EXPLORE_EVAL Namespace Reference

Classes
struct	explore_eval

Functions
void	finish (explore_eval &data)
void	output_example (vw &all, explore_eval &c, example &ec, multi_ex *ec_seq)
void	output_example_seq (vw &all, explore_eval &data, multi_ex &ec_seq)
void	finish_multiline_example (vw &all, explore_eval &data, multi_ex &ec_seq)
template<bool is_learn>
void	do_actual_learning (explore_eval &data, multi_learner &base, multi_ex &ec_seq)

Function Documentation

do_actual_learning()

template<bool is_learn>

void EXPLORE_EVAL::do_actual_learning	(	explore_eval &	data,
		multi_learner &	base,
		multi_ex &	ec_seq
	)

Definition at line 124 of file explore_eval.cc.

References EXPLORE_EVAL::explore_eval::_random_state, CB::cb_class::action, EXPLORE_EVAL::explore_eval::action_label, polylabel::cb, CB::label::costs, EXPLORE_EVAL::explore_eval::empty_label, EXPLORE_EVAL::explore_eval::fixed_multiplier, CB_ADF::get_observed_cost(), EXPLORE_EVAL::explore_eval::known_cost, example::l, EXPLORE_EVAL::explore_eval::multiplier, EXPLORE_EVAL::explore_eval::offset, CB::cb_class::probability, v_array< T >::size(), CB_ADF::test_adf_sequence(), EXPLORE_EVAL::explore_eval::update_count, EXPLORE_EVAL::explore_eval::violations, and example::weight.

{
  example* label_example = CB_ADF::test_adf_sequence(ec_seq);

  if (label_example != nullptr)  // extract label
  {
    data.action_label = label_example->l.cb;
    label_example->l.cb = data.empty_label;
  }
  multiline_learn_or_predict<false>(base, ec_seq, data.offset);

  if (label_example != nullptr)  // restore label
    label_example->l.cb = data.action_label;

  data.known_cost = CB_ADF::get_observed_cost(ec_seq);
  if (label_example != nullptr && is_learn)
  {
    ACTION_SCORE::action_scores& a_s = ec_seq[0]->pred.a_s;

    float action_probability = 0;
    for (size_t i = 0; i < a_s.size(); i++)
      if (data.known_cost.action == a_s[i].action)
        action_probability = a_s[i].score;

    float threshold = action_probability / data.known_cost.probability;

    if (!data.fixed_multiplier)
      data.multiplier = std::min(data.multiplier, 1 / threshold);
    else
      threshold *= data.multiplier;

    if (threshold > 1. + 1e-6)
      data.violations++;

    if (data._random_state->get_and_update_random() < threshold)
    {
      example* ec_found = nullptr;
      for (example*& ec : ec_seq)
      {
        if (ec->l.cb.costs.size() == 1 && ec->l.cb.costs[0].cost != FLT_MAX && ec->l.cb.costs[0].probability > 0)
          ec_found = ec;
        if (threshold > 1)
          ec->weight *= threshold;
      }
      ec_found->l.cb.costs[0].probability = action_probability;

      multiline_learn_or_predict<true>(base, ec_seq, data.offset);

      if (threshold > 1)
      {
        float inv_threshold = 1.f / threshold;
        for (auto& ec : ec_seq) ec->weight *= inv_threshold;
      }
      ec_found->l.cb.costs[0].probability = data.known_cost.probability;
      data.update_count++;
    }
  }
}

finish()

void EXPLORE_EVAL::finish

(

explore_eval &

data

)

Definition at line 36 of file explore_eval.cc.

References EXPLORE_EVAL::explore_eval::all, EXPLORE_EVAL::explore_eval::fixed_multiplier, EXPLORE_EVAL::explore_eval::multiplier, vw::quiet, vw::trace_message, EXPLORE_EVAL::explore_eval::update_count, and EXPLORE_EVAL::explore_eval::violations.

Referenced by explore_eval_setup().

{
  if (!data.all->quiet)
  {
    data.all->trace_message << "update count = " << data.update_count << std::endl;
    if (data.violations > 0)
      data.all->trace_message << "violation count = " << data.violations << std::endl;
    if (!data.fixed_multiplier)
      data.all->trace_message << "final multiplier = " << data.multiplier << std::endl;
  }
}

finish_multiline_example()

void EXPLORE_EVAL::finish_multiline_example	(	vw &	all,
		explore_eval &	data,
		multi_ex &	ec_seq
	)

Definition at line 113 of file explore_eval.cc.

References vw::final_prediction_sink, VW::finish_example(), CB_ADF::global_print_newline(), and output_example_seq().

{
  if (ec_seq.size() > 0)
  {
    output_example_seq(all, data, ec_seq);
    CB_ADF::global_print_newline(all.final_prediction_sink);
  }
  VW::finish_example(all, ec_seq);
}

output_example()

void EXPLORE_EVAL::output_example	(	vw &	all,
		explore_eval &	c,
		example &	ec,
		multi_ex *	ec_seq
	)

Definition at line 52 of file explore_eval.cc.

References polyprediction::a_s, polylabel::cb, CB::label::costs, CB::ec_is_example_header(), LEARNER::example_is_newline_not_header(), vw::final_prediction_sink, CB_ALGS::get_cost_estimate(), EXPLORE_EVAL::explore_eval::known_cost, example::l, loss(), example::pred, ACTION_SCORE::print_action_score(), vw::print_text, CB::print_update(), CB::cb_class::probability, vw::raw_prediction, vw::sd, v_array< T >::size(), example::tag, shared_data::update(), and example::weight.

Referenced by output_example_seq().

{
  if (example_is_newline_not_header(ec))
    return;

  size_t num_features = 0;

  float loss = 0.;
  ACTION_SCORE::action_scores preds = (*ec_seq)[0]->pred.a_s;

  for (size_t i = 0; i < (*ec_seq).size(); i++)
    if (!CB::ec_is_example_header(*(*ec_seq)[i]))
      num_features += (*ec_seq)[i]->num_features;

  bool labeled_example = true;
  if (c.known_cost.probability > 0)
  {
    for (uint32_t i = 0; i < preds.size(); i++)
    {
      float l = get_cost_estimate(&c.known_cost, preds[i].action);
      loss += l * preds[i].score;
    }
  }
  else
    labeled_example = false;

  bool holdout_example = labeled_example;
  for (size_t i = 0; i < ec_seq->size(); i++) holdout_example &= (*ec_seq)[i]->test_only;

  all.sd->update(holdout_example, labeled_example, loss, ec.weight, num_features);

  for (int sink : all.final_prediction_sink) print_action_score(sink, ec.pred.a_s, ec.tag);

  if (all.raw_prediction > 0)
  {
    std::string outputString;
    std::stringstream outputStringStream(outputString);
    v_array<CB::cb_class> costs = ec.l.cb.costs;

    for (size_t i = 0; i < costs.size(); i++)
    {
      if (i > 0)
        outputStringStream << ' ';
      outputStringStream << costs[i].action << ':' << costs[i].partial_prediction;
    }
    all.print_text(all.raw_prediction, outputStringStream.str(), ec.tag);
  }

  CB::print_update(all, !labeled_example, ec, ec_seq, true);
}

output_example_seq()

void EXPLORE_EVAL::output_example_seq	(	vw &	all,
		explore_eval &	data,
		multi_ex &	ec_seq
	)

Definition at line 103 of file explore_eval.cc.

References output_example(), vw::print_text, and vw::raw_prediction.

{
  if (ec_seq.size() > 0)
  {
    output_example(all, data, **(ec_seq.begin()), &(ec_seq));
    if (all.raw_prediction > 0)
      all.print_text(all.raw_prediction, "", ec_seq[0]->tag);
  }
}