void RateFree::setVariables(double *variables) {
if (getNDim() == 0) return;
int i;
// Modified by Thomas on 13 May 2015
// --start--
/*
variables[1] = prop[0];
for (i = 2; i < ncategory; i++)
variables[i] = variables[i-1] + prop[i-1];
*/
if (optimizing_params == 2) {
// proportions
for (i = 0; i < ncategory-1; i++)
variables[i+1] = prop[i] / prop[ncategory-1];
} else if (optimizing_params == 1) {
// rates
for (i = 0; i < ncategory-1; i++)
variables[i+1] = rates[i];
} else {
// both rates and weights
for (i = 0; i < ncategory-1; i++)
variables[i+1] = prop[i] / prop[ncategory-1];
for (i = 0; i < ncategory-1; i++)
variables[i+ncategory] = rates[i] / rates[ncategory-1];
}
}
bool RateGamma::getVariables(double *variables) {
if (getNDim() == 0) return false;
bool changed = (gamma_shape != variables[1]);
gamma_shape = variables[1];
if (changed)
computeRates();
return changed;
}
int getStrides( void* arr, int* strides ) {
npy_intp* np_dims = PyArray_STRIDES( arr );
int i;
for( i=0; i<getNDim( arr ); i++ ) {
strides[i] = (int)np_dims[i];
}
return 0;
}
int getSize( void* arr ) {
npy_intp* np_dims = PyArray_DIMS( arr );
int i, size = 1, nd = getNDim( arr );
for( i=0; i<nd; i++ ) {
size = size * (int)np_dims[i];
}
return size;
}
int getDims( void* arr, int* dims ) {
npy_intp* np_dims = PyArray_DIMS( arr );
int i;
for( i=0; i<getNDim( arr ); i++ ) {
dims[i] = (int)np_dims[i];
}
return 0;
}
double RateGammaInvar::optimizeParameters(double gradient_epsilon) {
int ndim = getNDim();
// return if nothing to be optimized
if (ndim == 0)
return phylo_tree->computeLikelihood();
if (verbose_mode >= VB_MED)
cout << "Optimizing " << name << " model parameters by " << optimize_alg << " algorithm..." << endl;
if (optimize_alg.find("EM_RR") != string::npos) {
return randomRestartOptimization(gradient_epsilon);
} else if (optimize_alg.find("Brent") != string::npos || phylo_tree->aln->frac_const_sites == 0.0 || isFixPInvar() || isFixGammaShape()) {
double lh = phylo_tree->computeLikelihood();
cur_optimize = 0;
double gamma_lh = RateGamma::optimizeParameters(gradient_epsilon);
ASSERT(gamma_lh >= lh-0.1);
cur_optimize = 1;
double invar_lh = -DBL_MAX;
invar_lh = RateInvar::optimizeParameters(gradient_epsilon);
ASSERT(invar_lh >= gamma_lh-0.1);
cur_optimize = 0;
return invar_lh;
} else if (optimize_alg.find("EM") != string::npos) {
return optimizeWithEM(gradient_epsilon);
} else if (optimize_alg.find("BFGS") != string::npos) {
//if (freq_type == FREQ_ESTIMATE) scaleStateFreq(false);
double *variables = new double[ndim+1];
double *upper_bound = new double[ndim+1];
double *lower_bound = new double[ndim+1];
bool *bound_check = new bool[ndim+1];
double score;
// by BFGS algorithm
setVariables(variables);
setBounds(lower_bound, upper_bound, bound_check);
score = -minimizeMultiDimen(variables, ndim, lower_bound, upper_bound, bound_check, max(gradient_epsilon, TOL_GAMMA_SHAPE));
getVariables(variables);
phylo_tree->clearAllPartialLH();
score = phylo_tree->computeLikelihood();
delete [] bound_check;
delete [] lower_bound;
delete [] upper_bound;
delete [] variables;
return score;
} else {
string errMsg = "Unknown optimization algorithm: " + optimize_alg;
outError(errMsg.c_str());
return 0.0;
}
}
double NGSRateCat::optimizeParameters(double epsilon) {
int ndim = getNDim();
// return if nothing to be optimized
if (ndim == 0) return 0.0;
cout << "Optimizing " << name << " model parameters..." << endl;
double *variables = new double[ndim+1];
double *upper_bound = new double[ndim+1];
double *lower_bound = new double[ndim+1];
bool *bound_check = new bool[ndim+1];
int i;
double score;
// by BFGS algorithm
setVariables(variables);
for (i = 1; i <= ndim; i++) {
//cout << variables[i] << endl;
lower_bound[i] = 1e-4;
upper_bound[i] = 100.0;
bound_check[i] = false;
}
for (i = ndim-ncategory+2; i <= ndim; i++)
upper_bound[i] = 1.0;
//packData(variables, lower_bound, upper_bound, bound_check);
score = -minimizeMultiDimen(variables, ndim, lower_bound, upper_bound, bound_check, max(epsilon, 1e-6));
getVariables(variables);
delete [] bound_check;
delete [] lower_bound;
delete [] upper_bound;
delete [] variables;
return score;
}
void RateFree::setBounds(double *lower_bound, double *upper_bound, bool *bound_check) {
if (getNDim() == 0) return;
int i;
if (optimizing_params == 2) {
// proportions
for (i = 1; i < ncategory; i++) {
lower_bound[i] = MIN_FREE_RATE_PROP;
upper_bound[i] = MAX_FREE_RATE_PROP;
bound_check[i] = false;
}
} else if (optimizing_params == 1){
// rates
for (i = 1; i < ncategory; i++) {
lower_bound[i] = MIN_FREE_RATE;
upper_bound[i] = MAX_FREE_RATE;
bound_check[i] = false;
}
} else {
// both weights and rates
for (i = 1; i < ncategory; i++) {
lower_bound[i] = MIN_FREE_RATE_PROP;
upper_bound[i] = MAX_FREE_RATE_PROP;
bound_check[i] = false;
}
for (i = 1; i < ncategory; i++) {
lower_bound[i+ncategory-1] = MIN_FREE_RATE;
upper_bound[i+ncategory-1] = MAX_FREE_RATE;
bound_check[i+ncategory-1] = false;
}
}
// for (i = ncategory; i <= 2*ncategory-2; i++) {
// lower_bound[i] = MIN_FREE_RATE;
// upper_bound[i] = MAX_FREE_RATE;
// bound_check[i] = false;
// }
}
void RateInvar::setBounds(double *lower_bound, double *upper_bound, bool *bound_check) {
if (getNDim() == 0) return;
lower_bound[1] = MIN_PINVAR;
upper_bound[1] = phylo_tree->aln->frac_const_sites;
bound_check[1] = false;
}
void RateGamma::setVariables(double *variables) {
if (getNDim() == 0) return;
variables[1] = gamma_shape;
}
void RateGamma::setBounds(double *lower_bound, double *upper_bound, bool *bound_check) {
if (getNDim() == 0) return;
lower_bound[1] = phylo_tree->params->min_gamma_shape;
upper_bound[1] = MAX_GAMMA_SHAPE;
bound_check[1] = false;
}
bool RateFree::getVariables(double *variables) {
if (getNDim() == 0) return false;
int i;
bool changed = false;
// Modified by Thomas on 13 May 2015
// --start--
/*
double *y = new double[2*ncategory+1];
double *z = y+ncategory+1;
// site proportions: y[0..c] <-> (0.0, variables[1..c-1], 1.0)
y[0] = 0; y[ncategory] = 1.0;
memcpy(y+1, variables+1, (ncategory-1) * sizeof(double));
std::sort(y+1, y+ncategory);
// category rates: z[0..c-1] <-> (variables[c..2*c-2], 1.0)
memcpy(z, variables+ncategory, (ncategory-1) * sizeof(double));
z[ncategory-1] = 1.0;
//std::sort(z, z+ncategory-1);
double sum = 0.0;
for (i = 0; i < ncategory; i++) {
prop[i] = (y[i+1]-y[i]);
sum += prop[i] * z[i];
}
for (i = 0; i < ncategory; i++) {
rates[i] = z[i] / sum;
}
delete [] y;
*/
double sum = 1.0;
if (optimizing_params == 2) {
// proportions
for (i = 0; i < ncategory-1; i++) {
sum += variables[i+1];
}
for (i = 0; i < ncategory-1; i++) {
changed |= (prop[i] != variables[i+1] / sum);
prop[i] = variables[i+1] / sum;
}
changed |= (prop[ncategory-1] != 1.0 / sum);
prop[ncategory-1] = 1.0 / sum;
// added by Thomas on Sept 10, 15
// update the values of rates, in order to
// maintain the sum of prop[i]*rates[i] = 1
// sum = 0;
// for (i = 0; i < ncategory; i++) {
// sum += prop[i] * rates[i];
// }
// for (i = 0; i < ncategory; i++) {
// rates[i] = rates[i] / sum;
// }
} else if (optimizing_params == 1) {
// rates
for (i = 0; i < ncategory-1; i++) {
changed |= (rates[i] != variables[i+1]);
rates[i] = variables[i+1];
}
// added by Thomas on Sept 10, 15
// need to normalize the values of rates, in order to
// maintain the sum of prop[i]*rates[i] = 1
// sum = 0;
// for (i = 0; i < ncategory; i++) {
// sum += prop[i] * rates[i];
// }
// for (i = 0; i < ncategory; i++) {
// rates[i] = rates[i] / sum;
// }
} else {
// both weights and rates
for (i = 0; i < ncategory-1; i++) {
sum += variables[i+1];
}
for (i = 0; i < ncategory-1; i++) {
changed |= (prop[i] != variables[i+1] / sum);
prop[i] = variables[i+1] / sum;
}
changed |= (prop[ncategory-1] != 1.0 / sum);
prop[ncategory-1] = 1.0 / sum;
// then rates
sum = prop[ncategory-1];
for (i = 0; i < ncategory-1; i++) {
sum += prop[i] * variables[i+ncategory];
}
for (i = 0; i < ncategory-1; i++) {
changed |= (rates[i] != variables[i+ncategory] / sum);
rates[i] = variables[i+ncategory] / sum;
}
changed |= (rates[ncategory-1] != 1.0 / sum);
rates[ncategory-1] = 1.0 / sum;
}
// --end--
return changed;
}
/**
optimize parameters. Default is to optimize gamma shape
@return the best likelihood
*/
double RateFree::optimizeParameters(double gradient_epsilon) {
int ndim = getNDim();
// return if nothing to be optimized
if (ndim == 0)
return phylo_tree->computeLikelihood();
if (verbose_mode >= VB_MED)
cout << "Optimizing " << name << " model parameters by " << optimize_alg << " algorithm..." << endl;
// TODO: turn off EM algorithm for +ASC model
if ((optimize_alg.find("EM") != string::npos && phylo_tree->getModelFactory()->unobserved_ptns.empty()))
if (fix_params == 0)
return optimizeWithEM();
//if (freq_type == FREQ_ESTIMATE) scaleStateFreq(false);
double *variables = new double[ndim+1];
double *upper_bound = new double[ndim+1];
double *lower_bound = new double[ndim+1];
bool *bound_check = new bool[ndim+1];
double score;
// score = optimizeWeights();
int left = 1, right = 2;
if (fix_params == 1) // fix proportions
right = 1;
if (optimize_alg.find("1-BFGS") != string::npos) {
left = 0;
right = 0;
}
// changed to Wi -> Ri by Thomas on Sept 11, 15
for (optimizing_params = right; optimizing_params >= left; optimizing_params--) {
ndim = getNDim();
// by BFGS algorithm
setVariables(variables);
setBounds(lower_bound, upper_bound, bound_check);
// if (optimizing_params == 2 && optimize_alg.find("-EM") != string::npos)
// score = optimizeWeights();
// else
if (optimize_alg.find("BFGS-B") != string::npos)
score = -L_BFGS_B(ndim, variables+1, lower_bound+1, upper_bound+1, max(gradient_epsilon, TOL_FREE_RATE));
else
score = -minimizeMultiDimen(variables, ndim, lower_bound, upper_bound, bound_check, max(gradient_epsilon, TOL_FREE_RATE));
getVariables(variables);
// sort the rates in increasing order
if (sorted_rates)
quicksort(rates, 0, ncategory-1, prop);
phylo_tree->clearAllPartialLH();
score = phylo_tree->computeLikelihood();
}
optimizing_params = 0;
delete [] bound_check;
delete [] lower_bound;
delete [] upper_bound;
delete [] variables;
return score;
}
