search_graph.cc

/*
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 "search_graph.h"
#include "vw.h"
#include "gd.h"
#include "vw_exception.h"
using namespace std;
/*
example format:

ALL NODES
ALL EDGES
<blank>
ALL NODES
ALL EDGES
<blank>

and so on

node lines look like normal vw examples with unary features:

label:weight |n features
label:weight |n features
...

they are *implicitly* labeled starting at 1. (note the namespace
needn't be called n.) if weight is omitted it is assumed to be 1.0.

edge lines look like:

n1 n2 n3 ... |e features
n1 n2 n3 ... |e features
...

here, n1 n2 n3 are integer node ids, starting at one. technically
these are hyperedges, since they can touch more than two nodes. in the
canonical representation, there will just be n1 and n2.

the only thing that differentiates edges from nodes is that edges have
>1 input.

edges can either be directed or undirected. the default is undirected,
in which case the order of nodes listed in the edge is irrelevant. if
you add --search_graph_directed then the edges will be interpreted as
directed edges. for simple edges that connect two nodes, such as "n1
n2 | ...", the direction will be n1-->n2. for hyperedges of size K
such as "n1 n2 n3 n4 | ...", the default assumption is that the edge
has K-1 sources and 1 sink, so {n1,n2,n3}-->n4 in this case. if you
desire hyper-edges with more than one sink, you can specify them by
separating the source nodes from the sink nodes with a "0". for
instance "n1 n2 n3 0 n4 | ..." is the same as "n1 n2 n3 n4 | ...", but
"n1 n2 0 n3 n4 | ..." means {n1,n2}-->{n3,n4}. in undirected mode, all
0s are ignored; in directed mode, all but the first are ignored.
*/

namespace GraphTask
{
Search::search_task task = { "graph", run, initialize, finish, setup, takedown };

struct task_data
{ // global data
  size_t num_loops;
  size_t K;  // number of labels, *NOT* including the +1 for 'unlabeled'
  size_t numN; // number of neighbor predictions (equals K+1 for undirected, or 2*(K+1) for directed)
  bool   use_structure;
  bool   separate_learners;
  bool   directed;

  // for adding new features
  uint64_t mask; // all->reg.weight_mask
  uint64_t multiplier;   // all.wpp << all.stride_shift
  size_t ss; // stride_shift
  size_t wpp;

  // per-example data
  uint32_t N;  // number of nodes
  uint32_t E;  // number of edges
  vector< vector<size_t> > adj;  // adj[n] is a vector of *edge example ids* that contain n
  vector<uint32_t> bfs;   // order of nodes to process
  vector<size_t>   pred;  // predictions
  example*cur_node;       // pointer to the current node for add_edge_features_fn
  float* neighbor_predictions;  // prediction on this neighbor for add_edge_features_fn
  weight_parameters* weights;
  uint32_t* confusion_matrix;
  float* true_counts;
  float true_counts_total;

};

inline bool example_is_test(polylabel&l) { return l.cs.costs.size() == 0; }

void initialize(Search::search& sch, size_t& num_actions, po::variables_map& vm)
{ task_data * D = new task_data();
  po::options_description sspan_opts("search graphtask options");
  sspan_opts.add_options()("search_graph_num_loops", po::value<size_t>(), "how many loops to run [def: 2]");
  sspan_opts.add_options()("search_graph_no_structure", "turn off edge features");
  sspan_opts.add_options()("search_graph_separate_learners", "use a different learner for each pass");
  sspan_opts.add_options()("search_graph_directed", "construct features based on directed graph semantics");
  sch.add_program_options(vm, sspan_opts);

  D->num_loops = 2;
  D->use_structure = true;
  D->directed = false;
  if (vm.count("search_graph_num_loops"))      D->num_loops = vm["search_graph_num_loops"].as<size_t>();
  if (vm.count("search_graph_no_structure"))   D->use_structure = false;
  if (vm.count("search_graph_separate_learners")) D->separate_learners = true;
  if (vm.count("search_graph_directed"))       D->directed = true;

  if (D->num_loops <= 1) { D->num_loops = 1; D->separate_learners = false; }

  D->K = num_actions;
  D->numN = (D->directed+1) * (D->K+1);
  cerr << "K=" << D->K << ", numN=" << D->numN << endl;
  D->neighbor_predictions = calloc_or_throw<float>(D->numN);

  D->confusion_matrix = calloc_or_throw<uint32_t>( (D->K+1)*(D->K+1) );
  D->true_counts = calloc_or_throw<float>(D->K+1);
  D->true_counts_total = (float)(D->K+1);
  for (size_t k=0; k<=D->K; k++) D->true_counts[k] = 1.;

  if (D->separate_learners) sch.set_num_learners(D->num_loops);

  sch.set_task_data<task_data>(D);
  sch.set_options( 0 ); // Search::AUTO_HAMMING_LOSS
  sch.set_label_parser( COST_SENSITIVE::cs_label, example_is_test );
}

void finish(Search::search& sch)
{ task_data * D = sch.get_task_data<task_data>();
  free(D->neighbor_predictions);
  free(D->confusion_matrix);
  free(D->true_counts);
  delete D;
}

inline bool example_is_edge(example*e) { return e->l.cs.costs.size() > 1; }

void run_bfs(task_data &D, vector<example*>& ec)
{ D.bfs.clear();
  vector<bool> touched;
  for (size_t n=0; n<D.N; n++) touched.push_back(false);

  touched[0] = true;
  D.bfs.push_back(0);

  size_t i = 0;
  while (D.bfs.size() < D.N)
  { while (i < D.bfs.size())
    { uint32_t n = D.bfs[i];
      for (size_t id : D.adj[n])
        for (size_t j=0; j<ec[id]->l.cs.costs.size(); j++)
        { uint32_t m = ec[id]->l.cs.costs[j].class_index;
          if ((m > 0) && (!touched[m-1]))
          { D.bfs.push_back(m-1);
            touched[m-1] = true;
          }
        }
      i++;
    }

    if (D.bfs.size() < D.N)
      // we finished a SCC, need to find another
      for (uint32_t n=0; n<D.N; n++)
        if (! touched[n])
        { touched[n] = true;
          D.bfs.push_back(n);
          break;
        }
  }
}

void setup(Search::search& sch, vector<example*>& ec)
{ task_data& D = *sch.get_task_data<task_data>();

  D.mask = sch.get_vw_pointer_unsafe().weights.mask();
  D.wpp  = sch.get_vw_pointer_unsafe().wpp;
  D.ss = sch.get_vw_pointer_unsafe().weights.stride_shift();
  D.multiplier = D.wpp << D.ss;
  D.weights = &sch.get_vw_pointer_unsafe().weights;

  D.N = 0;
  D.E = 0;
  for (size_t i=0; i<ec.size(); i++)
    if (example_is_edge(ec[i]))
      D.E++;
    else   // it's a node!
    { if (D.E > 0)
        THROW("error: got a node after getting edges!");

      D.N++;
      if (ec[i]->l.cs.costs.size() > 0)
      { D.true_counts[ec[i]->l.cs.costs[0].class_index] += 1.;
        D.true_counts_total += 1.;
      }
    }

  if ((D.N == 0) && (D.E > 0))
    THROW("error: got edges without any nodes (perhaps ring_size is too small?)!");

  D.adj = vector<vector<size_t>>(D.N, vector<size_t>(0));

  for (size_t i=D.N; i<ec.size(); i++)
  { for (size_t n=0; n<ec[i]->l.cs.costs.size(); n++)
    { if (ec[i]->l.cs.costs[n].class_index > D.N)
        THROW("error: edge source points to too large of a node id: " << (ec[i]->l.cs.costs[n].class_index) << " > " << D.N);
    }
    for (size_t n=0; n<ec[i]->l.cs.costs.size(); n++)
    { size_t nn = ec[i]->l.cs.costs[n].class_index;
      if ((nn > 0) && (((D.adj[nn-1].size() == 0) || (D.adj[nn-1][D.adj[nn-1].size()-1] != i)))) // don't allow dups
        D.adj[nn-1].push_back(i);
    }
  }

  run_bfs(D, ec);

  D.pred.clear();
  for (size_t n=0; n<D.N; n++)
    D.pred.push_back( D.K+1 );
}

void takedown(Search::search& sch, vector<example*>& /*ec*/)
{ task_data& D = *sch.get_task_data<task_data>();
  D.bfs.clear();
  D.pred.clear();
  for (auto x : D.adj) x.clear();
  D.adj.clear();
}

void add_edge_features_group_fn(task_data&D, float fv, uint64_t fx)
{ example*node = D.cur_node;
  uint64_t fx2 = fx / (uint64_t)D.multiplier;
  for (size_t k=0; k<D.numN; k++)
    { if (D.neighbor_predictions[k] == 0.) continue;
      node->feature_space[neighbor_namespace].push_back(fv * D.neighbor_predictions[k], (uint64_t)(( fx2 + 348919043 * k ) * D.multiplier) & (uint64_t)D.mask);
    }
}

void add_edge_features_single_fn(task_data&D, float fv, uint64_t fx)
{ example*node = D.cur_node;
  features& fs = node->feature_space[neighbor_namespace];
  uint64_t fx2 = fx / (uint64_t)D.multiplier;
  size_t k = (size_t) D.neighbor_predictions[0];
  fs.push_back(fv, (uint32_t)(( fx2 + 348919043 * k ) * D.multiplier) & (uint64_t)D.mask);
}

void add_edge_features(Search::search&sch, task_data&D, size_t n, vector<example*>&ec)
{ D.cur_node = ec[n];

  for (size_t i : D.adj[n])
  { for (size_t k=0; k<D.numN; k++) D.neighbor_predictions[k] = 0.;

    float pred_total = 0.;
    uint32_t last_pred = 0;
    if (D.use_structure)
    { bool n_in_sink = true;
      if (D.directed)
        for (size_t j=0; j<ec[i]->l.cs.costs.size()-1; j++)
        { size_t m = ec[i]->l.cs.costs[j].class_index;
          if (m == 0) break;
          if (m-1 == n) { n_in_sink = false; break; }
        }

      bool m_in_sink = false;
      for (size_t j=0; j<ec[i]->l.cs.costs.size(); j++)
      { size_t m = ec[i]->l.cs.costs[j].class_index;
        if (m == 0) { m_in_sink = true; continue; }
        if (j == ec[i]->l.cs.costs.size()-1) m_in_sink = true;
        m--;
        if (m == n) continue;
        size_t other_side = (D.directed && (n_in_sink != m_in_sink)) ? (D.K+1) : 0;
        D.neighbor_predictions[ D.pred[m]-1 + other_side ] += 1.;
        pred_total += 1.;
        last_pred = (uint32_t)D.pred[m]-1 + (uint32_t)other_side;
      }
    }
    else
    { D.neighbor_predictions[0] += 1.;
      pred_total += 1.;
      last_pred = 0;
    }

    if (pred_total == 0.) continue;
    //cerr << n << ':' << i << " -> ["; for (size_t k=0; k<D.numN; k++) cerr << ' ' << D.neighbor_predictions[k]; cerr << " ]" << endl;
    for (size_t k=0; k<D.numN; k++) D.neighbor_predictions[k] /= pred_total;
    example&edge = *ec[i];

    if (pred_total <= 1.)    // single edge
    { D.neighbor_predictions[0] = (float)last_pred;
      GD::foreach_feature<task_data,uint64_t,add_edge_features_single_fn>(sch.get_vw_pointer_unsafe(), edge, D);
    }
    else   // lots of edges
      GD::foreach_feature<task_data,uint64_t,add_edge_features_group_fn>(sch.get_vw_pointer_unsafe(), edge, D);
  }
  ec[n]->indices.push_back(neighbor_namespace);
  ec[n]->total_sum_feat_sq += ec[n]->feature_space[neighbor_namespace].sum_feat_sq;
  ec[n]->num_features += ec[n]->feature_space[neighbor_namespace].size();

  vw& all = sch.get_vw_pointer_unsafe();
  for (string& i : all.pairs)
  { int i0 = (int)i[0];
    int i1 = (int)i[1];
    if ((i0 == (int)neighbor_namespace) || (i1 == (int)neighbor_namespace))
    { ec[n]->num_features      += ec[n]->feature_space[i0].size() * ec[n]->feature_space[i1].size();
      ec[n]->total_sum_feat_sq += ec[n]->feature_space[i0].sum_feat_sq*ec[n]->feature_space[i1].sum_feat_sq;
    }
  }

}

void del_edge_features(task_data&/*D*/, uint32_t n, vector<example*>&ec)
{ ec[n]->indices.pop();
  features& fs = ec[n]->feature_space[neighbor_namespace];
  ec[n]->total_sum_feat_sq -= fs.sum_feat_sq;
  ec[n]->num_features -= fs.size();
  fs.erase();
}

#define IDX(i,j) ( (i) * (D.K+1) + j )

float macro_f(task_data& D)
{ float total_f1 = 0.;
  float count_f1 = 0.;
  for (size_t k=1; k<=D.K; k++)
  { float trueC = 0.;
    float predC = 0.;
    for (size_t j=1; j<=D.K; j++)
    { trueC += (float)D.confusion_matrix[ IDX(k,j) ];
      predC += (float)D.confusion_matrix[ IDX(j,k) ];
    }
    if (trueC == 0) continue;
    float correctC = (float)D.confusion_matrix[ IDX(k,k) ];
    count_f1++;
    if (correctC > 0)
    { float pre = correctC / predC;
      float rec = correctC / trueC;
      total_f1 += 2 * pre * rec / (pre + rec);
    }
  }
  return total_f1 / count_f1;
}

void run(Search::search& sch, vector<example*>& ec)
{ task_data& D = *sch.get_task_data<task_data>();
  float loss_val = 0.5f / (float)D.num_loops;
  for (size_t n=0; n<D.N; n++) D.pred[n] = D.K+1;

  for (size_t loop=0; loop<D.num_loops; loop++)
  { bool last_loop = loop == (D.num_loops-1);
    int start = 0; int end = D.N; int step = 1;
    if (loop % 2 == 1) { start = D.N-1; end=-1; step = -1; } // go inward on odd loops
    for (int n_id = start; n_id != end; n_id += step)
    { uint32_t n = D.bfs[n_id];
      uint32_t k = (ec[n]->l.cs.costs.size() > 0) ? ec[n]->l.cs.costs[0].class_index : 0;

      bool add_features = /* D.use_structure && */ sch.predictNeedsExample();
      //add_features = false;

      if (add_features) add_edge_features(sch, D, n, ec);
      Search::predictor P = Search::predictor(sch, n+1);
      P.set_input(*ec[n]);
      if (false && (k > 0))
      { float min_count = 1e12f;
        for (size_t k2=1; k2<=D.K; k2++)
          min_count = min(min_count, D.true_counts[k2]);
        float w = min_count / D.true_counts[k];
        //float w = D.true_counts_total / D.true_counts[k] / (float)(D.K);
        P.set_weight( w );
        //cerr << "w = " << D.true_counts_total / D.true_counts[k] / (float)(D.K) << endl;
        //P.set_weight( D.true_counts_total / D.true_counts[k] / (float)(D.K) );
      }
      if (D.separate_learners) P.set_learner_id(loop);
      if (k > 0) // for test examples
        P.set_oracle(k);
      // add all the conditioning
      for (size_t i=0; i<D.adj[n].size(); i++)
      { for (size_t j=0; j<ec[i]->l.cs.costs.size(); j++)
        { uint32_t m = ec[i]->l.cs.costs[j].class_index;
          if (m == 0) continue;
          m--;
          if (m == n) continue;
          P.add_condition(m+1, 'e');
        }
      }

      // make the prediction
      D.pred[n] = P.predict();
      if (ec[n]->l.cs.costs.size() > 0) // for test examples
        sch.loss((ec[n]->l.cs.costs[0].class_index == D.pred[n]) ? 0.f : (last_loop ? 0.5f : loss_val));

      if (add_features) del_edge_features(D, n, ec);
    }
  }

  for (uint32_t n=0; n<D.N; n++)
    D.confusion_matrix[ IDX( ec[n]->l.cs.costs[0].class_index, D.pred[n] ) ] ++;
  sch.loss( 1.f - macro_f(D) );

  if (sch.output().good())
    for (uint32_t n=0; n<D.N; n++)
      sch.output() << D.pred[n] << ' ';
}
}