diff options
author | Kenneth Heafield <github@kheafield.com> | 2012-09-13 04:04:17 -0700 |
---|---|---|
committer | Kenneth Heafield <github@kheafield.com> | 2012-09-13 04:04:17 -0700 |
commit | 8d0baa1dfb5c18f5be45b4d12b3f925dc04afc48 (patch) | |
tree | af96b81bfa50321d36a02bc036eabfeb29c32538 /decoder/lazy.cc | |
parent | a548c46f3d636093d26f44d5b7642c98def6f351 (diff) | |
parent | 816f33174b2c2938a3df9c75b2834da6f5184a3e (diff) |
Merge branch 'master' of github.com:kpu/cdec
Diffstat (limited to 'decoder/lazy.cc')
-rw-r--r-- | decoder/lazy.cc | 78 |
1 files changed, 57 insertions, 21 deletions
diff --git a/decoder/lazy.cc b/decoder/lazy.cc index f5b61c75..4776c1b8 100644 --- a/decoder/lazy.cc +++ b/decoder/lazy.cc @@ -1,15 +1,23 @@ #include "hg.h" #include "lazy.h" +#include "fdict.h" #include "tdict.h" #include "lm/enumerate_vocab.hh" #include "lm/model.hh" +#include "search/config.hh" +#include "search/context.hh" #include "search/edge.hh" #include "search/vertex.hh" +#include "search/vertex_generator.hh" #include "util/exception.hh" +#include <boost/scoped_ptr.hpp> #include <boost/scoped_array.hpp> +#include <iostream> +#include <vector> + namespace { struct MapVocab : public lm::EnumerateVocab { @@ -19,13 +27,13 @@ struct MapVocab : public lm::EnumerateVocab { // Do not call after Lookup. void Add(lm::WordIndex index, const StringPiece &str) { const WordID cdec_id = TD::Convert(str.as_string()); - if (cdec_id >= out_->size()) out_.resize(cdec_id + 1); + if (cdec_id >= out_.size()) out_.resize(cdec_id + 1); out_[cdec_id] = index; } // Assumes Add has been called and will never be called again. lm::WordIndex FromCDec(WordID id) const { - return out_[out.size() > id ? id : 0]; + return out_[out_.size() > id ? id : 0]; } private: @@ -34,44 +42,50 @@ struct MapVocab : public lm::EnumerateVocab { class LazyBase { public: - LazyBase() {} + LazyBase(const std::vector<weight_t> &weights) : + cdec_weights_(weights), + config_(search::Weights(weights[FD::Convert("KLanguageModel")], weights[FD::Convert("KLanguageModel_OOV")], weights[FD::Convert("WordPenalty")]), 1000) {} virtual ~LazyBase() {} virtual void Search(const Hypergraph &hg) const = 0; - static LazyBase *Load(const char *model_file); + static LazyBase *Load(const char *model_file, const std::vector<weight_t> &weights); protected: - lm::ngram::Config GetConfig() const { + lm::ngram::Config GetConfig() { lm::ngram::Config ret; ret.enumerate_vocab = &vocab_; return ret; } MapVocab vocab_; + + const std::vector<weight_t> &cdec_weights_; + + const search::Config config_; }; template <class Model> class Lazy : public LazyBase { public: - explicit Lazy(const char *model_file) : m_(model_file, GetConfig()) {} + Lazy(const char *model_file, const std::vector<weight_t> &weights) : LazyBase(weights), m_(model_file, GetConfig()) {} void Search(const Hypergraph &hg) const; private: - void ConvertEdge(const Context<Model> &context, bool final, search::Vertex *vertices, const Hypergraph::Edge &in, search::Edge &out) const; + void ConvertEdge(const search::Context<Model> &context, bool final, search::Vertex *vertices, const Hypergraph::Edge &in, search::Edge &out) const; const Model m_; }; -static LazyBase *LazyBase::Load(const char *model_file) { +LazyBase *LazyBase::Load(const char *model_file, const std::vector<weight_t> &weights) { lm::ngram::ModelType model_type; - if (!lm::ngram::RecognizeBinary(lm_name, model_type)) model_type = lm::ngram::PROBING; + if (!lm::ngram::RecognizeBinary(model_file, model_type)) model_type = lm::ngram::PROBING; switch (model_type) { case lm::ngram::PROBING: - return new Lazy<lm::ngram::ProbingModel>(model_file); + return new Lazy<lm::ngram::ProbingModel>(model_file, weights); case lm::ngram::REST_PROBING: - return new Lazy<lm::ngram::RestProbingModel>(model_file); + return new Lazy<lm::ngram::RestProbingModel>(model_file, weights); default: UTIL_THROW(util::Exception, "Sorry this lm type isn't supported yet."); } @@ -80,25 +94,41 @@ static LazyBase *LazyBase::Load(const char *model_file) { template <class Model> void Lazy<Model>::Search(const Hypergraph &hg) const { boost::scoped_array<search::Vertex> out_vertices(new search::Vertex[hg.nodes_.size()]); boost::scoped_array<search::Edge> out_edges(new search::Edge[hg.edges_.size()]); + + search::Context<Model> context(config_, m_); + for (unsigned int i = 0; i < hg.nodes_.size(); ++i) { - search::Vertex *out_vertex = out_vertices[i]; + search::Vertex &out_vertex = out_vertices[i]; const Hypergraph::EdgesVector &down_edges = hg.nodes_[i].in_edges_; - for (unsigned int j = 0; j < edges.size(); ++j) { + for (unsigned int j = 0; j < down_edges.size(); ++j) { unsigned int edge_index = down_edges[j]; - const Hypergraph::Edge &in_edge = hg.edges_[edge_index]; - search::Edge &out_edge = out_edges[edge_index]; + ConvertEdge(context, i == hg.nodes_.size() - 1, out_vertices.get(), hg.edges_[edge_index], out_edges[edge_index]); + out_vertex.Add(out_edges[edge_index]); } + out_vertex.FinishedAdding(); + search::VertexGenerator(context, out_vertex); + } + search::PartialVertex top = out_vertices[hg.nodes_.size() - 1].RootPartial(); + if (top.Empty()) { + std::cout << "NO PATH FOUND"; + } else { + search::PartialVertex continuation; + while (!top.Complete()) { + top.Split(continuation); + top = continuation; + } + std::cout << top.End().Bound() << std::endl; } } // TODO: get weights into here somehow. -template <class Model> void Lazy<Model>::ConvertEdge(const Context<Model> &context, bool final, search::Vertices *vertices, const Hypergraph::Edge &in, search::Edge &out) const { - const std::vector<WordID> &e = in_edge.rule_->e(); +template <class Model> void Lazy<Model>::ConvertEdge(const search::Context<Model> &context, bool final, search::Vertex *vertices, const Hypergraph::Edge &in, search::Edge &out) const { + const std::vector<WordID> &e = in.rule_->e(); std::vector<lm::WordIndex> words; unsigned int terminals = 0; for (std::vector<WordID>::const_iterator word = e.begin(); word != e.end(); ++word) { if (*word <= 0) { - out.Add(vertices[edge.tail_nodes_[-*word]]); + out.Add(vertices[in.tail_nodes_[-*word]]); words.push_back(lm::kMaxWordIndex); } else { ++terminals; @@ -110,13 +140,19 @@ template <class Model> void Lazy<Model>::ConvertEdge(const Context<Model> &conte words.push_back(m_.GetVocabulary().EndSentence()); } - float additive = edge.rule_->GetFeatureValues().dot(weight_vector); + float additive = in.rule_->GetFeatureValues().dot(cdec_weights_); + additive -= terminals * context.GetWeights().WordPenalty() * static_cast<float>(terminals) / M_LN10; out.InitRule().Init(context, additive, words, final); } -} // namespace +boost::scoped_ptr<LazyBase> AwfulGlobalLazy; -void PassToLazy(const Hypergraph &hg) { +} // namespace +void PassToLazy(const Hypergraph &hg, const std::vector<weight_t> &weights) { + if (!AwfulGlobalLazy.get()) { + AwfulGlobalLazy.reset(LazyBase::Load("lm", weights)); + } + AwfulGlobalLazy->Search(hg); } |