VW::cb_sample_data Struct Reference

Public Member Functions
	cb_sample_data (std::shared_ptr< rand_state > &random_state)
	cb_sample_data (std::shared_ptr< rand_state > &&random_state)
template<bool is_learn>
void	learn_or_predict (multi_learner &base, multi_ex &examples)

Private Attributes
std::shared_ptr< rand_state >	_random_state

Detailed Description

Definition at line 14 of file cb_sample.cc.

Constructor & Destructor Documentation

cb_sample_data() [1/2]

VW::cb_sample_data::cb_sample_data ( std::shared_ptr< rand_state > &  random_state )

inlineexplicit

Definition at line 16 of file cb_sample.cc.

: _random_state(random_state) {}

cb_sample_data() [2/2]

VW::cb_sample_data::cb_sample_data ( std::shared_ptr< rand_state > &&  random_state )

inlineexplicit

Definition at line 17 of file cb_sample.cc.

: _random_state(random_state) {}

Member Function Documentation

learn_or_predict()

template<bool is_learn>

void VW::cb_sample_data::learn_or_predict ( multi_learner &  base, multi_ex &  examples  )

inline

Definition at line 20 of file cb_sample.cc.

References _UNUSED, substring::begin, ACTION_SCORE::begin_scores(), ACTION_SCORE::end_scores(), S_EXPLORATION_OK, exploration::sample_after_normalizing(), substring_len(), exploration::swap_chosen(), and uniform_hash().

Referenced by learn_or_predict().

  {
    multiline_learn_or_predict<is_learn>(base, examples, examples[0]->ft_offset);

    auto action_scores = examples[0]->pred.a_s;
    uint32_t chosen_action = -1;

    int labelled_action = -1;
    // Find that chosen action in the learning case, skip the shared example.
    auto it = std::find_if(examples.begin(), examples.end(), [](example *item) { return !item->l.cb.costs.empty(); });
    if (it != examples.end())
    {
      labelled_action = std::distance(examples.begin(), it);
    }

    // If we are learning and have a label, then take that action as the chosen action. Otherwise sample the
    // distribution.
    if (is_learn && labelled_action != -1)
    {
      // Find where the labelled action is in the final prediction to determine if swapping needs to occur.
      // This only matters if the prediction decided to explore, but the same output should happen for the learn case.
      for (size_t i = 0; i < action_scores.size(); i++)
      {
        auto &a_s = action_scores[i];
        if (a_s.action == static_cast<uint32_t>(labelled_action))
        {
          chosen_action = static_cast<uint32_t>(i);
          break;
        }
      }
    }
    else
    {
      bool tag_provided_seed = false;
      uint64_t seed = _random_state->get_current_state();
      if (!examples[0]->tag.empty())
      {
        const std::string SEED_IDENTIFIER = "seed=";
        if (strncmp(examples[0]->tag.begin(), SEED_IDENTIFIER.c_str(), SEED_IDENTIFIER.size()) == 0 &&
            examples[0]->tag.size() > SEED_IDENTIFIER.size())
        {
          substring tag_seed{examples[0]->tag.begin() + 5, examples[0]->tag.begin() + examples[0]->tag.size()};
          seed = uniform_hash(tag_seed.begin, substring_len(tag_seed), 0);
          tag_provided_seed = true;
        }
      }

      // Sampling is done after the base learner has generated a pdf.
      auto result = exploration::sample_after_normalizing(
          seed, ACTION_SCORE::begin_scores(action_scores), ACTION_SCORE::end_scores(action_scores), chosen_action);
      assert(result == S_EXPLORATION_OK);
      _UNUSED(result);

      // Update the seed state in place if it was used for this example.
      if (!tag_provided_seed)
      {
        _random_state->get_and_update_random();
      }
    }

    auto result = exploration::swap_chosen(action_scores.begin(), action_scores.end(), chosen_action);
    assert(result == S_EXPLORATION_OK);

    _UNUSED(result);
  }

Member Data Documentation

_random_state

std::shared_ptr<rand_state> VW::cb_sample_data::_random_state

private

Definition at line 87 of file cb_sample.cc.

---

The documentation for this struct was generated from the following file:

• /mnt/c/w/linux/vowpal_wabbit/vowpalwabbit/cb_sample.cc