summaryrefslogtreecommitdiff
path: root/decoder/hg.h
blob: 53b5a53da1bea2541644367085870ff7da23e64e (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
#ifndef _HG_H_
#define _HG_H_

#include <string>
#include <vector>

#include "small_vector.h"
#include "sparse_vector.h"
#include "wordid.h"
#include "trule.h"
#include "prob.h"

// if you define this, edges_ will be sorted
// (normally, just nodes_ are - root must be nodes_.back()), but this can be quite
// slow
#undef HG_EDGES_TOPO_SORTED

// class representing an acyclic hypergraph
//  - edges have 1 head, 0..n tails
class Hypergraph {
 public:
  Hypergraph() : is_linear_chain_(false) {}

  // SmallVector is a fast, small vector<int> implementation for sizes <= 2
  typedef SmallVectorInt TailNodeVector;

  // TODO get rid of cat_?
  // TODO keep cat_ and add span and/or state? :)
  struct Node {
    Node() : id_(), cat_(), promise(1) {}
    int id_; // equal to this object's position in the nodes_ vector
    WordID cat_;  // non-terminal category if <0, 0 if not set
    std::vector<int> in_edges_;   // contents refer to positions in edges_
    std::vector<int> out_edges_;  // contents refer to positions in edges_
    double promise; // set in global pruning; in [0,infty) so that mean is 1.  use: e.g. scale cube poplimit
  };

  // TODO get rid of edge_prob_? (can be computed on the fly as the dot
  // product of the weight vector and the feature values)
  struct Edge {
    Edge() : i_(-1), j_(-1), prev_i_(-1), prev_j_(-1) {}
    inline int Arity() const { return tail_nodes_.size(); }
    int head_node_;               // refers to a position in nodes_
    TailNodeVector tail_nodes_;   // contents refer to positions in nodes_
    TRulePtr rule_;
    SparseVector<double> feature_values_;
    prob_t edge_prob_;             // dot product of weights and feat_values
    int id_;   // equal to this object's position in the edges_ vector

    // span info. typically, i_ and j_ refer to indices in the source sentence
    // if a synchronous parse has been executed i_ and j_ will refer to indices
    // in the target sentence / lattice and prev_i_ prev_j_ will refer to
    // positions in the source.  Note: it is up to the translator implementation
    // to properly set these values.  For some models (like the Forest-input
    // phrase based model) it may not be straightforward to do.  if these values
    // are not properly set, most things will work but alignment and any features
    // that depend on them will be broken.
    short int i_;
    short int j_;
    short int prev_i_;
    short int prev_j_;
  };

  // returns edge with rule_.IsGoal, returns 0 if none found.  otherwise gives best edge_prob_ - note: I don't think edge_prob_ is viterbi cumulative, so this would just be the best local probability.
  Edge const* ViterbiGoalEdge() const;

  void swap(Hypergraph& other) {
    other.nodes_.swap(nodes_);
    std::swap(is_linear_chain_, other.is_linear_chain_);
    other.edges_.swap(edges_);
  }

  void ResizeNodes(int size) {
    nodes_.resize(size);
    for (int i = 0; i < size; ++i) nodes_[i].id_ = i;
  }

  // reserves space in the nodes vector to prevent memory locations
  // from changing
  void ReserveNodes(size_t n, size_t e = 0) {
    nodes_.reserve(n);
    if (e) edges_.reserve(e);
  }

  Edge* AddEdge(const TRulePtr& rule, const TailNodeVector& tail) {
    edges_.push_back(Edge());
    Edge* edge = &edges_.back();
    edge->rule_ = rule;
    edge->tail_nodes_ = tail;
    edge->id_ = edges_.size() - 1;
    for (int i = 0; i < edge->tail_nodes_.size(); ++i)
      nodes_[edge->tail_nodes_[i]].out_edges_.push_back(edge->id_);
    return edge;
  }

  Node* AddNode(const WordID& cat) {
    nodes_.push_back(Node());
    nodes_.back().cat_ = cat;
    nodes_.back().id_ = nodes_.size() - 1;
    return &nodes_.back();
  }

  void ConnectEdgeToHeadNode(const int edge_id, const int head_id) {
    edges_[edge_id].head_node_ = head_id;
    nodes_[head_id].in_edges_.push_back(edge_id);
  }

  // TODO remove this - use the version that takes indices
  void ConnectEdgeToHeadNode(Edge* edge, Node* head) {
    edge->head_node_ = head->id_;
    head->in_edges_.push_back(edge->id_);
  }

  // merge the goal node from other with this goal node
  void Union(const Hypergraph& other);

  void PrintGraphviz() const;

  // compute the total number of paths in the forest
  double NumberOfPaths() const;

  // BEWARE. this assumes that the source and target language
  // strings are identical and that there are no loops.
  // It assumes a bunch of other things about where the
  // epsilons will be.  It tries to assert failure if you
  // break these assumptions, but it may not.
  // TODO - make this work
  void EpsilonRemove(WordID eps);

  // multiple the weights vector by the edge feature vector
  // (inner product) to set the edge probabilities
  template <typename V>
  void Reweight(const V& weights) {
    for (int i = 0; i < edges_.size(); ++i) {
      Edge& e = edges_[i];
      e.edge_prob_.logeq(e.feature_values_.dot(weights));
    }
  }

  typedef std::vector<prob_t> EdgeProbs;
  typedef std::vector<bool> EdgeMask;

  // computes inside and outside scores for each
  // edge in the hypergraph
  // alpha->size = edges_.size = beta->size
  // returns inside prob of goal node
  prob_t ComputeEdgePosteriors(double scale,
                               EdgeProbs* posts) const;

  // find the score of the very best path passing through each edge
  prob_t ComputeBestPathThroughEdges(EdgeProbs* posts) const;

  // create a new hypergraph consisting only of the nodes / edges
  // in the Viterbi derivation of this hypergraph
  // if edges is set, use the EdgeSelectEdgeWeightFunction
  Hypergraph* CreateViterbiHypergraph(const EdgeMask* edges = NULL) const;

  // move weights as near to the source as possible, resulting in a
  // stochastic automaton.  ONLY FUNCTIONAL FOR *LATTICES*.
  // See M. Mohri and M. Riley. A Weight Pushing Algorithm for Large
  //   Vocabulary Speech Recognition. 2001.
  // the log semiring (NOT tropical) is used
  void PushWeightsToSource(double scale = 1.0);
  // same, except weights are pushed to the goal, works for HGs,
  // not just lattices
  void PushWeightsToGoal(double scale = 1.0);

  void SortInEdgesByEdgeWeights();

  void PruneUnreachable(int goal_node_id); // DEPRECATED

  void RemoveNoncoaccessibleStates(int goal_node_id = -1);

  // remove edges from the hypergraph if prune_edge[edge_id] is true
  // note: if run_inside_algorithm is false, then consumers may be unhappy if you pruned nodes that are built on by nodes that are kept.
  void PruneEdges(const EdgeMask& prune_edge, bool run_inside_algorithm = false);

  /// drop edge i if edge_margin[i] < prune_below, unless preserve_mask[i]
  void MarginPrune(EdgeProbs const& edge_margin,prob_t prune_below,EdgeMask const* preserve_mask=0,bool safe_inside=false,bool verbose=false);
  // promise[i]=((max_marginal[i]/viterbi)^power).todouble.  if normalize, ensure that avg promise is 1.
  typedef EdgeProbs NodeProbs;
  void SetPromise(NodeProbs const& inside,NodeProbs const& outside, double power=1, bool normalize=true);

  //TODO: in my opinion, looking at the ratio of logprobs (features \dot weights) rather than the absolute difference generalizes more nicely across sentence lengths and weight vectors that are constant multiples of one another.  at least make that an option.  i worked around this a little in cdec by making "beam alpha per source word" but that's not helping with different tuning runs.  this would also make me more comfortable about allocating promise

  // beam_alpha=0 means don't beam prune, otherwise drop things that are e^beam_alpha times worse than best -   // prunes any edge whose prob_t on the best path taking that edge is more than e^alpha times
  //density=0 means don't density prune:   // for density>=1.0, keep this many times the edges needed for the 1best derivation
  // worse than the score of the global best past (or the highest edge posterior)
  // scale is for use_sum_prod_semiring (sharpens distribution?)
  // promise_power is for a call to SetPromise (no call happens if power=0)
  // returns true if density pruning was tighter than beam
  // safe_inside would be a redundant anti-rounding error second bottom-up reachability before actually removing edges, to prevent stranded edges.  shouldn't be needed - if the hyperedges occur in defined-before-use (all edges with head h occur before h is used as a tail) order, then a grace margin for keeping edges that starts leniently and becomes more forbidding will make it impossible for this to occur, i.e. safe_inside=true is not needed.
  bool PruneInsideOutside(double beam_alpha,double density,const EdgeMask* preserve_mask = NULL,const bool use_sum_prod_semiring=false, const double scale=1,double promise_power=0,bool safe_inside=false);

  // legacy:
  void DensityPruneInsideOutside(const double scale, const bool use_sum_prod_semiring, const double density,const EdgeMask* preserve_mask = NULL) {
    PruneInsideOutside(0,density,preserve_mask,use_sum_prod_semiring,scale);
  }

  // legacy:
  void BeamPruneInsideOutside(const double scale, const bool use_sum_prod_semiring, const double alpha,const EdgeMask* preserve_mask = NULL) {
    PruneInsideOutside(alpha,0,preserve_mask,use_sum_prod_semiring,scale);
  }

  // report nodes, edges, paths
  std::string stats(std::string const& name="forest") const;

  void clear() {
    nodes_.clear();
    edges_.clear();
  }

  inline size_t NumberOfEdges() const { return edges_.size(); }
  inline size_t NumberOfNodes() const { return nodes_.size(); }
  inline bool empty() const { return nodes_.empty(); }

  // linear chains can be represented in a number of ways in a hypergraph,
  // we define them to consist only of lexical translations and monotonic rules
  inline bool IsLinearChain() const { return is_linear_chain_; }
  bool is_linear_chain_;

  // nodes_ is sorted in topological order
  std::vector<Node> nodes_;
  // edges_ is not guaranteed to be in any particular order
  std::vector<Edge> edges_;

  // reorder nodes_ so they are in topological order
  // source nodes at 0 sink nodes at size-1
  void TopologicallySortNodesAndEdges(int goal_idx,
                                      const EdgeMask* prune_edges = NULL);
 private:
  Hypergraph(int num_nodes, int num_edges, bool is_lc) : is_linear_chain_(is_lc), nodes_(num_nodes), edges_(num_edges) {}

  static TRulePtr kEPSRule;
  static TRulePtr kUnaryRule;
};

// common WeightFunctions, map an edge -> WeightType
// for generic Viterbi/Inside algorithms
struct EdgeProb {
  typedef prob_t Weight;
  inline const prob_t& operator()(const Hypergraph::Edge& e) const { return e.edge_prob_; }
};

struct EdgeSelectEdgeWeightFunction {
  typedef prob_t Weight;
  typedef std::vector<bool> EdgeMask;
  EdgeSelectEdgeWeightFunction(const EdgeMask& v) : v_(v) {}
  inline prob_t operator()(const Hypergraph::Edge& e) const {
    if (v_[e.id_]) return prob_t::One();
    else return prob_t::Zero();
  }
 private:
  const EdgeMask& v_;
};

struct ScaledEdgeProb {
  ScaledEdgeProb(const double& alpha) : alpha_(alpha) {}
  inline prob_t operator()(const Hypergraph::Edge& e) const { return e.edge_prob_.pow(alpha_); }
  const double alpha_;
  typedef prob_t Weight;
};

// see Li (2010), Section 3.2.2-- this is 'x_e = p_e*r_e'
struct EdgeFeaturesAndProbWeightFunction {
  typedef SparseVector<prob_t> Weight;
  typedef Weight Result; //TODO: change Result->Weight everywhere?
  inline const Weight operator()(const Hypergraph::Edge& e) const {
    SparseVector<prob_t> res;
    for (SparseVector<double>::const_iterator it = e.feature_values_.begin();
         it != e.feature_values_.end(); ++it)
      res.set_value(it->first, prob_t(it->second) * e.edge_prob_);
    return res;
  }
};

struct TransitionCountWeightFunction {
  typedef double Weight;
  inline double operator()(const Hypergraph::Edge& e) const { (void)e; return 1.0; }
};

#endif