int main(int argc, char const *argv[])
{
std::string FILE="/homes/kgori/scratch/basic_pll/GGPS1.phy";
std::string PART="/homes/kgori/scratch/basic_pll/GGPS1.partitions.txt";
std::string TREE="/homes/kgori/scratch/basic_pll/GGPS1.tree";
auto inst = make_unique<pll>(FILE, PART, TREE, 1, 123);
std::cout << inst->get_likelihood() << std::endl;
std::cout << inst->get_partition_name(0) << std::endl;
std::cout << inst->get_model_name(0) << std::endl;
std::cout << inst->get_epsilon() << std::endl;
std::cout << inst->get_alpha(0) << std::endl;
std::cout << inst->get_number_of_partitions() << std::endl;
std::cout << inst->get_tree() << std::endl;
std::cout << inst->get_frac_change() << std::endl;
inst->set_tree("((Ptro:0.00147084048849247711,Ppan:0.00106294370976534763):0.00444598321816729036,(Cjac:0.05157798212603657839,(Csab:0.00440006365327790666,(Mmul:0.00652936575529242547,Panu:0.00101047194512476272):0.00381049900890796569):0.01359968787254225639):0.01375788728353777995,Hsap:0.00674547067186638278):0.0;");
inst->set_epsilon(0.001);
inst->set_alpha(2, 0, true);
auto ef = inst->get_empirical_frequencies();
inst->optimise(true, true, true, true);
inst->optimise_tree_search(true);
std::cout << inst->get_likelihood() << std::endl;
std::cout << inst->get_tree() << std::endl;
return 0;
}
void update_delta(const Expression& gradient, int index) {
if (index >= _delta.size() || index >= _d2.size() || index >= _delta2.size()) {
_delta.resize(index + 1);
_delta2.resize(index + 1);
_d2.resize(index + 1);
}
if (!_delta[index] || !_delta2[index] || !_d2[index]) {
_delta[index] = boost::make_shared<vector_t>(zeros<value_t>(gradient.count()));
_delta2[index] = boost::make_shared<vector_t>(zeros<value_t>(gradient.count()));
_d2[index] = boost::make_shared<vector_t>(zeros<value_t>(gradient.count()));
}
// Perform adadelta updates
*_d2[index] = get_decay_rate() * *_d2[index] + (1.0 - get_decay_rate()) * reshape(gradient, _d2[index]->size()) * reshape(gradient, _d2[index]->size());
*_delta[index] = -sqrt(*_delta2[index] + get_epsilon()) / sqrt(*_d2[index] + get_epsilon()) * reshape(gradient, _d2[index]->size());
*_delta2[index] = get_decay_rate() * *_delta2[index] + (1.0 - get_decay_rate()) * *_delta[index] * *_delta[index];
}
double MultivariateFNormalSufficient::get_mean_square_residuals() const
{
//std::cout << "P " << std::endl << P_ << std::endl;
//std::cout << "epsilon " << std::endl << get_epsilon() << std::endl;
VectorXd Peps(get_Peps());
VectorXd epsilon(get_epsilon());
double dist = epsilon.transpose()*Peps;
LOG( "MVN: get_mean_square_residuals = " << dist << std::endl);
return dist;
}
VectorXd MultivariateFNormalSufficient::get_Peps() const
{
if (!flag_Peps_)
{
////Peps
LOG( "MVN: solving for P*epsilon" << std::endl);
const_cast<MultivariateFNormalSufficient *>(this)
->set_Peps(get_ldlt().solve(get_epsilon()));
}
return Peps_;
}
double __declspec(dllexport) WINAPI _get_epsilon(lprec *lp)
{
double ret;
if (lp != NULL) {
freebuferror();
ret = (double) get_epsilon(lp);
}
else
ret = 0;
return(ret);
}
VectorXd MultivariateFNormalSufficient::get_Peps() const
{
if (!flag_Peps_)
{
////Peps
timer_.start(SOLVE);
IMP_LOG(TERSE, "MVN: solving for P*epsilon" << std::endl);
const_cast<MultivariateFNormalSufficient *>(this)
->set_Peps(get_ldlt().solve(get_epsilon()));
timer_.stop(SOLVE);
}
return Peps_;
}
void update_delta(const Expression& gradient, int index) {
if (index >= _delta.size()) {
_delta.resize(index + 1);
}
if (!_delta[index]) {
_delta[index] = boost::make_shared<vector_t>(zeros<value_t>(gradient.count()));
}
// Block-wise normalize the gradient
vector_t norm_grad = reshape(gradient, _delta[index]->size()) / (sqrt(dot(reshape(gradient, _delta[index]->size()), reshape(gradient, _delta[index]->size()))) + get_epsilon());
// Calculate running averages
mg = get_decay_rate() * mg + (1.0 - get_decay_rate()) * norm_grad;
gamma_nume_sqr = gamma_nume_sqr * (1 - 1 / taus_x_t) + (norm_grad - old_grad) * (old_grad - mg) * (norm_grad - old_grad) * (old_grad - mg) / taus_x_t;
gamma_deno_sqr = gamma_dneo_sqr * (1 - 1 / taus_x_t) + (mg - norm_grad) * (old_grad - mg) * (mg - norm_grad) * (old_grad - mg) / taus_x_t;
_gamma = sqrt(_gamma_nume) / (sqrt(_gamma_deno_sqr) + get_epsilon());
// Update delta with momentum and epsilon parameter
*_delta[index] = _gamma * mg;
}
void update_delta(const Expression& gradient, int index) {
if (index >= _delta.size() || index >= _old_grad.size() || index >= _g.size() || index >= _g2.size() ||
index >= _h.size() || index >= _h2.size() || index >= _tau.size() || index >= _current_iteration.size())
{
_delta.resize(index + 1);
_old_grad.resize(index + 1);
_g.resize(index + 1);
_g2.resize(index + 1);
_h.resize(index + 1);
_h2.resize(index + 1);
_tau.resize(index + 1);
_current_iteration.resize(index + 1);
}
if (!_delta[index] || !_old_grad[index] || !_g[index] || !_g2[index] || !_h[index] || !_h2[index] || !_tau[index]) {
_delta[index] = boost::make_shared<vector_t>(zeros<value_t>(gradient.count()));
_old_grad[index] = boost::make_shared<vector_t>(zeros<value_t>(gradient.count()));
_g[index] = boost::make_shared<vector_t>(zeros<value_t>(gradient.count()));
_g2[index] = boost::make_shared<vector_t>(zeros<value_t>(gradient.count()));
_h[index] = boost::make_shared<vector_t>(zeros<value_t>(gradient.count()));
_h2[index] = boost::make_shared<vector_t>(zeros<value_t>(gradient.count()));
_tau[index] = boost::make_shared<vector_t>(ones<value_t>(gradient.count()));
_current_iteration[index] = 0;
}
++_current_iteration[index];
// Detect outlier
// if (abs(reshape(gradient, _delta[index]->size()) - *_g[index]) > 2 * sqrt(*_g2[index] - *_g[index] * *_g[index]) ||
// abs(abs((reshape(gradient, _delta[index]->size()) - *_old_grad[index]) / (*_delta[index] + get_epsilon())) - *_h[index]) > 2 * sqrt(*_h2[index] - *_h[index] * *_h[index]))
// {
// *_tau[index] = *_tau[index] + 1.0;
// }
*_tau[index] = *_tau[index] + abs(reshape(gradient, _delta[index]->size()) - *_g[index]) > 2 * sqrt(*_g2[index] - *_g[index] * *_g[index]);
// Update moving averages
*_g[index] = (1.0 - 1.0 / *_tau[index]) * *_g[index] + 1.0 / *_tau[index] * reshape(gradient, _delta[index]->size());
*_g2[index] = (1.0 - 1.0 / *_tau[index]) * *_g2[index] + 1.0 / *_tau[index] * reshape(gradient, _delta[index]->size()) * reshape(gradient, _delta[index]->size());
if (_current_iteration[index] > 1) {
// Calculate h and do h updates
// diag(H) = abs((reshape(gradient, _delta[index]->size()) - *_old_grad[index]) / (*_delta[index] + get_epsilon()));
*_h[index] = (1.0 - 1.0 / *_tau[index]) * *_h[index] + 1.0 / *_tau[index] * abs((reshape(gradient, _delta[index]->size()) - *_old_grad[index]) / (*_delta[index] + get_epsilon()));
*_h2[index] = (1.0 - 1.0 / *_tau[index]) * *_h[index] + 1.0 / *_tau[index] * ((reshape(gradient, _delta[index]->size()) - *_old_grad[index]) / (*_delta[index] + get_epsilon())) * ((reshape(gradient, _delta[index]->size()) - *_old_grad[index]) / (*_delta[index] + get_epsilon()));
// Initialization phase -> multiply with C where C = D/10
if (_current_iteration[index] == 2) {
*_g2[index] = *_g2[index] * get_c();
*_h[index] = *_h[index] * get_c();
*_h2[index] = *_h2[index] * get_c();
}
*_delta[index] = -*_h[index] * *_g[index] * *_g[index] / (*_h2[index] * *_g2[index] + get_epsilon()) * reshape(gradient, _delta[index]->size());
} else {
*_delta[index] = get_epsilon() * *_g[index];
}
*_tau[index] = (1.0 - *_g[index] * *_g[index] / (*_g2[index] + get_epsilon())) * *_tau[index] + 1;
*_old_grad[index] = reshape(gradient, _delta[index]->size());
}
void update_delta(const Expression& gradient, int index) {
if (index >= _delta.size() || index >= _first_moment.size() || index >= _second_moment.size() || index >= _current_iteration.size()) {
_delta.resize(index + 1);
_first_moment.resize(index + 1);
_second_moment.resize(index + 1);
_current_iteration.resize(index + 1);
}
if (!_delta[index] || !_first_moment[index] || !_second_moment[index]) {
_delta[index] = boost::make_shared<vector_t>(zeros<value_t>(gradient.count()));
_first_moment[index] = boost::make_shared<vector_t>(zeros<value_t>(gradient.count()));
_second_moment[index] = boost::make_shared<vector_t>(zeros<value_t>(gradient.count()));
_current_iteration[index] = 0;
}
++_current_iteration[index];
// Perform adadelta updates
*_first_moment[index] = get_beta1() * reshape(gradient, _delta[index]->size()) + (1.0 - get_beta1()) * *_first_moment[index];
*_second_moment[index] = get_beta2() * reshape(gradient, _delta[index]->size()) * reshape(gradient, _delta[index]->size()) + (1.0 - get_beta2()) * *_second_moment[index];
*_delta[index] = -get_alpha() * *_first_moment[index] / (value_t(1) - ::pow(value_t(1) - get_beta1(), _current_iteration[index])) /
(sqrt(*_second_moment[index] / (value_t(1) - ::pow(value_t(1) - get_beta2(), _current_iteration[index]))) + get_epsilon());
}