bfgs.h File Reference

Go to the source code of this file.

Functions
LEARNER::base_learner *	bfgs_setup (VW::config::options_i &options, vw &all)

Function Documentation

bfgs_setup()

LEARNER::base_learner* bfgs_setup	(	VW::config::options_i &	options,
		vw &	all
	)

Definition at line 1093 of file bfgs.cc.

References VW::config::option_group_definition::add(), VW::config::options_i::add_and_parse(), vw::audit, vw::bfgs, LEARNER::end_pass(), VW::config::options_i::get_typed_option(), vw::hash_inv, vw::hessian_on, shared_data::holdout_best_loss, vw::holdout_set_off, init_driver(), LEARNER::init_learner(), LEARNER::make_base(), VW::config::make_option(), vw::numpasses, vw::quiet, save_load(), vw::sd, parameters::stride(), parameters::stride_shift(), THROW, vw::training, and vw::weights.

Referenced by parse_reductions().

{
  auto b = scoped_calloc_or_throw<bfgs>();
  bool conjugate_gradient = false;
  bool bfgs_option = false;
  option_group_definition bfgs_outer_options("LBFGS and Conjugate Gradient options");
  bfgs_outer_options.add(
      make_option("conjugate_gradient", conjugate_gradient).keep().help("use conjugate gradient based optimization"));

  option_group_definition bfgs_inner_options("LBFGS and Conjugate Gradient options");
  bfgs_inner_options.add(make_option("bfgs", bfgs_option).keep().help("use conjugate gradient based optimization"));
  bfgs_inner_options.add(make_option("hessian_on", all.hessian_on).help("use second derivative in line search"));
  bfgs_inner_options.add(make_option("mem", b->m).default_value(15).help("memory in bfgs"));
  bfgs_inner_options.add(
      make_option("termination", b->rel_threshold).default_value(0.001f).help("Termination threshold"));

  options.add_and_parse(bfgs_outer_options);
  if (!conjugate_gradient)
  {
    options.add_and_parse(bfgs_inner_options);
    if (!bfgs_option)
    {
      return nullptr;
    }
  }

  b->all = &all;
  b->wolfe1_bound = 0.01;
  b->first_hessian_on = true;
  b->first_pass = true;
  b->gradient_pass = true;
  b->preconditioner_pass = true;
  b->backstep_on = false;
  b->final_pass = all.numpasses;
  b->no_win_counter = 0;

  if (!all.holdout_set_off)
  {
    all.sd->holdout_best_loss = FLT_MAX;
    b->early_stop_thres = options.get_typed_option<size_t>("early_terminate").value();
  }

  if (b->m == 0)
    all.hessian_on = true;

  if (!all.quiet)
  {
    if (b->m > 0)
      b->all->trace_message << "enabling BFGS based optimization ";
    else
      b->all->trace_message << "enabling conjugate gradient optimization via BFGS ";
    if (all.hessian_on)
      b->all->trace_message << "with curvature calculation" << std::endl;
    else
      b->all->trace_message << "**without** curvature calculation" << std::endl;
  }

  if (all.numpasses < 2 && all.training)
    THROW("you must make at least 2 passes to use BFGS");

  all.bfgs = true;
  all.weights.stride_shift(2);

  void (*learn_ptr)(bfgs&, base_learner&, example&) = nullptr;
  if (all.audit)
    learn_ptr = learn<true>;
  else
    learn_ptr = learn<false>;

  learner<bfgs, example>* l;
  if (all.audit || all.hash_inv)
    l = &init_learner(b, learn_ptr, predict<true>, all.weights.stride());
  else
    l = &init_learner(b, learn_ptr, predict<false>, all.weights.stride());

  l->set_save_load(save_load);
  l->set_init_driver(init_driver);
  l->set_end_pass(end_pass);

  return make_base(*l);
}