/* Normalize a given vector in 3D. */
void normalize(GLdouble *v){
GLdouble norm = calculate_norm(v[0], v[1], v[2]);
v[0] = v[0] / norm;
v[1] = v[1] / norm;
v[2] = v[2] / norm;
}
std::set<int>& DiscontinuityDetector::get_discontinuous_element_ids(double threshold)
{
Element* e;
for_all_active_elements(e, mesh) {
for(int edge_i = 0; edge_i < e->get_num_surf(); edge_i++)
if(calculate_relative_flow_direction(e, edge_i) < -1e-3 && !e->en[edge_i]->bnd) {
double jump = calculate_jumps(e, edge_i);
double diameter_indicator = std::pow(e->get_diameter(), (H2D_GET_H_ORDER(spaces[0]->get_element_order(e->id)) + 1) / 2);
double edge_length = std::sqrt(std::pow(e->vn[(edge_i + 1) % 4]->x - e->vn[edge_i]->x, 2) + std::pow(e->vn[(edge_i + 1) % 4]->y - e->vn[edge_i]->y, 2));
double norm = calculate_norm(e, edge_i);
double discontinuity_detector = jump / (diameter_indicator * edge_length * norm);
if(discontinuity_detector > threshold)
discontinuous_element_ids.insert(e->id);
}
}
return discontinuous_element_ids;
};
void TwostepTimeSolver::solve()
{
libmesh_assert(core_time_solver.get());
// The core_time_solver will handle any first_solve actions
first_solve = false;
// We may have to repeat timesteps entirely if our error is bad
// enough
bool max_tolerance_met = false;
// Calculating error values each time
Real single_norm(0.), double_norm(0.), error_norm(0.),
relative_error(0.);
while (!max_tolerance_met)
{
// If we've been asked to reduce deltat if necessary, make sure
// the core timesolver does so
core_time_solver->reduce_deltat_on_diffsolver_failure =
this->reduce_deltat_on_diffsolver_failure;
if (!quiet)
{
libMesh::out << "\n === Computing adaptive timestep === "
<< std::endl;
}
// Use the double-length timestep first (so the
// old_nonlinear_solution won't have to change)
core_time_solver->solve();
// Save a copy of the double-length nonlinear solution
// and the old nonlinear solution
std::unique_ptr<NumericVector<Number>> double_solution =
_system.solution->clone();
std::unique_ptr<NumericVector<Number>> old_solution =
_system.get_vector("_old_nonlinear_solution").clone();
double_norm = calculate_norm(_system, *double_solution);
if (!quiet)
{
libMesh::out << "Double norm = " << double_norm << std::endl;
}
// Then reset the initial guess for our single-length calcs
*(_system.solution) = _system.get_vector("_old_nonlinear_solution");
// Call two single-length timesteps
// Be sure that the core_time_solver does not change the
// timestep here. (This is unlikely because it just succeeded
// with a timestep twice as large!)
// FIXME: even if diffsolver failure is unlikely, we ought to
// do *something* if it happens
core_time_solver->reduce_deltat_on_diffsolver_failure = 0;
Real old_time = _system.time;
Real old_deltat = _system.deltat;
_system.deltat *= 0.5;
core_time_solver->solve();
core_time_solver->advance_timestep();
core_time_solver->solve();
single_norm = calculate_norm(_system, *_system.solution);
if (!quiet)
{
libMesh::out << "Single norm = " << single_norm << std::endl;
}
// Reset the core_time_solver's reduce_deltat... value.
core_time_solver->reduce_deltat_on_diffsolver_failure =
this->reduce_deltat_on_diffsolver_failure;
// But then back off just in case our advance_timestep() isn't
// called.
// FIXME: this probably doesn't work with multistep methods
_system.get_vector("_old_nonlinear_solution") = *old_solution;
_system.time = old_time;
_system.deltat = old_deltat;
// Find the relative error
*double_solution -= *(_system.solution);
error_norm = calculate_norm(_system, *double_solution);
relative_error = error_norm / _system.deltat /
std::max(double_norm, single_norm);
// If the relative error makes no sense, we're done
if (!double_norm && !single_norm)
return;
if (!quiet)
{
libMesh::out << "Error norm = " << error_norm << std::endl;
libMesh::out << "Local relative error = "
<< (error_norm /
std::max(double_norm, single_norm))
<< std::endl;
libMesh::out << "Global relative error = "
<< (error_norm / _system.deltat /
std::max(double_norm, single_norm))
<< std::endl;
libMesh::out << "old delta t = " << _system.deltat << std::endl;
}
// If our upper tolerance is negative, that means we want to set
// it based on the first successful time step
if (this->upper_tolerance < 0)
this->upper_tolerance = -this->upper_tolerance * relative_error;
// If we haven't met our upper error tolerance, we'll have to
// repeat this timestep entirely
if (this->upper_tolerance && relative_error > this->upper_tolerance)
{
// Reset the initial guess for our next try
*(_system.solution) =
_system.get_vector("_old_nonlinear_solution");
// Chop delta t in half
_system.deltat /= 2.;
if (!quiet)
{
libMesh::out << "Failed to meet upper error tolerance"
<< std::endl;
libMesh::out << "Retrying with delta t = "
<< _system.deltat << std::endl;
}
}
else
max_tolerance_met = true;
}
// Otherwise, compare the relative error to the tolerance
// and adjust deltat
last_deltat = _system.deltat;
// If our target tolerance is negative, that means we want to set
// it based on the first successful time step
if (this->target_tolerance < 0)
this->target_tolerance = -this->target_tolerance * relative_error;
const Real global_shrink_or_growth_factor =
std::pow(this->target_tolerance / relative_error,
static_cast<Real>(1. / core_time_solver->error_order()));
const Real local_shrink_or_growth_factor =
std::pow(this->target_tolerance /
(error_norm/std::max(double_norm, single_norm)),
static_cast<Real>(1. / (core_time_solver->error_order()+1.)));
if (!quiet)
{
libMesh::out << "The global growth/shrink factor is: "
<< global_shrink_or_growth_factor << std::endl;
libMesh::out << "The local growth/shrink factor is: "
<< local_shrink_or_growth_factor << std::endl;
}
// The local s.o.g. factor is based on the expected **local**
// truncation error for the timestepping method, the global
// s.o.g. factor is based on the method's **global** truncation
// error. You can shrink/grow the timestep to attempt to satisfy
// either a global or local time-discretization error tolerance.
Real shrink_or_growth_factor =
this->global_tolerance ? global_shrink_or_growth_factor :
local_shrink_or_growth_factor;
if (this->max_growth && this->max_growth < shrink_or_growth_factor)
{
if (!quiet && this->global_tolerance)
{
libMesh::out << "delta t is constrained by max_growth" << std::endl;
}
shrink_or_growth_factor = this->max_growth;
}
_system.deltat *= shrink_or_growth_factor;
// Restrict deltat to max-allowable value if necessary
if ((this->max_deltat != 0.0) && (_system.deltat > this->max_deltat))
{
if (!quiet)
{
libMesh::out << "delta t is constrained by maximum-allowable delta t."
<< std::endl;
}
_system.deltat = this->max_deltat;
}
// Restrict deltat to min-allowable value if necessary
if ((this->min_deltat != 0.0) && (_system.deltat < this->min_deltat))
{
if (!quiet)
{
libMesh::out << "delta t is constrained by minimum-allowable delta t."
<< std::endl;
}
_system.deltat = this->min_deltat;
}
if (!quiet)
{
libMesh::out << "new delta t = " << _system.deltat << std::endl;
}
}
int main(int argc, char *argv[]) {
struct RNAfold_args_info args_info;
char *string, *input_string, *structure=NULL, *cstruc=NULL;
char fname[80], ffname[80], gfname[80], *ParamFile=NULL;
char *ns_bases=NULL, *c;
int i, j, ii, jj, mu, length, l, sym, r, pf=0, noconv=0;
unsigned int input_type;
double energy, min_en, kT, sfact=1.07;
int doMEA=0, circular = 0, N;
char *pf_struc;
double dist;
plist *pl;
FILE * filehandle;
FILE * statsfile;
char* line;
double tau = 0.01; /* Variance of energy parameters */
double sigma = 0.01; /* Variance of experimental constraints */
double *gradient; /* Gradient for steepest descent search
epsilon[i+1]= epsilon[i] - gradient *
step_size */
double initial_step_size = 0.5; /* Initial step size for steepest
descent search */
double step_size;
double D; /* Discrepancy (i.e. value of objective
function) for the current
prediction */
int iteration, max_iteration = 2000; /* Current and maximum number of
iterations after which
algorithm stops */
double precision = 0.1; /* cutoff used for stop conditions */
double tolerance = 0.1; /* Parameter used by various GSL minimizers */
int method_id = 1; /* Method to use for minimization, 0 and 1
are custom steepest descent, the rest
are GSL implementations (see below)*/
int initial_guess_method = 0;
int sample_N = 1000;
double *prev_epsilon;
double *prev_gradient;
double DD, prev_D, sum, norm;
int status;
double* gradient_numeric;
double* gradient_numeric_gsl;
/* Minimizer vars */
const gsl_multimin_fdfminimizer_type *T;
gsl_multimin_fdfminimizer *minimizer;
gsl_vector *minimizer_x;
gsl_vector *minimizer_g;
gsl_multimin_function_fdf minimizer_func;
minimizer_pars_struct minimizer_pars;
char *constraints;
char outfile[256];
char constraints_file[256];
char epsilon_file[256];
FILE* fh;
double last_non_nan_lnQ;
pf_overflow = 0;
pf_underflow = 0;
dangles=2;
do_backtrack = 1;
string = NULL;
noPS = 0;
outfile[0]='\0';
epsilon_file[0]='\0';
strcpy(psDir, "dotplots");
if(RNAfold_cmdline_parser (argc, argv, &args_info) != 0) exit(1);
/* RNAbpfold specific options */
if (args_info.tau_given) tau = args_info.tau_arg;
if (args_info.sigma_given) sigma = args_info.sigma_arg;
if (args_info.precision_given) precision = args_info.precision_arg;
if (args_info.step_given) initial_step_size = args_info.step_arg;
if (args_info.maxN_given) max_iteration = args_info.maxN_arg;
if (args_info.minimization_given) method_id = args_info.minimization_arg;
if (args_info.init_given) initial_guess_method = args_info.init_arg;
if (args_info.tolerance_given) tolerance = args_info.tolerance_arg;
if (args_info.outfile_given) strcpy(outfile, args_info.outfile_arg);
if (args_info.constraints_given) strcpy(constraints_file, args_info.constraints_arg);
if (args_info.epsilon_given) strcpy(epsilon_file, args_info.epsilon_arg);
if (args_info.sampleGradient_given) sample_conditionals=1;
if (args_info.hybridGradient_given) {
sample_conditionals=1;
hybrid_conditionals=1;
}
if (args_info.numericalGradient_given) numerical=1;
if (args_info.sampleStructure_given) sample_structure=1;
if (args_info.psDir_given) strcpy(psDir, args_info.psDir_arg);
if (args_info.sparsePS_given) sparsePS=args_info.sparsePS_arg;
if (args_info.gridSearch_given) grid_search = 1;
/* Generic RNAfold options */
if (args_info.temp_given) temperature = args_info.temp_arg;
if (args_info.reference_given) fold_constrained=1;
if (args_info.noTetra_given) tetra_loop=0;
if (args_info.dangles_given) dangles = args_info.dangles_arg;
if (args_info.noLP_given) noLonelyPairs = 1;
if (args_info.noGU_given) noGU = 1;
if (args_info.noClosingGU_given) no_closingGU = 1;
if (args_info.noconv_given) noconv = 1;
if (args_info.energyModel_given) energy_set = args_info.energyModel_arg;
if (args_info.paramFile_given) ParamFile = strdup(args_info.paramFile_arg);
if (args_info.nsp_given) ns_bases = strdup(args_info.nsp_arg);
if (args_info.pfScale_given) sfact = args_info.pfScale_arg;
if (args_info.noPS_given) noPS=1;
/* Create postscript directory */
if (!noPS) {
struct stat stat_p;
if (stat (psDir, &stat_p) != 0) {
if (mkdir(psDir, S_IRWXU|S_IROTH|S_IRGRP ) !=0) {
fprintf(stderr, "WARNING: Could not create directory: %s", psDir);
}
}
}
if (ParamFile != NULL) {
read_parameter_file(ParamFile);
}
if (ns_bases != NULL) {
nonstandards = space(33);
c=ns_bases;
i=sym=0;
if (*c=='-') {
sym=1;
c++;
}
while (*c!='\0') {
if (*c!=',') {
nonstandards[i++]=*c++;
nonstandards[i++]=*c;
if ((sym)&&(*c!=*(c-1))) {
nonstandards[i++]=*c;
nonstandards[i++]=*(c-1);
}
}
c++;
}
}
/*Read sequence*/
fname[0] = '\0';
while((input_type = get_input_line(&input_string, 0)) & VRNA_INPUT_FASTA_HEADER) {
(void) sscanf(input_string, "%42s", fname);
free(input_string);
}
length = (int) strlen(input_string);
string = strdup(input_string);
free(input_string);
structure = (char *) space((unsigned) length+1);
/* For testing purpose pass dot bracket structure of reference structure via -C */
if (fold_constrained) {
input_type = get_input_line(&input_string, VRNA_INPUT_NOSKIP_COMMENTS);
if(input_type & VRNA_INPUT_QUIT) {
exit(1);
}
else if((input_type & VRNA_INPUT_MISC) && (strlen(input_string) > 0)) {
cstruc = strdup(input_string);
free(input_string);
}
else warn_user("-C was given but reference structure is missing");
}
if(noconv) {
str_RNA2RNA(string);
} else {
str_DNA2RNA(string);
}
/* Allocating space */
epsilon = (double *) space(sizeof(double)*(length+1));
exp_pert = (double **)space(sizeof(double *)*(length+1));
perturbations = (double **)space(sizeof(double *)*(length+1));
prev_epsilon = (double *) space(sizeof(double)*(length+1));
gradient = (double *) space(sizeof(double)*(length+1));
gradient_numeric = (double *) space(sizeof(double)*(length+1));
gradient_numeric_gsl = (double *) space(sizeof(double)*(length+1));
prev_gradient = (double *) space(sizeof(double)*(length+1));
q_unpaired = (double *) space(sizeof(double)*(length+1));
p_unpaired_cond = (double **)space(sizeof(double *)*(length+1));
p_unpaired_cond_sampled = (double **)space(sizeof(double *)*(length+1));
p_pp = (double **)space(sizeof(double *)*(length+1));
p_unpaired = (double *) space(sizeof(double)*(length+1));
p_unpaired_tmp = (double *) space(sizeof(double)*(length+1));
for (i=0; i <= length; i++) {
epsilon[i] = gradient[i] = q_unpaired[i] = 0.0;
p_unpaired_cond[i] = (double *) space(sizeof(double)*(length+1));
p_unpaired_cond_sampled[i] = (double *) space(sizeof(double)*(length+1));
p_pp[i] = (double *) space(sizeof(double)*(length+1));
exp_pert[i] = (double *) space(sizeof(double)*(length+1));
perturbations[i] = (double *) space(sizeof(double)*(length+1));
for (j=0; j <= length; j++) {
p_pp[i][j]=p_unpaired_cond[i][j] = 0.0;
p_unpaired_cond_sampled[i][j] = 0.0;
}
}
/*** If file with perturbation vector epsilon is given we fold using
this epsilon and are done ***/
if (args_info.epsilon_given) {
plist *pl, *pl1,*pl2;
filehandle = fopen (epsilon_file,"r");
if (filehandle == NULL) {
nrerror("Could not open file with perturbation vector.");
}
i=1;
while (1) {
double t;
line = get_line(filehandle);
if (line == NULL) break;
if (i>length) nrerror("Too many values in perturbation vector file.");
if (sscanf(line, "%lf", &epsilon[i]) !=1) {
nrerror("Error while reading perturbation vector file.");
}
i++;
}
if (i-1 != length) {
nrerror("Too few values in perturbation vector file.");
}
init_pf_fold(length);
pf_fold_pb(string, NULL);
sprintf(fname,"%s/dot.ps", psDir);
pl1 = make_plist(length, 1e-5);
(void) PS_dot_plot_list_epsilon(string, fname, NULL, pl1, epsilon, "");
exit(0);
}
/*** Get constraints from reference structure or from external file ***/
/* Structure was given by -C */
if (fold_constrained) {
for (i=0; i<length; i++) {
if (cstruc[i] == '(' || cstruc[i] == ')') {
q_unpaired[i+1] = 0.0;
} else {
q_unpaired[i+1] = 1.0;
}
}
/*Read constraints from file*/
} else {
filehandle = fopen (constraints_file,"r");
if (filehandle == NULL) {
nrerror("No constraints given as dot bracket or wrong file name");
}
i=1;
while (1) {
double t;
line = get_line(filehandle);
if (line == NULL) break;
if (i>length) nrerror("Too many values in constraints.dat");
if (sscanf(line, "%lf", &q_unpaired[i]) !=1) {
nrerror("Error while reading constraints.dat");
}
i++;
}
if (i-1 != length) {
nrerror("Too few values in constraints.dat");
}
}
/* Create file handle */
if (outfile[0] !='\0') {
statsfile = fopen (outfile,"w");
} else {
statsfile = fopen ("stats.dat","w");
}
setvbuf(statsfile, NULL, _IONBF, 0);
if (!grid_search) {
fprintf(statsfile, "Iteration\tDiscrepancy\tNorm\tdfCount\tMEA\tSampled_structure\tSampled_energy\tSampled_distance\tEpsilon\ttimestamp\n");
} else {
/* If we do a grid search we have a different output. */
fprintf(statsfile, "Dummy\tm\tb\tdummy\tMEA\tepsilon\n");
}
if (statsfile == NULL) {
nrerror("Could not open stats.dat for writing.");
}
fprintf(stderr, "tau^2 = %.4f; sigma^2 = %.4f; precision = %.4f; tolerance = %.4f; step-size: %.4f\n\n",
tau, sigma, precision, tolerance, initial_step_size);
st_back=1;
min_en = fold(string, structure);
(void) fflush(stdout);
if (length>2000) free_arrays();
pf_struc = (char *) space((unsigned) length+1);
kT = (temperature+273.15)*1.98717/1000.; /* in Kcal */
pf_scale = exp(-(sfact*min_en)/kT/length);
/* Set up minimizer */
minimizer_x = gsl_vector_alloc (length+1);
minimizer_g = gsl_vector_alloc (length+1);
for (i=0; i <= length; i++) {
epsilon[i] = 0.0;
gsl_vector_set (minimizer_g, i, 0.0);
gsl_vector_set (minimizer_x, i, epsilon[i]);
}
minimizer_pars.length=length;
minimizer_pars.seq = string;
minimizer_pars.tau=tau;
minimizer_pars.sigma=sigma;
minimizer_pars.kT=kT;
minimizer_func.n = length+1;
minimizer_func.f = calculate_f;
minimizer_func.df = numerical ? calculate_df_numerically: calculate_df;
minimizer_func.fdf = calculate_fdf;
minimizer_func.params = &minimizer_pars;
//min_en = fold_pb(string, structure);
//fprintf(stderr, "%f", min_en);
//exit(0);
/* Calling test functions for debugging */
for (i=1; i <= length; i++) {
if (i%2==0) {
epsilon[i] = +0.2*i;
} else {
epsilon[i] = -0.2*i;
}
}
//test_folding(string, length);
/* //test_stochastic_backtracking(string, length); */
/* //test_gradient(minimizer_func, minimizer_pars); */
/* //test_gradient_sampling(minimizer_func, minimizer_pars); */
//exit(1);
count_df_evaluations=0;
/* Initial guess for epsilon */
if (initial_guess_method !=0 && initial_guess_method !=3) {
/* Vars for inital guess methods */
double m,b;
double* curr_epsilon;
double* best_epsilon;
double best_m, best_b, best_scale;
double curr_D;
double min_D = 999999999.0;
double inc = +0.25;
double cut;
if (initial_guess_method == 1) fprintf(stderr, "Mathew's constant perturbations\n");
if (initial_guess_method == 2) fprintf(stderr, "Perturbations proportional to q-p\n");
curr_epsilon = (double *) space(sizeof(double)*(length+1));
best_epsilon = (double *) space(sizeof(double)*(length+1));
last_non_nan_lnQ = min_en;
// Calculate p_unpaired for unperturbed state which we need later
// for the proportinal method
if (initial_guess_method == 2) {
init_pf_fold(length);
for (i=0; i <= length; i++) {
epsilon[i] = 0.0;
}
pf_fold_pb(string, NULL);
for (i = 1; i < length; i++) {
for (j = i+1; j<= length; j++) {
p_pp[i][j]=p_pp[j][i]=pr[iindx[i]-j];
}
}
get_pair_prob_vector(p_pp, p_unpaired_tmp, length, 1);
free_pf_arrays();
}
/* We do the same grid search as in the Mathews paper Fig. 4*/
for (m=0.25; m <=7.0; m+=0.25) {
// Weird way of writing this inner loop for the grid search. We
// traverse the grid without big jumps in the parameters to make
// sure that the updated scaling factor is accurate all the time.
inc*=-1;
for (b = inc < 0.0 ? 0.0 : -3.0; inc < 0.0 ? b >= -3.0 : b<= 0.0 ; b+=inc) {
// calculate cut point with x-axis and skip parameter pairs
// which give a cut point outside the range of
// q_unpaired (0 to 1). They gave frequently overflows and the
// idea is that we both want positive and negative perturbations
cut = exp( (-1) * b / m ) - 1;
fprintf(stderr, "\nm = %.2f, b = %.2f, cut=%.2f\n", m, b, cut);
if (cut > 1.0 || cut < 0.01) {
fprintf(stderr, "\nSkipping m = %.2f, b = %.2f\n", m, b);
continue;
}
/* Mathew's constant perturbations */
if (initial_guess_method == 1) {
for (i=0; i <= length; i++) {
/* We add epsilon to unpaired regions (as opposed to
paired regions as in the Mathews paper) so we multiply
by -1; if missing data we set it to 0.0 */
if (q_unpaired[i] < -0.5) {
curr_epsilon[i] = 0.0;
} else {
curr_epsilon[i] = (m *(log(q_unpaired[i]+1))+b) *(-1);
}
gsl_vector_set (minimizer_x, i, curr_epsilon[i]);
}
/* Perturbations proportional to q-p */
} else {
for (i=0; i <= length; i++) {
curr_epsilon[i] = (m *(log(q_unpaired[i]+1)-log(p_unpaired_tmp[i]+1))+ b ) * (-1);
gsl_vector_set (minimizer_x, i, curr_epsilon[i]);
}
}
// Repeat and adjust scaling factor until we get result without over-/underflows
do {
// First we use default scaling factor
if (pf_underflow == 0 && pf_overflow == 0) {
sfact = 1.070;
}
if (pf_underflow) {
sfact *= 0.8;
fprintf(stderr,"Underflow, adjusting sfact to %.4f\n", sfact );
}
if (pf_overflow) {
sfact *= 1.2;
fprintf(stderr,"Overflow, adjusting sfact to %.4f\n", sfact );
}
pf_scale = exp(-(sfact*last_non_nan_lnQ)/kT/length);
//fprintf(stderr,"Scaling factor is now: %.4e\n", pf_scale);
curr_D = calculate_f(minimizer_x, (void*)&minimizer_pars);
if (!isnan(last_lnQ)) last_non_nan_lnQ = last_lnQ;
// Give up when even extreme scaling does not give results
// (for some reason I could not get rid of overflows even with high scaling factors)
if (sfact < 0.1 || sfact > 2.0) break;
} while (pf_underflow == 1 || pf_overflow == 1);
// We have not given up so everything is ok now
if (!(sfact < 0.1 || sfact > 2.0)) {
if (curr_D < min_D) {
min_D = curr_D;
for (i=0; i <= length; i++) {
best_epsilon[i] = curr_epsilon[i];
}
best_m = m;
best_b = b;
best_scale = pf_scale;
}
/*If we are interested in the grid search we misuse the
print_stats function and report m and b together with MEA*/
if (grid_search) {
for (i=0; i <= length; i++) {
epsilon[i] = curr_epsilon[i];
}
print_stats(statsfile, string, cstruc, length, 0, 0, m, 0.0, b, 0);
}
fprintf(stderr, "curr D: %.2f, minimum D: %.2f\n", curr_D, min_D);
// Adjust pf_scale with default scaling factor but lnQ from
// previous step
sfact = 1.070;
pf_scale = exp(-(sfact*last_lnQ)/kT/length);
} else {
sfact = 1.070;
fprintf(stderr, "Skipping m = %.2f, b = %.2f; did not get stable result.\n", m, b);
}
} // for b
} // for m
fprintf(stderr, "Minimum found: m=%.2f, b=%.2f: %.2f\n", best_m, best_b, min_D);
for (i=0; i <= length; i++) {
epsilon[i] = best_epsilon[i];
gsl_vector_set (minimizer_x, i, best_epsilon[i]);
}
pf_scale = best_scale;
}
if (initial_guess_method == 3) {
srand((unsigned)time(0));
for (i=0; i <= length; i++) {
double r = (double)rand()/(double)RAND_MAX * 4 - 2;
epsilon[i] = r;
gsl_vector_set (minimizer_x, i, epsilon[i]);
}
}
/* If we just wanted a grid search we are done now. */
if (grid_search) {
exit(0);
}
prev_D = calculate_f(minimizer_x, (void*)&minimizer_pars);
print_stats(statsfile, string, cstruc, length, 0 , count_df_evaluations , prev_D, -1.0, 0.0,1);
/* GSL minimization */
if (method_id !=0) {
if (method_id > 2) {
char name[100];
// Available algorithms
// 3 gsl_multimin_fdfminimizer_conjugate_fr
// 4 gsl_multimin_fdfminimizer_conjugate_pr
// 5 gsl_multimin_fdfminimizer_vector_bfgs
// 6 gsl_multimin_fdfminimizer_vector_bfgs2
// 7 gsl_multimin_fdfminimizer_steepest_descent
// http://www.gnu.org/software/gsl/manual/html_node/Multimin-Algorithms-with-Derivatives.html
switch (method_id) {
case 2:
minimizer = gsl_multimin_fdfminimizer_alloc (gsl_multimin_fdfminimizer_conjugate_fr, length+1);
strcpy(name, "Fletcher-Reeves conjugate gradient");
break;
case 3:
minimizer = gsl_multimin_fdfminimizer_alloc (gsl_multimin_fdfminimizer_conjugate_pr, length+1);
strcpy(name, "Polak-Ribiere conjugate gradient");
break;
case 4:
minimizer = gsl_multimin_fdfminimizer_alloc ( gsl_multimin_fdfminimizer_vector_bfgs, length+1);
strcpy(name, "Broyden-Fletcher-Goldfarb-Shanno");
break;
case 5:
minimizer = gsl_multimin_fdfminimizer_alloc ( gsl_multimin_fdfminimizer_vector_bfgs2, length+1);
strcpy(name, "Broyden-Fletcher-Goldfarb-Shanno (improved version)");
break;
case 6:
minimizer = gsl_multimin_fdfminimizer_alloc (gsl_multimin_fdfminimizer_steepest_descent, length+1);
strcpy(name, "Gradient descent (GSL implmementation)");
break;
}
fprintf(stderr, "Starting minimization via GSL implementation of %s...\n\n", name);
// The last two parmeters are step size and tolerance (with
// different meaning for different algorithms
gsl_multimin_fdfminimizer_set (minimizer, &minimizer_func, minimizer_x, initial_step_size, tolerance);
iteration = 1;
do {
status = gsl_multimin_fdfminimizer_iterate (minimizer);
D = minimizer->f;
norm = gsl_blas_dnrm2(minimizer->gradient);
print_stats(statsfile, string, cstruc, length,iteration, count_df_evaluations, D, prev_D, norm, iteration%sparsePS == 0);
prev_D = D;
if (status) {
fprintf(stderr, "An unexpected error has occured in the iteration (status:%i)\n", status);
break;
}
status = gsl_multimin_test_gradient (minimizer->gradient, precision);
if (status == GSL_SUCCESS) fprintf(stderr, "Minimum found stopping.\n");
iteration++;
} while (status == GSL_CONTINUE && iteration < max_iteration);
gsl_multimin_fdfminimizer_free (minimizer);
gsl_vector_free (minimizer_x);
/* Custom implementation of steepest descent */
} else {
if (method_id == 1) {
fprintf(stderr, "Starting custom implemented steepest descent search...\n\n");
} else {
fprintf(stderr, "Starting custom implemented steepest descent search with Barzilai Borwein step size...\n\n");
}
iteration = 0;
D = 0.0;
while (iteration++ < max_iteration) {
for (i=1; i <= length; i++) {
gsl_vector_set (minimizer_x, i, epsilon[i]);
}
D = calculate_f(minimizer_x, (void*)&minimizer_pars);
if (numerical) {
calculate_df_numerically(minimizer_x, (void*)&minimizer_pars, minimizer_g);
} else {
calculate_df(minimizer_x, (void*)&minimizer_pars, minimizer_g);
}
for (i=1; i <= length; i++) {
gradient[i] = gsl_vector_get (minimizer_g, i);
}
// Do line search
fprintf(stderr, "\nLine search:\n");
// After the first iteration, use Barzilai-Borwain (1988) step size (currently turned off)
if (iteration>1 && method_id==2) {
double denominator=0.0;
double numerator=0.0;
for (i=1; i <= length; i++) {
numerator += (epsilon[i]-prev_epsilon[i]) * (gradient[i]-prev_gradient[i]);
denominator+=(gradient[i]-prev_gradient[i]) * (gradient[i]-prev_gradient[i]);
}
step_size = numerator / denominator;
norm =1.0;
} else {
// Use step sized given by the user (normalize it first)
step_size = initial_step_size / calculate_norm(gradient, length);
}
for (i=1; i <= length; i++) {
prev_epsilon[i] = epsilon[i];
prev_gradient[i] = gradient[i];
}
do {
for (mu=1; mu <= length; mu++) {
epsilon[mu] = prev_epsilon[mu] - step_size * gradient[mu];
}
for (i=1; i <= length; i++) {
gsl_vector_set (minimizer_x, i, epsilon[i]);
}
DD = calculate_f(minimizer_x, (void*)&minimizer_pars);
if (step_size > 0.0001) {
fprintf(stderr, "Old D: %.4f; New D: %.4f; Step size: %.4f\n", D, DD, step_size);
} else {
fprintf(stderr, "Old D: %.4f; New D: %.4f; Step size: %.4e\n", D, DD, step_size);
}
step_size /= 2;
} while (step_size > 1e-12 && DD > D);
norm = calculate_norm(gradient,length);
if (DD > D) {
fprintf(stderr, "Line search did not improve D in iteration %i. Stop.\n", iteration);
if (hybrid_conditionals) {
sample_conditionals=0;
} else {
break;
}
}
print_stats(statsfile, string, cstruc, length,iteration, count_df_evaluations, DD, prev_D, norm, iteration%sparsePS == 0);
if (norm<precision && iteration>1) {
fprintf(stderr, "Minimum found stopping.\n");
break;
}
prev_D = DD;
}
}
/* Force last dotplot to be printed */
print_stats(statsfile, string, cstruc, length,iteration, count_df_evaluations, DD, prev_D, norm, 1);
}
free(pf_struc);
if (cstruc!=NULL) free(cstruc);
(void) fflush(stdout);
free(string);
free(structure);
RNAfold_cmdline_parser_free (&args_info);
return 0;
}
