interact.cc

Go to the documentation of this file.

/*
Copyright (c) by respective owners including Yahoo!, Microsoft, and
individual contributors. All rights reserved.  Released under a BSD (revised)
license as described in the file LICENSE.
 */
#include <sstream>
#include <cfloat>
#include "reductions.h"
#include "v_array.h"

using namespace VW::config;

struct interact
{
  unsigned char n1, n2;  // namespaces to interact
  features feat_store;
  vw* all;
  float n1_feat_sq;
  float total_sum_feat_sq;
  size_t num_features;

  ~interact() { feat_store.delete_v(); }
};

bool contains_valid_namespaces(vw& all, features& f_src1, features& f_src2, interact& in)
{
  // first feature must be 1 so we're sure that the anchor feature is present
  if (f_src1.size() == 0 || f_src2.size() == 0)
    return false;

  if (f_src1.values[0] != 1)
  {
    all.trace_message << "Namespace '" << (char)in.n1 << "' misses anchor feature with value 1";
    return false;
  }

  if (f_src2.values[0] != 1)
  {
    all.trace_message << "Namespace '" << (char)in.n2 << "' misses anchor feature with value 1";
    return false;
  }

  return true;
}

void multiply(features& f_dest, features& f_src2, interact& in)
{
  f_dest.clear();
  features& f_src1 = in.feat_store;
  vw* all = in.all;
  uint64_t weight_mask = all->weights.mask();
  uint64_t base_id1 = f_src1.indicies[0] & weight_mask;
  uint64_t base_id2 = f_src2.indicies[0] & weight_mask;

  f_dest.push_back(f_src1.values[0] * f_src2.values[0], f_src1.indicies[0]);

  uint64_t prev_id1 = 0;
  uint64_t prev_id2 = 0;

  for (size_t i1 = 1, i2 = 1; i1 < f_src1.size() && i2 < f_src2.size();)
  {
    // calculating the relative offset from the namespace offset used to match features
    uint64_t cur_id1 = (uint64_t)(((f_src1.indicies[i1] & weight_mask) - base_id1) & weight_mask);
    uint64_t cur_id2 = (uint64_t)(((f_src2.indicies[i2] & weight_mask) - base_id2) & weight_mask);

    // checking for sorting requirement
    if (cur_id1 < prev_id1)
    {
      std::cout << "interact features are out of order: " << cur_id1 << " < " << prev_id1 << ". Skipping features."
                << std::endl;
      return;
    }

    if (cur_id2 < prev_id2)
    {
      std::cout << "interact features are out of order: " << cur_id2 << " < " << prev_id2 << ". Skipping features."
                << std::endl;
      return;
    }

    if (cur_id1 == cur_id2)
    {
      f_dest.push_back(f_src1.values[i1] * f_src2.values[i2], f_src1.indicies[i1]);
      i1++;
      i2++;
    }
    else if (cur_id1 < cur_id2)
      i1++;
    else
      i2++;
    prev_id1 = cur_id1;
    prev_id2 = cur_id2;
  }
}

template <bool is_learn, bool print_all>
void predict_or_learn(interact& in, LEARNER::single_learner& base, example& ec)
{
  features& f1 = ec.feature_space[in.n1];
  features& f2 = ec.feature_space[in.n2];

  if (!contains_valid_namespaces(*in.all, f1, f2, in))
  {
    if (is_learn)
      base.learn(ec);
    else
      base.predict(ec);

    return;
  }

  in.num_features = ec.num_features;
  in.total_sum_feat_sq = ec.total_sum_feat_sq;
  ec.total_sum_feat_sq -= f1.sum_feat_sq;
  ec.total_sum_feat_sq -= f2.sum_feat_sq;
  ec.num_features -= f1.size();
  ec.num_features -= f2.size();

  in.feat_store.deep_copy_from(f1);

  multiply(f1, f2, in);
  ec.total_sum_feat_sq += f1.sum_feat_sq;
  ec.num_features += f1.size();

  /*for(uint64_t i = 0;i < f1.size();i++)
    std::cout<<f1[i].weight_index<<":"<<f1[i].x<<" ";
    std::cout<< std::endl;*/

  // remove 2nd namespace
  int n2_i = -1;
  for (size_t i = 0; i < ec.indices.size(); i++)
  {
    if (ec.indices[i] == in.n2)
    {
      n2_i = (int)i;
      memmove(&ec.indices[n2_i], &ec.indices[n2_i + 1], sizeof(unsigned char) * (ec.indices.size() - n2_i - 1));
      ec.indices.decr();
      break;
    }
  }

  base.predict(ec);
  if (is_learn)
    base.learn(ec);

  // re-insert namespace into the right position
  ec.indices.incr();
  memmove(&ec.indices[n2_i + 1], &ec.indices[n2_i], sizeof(unsigned char) * (ec.indices.size() - n2_i - 1));
  ec.indices[n2_i] = in.n2;

  f1.deep_copy_from(in.feat_store);
  ec.total_sum_feat_sq = in.total_sum_feat_sq;
  ec.num_features = in.num_features;
}

LEARNER::base_learner* interact_setup(options_i& options, vw& all)
{
  std::string s;
  option_group_definition new_options("Interact via elementwise multiplication");
  new_options.add(
      make_option("interact", s).keep().help("Put weights on feature products from namespaces <n1> and <n2>"));
  options.add_and_parse(new_options);

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

  if (s.length() != 2)
  {
    std::cerr << "Need two namespace arguments to interact: " << s << " won't do EXITING\n";
    return nullptr;
  }

  auto data = scoped_calloc_or_throw<interact>();

  data->n1 = (unsigned char)s[0];
  data->n2 = (unsigned char)s[1];
  if (!all.quiet)
    std::cerr << "Interacting namespaces " << data->n1 << " and " << data->n2 << std::endl;
  data->all = &all;

  LEARNER::learner<interact, example>* l;
  l = &LEARNER::init_learner(
      data, as_singleline(setup_base(options, all)), predict_or_learn<true, true>, predict_or_learn<false, true>, 1);

  return make_base(*l);
}