diff options
Diffstat (limited to 'klm/search')
-rw-r--r-- | klm/search/Jamfile | 5 | ||||
-rw-r--r-- | klm/search/arity.hh | 8 | ||||
-rw-r--r-- | klm/search/config.hh | 25 | ||||
-rw-r--r-- | klm/search/context.hh | 65 | ||||
-rw-r--r-- | klm/search/edge.hh | 31 | ||||
-rw-r--r-- | klm/search/edge_generator.cc | 120 | ||||
-rw-r--r-- | klm/search/edge_generator.hh | 58 | ||||
-rw-r--r-- | klm/search/edge_queue.cc | 25 | ||||
-rw-r--r-- | klm/search/edge_queue.hh | 73 | ||||
-rw-r--r-- | klm/search/final.hh | 39 | ||||
-rw-r--r-- | klm/search/note.hh | 12 | ||||
-rw-r--r-- | klm/search/rule.cc | 43 | ||||
-rw-r--r-- | klm/search/rule.hh | 20 | ||||
-rw-r--r-- | klm/search/source.hh | 48 | ||||
-rw-r--r-- | klm/search/types.hh | 18 | ||||
-rw-r--r-- | klm/search/vertex.cc | 48 | ||||
-rw-r--r-- | klm/search/vertex.hh | 158 | ||||
-rw-r--r-- | klm/search/vertex_generator.cc | 83 | ||||
-rw-r--r-- | klm/search/vertex_generator.hh | 59 | ||||
-rw-r--r-- | klm/search/weights.cc | 71 | ||||
-rw-r--r-- | klm/search/weights.hh | 52 | ||||
-rw-r--r-- | klm/search/weights_test.cc | 38 |
22 files changed, 1099 insertions, 0 deletions
diff --git a/klm/search/Jamfile b/klm/search/Jamfile new file mode 100644 index 00000000..e8b14363 --- /dev/null +++ b/klm/search/Jamfile @@ -0,0 +1,5 @@ +lib search : weights.cc vertex.cc vertex_generator.cc edge_queue.cc edge_generator.cc rule.cc ../lm//kenlm ../util//kenutil /top//boost_system : : : <include>.. ; + +import testing ; + +unit-test weights_test : weights_test.cc search /top//boost_unit_test_framework ; diff --git a/klm/search/arity.hh b/klm/search/arity.hh new file mode 100644 index 00000000..09c2c671 --- /dev/null +++ b/klm/search/arity.hh @@ -0,0 +1,8 @@ +#ifndef SEARCH_ARITY__ +#define SEARCH_ARITY__ +namespace search { + +const unsigned int kMaxArity = 2; + +} // namespace search +#endif // SEARCH_ARITY__ diff --git a/klm/search/config.hh b/klm/search/config.hh new file mode 100644 index 00000000..ef8e2354 --- /dev/null +++ b/klm/search/config.hh @@ -0,0 +1,25 @@ +#ifndef SEARCH_CONFIG__ +#define SEARCH_CONFIG__ + +#include "search/weights.hh" +#include "util/string_piece.hh" + +namespace search { + +class Config { + public: + Config(const Weights &weights, unsigned int pop_limit) : + weights_(weights), pop_limit_(pop_limit) {} + + const Weights &GetWeights() const { return weights_; } + + unsigned int PopLimit() const { return pop_limit_; } + + private: + Weights weights_; + unsigned int pop_limit_; +}; + +} // namespace search + +#endif // SEARCH_CONFIG__ diff --git a/klm/search/context.hh b/klm/search/context.hh new file mode 100644 index 00000000..27940053 --- /dev/null +++ b/klm/search/context.hh @@ -0,0 +1,65 @@ +#ifndef SEARCH_CONTEXT__ +#define SEARCH_CONTEXT__ + +#include "lm/model.hh" +#include "search/config.hh" +#include "search/final.hh" +#include "search/types.hh" +#include "search/vertex.hh" +#include "util/exception.hh" + +#include <boost/pool/object_pool.hpp> +#include <boost/ptr_container/ptr_vector.hpp> + +#include <vector> + +namespace search { + +class Weights; + +class ContextBase { + public: + explicit ContextBase(const Config &config) : pop_limit_(config.PopLimit()), weights_(config.GetWeights()) {} + + Final *NewFinal() { + Final *ret = final_pool_.construct(); + assert(ret); + return ret; + } + + VertexNode *NewVertexNode() { + VertexNode *ret = vertex_node_pool_.construct(); + assert(ret); + return ret; + } + + void DeleteVertexNode(VertexNode *node) { + vertex_node_pool_.destroy(node); + } + + unsigned int PopLimit() const { return pop_limit_; } + + const Weights &GetWeights() const { return weights_; } + + private: + boost::object_pool<Final> final_pool_; + boost::object_pool<VertexNode> vertex_node_pool_; + + unsigned int pop_limit_; + + const Weights &weights_; +}; + +template <class Model> class Context : public ContextBase { + public: + Context(const Config &config, const Model &model) : ContextBase(config), model_(model) {} + + const Model &LanguageModel() const { return model_; } + + private: + const Model &model_; +}; + +} // namespace search + +#endif // SEARCH_CONTEXT__ diff --git a/klm/search/edge.hh b/klm/search/edge.hh new file mode 100644 index 00000000..77ab0ade --- /dev/null +++ b/klm/search/edge.hh @@ -0,0 +1,31 @@ +#ifndef SEARCH_EDGE__ +#define SEARCH_EDGE__ + +#include "lm/state.hh" +#include "search/arity.hh" +#include "search/rule.hh" +#include "search/types.hh" +#include "search/vertex.hh" + +#include <queue> + +namespace search { + +struct PartialEdge { + Score score; + // Terminals + lm::ngram::ChartState between[kMaxArity + 1]; + // Non-terminals + PartialVertex nt[kMaxArity]; + + const lm::ngram::ChartState &CompletedState() const { + return between[0]; + } + + bool operator<(const PartialEdge &other) const { + return score < other.score; + } +}; + +} // namespace search +#endif // SEARCH_EDGE__ diff --git a/klm/search/edge_generator.cc b/klm/search/edge_generator.cc new file mode 100644 index 00000000..56239dfb --- /dev/null +++ b/klm/search/edge_generator.cc @@ -0,0 +1,120 @@ +#include "search/edge_generator.hh" + +#include "lm/left.hh" +#include "lm/partial.hh" +#include "search/context.hh" +#include "search/vertex.hh" +#include "search/vertex_generator.hh" + +#include <numeric> + +namespace search { + +EdgeGenerator::EdgeGenerator(PartialEdge &root, unsigned char arity, Note note) : arity_(arity), note_(note) { +/* for (unsigned char i = 0; i < edge.Arity(); ++i) { + root.nt[i] = edge.GetVertex(i).RootPartial(); + } + for (unsigned char i = edge.Arity(); i < 2; ++i) { + root.nt[i] = kBlankPartialVertex; + }*/ + generate_.push(&root); + top_score_ = root.score; +} + +namespace { + +template <class Model> float FastScore(const Context<Model> &context, unsigned char victim, unsigned char arity, const PartialEdge &previous, PartialEdge &update) { + memcpy(update.between, previous.between, sizeof(lm::ngram::ChartState) * (arity + 1)); + + float ret = 0.0; + lm::ngram::ChartState *before, *after; + if (victim == 0) { + before = &update.between[0]; + after = &update.between[(arity == 2 && previous.nt[1].Complete()) ? 2 : 1]; + } else { + assert(victim == 1); + assert(arity == 2); + before = &update.between[previous.nt[0].Complete() ? 0 : 1]; + after = &update.between[2]; + } + const lm::ngram::ChartState &previous_reveal = previous.nt[victim].State(); + const PartialVertex &update_nt = update.nt[victim]; + const lm::ngram::ChartState &update_reveal = update_nt.State(); + float just_after = 0.0; + if ((update_reveal.left.length > previous_reveal.left.length) || (update_reveal.left.full && !previous_reveal.left.full)) { + just_after += lm::ngram::RevealAfter(context.LanguageModel(), before->left, before->right, update_reveal.left, previous_reveal.left.length); + } + if ((update_reveal.right.length > previous_reveal.right.length) || (update_nt.RightFull() && !previous.nt[victim].RightFull())) { + ret += lm::ngram::RevealBefore(context.LanguageModel(), update_reveal.right, previous_reveal.right.length, update_nt.RightFull(), after->left, after->right); + } + if (update_nt.Complete()) { + if (update_reveal.left.full) { + before->left.full = true; + } else { + assert(update_reveal.left.length == update_reveal.right.length); + ret += lm::ngram::Subsume(context.LanguageModel(), before->left, before->right, after->left, after->right, update_reveal.left.length); + } + if (victim == 0) { + update.between[0].right = after->right; + } else { + update.between[2].left = before->left; + } + } + return previous.score + (ret + just_after) * context.GetWeights().LM(); +} + +} // namespace + +template <class Model> PartialEdge *EdgeGenerator::Pop(Context<Model> &context, boost::pool<> &partial_edge_pool) { + assert(!generate_.empty()); + PartialEdge &top = *generate_.top(); + generate_.pop(); + unsigned int victim = 0; + unsigned char lowest_length = 255; + for (unsigned char i = 0; i != arity_; ++i) { + if (!top.nt[i].Complete() && top.nt[i].Length() < lowest_length) { + lowest_length = top.nt[i].Length(); + victim = i; + } + } + if (lowest_length == 255) { + // All states report complete. + top.between[0].right = top.between[arity_].right; + // Now top.between[0] is the full edge state. + top_score_ = generate_.empty() ? -kScoreInf : generate_.top()->score; + return ⊤ + } + + unsigned int stay = !victim; + PartialEdge &continuation = *static_cast<PartialEdge*>(partial_edge_pool.malloc()); + float old_bound = top.nt[victim].Bound(); + // The alternate's score will change because alternate.nt[victim] changes. + bool split = top.nt[victim].Split(continuation.nt[victim]); + // top is now the alternate. + + continuation.nt[stay] = top.nt[stay]; + continuation.score = FastScore(context, victim, arity_, top, continuation); + // TODO: dedupe? + generate_.push(&continuation); + + if (split) { + // We have an alternate. + top.score += top.nt[victim].Bound() - old_bound; + // TODO: dedupe? + generate_.push(&top); + } else { + partial_edge_pool.free(&top); + } + + top_score_ = generate_.top()->score; + return NULL; +} + +template PartialEdge *EdgeGenerator::Pop(Context<lm::ngram::RestProbingModel> &context, boost::pool<> &partial_edge_pool); +template PartialEdge *EdgeGenerator::Pop(Context<lm::ngram::ProbingModel> &context, boost::pool<> &partial_edge_pool); +template PartialEdge *EdgeGenerator::Pop(Context<lm::ngram::TrieModel> &context, boost::pool<> &partial_edge_pool); +template PartialEdge *EdgeGenerator::Pop(Context<lm::ngram::QuantTrieModel> &context, boost::pool<> &partial_edge_pool); +template PartialEdge *EdgeGenerator::Pop(Context<lm::ngram::ArrayTrieModel> &context, boost::pool<> &partial_edge_pool); +template PartialEdge *EdgeGenerator::Pop(Context<lm::ngram::QuantArrayTrieModel> &context, boost::pool<> &partial_edge_pool); + +} // namespace search diff --git a/klm/search/edge_generator.hh b/klm/search/edge_generator.hh new file mode 100644 index 00000000..875ccc5e --- /dev/null +++ b/klm/search/edge_generator.hh @@ -0,0 +1,58 @@ +#ifndef SEARCH_EDGE_GENERATOR__ +#define SEARCH_EDGE_GENERATOR__ + +#include "search/edge.hh" +#include "search/note.hh" + +#include <boost/pool/pool.hpp> +#include <boost/unordered_map.hpp> + +#include <functional> +#include <queue> + +namespace lm { +namespace ngram { +class ChartState; +} // namespace ngram +} // namespace lm + +namespace search { + +template <class Model> class Context; + +class VertexGenerator; + +struct PartialEdgePointerLess : std::binary_function<const PartialEdge *, const PartialEdge *, bool> { + bool operator()(const PartialEdge *first, const PartialEdge *second) const { + return *first < *second; + } +}; + +class EdgeGenerator { + public: + EdgeGenerator(PartialEdge &root, unsigned char arity, Note note); + + Score TopScore() const { + return top_score_; + } + + Note GetNote() const { + return note_; + } + + // Pop. If there's a complete hypothesis, return it. Otherwise return NULL. + template <class Model> PartialEdge *Pop(Context<Model> &context, boost::pool<> &partial_edge_pool); + + private: + Score top_score_; + + unsigned char arity_; + + typedef std::priority_queue<PartialEdge*, std::vector<PartialEdge*>, PartialEdgePointerLess> Generate; + Generate generate_; + + Note note_; +}; + +} // namespace search +#endif // SEARCH_EDGE_GENERATOR__ diff --git a/klm/search/edge_queue.cc b/klm/search/edge_queue.cc new file mode 100644 index 00000000..e3ae6ebf --- /dev/null +++ b/klm/search/edge_queue.cc @@ -0,0 +1,25 @@ +#include "search/edge_queue.hh" + +#include "lm/left.hh" +#include "search/context.hh" + +#include <stdint.h> + +namespace search { + +EdgeQueue::EdgeQueue(unsigned int pop_limit_hint) : partial_edge_pool_(sizeof(PartialEdge), pop_limit_hint * 2) { + take_ = static_cast<PartialEdge*>(partial_edge_pool_.malloc()); +} + +/*void EdgeQueue::AddEdge(PartialEdge &root, unsigned char arity, Note note) { + // Ignore empty edges. + for (unsigned char i = 0; i < edge.Arity(); ++i) { + PartialVertex root(edge.GetVertex(i).RootPartial()); + if (root.Empty()) return; + total_score += root.Bound(); + } + PartialEdge &allocated = *static_cast<PartialEdge*>(partial_edge_pool_.malloc()); + allocated.score = total_score; +}*/ + +} // namespace search diff --git a/klm/search/edge_queue.hh b/klm/search/edge_queue.hh new file mode 100644 index 00000000..187eaed7 --- /dev/null +++ b/klm/search/edge_queue.hh @@ -0,0 +1,73 @@ +#ifndef SEARCH_EDGE_QUEUE__ +#define SEARCH_EDGE_QUEUE__ + +#include "search/edge.hh" +#include "search/edge_generator.hh" +#include "search/note.hh" + +#include <boost/pool/pool.hpp> +#include <boost/pool/object_pool.hpp> + +#include <queue> + +namespace search { + +template <class Model> class Context; + +class EdgeQueue { + public: + explicit EdgeQueue(unsigned int pop_limit_hint); + + PartialEdge &InitializeEdge() { + return *take_; + } + + void AddEdge(unsigned char arity, Note note) { + generate_.push(edge_pool_.construct(*take_, arity, note)); + take_ = static_cast<PartialEdge*>(partial_edge_pool_.malloc()); + } + + bool Empty() const { return generate_.empty(); } + + /* Generate hypotheses and send them to output. Normally, output is a + * VertexGenerator, but the decoder may want to route edges to different + * vertices i.e. if they have different LHS non-terminal labels. + */ + template <class Model, class Output> void Search(Context<Model> &context, Output &output) { + int to_pop = context.PopLimit(); + while (to_pop > 0 && !generate_.empty()) { + EdgeGenerator *top = generate_.top(); + generate_.pop(); + PartialEdge *ret = top->Pop(context, partial_edge_pool_); + if (ret) { + output.NewHypothesis(*ret, top->GetNote()); + --to_pop; + if (top->TopScore() != -kScoreInf) { + generate_.push(top); + } + } else { + generate_.push(top); + } + } + output.FinishedSearch(); + } + + private: + boost::object_pool<EdgeGenerator> edge_pool_; + + struct LessByTopScore : public std::binary_function<const EdgeGenerator *, const EdgeGenerator *, bool> { + bool operator()(const EdgeGenerator *first, const EdgeGenerator *second) const { + return first->TopScore() < second->TopScore(); + } + }; + + typedef std::priority_queue<EdgeGenerator*, std::vector<EdgeGenerator*>, LessByTopScore> Generate; + Generate generate_; + + boost::pool<> partial_edge_pool_; + + PartialEdge *take_; +}; + +} // namespace search +#endif // SEARCH_EDGE_QUEUE__ diff --git a/klm/search/final.hh b/klm/search/final.hh new file mode 100644 index 00000000..1b3092ac --- /dev/null +++ b/klm/search/final.hh @@ -0,0 +1,39 @@ +#ifndef SEARCH_FINAL__ +#define SEARCH_FINAL__ + +#include "search/arity.hh" +#include "search/note.hh" +#include "search/types.hh" + +#include <boost/array.hpp> + +namespace search { + +class Final { + public: + typedef boost::array<const Final*, search::kMaxArity> ChildArray; + + void Reset(Score bound, Note note, const Final &left, const Final &right) { + bound_ = bound; + note_ = note; + children_[0] = &left; + children_[1] = &right; + } + + const ChildArray &Children() const { return children_; } + + Note GetNote() const { return note_; } + + Score Bound() const { return bound_; } + + private: + Score bound_; + + Note note_; + + ChildArray children_; +}; + +} // namespace search + +#endif // SEARCH_FINAL__ diff --git a/klm/search/note.hh b/klm/search/note.hh new file mode 100644 index 00000000..50bed06e --- /dev/null +++ b/klm/search/note.hh @@ -0,0 +1,12 @@ +#ifndef SEARCH_NOTE__ +#define SEARCH_NOTE__ + +namespace search { + +union Note { + const void *vp; +}; + +} // namespace search + +#endif // SEARCH_NOTE__ diff --git a/klm/search/rule.cc b/klm/search/rule.cc new file mode 100644 index 00000000..5b00207e --- /dev/null +++ b/klm/search/rule.cc @@ -0,0 +1,43 @@ +#include "search/rule.hh" + +#include "search/context.hh" +#include "search/final.hh" + +#include <ostream> + +#include <math.h> + +namespace search { + +template <class Model> float ScoreRule(const Context<Model> &context, const std::vector<lm::WordIndex> &words, bool prepend_bos, lm::ngram::ChartState *writing) { + unsigned int oov_count = 0; + float prob = 0.0; + const Model &model = context.LanguageModel(); + const lm::WordIndex oov = model.GetVocabulary().NotFound(); + for (std::vector<lm::WordIndex>::const_iterator word = words.begin(); ; ++word) { + lm::ngram::RuleScore<Model> scorer(model, *(writing++)); + // TODO: optimize + if (prepend_bos && (word == words.begin())) { + scorer.BeginSentence(); + } + for (; ; ++word) { + if (word == words.end()) { + prob += scorer.Finish(); + return static_cast<float>(oov_count) * context.GetWeights().OOV() + prob * context.GetWeights().LM(); + } + if (*word == kNonTerminal) break; + if (*word == oov) ++oov_count; + scorer.Terminal(*word); + } + prob += scorer.Finish(); + } +} + +template float ScoreRule(const Context<lm::ngram::RestProbingModel> &model, const std::vector<lm::WordIndex> &words, bool prepend_bos, lm::ngram::ChartState *writing); +template float ScoreRule(const Context<lm::ngram::ProbingModel> &model, const std::vector<lm::WordIndex> &words, bool prepend_bos, lm::ngram::ChartState *writing); +template float ScoreRule(const Context<lm::ngram::TrieModel> &model, const std::vector<lm::WordIndex> &words, bool prepend_bos, lm::ngram::ChartState *writing); +template float ScoreRule(const Context<lm::ngram::QuantTrieModel> &model, const std::vector<lm::WordIndex> &words, bool prepend_bos, lm::ngram::ChartState *writing); +template float ScoreRule(const Context<lm::ngram::ArrayTrieModel> &model, const std::vector<lm::WordIndex> &words, bool prepend_bos, lm::ngram::ChartState *writing); +template float ScoreRule(const Context<lm::ngram::QuantArrayTrieModel> &model, const std::vector<lm::WordIndex> &words, bool prepend_bos, lm::ngram::ChartState *writing); + +} // namespace search diff --git a/klm/search/rule.hh b/klm/search/rule.hh new file mode 100644 index 00000000..0ce2794d --- /dev/null +++ b/klm/search/rule.hh @@ -0,0 +1,20 @@ +#ifndef SEARCH_RULE__ +#define SEARCH_RULE__ + +#include "lm/left.hh" +#include "lm/word_index.hh" +#include "search/types.hh" + +#include <vector> + +namespace search { + +template <class Model> class Context; + +const lm::WordIndex kNonTerminal = lm::kMaxWordIndex; + +template <class Model> float ScoreRule(const Context<Model> &context, const std::vector<lm::WordIndex> &words, bool prepend_bos, lm::ngram::ChartState *state_out); + +} // namespace search + +#endif // SEARCH_RULE__ diff --git a/klm/search/source.hh b/klm/search/source.hh new file mode 100644 index 00000000..11839f7b --- /dev/null +++ b/klm/search/source.hh @@ -0,0 +1,48 @@ +#ifndef SEARCH_SOURCE__ +#define SEARCH_SOURCE__ + +#include "search/types.hh" + +#include <assert.h> +#include <vector> + +namespace search { + +template <class Final> class Source { + public: + Source() : bound_(kScoreInf) {} + + Index Size() const { + return final_.size(); + } + + Score Bound() const { + return bound_; + } + + const Final &operator[](Index index) const { + return *final_[index]; + } + + Score ScoreOrBound(Index index) const { + return Size() > index ? final_[index]->Total() : Bound(); + } + + protected: + void AddFinal(const Final &store) { + final_.push_back(&store); + } + + void SetBound(Score to) { + assert(to <= bound_ + 0.001); + bound_ = to; + } + + private: + std::vector<const Final *> final_; + + Score bound_; +}; + +} // namespace search +#endif // SEARCH_SOURCE__ diff --git a/klm/search/types.hh b/klm/search/types.hh new file mode 100644 index 00000000..9726379f --- /dev/null +++ b/klm/search/types.hh @@ -0,0 +1,18 @@ +#ifndef SEARCH_TYPES__ +#define SEARCH_TYPES__ + +#include <cmath> + +namespace search { + +typedef float Score; +const Score kScoreInf = INFINITY; + +// This could have been an enum but gcc wants 4 bytes. +typedef bool ExtendDirection; +const ExtendDirection kExtendLeft = 0; +const ExtendDirection kExtendRight = 1; + +} // namespace search + +#endif // SEARCH_TYPES__ diff --git a/klm/search/vertex.cc b/klm/search/vertex.cc new file mode 100644 index 00000000..cc53c0dd --- /dev/null +++ b/klm/search/vertex.cc @@ -0,0 +1,48 @@ +#include "search/vertex.hh" + +#include "search/context.hh" + +#include <algorithm> +#include <functional> + +#include <assert.h> + +namespace search { + +namespace { + +struct GreaterByBound : public std::binary_function<const VertexNode *, const VertexNode *, bool> { + bool operator()(const VertexNode *first, const VertexNode *second) const { + return first->Bound() > second->Bound(); + } +}; + +} // namespace + +void VertexNode::SortAndSet(ContextBase &context, VertexNode **parent_ptr) { + if (Complete()) { + assert(end_); + assert(extend_.empty()); + bound_ = end_->Bound(); + return; + } + if (extend_.size() == 1 && parent_ptr) { + *parent_ptr = extend_[0]; + extend_[0]->SortAndSet(context, parent_ptr); + context.DeleteVertexNode(this); + return; + } + for (std::vector<VertexNode*>::iterator i = extend_.begin(); i != extend_.end(); ++i) { + (*i)->SortAndSet(context, &*i); + } + std::sort(extend_.begin(), extend_.end(), GreaterByBound()); + bound_ = extend_.front()->Bound(); +} + +namespace { +VertexNode kBlankVertexNode; +} // namespace + +PartialVertex kBlankPartialVertex(kBlankVertexNode); + +} // namespace search diff --git a/klm/search/vertex.hh b/klm/search/vertex.hh new file mode 100644 index 00000000..e1a9ad11 --- /dev/null +++ b/klm/search/vertex.hh @@ -0,0 +1,158 @@ +#ifndef SEARCH_VERTEX__ +#define SEARCH_VERTEX__ + +#include "lm/left.hh" +#include "search/final.hh" +#include "search/types.hh" + +#include <boost/unordered_set.hpp> + +#include <queue> +#include <vector> + +#include <stdint.h> + +namespace search { + +class ContextBase; + +class VertexNode { + public: + VertexNode() : end_(NULL) {} + + void InitRoot() { + extend_.clear(); + state_.left.full = false; + state_.left.length = 0; + state_.right.length = 0; + right_full_ = false; + bound_ = -kScoreInf; + end_ = NULL; + } + + lm::ngram::ChartState &MutableState() { return state_; } + bool &MutableRightFull() { return right_full_; } + + void AddExtend(VertexNode *next) { + extend_.push_back(next); + } + + void SetEnd(Final *end) { end_ = end; } + + Final &MutableEnd() { return *end_; } + + void SortAndSet(ContextBase &context, VertexNode **parent_pointer); + + // Should only happen to a root node when the entire vertex is empty. + bool Empty() const { + return !end_ && extend_.empty(); + } + + bool Complete() const { + return end_; + } + + const lm::ngram::ChartState &State() const { return state_; } + bool RightFull() const { return right_full_; } + + Score Bound() const { + return bound_; + } + + unsigned char Length() const { + return state_.left.length + state_.right.length; + } + + // May be NULL. + const Final *End() const { return end_; } + + const VertexNode &operator[](size_t index) const { + return *extend_[index]; + } + + size_t Size() const { + return extend_.size(); + } + + private: + std::vector<VertexNode*> extend_; + + lm::ngram::ChartState state_; + bool right_full_; + + Score bound_; + Final *end_; +}; + +class PartialVertex { + public: + PartialVertex() {} + + explicit PartialVertex(const VertexNode &back) : back_(&back), index_(0) {} + + bool Empty() const { return back_->Empty(); } + + bool Complete() const { return back_->Complete(); } + + const lm::ngram::ChartState &State() const { return back_->State(); } + bool RightFull() const { return back_->RightFull(); } + + Score Bound() const { return Complete() ? back_->End()->Bound() : (*back_)[index_].Bound(); } + + unsigned char Length() const { return back_->Length(); } + + bool HasAlternative() const { + return index_ + 1 < back_->Size(); + } + + // Split into continuation and alternative, rendering this the alternative. + bool Split(PartialVertex &continuation) { + assert(!Complete()); + continuation.back_ = &((*back_)[index_]); + continuation.index_ = 0; + if (index_ + 1 < back_->Size()) { + ++index_; + return true; + } + return false; + } + + const Final &End() const { + return *back_->End(); + } + + private: + const VertexNode *back_; + unsigned int index_; +}; + +extern PartialVertex kBlankPartialVertex; + +class Vertex { + public: + Vertex() {} + + PartialVertex RootPartial() const { return PartialVertex(root_); } + + const Final *BestChild() const { + PartialVertex top(RootPartial()); + if (top.Empty()) { + return NULL; + } else { + PartialVertex continuation; + while (!top.Complete()) { + top.Split(continuation); + top = continuation; + } + return &top.End(); + } + } + + private: + friend class VertexGenerator; + + VertexNode root_; +}; + +} // namespace search +#endif // SEARCH_VERTEX__ diff --git a/klm/search/vertex_generator.cc b/klm/search/vertex_generator.cc new file mode 100644 index 00000000..d94e6e06 --- /dev/null +++ b/klm/search/vertex_generator.cc @@ -0,0 +1,83 @@ +#include "search/vertex_generator.hh" + +#include "lm/left.hh" +#include "search/context.hh" +#include "search/edge.hh" + +#include <stdint.h> + +namespace search { + +VertexGenerator::VertexGenerator(ContextBase &context, Vertex &gen) : context_(context), gen_(gen) { + gen.root_.InitRoot(); + root_.under = &gen.root_; +} + +namespace { +const uint64_t kCompleteAdd = static_cast<uint64_t>(-1); +} // namespace + +void VertexGenerator::NewHypothesis(const PartialEdge &partial, Note note) { + const lm::ngram::ChartState &state = partial.CompletedState(); + std::pair<Existing::iterator, bool> got(existing_.insert(std::pair<uint64_t, Final*>(hash_value(state), NULL))); + if (!got.second) { + // Found it already. + Final &exists = *got.first->second; + if (exists.Bound() < partial.score) { + exists.Reset(partial.score, note, partial.nt[0].End(), partial.nt[1].End()); + } + return; + } + unsigned char left = 0, right = 0; + Trie *node = &root_; + while (true) { + if (left == state.left.length) { + node = &FindOrInsert(*node, kCompleteAdd - state.left.full, state, left, true, right, false); + for (; right < state.right.length; ++right) { + node = &FindOrInsert(*node, state.right.words[right], state, left, true, right + 1, false); + } + break; + } + node = &FindOrInsert(*node, state.left.pointers[left], state, left + 1, false, right, false); + left++; + if (right == state.right.length) { + node = &FindOrInsert(*node, kCompleteAdd - state.left.full, state, left, false, right, true); + for (; left < state.left.length; ++left) { + node = &FindOrInsert(*node, state.left.pointers[left], state, left + 1, false, right, true); + } + break; + } + node = &FindOrInsert(*node, state.right.words[right], state, left, false, right + 1, false); + right++; + } + + node = &FindOrInsert(*node, kCompleteAdd - state.left.full, state, state.left.length, true, state.right.length, true); + got.first->second = CompleteTransition(*node, state, note, partial); +} + +VertexGenerator::Trie &VertexGenerator::FindOrInsert(VertexGenerator::Trie &node, uint64_t added, const lm::ngram::ChartState &state, unsigned char left, bool left_full, unsigned char right, bool right_full) { + VertexGenerator::Trie &next = node.extend[added]; + if (!next.under) { + next.under = context_.NewVertexNode(); + lm::ngram::ChartState &writing = next.under->MutableState(); + writing = state; + writing.left.full &= left_full && state.left.full; + next.under->MutableRightFull() = right_full && state.left.full; + writing.left.length = left; + writing.right.length = right; + node.under->AddExtend(next.under); + } + return next; +} + +Final *VertexGenerator::CompleteTransition(VertexGenerator::Trie &starter, const lm::ngram::ChartState &state, Note note, const PartialEdge &partial) { + VertexNode &node = *starter.under; + assert(node.State().left.full == state.left.full); + assert(!node.End()); + Final *final = context_.NewFinal(); + final->Reset(partial.score, note, partial.nt[0].End(), partial.nt[1].End()); + node.SetEnd(final); + return final; +} + +} // namespace search diff --git a/klm/search/vertex_generator.hh b/klm/search/vertex_generator.hh new file mode 100644 index 00000000..6b98da3e --- /dev/null +++ b/klm/search/vertex_generator.hh @@ -0,0 +1,59 @@ +#ifndef SEARCH_VERTEX_GENERATOR__ +#define SEARCH_VERTEX_GENERATOR__ + +#include "search/note.hh" +#include "search/vertex.hh" + +#include <boost/unordered_map.hpp> + +#include <queue> + +namespace lm { +namespace ngram { +class ChartState; +} // namespace ngram +} // namespace lm + +namespace search { + +class ContextBase; +class Final; +struct PartialEdge; + +class VertexGenerator { + public: + VertexGenerator(ContextBase &context, Vertex &gen); + + void NewHypothesis(const PartialEdge &partial, Note note); + + void FinishedSearch() { + root_.under->SortAndSet(context_, NULL); + } + + const Vertex &Generating() const { return gen_; } + + private: + // Parallel structure to VertexNode. + struct Trie { + Trie() : under(NULL) {} + + VertexNode *under; + boost::unordered_map<uint64_t, Trie> extend; + }; + + Trie &FindOrInsert(Trie &node, uint64_t added, const lm::ngram::ChartState &state, unsigned char left, bool left_full, unsigned char right, bool right_full); + + Final *CompleteTransition(Trie &node, const lm::ngram::ChartState &state, Note note, const PartialEdge &partial); + + ContextBase &context_; + + Vertex &gen_; + + Trie root_; + + typedef boost::unordered_map<uint64_t, Final*> Existing; + Existing existing_; +}; + +} // namespace search +#endif // SEARCH_VERTEX_GENERATOR__ diff --git a/klm/search/weights.cc b/klm/search/weights.cc new file mode 100644 index 00000000..d65471ad --- /dev/null +++ b/klm/search/weights.cc @@ -0,0 +1,71 @@ +#include "search/weights.hh" +#include "util/tokenize_piece.hh" + +#include <cstdlib> + +namespace search { + +namespace { +struct Insert { + void operator()(boost::unordered_map<std::string, search::Score> &map, StringPiece name, search::Score score) const { + std::string copy(name.data(), name.size()); + map[copy] = score; + } +}; + +struct DotProduct { + search::Score total; + DotProduct() : total(0.0) {} + + void operator()(const boost::unordered_map<std::string, search::Score> &map, StringPiece name, search::Score score) { + boost::unordered_map<std::string, search::Score>::const_iterator i(FindStringPiece(map, name)); + if (i != map.end()) + total += score * i->second; + } +}; + +template <class Map, class Op> void Parse(StringPiece text, Map &map, Op &op) { + for (util::TokenIter<util::SingleCharacter, true> spaces(text, ' '); spaces; ++spaces) { + util::TokenIter<util::SingleCharacter> equals(*spaces, '='); + UTIL_THROW_IF(!equals, WeightParseException, "Bad weight token " << *spaces); + StringPiece name(*equals); + UTIL_THROW_IF(!++equals, WeightParseException, "Bad weight token " << *spaces); + char *end; + // Assumes proper termination. + double value = std::strtod(equals->data(), &end); + UTIL_THROW_IF(end != equals->data() + equals->size(), WeightParseException, "Failed to parse weight" << *equals); + UTIL_THROW_IF(++equals, WeightParseException, "Too many equals in " << *spaces); + op(map, name, value); + } +} + +} // namespace + +Weights::Weights(StringPiece text) { + Insert op; + Parse<Map, Insert>(text, map_, op); + lm_ = Steal("LanguageModel"); + oov_ = Steal("OOV"); + word_penalty_ = Steal("WordPenalty"); +} + +Weights::Weights(Score lm, Score oov, Score word_penalty) : lm_(lm), oov_(oov), word_penalty_(word_penalty) {} + +search::Score Weights::DotNoLM(StringPiece text) const { + DotProduct dot; + Parse<const Map, DotProduct>(text, map_, dot); + return dot.total; +} + +float Weights::Steal(const std::string &str) { + Map::iterator i(map_.find(str)); + if (i == map_.end()) { + return 0.0; + } else { + float ret = i->second; + map_.erase(i); + return ret; + } +} + +} // namespace search diff --git a/klm/search/weights.hh b/klm/search/weights.hh new file mode 100644 index 00000000..df1c419f --- /dev/null +++ b/klm/search/weights.hh @@ -0,0 +1,52 @@ +// For now, the individual features are not kept. +#ifndef SEARCH_WEIGHTS__ +#define SEARCH_WEIGHTS__ + +#include "search/types.hh" +#include "util/exception.hh" +#include "util/string_piece.hh" + +#include <boost/unordered_map.hpp> + +#include <string> + +namespace search { + +class WeightParseException : public util::Exception { + public: + WeightParseException() {} + ~WeightParseException() throw() {} +}; + +class Weights { + public: + // Parses weights, sets lm_weight_, removes it from map_. + explicit Weights(StringPiece text); + + // Just the three scores we care about adding. + Weights(Score lm, Score oov, Score word_penalty); + + Score DotNoLM(StringPiece text) const; + + Score LM() const { return lm_; } + + Score OOV() const { return oov_; } + + Score WordPenalty() const { return word_penalty_; } + + // Mostly for testing. + const boost::unordered_map<std::string, Score> &GetMap() const { return map_; } + + private: + float Steal(const std::string &str); + + typedef boost::unordered_map<std::string, Score> Map; + + Map map_; + + Score lm_, oov_, word_penalty_; +}; + +} // namespace search + +#endif // SEARCH_WEIGHTS__ diff --git a/klm/search/weights_test.cc b/klm/search/weights_test.cc new file mode 100644 index 00000000..4811ff06 --- /dev/null +++ b/klm/search/weights_test.cc @@ -0,0 +1,38 @@ +#include "search/weights.hh" + +#define BOOST_TEST_MODULE WeightTest +#include <boost/test/unit_test.hpp> +#include <boost/test/floating_point_comparison.hpp> + +namespace search { +namespace { + +#define CHECK_WEIGHT(value, string) \ + i = parsed.find(string); \ + BOOST_REQUIRE(i != parsed.end()); \ + BOOST_CHECK_CLOSE((value), i->second, 0.001); + +BOOST_AUTO_TEST_CASE(parse) { + // These are not real feature weights. + Weights w("rarity=0 phrase-SGT=0 phrase-TGS=9.45117 lhsGrhs=0 lexical-SGT=2.33833 lexical-TGS=-28.3317 abstract?=0 LanguageModel=3 lexical?=1 glue?=5"); + const boost::unordered_map<std::string, search::Score> &parsed = w.GetMap(); + boost::unordered_map<std::string, search::Score>::const_iterator i; + CHECK_WEIGHT(0.0, "rarity"); + CHECK_WEIGHT(0.0, "phrase-SGT"); + CHECK_WEIGHT(9.45117, "phrase-TGS"); + CHECK_WEIGHT(2.33833, "lexical-SGT"); + BOOST_CHECK(parsed.end() == parsed.find("lm")); + BOOST_CHECK_CLOSE(3.0, w.LM(), 0.001); + CHECK_WEIGHT(-28.3317, "lexical-TGS"); + CHECK_WEIGHT(5.0, "glue?"); +} + +BOOST_AUTO_TEST_CASE(dot) { + Weights w("rarity=0 phrase-SGT=0 phrase-TGS=9.45117 lhsGrhs=0 lexical-SGT=2.33833 lexical-TGS=-28.3317 abstract?=0 LanguageModel=3 lexical?=1 glue?=5"); + BOOST_CHECK_CLOSE(9.45117 * 3.0, w.DotNoLM("phrase-TGS=3.0"), 0.001); + BOOST_CHECK_CLOSE(9.45117 * 3.0, w.DotNoLM("phrase-TGS=3.0 LanguageModel=10"), 0.001); + BOOST_CHECK_CLOSE(9.45117 * 3.0 + 28.3317 * 17.4, w.DotNoLM("rarity=5 phrase-TGS=3.0 LanguageModel=10 lexical-TGS=-17.4"), 0.001); +} + +} // namespace +} // namespace search |