nn.h File Reference

Go to the source code of this file.

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

Function Documentation

nn_setup()

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

Definition at line 417 of file nn.cc.

References VW::config::option_group_definition::add(), add(), VW::config::options_i::add_and_parse(), LEARNER::as_singleline(), LEARNER::end_pass(), finish_example(), vw::get_random_state(), getLossFunction(), LEARNER::init_learner(), LEARNER::make_base(), VW::config::make_option(), multipredict(), nn::multitask, vw::quiet, vw::random_seed, LEARNER::learner< T, E >::set_end_pass(), LEARNER::learner< T, E >::set_finish_example(), LEARNER::learner< T, E >::set_multipredict(), setup_base(), vw::training, and VW::config::options_i::was_supplied().

Referenced by parse_reductions().

{
  auto n = scoped_calloc_or_throw<nn>();
  bool meanfield = false;
  option_group_definition new_options("Neural Network");
  new_options.add(make_option("nn", n->k).keep().help("Sigmoidal feedforward network with <k> hidden units"))
      .add(make_option("inpass", n->inpass)
               .keep()
               .help("Train or test sigmoidal feedforward network with input passthrough."))
      .add(make_option("multitask", n->multitask).keep().help("Share hidden layer across all reduced tasks."))
      .add(make_option("dropout", n->dropout).keep().help("Train or test sigmoidal feedforward network using dropout."))
      .add(make_option("meanfield", meanfield).help("Train or test sigmoidal feedforward network using mean field."));
  options.add_and_parse(new_options);

  if (!options.was_supplied("nn"))
    return nullptr;

  n->all = &all;
  n->_random_state = all.get_random_state();

  if (n->multitask && !all.quiet)
    std::cerr << "using multitask sharing for neural network " << (all.training ? "training" : "testing") << std::endl;

  if (options.was_supplied("meanfield"))
  {
    n->dropout = false;
    if (!all.quiet)
      std::cerr << "using mean field for neural network " << (all.training ? "training" : "testing") << std::endl;
  }

  if (n->dropout && !all.quiet)
    std::cerr << "using dropout for neural network " << (all.training ? "training" : "testing") << std::endl;

  if (n->inpass && !all.quiet)
    std::cerr << "using input passthrough for neural network " << (all.training ? "training" : "testing") << std::endl;

  n->finished_setup = false;
  n->squared_loss = getLossFunction(all, "squared", 0);

  n->xsubi = all.random_seed;

  n->save_xsubi = n->xsubi;

  n->hidden_units = calloc_or_throw<float>(n->k);
  n->dropped_out = calloc_or_throw<bool>(n->k);
  n->hidden_units_pred = calloc_or_throw<polyprediction>(n->k);
  n->hiddenbias_pred = calloc_or_throw<polyprediction>(n->k);

  auto base = as_singleline(setup_base(options, all));
  n->increment = base->increment;  // Indexing of output layer is odd.
  nn& nv = *n.get();
  learner<nn, example>& l =
      init_learner(n, base, predict_or_learn_multi<true, true>, predict_or_learn_multi<false, true>, n->k + 1);
  if (nv.multitask)
    l.set_multipredict(multipredict);
  l.set_finish_example(finish_example);
  l.set_end_pass(end_pass);

  return make_base(l);
}