summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorKenneth Heafield <github@kheafield.com>2012-09-13 04:04:17 -0700
committerKenneth Heafield <github@kheafield.com>2012-09-13 04:04:17 -0700
commit7f809684f51d4e520b506c74ed2b475d611eee24 (patch)
tree9f73d5e50859f7c15d5279d4494a3fca304d96f2
parent5258e0355a3fe54ce29877f5b2a8d5cd5f1737ca (diff)
parent8505fdfdf0bc4ce9acec42e1980a2fdd4f254109 (diff)
Merge branch 'master' of github.com:kpu/cdec
-rw-r--r--decoder/Jamfile2
-rw-r--r--decoder/decoder.cc4
-rw-r--r--decoder/lazy.cc78
-rw-r--r--decoder/lazy.h5
-rw-r--r--klm/search/config.hh6
-rw-r--r--klm/search/weights.cc2
-rw-r--r--klm/search/weights.hh17
7 files changed, 82 insertions, 32 deletions
diff --git a/decoder/Jamfile b/decoder/Jamfile
index da02d063..d778dc7f 100644
--- a/decoder/Jamfile
+++ b/decoder/Jamfile
@@ -58,10 +58,12 @@ lib decoder :
rescore_translator.cc
hg_remove_eps.cc
hg_union.cc
+ lazy.cc
$(glc)
..//utils
..//mteval
../klm/lm//kenlm
+ ../klm/search//search
..//boost_program_options
: <include>.
: :
diff --git a/decoder/decoder.cc b/decoder/decoder.cc
index a69a6d05..3a410cf2 100644
--- a/decoder/decoder.cc
+++ b/decoder/decoder.cc
@@ -38,6 +38,7 @@
#include "sampler.h"
#include "forest_writer.h" // TODO this section should probably be handled by an Observer
+#include "lazy.h"
#include "hg_io.h"
#include "aligner.h"
@@ -832,6 +833,9 @@ bool DecoderImpl::Decode(const string& input, DecoderObserver* o) {
if (conf.count("show_target_graph"))
HypergraphIO::WriteTarget(conf["show_target_graph"].as<string>(), sent_id, forest);
+ if (conf.count("lazy_search"))
+ PassToLazy(forest, CurrentWeightVector());
+
for (int pass = 0; pass < rescoring_passes.size(); ++pass) {
const RescoringPass& rp = rescoring_passes[pass];
const vector<weight_t>& cur_weights = *rp.weight_vector;
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);
}
diff --git a/decoder/lazy.h b/decoder/lazy.h
index aecd030d..3e71a3b0 100644
--- a/decoder/lazy.h
+++ b/decoder/lazy.h
@@ -1,8 +1,11 @@
#ifndef _LAZY_H_
#define _LAZY_H_
+#include "weights.h"
+#include <vector>
+
class Hypergraph;
-void PassToLazy(const Hypergraph &hg);
+void PassToLazy(const Hypergraph &hg, const std::vector<weight_t> &weights);
#endif // _LAZY_H_
diff --git a/klm/search/config.hh b/klm/search/config.hh
index e21e4b7c..ef8e2354 100644
--- a/klm/search/config.hh
+++ b/klm/search/config.hh
@@ -8,15 +8,15 @@ namespace search {
class Config {
public:
- Config(StringPiece weight_str, unsigned int pop_limit) :
- weights_(weight_str), pop_limit_(pop_limit) {}
+ 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:
- search::Weights weights_;
+ Weights weights_;
unsigned int pop_limit_;
};
diff --git a/klm/search/weights.cc b/klm/search/weights.cc
index 82ff3f12..d65471ad 100644
--- a/klm/search/weights.cc
+++ b/klm/search/weights.cc
@@ -49,6 +49,8 @@ Weights::Weights(StringPiece text) {
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);
diff --git a/klm/search/weights.hh b/klm/search/weights.hh
index 4a4388c7..df1c419f 100644
--- a/klm/search/weights.hh
+++ b/klm/search/weights.hh
@@ -23,25 +23,28 @@ class Weights {
// Parses weights, sets lm_weight_, removes it from map_.
explicit Weights(StringPiece text);
- search::Score DotNoLM(StringPiece text) const;
+ // Just the three scores we care about adding.
+ Weights(Score lm, Score oov, Score word_penalty);
- search::Score LM() const { return lm_; }
+ Score DotNoLM(StringPiece text) const;
- search::Score OOV() const { return oov_; }
+ Score LM() const { return lm_; }
- search::Score WordPenalty() const { return word_penalty_; }
+ Score OOV() const { return oov_; }
+
+ Score WordPenalty() const { return word_penalty_; }
// Mostly for testing.
- const boost::unordered_map<std::string, search::Score> &GetMap() const { return map_; }
+ const boost::unordered_map<std::string, Score> &GetMap() const { return map_; }
private:
float Steal(const std::string &str);
- typedef boost::unordered_map<std::string, search::Score> Map;
+ typedef boost::unordered_map<std::string, Score> Map;
Map map_;
- search::Score lm_, oov_, word_penalty_;
+ Score lm_, oov_, word_penalty_;
};
} // namespace search