MWT Namespace Reference

Classes
struct	mwt
struct	policy_data

Functions
bool	observed_cost (CB::cb_class *cl)
CB::cb_class *	get_observed_cost (CB::label &ld)
void	value_policy (mwt &c, float val, uint64_t index)
template<bool learn, bool exclude, bool is_learn>
void	predict_or_learn (mwt &c, single_learner &base, example &ec)
void	print_scalars (int f, v_array< float > &scalars, v_array< char > &tag)
void	finish_example (vw &all, mwt &c, example &ec)
void	save_load (mwt &c, io_buf &model_file, bool read, bool text)
void	delete_scalars (void *v)

Function Documentation

delete_scalars()

void MWT::delete_scalars ( void *  v )

inline

Definition at line 37 of file example.h.

References v_array< T >::delete_v().

Referenced by lda_setup(), mwt_setup(), and oaa_setup().

{
  v_array<float>* preds = (v_array<float>*)v;
  preds->delete_v();
}

finish_example()

void MWT::finish_example	(	vw &	all,
		mwt &	c,
		example &	ec
	)

Definition at line 175 of file mwt.cc.

References vw::final_prediction_sink, VW::finish_example(), CB_ALGS::get_cost_estimate(), MWT::mwt::learn, loss(), polyprediction::multiclass, example::num_features, MWT::mwt::observation, example::pred, print_scalars(), CB::print_update(), polyprediction::scalars, vw::sd, example::tag, example::test_only, and shared_data::update().

Referenced by mwt_setup().

{
  float loss = 0.;
  if (c.learn)
    if (c.observation != nullptr)
      loss = get_cost_estimate(c.observation, (uint32_t)ec.pred.scalars[0]);
  all.sd->update(ec.test_only, c.observation != nullptr, loss, 1.f, ec.num_features);

  for (int sink : all.final_prediction_sink) print_scalars(sink, ec.pred.scalars, ec.tag);

  if (c.learn)
  {
    v_array<float> temp = ec.pred.scalars;
    ec.pred.multiclass = (uint32_t)temp[0];
    CB::print_update(all, c.observation != nullptr, ec, nullptr, false);
    ec.pred.scalars = temp;
  }
  VW::finish_example(all, ec);
}

get_observed_cost()

CB::cb_class* MWT::get_observed_cost

(

CB::label &

ld

)

Definition at line 57 of file mwt.cc.

References CB::label::costs, and observed_cost().

Referenced by predict_or_learn().

{
  for (auto& cl : ld.costs)
    if (observed_cost(&cl))
      return &cl;
  return nullptr;
}

observed_cost()

bool MWT::observed_cost ( CB::cb_class *  cl )

inline

Definition at line 48 of file mwt.cc.

References CB::cb_class::cost, and CB::cb_class::probability.

Referenced by get_observed_cost().

{
  // cost observed for this action if it has non zero probability and cost != FLT_MAX
  if (cl != nullptr)
    if (cl->cost != FLT_MAX && cl->probability > .0)
      return true;
  return false;
}

predict_or_learn()

template<bool learn, bool exclude, bool is_learn>

void MWT::predict_or_learn	(	mwt &	c,
		single_learner &	base,
		example &	ec
	)

Definition at line 83 of file mwt.cc.

References MWT::mwt::all, c, polylabel::cb, v_array< T >::clear(), features::clear(), v_array< T >::empty(), MWT::mwt::evals, f, MWT::mwt::feature_space, example_predict::feature_space, CB_ALGS::get_cost_estimate(), get_observed_cost(), MWT::mwt::indices, example_predict::indices, example::l, LEARNER::learner< T, E >::learn(), learn(), parameters::mask(), polyprediction::multiclass, MWT::mwt::namespaces, MWT::mwt::num_classes, MWT::mwt::observation, MWT::mwt::policies, v_array< T >::pop(), example::pred, LEARNER::learner< T, E >::predict(), v_array< T >::push_back(), features::push_back(), polyprediction::scalars, stride_shift(), parameters::stride_shift(), MWT::mwt::total, and vw::weights.

{
  c.observation = get_observed_cost(ec.l.cb);

  if (c.observation != nullptr)
  {
    c.total++;
    // For each nonzero feature in observed namespaces, check it's value.
    for (unsigned char ns : ec.indices)
      if (c.namespaces[ns])
        GD::foreach_feature<mwt, value_policy>(c.all, ec.feature_space[ns], c);
    for (uint64_t policy : c.policies)
    {
      c.evals[policy].cost += get_cost_estimate(c.observation, c.evals[policy].action);
      c.evals[policy].action = 0;
    }
  }
  if (exclude || learn)
  {
    c.indices.clear();
    uint32_t stride_shift = c.all->weights.stride_shift();
    uint64_t weight_mask = c.all->weights.mask();
    for (unsigned char ns : ec.indices)
      if (c.namespaces[ns])
      {
        c.indices.push_back(ns);
        if (learn)
        {
          c.feature_space[ns].clear();
          for (features::iterator& f : ec.feature_space[ns])
          {
            uint64_t new_index = ((f.index() & weight_mask) >> stride_shift) * c.num_classes + (uint64_t)f.value();
            c.feature_space[ns].push_back(1, new_index << stride_shift);
          }
        }
        std::swap(c.feature_space[ns], ec.feature_space[ns]);
      }
  }

  // modify the predictions to use a vector with a score for each evaluated feature.
  v_array<float> preds = ec.pred.scalars;

  if (learn)
  {
    if (is_learn)
      base.learn(ec);
    else
      base.predict(ec);
  }

  if (exclude || learn)
    while (!c.indices.empty())
    {
      unsigned char ns = c.indices.pop();
      std::swap(c.feature_space[ns], ec.feature_space[ns]);
    }

  // modify the predictions to use a vector with a score for each evaluated feature.
  preds.clear();
  if (learn)
    preds.push_back((float)ec.pred.multiclass);
  for (uint64_t index : c.policies) preds.push_back((float)c.evals[index].cost / (float)c.total);

  ec.pred.scalars = preds;
}

print_scalars()

void MWT::print_scalars	(	int	f,
		v_array< float > &	scalars,
		v_array< char > &	tag
	)

Definition at line 149 of file mwt.cc.

References v_array< T >::size(), and io_buf::write_file_or_socket().

Referenced by finish_example(), and return_example().

{
  if (f >= 0)
  {
    std::stringstream ss;

    for (size_t i = 0; i < scalars.size(); i++)
    {
      if (i > 0)
        ss << ' ';
      ss << scalars[i];
    }
    for (size_t i = 0; i < tag.size(); i++)
    {
      if (i == 0)
        ss << ' ';
      ss << tag[i];
    }
    ss << '\n';
    ssize_t len = ss.str().size();
    ssize_t t = io_buf::write_file_or_socket(f, ss.str().c_str(), (unsigned int)len);
    if (t != len)
      std::cerr << "write error: " << strerror(errno) << std::endl;
  }
}

save_load()

void MWT::save_load	(	mwt &	c,
		io_buf &	model_file,
		bool	read,
		bool	text
	)

Definition at line 195 of file mwt.cc.

References MWT::policy_data::action, v_array< T >::begin(), bin_text_read_write_fixed_validated(), MWT::policy_data::cost, v_array< T >::empty(), v_array< T >::end(), MWT::mwt::evals, io_buf::files, MWT::mwt::policies, v_array< T >::resize(), v_array< T >::size(), and MWT::mwt::total.

Referenced by mwt_setup().

{
  if (model_file.files.empty())
    return;

  std::stringstream msg;

  // total
  msg << "total: " << c.total;
  bin_text_read_write_fixed_validated(model_file, (char*)&c.total, sizeof(c.total), "", read, msg, text);

  // policies
  size_t policies_size = c.policies.size();
  bin_text_read_write_fixed_validated(model_file, (char*)&policies_size, sizeof(policies_size), "", read, msg, text);

  if (read)
  {
    c.policies.resize(policies_size);
    c.policies.end() = c.policies.begin() + policies_size;
  }
  else
  {
    msg << "policies: ";
    for (feature_index& policy : c.policies) msg << policy << " ";
  }

  bin_text_read_write_fixed_validated(
      model_file, (char*)c.policies.begin(), policies_size * sizeof(feature_index), "", read, msg, text);

  // c.evals is already initialized nicely to the same size as the regressor.
  for (feature_index& policy : c.policies)
  {
    policy_data& pd = c.evals[policy];
    if (read)
      msg << "evals: " << policy << ":" << pd.action << ":" << pd.cost << " ";
    bin_text_read_write_fixed_validated(model_file, (char*)&c.evals[policy], sizeof(policy_data), "", read, msg, text);
  }
}

value_policy()

void MWT::value_policy	(	mwt &	c,
		float	val,
		uint64_t	index
	)

Definition at line 65 of file mwt.cc.

References MWT::mwt::all, MWT::mwt::evals, parameters::mask(), MWT::mwt::policies, v_array< T >::push_back(), parameters::stride_shift(), and vw::weights.

{
  if (val < 0 || floor(val) != val)
    std::cout << "error " << val << " is not a valid action " << std::endl;

  uint32_t value = (uint32_t)val;
  uint64_t new_index = (index & c.all->weights.mask()) >> c.all->weights.stride_shift();

  if (!c.evals[new_index].seen)
  {
    c.evals[new_index].seen = true;
    c.policies.push_back(new_index);
  }

  c.evals[new_index].action = value;
}