don't do color test at all.
color_scale = 0.0;
}
const auto da = a_a[index] - b_a[index];
const auto db = a_b[index] - b_b[index];
const auto delta_e = (da * da + db * db) * color_scale;
error_sum += delta_e;
if (delta_e > factor)
{
pass = false;
}
}
if (not pass)
{
pixels_failed++;
if (args.image_difference_)
{
args.image_difference_->set(255, 0, 0, 255, index);
}
}
else
{
if (args.image_difference_)
{
args.image_difference_->set(0, 0, 0, 255, index);
}
}
}
}
const auto error_sum_buff =
std::to_string(error_sum) + " error sum\n";
const auto different =
std::to_string(pixels_failed) + " pixels are different\n";
// Always output image difference if requested.
if (args.image_difference_)
{
args.image_difference_->write_to_file(args.image_difference_->get_name());
args.error_string_ += "Wrote difference image to ";
args.error_string_ += args.image_difference_->get_name();
args.error_string_ += "\n";
}
if (pixels_failed < args.threshold_pixels_)
{
args.error_string_ = "Images are perceptually indistinguishable\n";
args.error_string_ += different;
return true;
}
args.error_string_ = "Images are visibly different\n";
args.error_string_ += different;
if (args.sum_errors_)
{
args.error_string_ += error_sum_buff;
}
return false;
}
void evolution(axl_network *mysys, int steps, int seed)
{
int i, j, step;
int n = mysys->nagents;
int *neighbors;
axl_mass_media *mm;
double b_total;
double random;
neighbors = (int *)malloc(sizeof(int) * n);
srand(seed);
for(step = 0; step < steps; step++)
{
for(i = 0; i < n; i++)
{
if(mysys->agent[i].degree != 0)
{
j = rand() % mysys->agent[i].degree;
neighbors[i] = mysys->agent[i].neighbors[j];
}
else
neighbors[i] = -(mysys->nmm) - 1;
}
if(mysys->nmm > 1)
{
b_total = 0.00;
for(i = 0; i < mysys->nmm; i++)
b_total += mysys->mass_media[i].b;
if(b_total > 0.00)
{
select_mass_media_to_interact(*mysys, neighbors, rand());
adaptation_of_medias(mysys);
}
}
else
{
for(i = 0; i < n; i++)
{
random = ((double)rand()) / RAND_MAX;
if(random < mysys->mass_media[0].b)
neighbors[i] = -1;
}
for(i = 0; i < mysys->mass_media[0].f; i++)
{
mm = &(mysys->mass_media[0]);
adaptation(mm, *mysys, i);
}
}
synchronous_updating(mysys, neighbors, rand());
if(mysys->noise > 0.00)
noise(mysys, rand());
}
free(neighbors);
return;
}
