void w_transform(wstr *in, wtrans_func f)
{
wstr t, *out = w_new();
f(in, out);
t = *in; *in = *out; *out = t;
w_del(out);
}
wstr *w_make(wchar_t *s)
{
int i, len = wcslen(s);
wstr *w = w_new();
for (i = 0; i < len; i++) w_append(w, s[i]);
return w;
}
static Weight *maw_new(Query *query, Searcher *searcher)
{
Weight *self = w_new(Weight, query);
self->scorer = &masc_new;
self->explain = &maw_explain;
self->to_s = &maw_to_s;
self->similarity = query->get_similarity(query, searcher);
self->idf = 1.0;
return self;
}
static Weight *fqw_new(Query *query, Weight *sub_weight, Similarity *sim)
{
Weight *self = w_new(FilteredQueryWeight, query);
FQW(self)->sub_weight = sub_weight;
self->get_value = &fqw_get_value;
self->normalize = &fqw_normalize;
self->scorer = &fqw_scorer;
self->explain = &fqw_explain;
self->to_s = &fqw_to_s;
self->destroy = &fqw_destroy;
self->sum_of_squared_weights = &fqw_sum_of_squared_weights;
self->similarity = sim;
self->idf = 1.0;
self->value = sub_weight->value;
return self;
}
Eigen::VectorXd l2r_l1hinge_spdc::train_warm_start(const Eigen::VectorXd &alp) {
set_alpha(alp);
alpha_ = Eigen::VectorXd::Constant(num_ins_, C);
// set_w_by_alpha(alpha_);
double alpha_old = 0.0;
std::default_random_engine g;
std::uniform_int_distribution<> uni_dist(0, num_ins_ - 1);
const auto ins_is_begin_it = std::begin(active_index_);
w_ = Eigen::VectorXd::Zero(num_fea_);
spdc_w_ = w_;
auto random_it = std::next(ins_is_begin_it, uni_dist(g));
gamma_ = 0.1;
double tau_ = std::sqrt(gamma_ / (1.0 * (1.0 / C))) / R_;
double sigma_ = std::sqrt(((1.0 / C) * 1.0) / gamma_) / R_;
double theta_ =
1.0 - 1.0 / ((1.0 / C) + R_ * std::sqrt((1.0 / C) / (1.0 * gamma_)));
// double tau_ = std::sqrt(1.0) / (2.0 * R_);
// double sigma_ = std::sqrt(1.0 / 1.0) / (2.0 * R_);
// double theta_ = 1.0 - 1.0 / (1.0 + (R_ / gamma_) * std::sqrt(1.0 / 1.0));
const double sig_gam = -sigma_ * gamma_ - 1.0;
int i = 0, itr_ = 0;
double delta_alpha_i = 0.0, yi = 0.0, beta_i = 0.0, alpha_i_new = 0.0,
eta = 0.0;
Eigen::VectorXd delta_v(num_fea_), w_new(num_fea_);
calculate_duality_gap(true, true);
std::cout << tau_ << " " << sigma_ << " " << theta_ << " " << R_ << std::endl;
std::cout << " start optimization " << max_iteration << std::endl;
for (itr_ = 1; itr_ < max_iteration && duality_gap_ > 1e-6; ++itr_) {
for (int ir = 0; ir < num_ins_; ++ir) {
random_it = std::next(ins_is_begin_it, uni_dist(g));
i = *random_it;
yi = y_[i];
// beta_i = -(sigma_ * (yi * (x_.row(i) * spdc_w_)(0) - 1.0)) + alpha_[i];
beta_i = (sigma_ * (yi * (x_.row(i) * spdc_w_)(0) - 1.0) - alpha_[i]) /
(sig_gam);
alpha_i_new = std::min(C, std::max(0.0, beta_i));
delta_alpha_i = alpha_i_new - alpha_[i];
alpha_[i] = alpha_i_new;
delta_v = (yi * delta_alpha_i) * x_.row(i);
// w_new = (1.0 / (1.0 + tau_)) * (w_ - tau_ * (za_ - delta_v));
for (int j = 0; j < num_fea_; ++j) {
eta = (1.0 / (1.0 + tau_)) * (w_[j] + tau_ * (za_[j] + delta_v[j]));
spdc_w_[j] = eta + theta_ * (eta - w_[j]);
w_[j] = eta;
}
// spdc_w_ = w_new + theta_ * (w_new - w_);
// w_ = w_new;
// for (srm_iit it(x_, i); it; ++it)
// za_[it.index()] += delta_v[it.index()];
za_ += delta_v;
}
// if (itr_) {
calculate_duality_gap(true, true);
std::cout << itr_ << " optimization end gap : " << duality_gap_ << " "
<< primal_obj_value_ << " " << dual_obj_value_ << std::endl;
}
std::cout << itr_ << " optimization end gap : " << duality_gap_ << " "
<< primal_obj_value_ << " " << dual_obj_value_ << std::endl;
std::cout << w_.transpose() << std::endl;
return w_;
}
