From a47bbc78b3d38ea998b2d484470061140142048d Mon Sep 17 00:00:00 2001 From: Patrick Simianer Date: Mon, 18 Jun 2012 17:26:33 +0200 Subject: extract_rules cdec param --- .gitignore | 1 + decoder/decoder.cc | 22 +++++++++++++++------- decoder/viterbi.cc | 12 +++++++++++- decoder/viterbi.h | 5 ++++- dtrain/README.md | 22 ---------------------- 5 files changed, 31 insertions(+), 31 deletions(-) diff --git a/.gitignore b/.gitignore index 27c6a739..943e6dc5 100644 --- a/.gitignore +++ b/.gitignore @@ -128,6 +128,7 @@ decoder/rule_lexer.cc training/atools training/collapse_weights training/lbfgs_test +training/libtraining.a training/mr_optimize_reduce training/mr_em_adapted_reduce training/mr_em_map_adapter diff --git a/decoder/decoder.cc b/decoder/decoder.cc index cbb97a0d..333f0fb6 100644 --- a/decoder/decoder.cc +++ b/decoder/decoder.cc @@ -3,6 +3,7 @@ #include #include #include +#include #include "program_options.h" #include "stringlib.h" @@ -187,8 +188,8 @@ struct DecoderImpl { } void SetId(int next_sent_id) { sent_id = next_sent_id - 1; } - void forest_stats(Hypergraph &forest,string name,bool show_tree,bool show_deriv=false) { - cerr << viterbi_stats(forest,name,true,show_tree,show_deriv); + void forest_stats(Hypergraph &forest,string name,bool show_tree,bool show_deriv=false, bool extract_rules=false, boost::shared_ptr extract_file = boost::make_shared()) { + cerr << viterbi_stats(forest,name,true,show_tree,show_deriv,extract_rules, extract_file); cerr << endl; } @@ -424,7 +425,7 @@ DecoderImpl::DecoderImpl(po::variables_map& conf, int argc, char** argv, istream ("tagger_tagset,t", po::value(), "(Tagger) file containing tag set") ("csplit_output_plf", "(Compound splitter) Output lattice in PLF format") ("csplit_preserve_full_word", "(Compound splitter) Always include the unsegmented form in the output lattice") - ("extract_rules", po::value(), "Extract the rules used in translation (de-duped) to this file") + ("extract_rules", po::value(), "Extract the rules used in translation (not de-duped!) to a file in this directory") ("show_derivations", po::value(), "Directory to print the derivation structures to") ("graphviz","Show (constrained) translation forest in GraphViz format") ("max_translation_beam,x", po::value(), "Beam approximation to get max translation from the chart") @@ -570,6 +571,11 @@ DecoderImpl::DecoderImpl(po::variables_map& conf, int argc, char** argv, istream // cube pruning pop-limit: we may want to configure this on a per-pass basis pop_limit = conf["cubepruning_pop_limit"].as(); + if (conf.count("extract_rules")) { + if (!DirectoryExists(conf["extract_rules"].as())) + MkDirP(conf["extract_rules"].as()); + } + // determine the number of rescoring/pruning/weighting passes configured const int MAX_PASSES = 3; for (int pass = 0; pass < MAX_PASSES; ++pass) { @@ -712,9 +718,11 @@ DecoderImpl::DecoderImpl(po::variables_map& conf, int argc, char** argv, istream cfg_options.Validate(); #endif - if (conf.count("extract_rules")) - extract_file.reset(new WriteFile(str("extract_rules",conf))); - + if (conf.count("extract_rules")) { + stringstream ss; + ss << sent_id; + extract_file.reset(new WriteFile(str("extract_rules",conf)+"/"+ss.str())); + } combine_size = conf["combine_size"].as(); if (combine_size < 1) combine_size = 1; sent_id = -1; @@ -851,7 +859,7 @@ bool DecoderImpl::Decode(const string& input, DecoderObserver* o) { #endif forest.swap(rescored_forest); forest.Reweight(cur_weights); - if (!SILENT) forest_stats(forest," " + passtr +" forest",show_tree_structure,oracle.show_derivation); + if (!SILENT) forest_stats(forest," " + passtr +" forest",show_tree_structure,oracle.show_derivation, conf.count("extract_rules"), extract_file); } if (conf.count("show_partition")) { diff --git a/decoder/viterbi.cc b/decoder/viterbi.cc index 9d19914b..1b9c6665 100644 --- a/decoder/viterbi.cc +++ b/decoder/viterbi.cc @@ -5,11 +5,12 @@ #include #include "hg.h" + //#define DEBUG_VITERBI_SORT using namespace std; -std::string viterbi_stats(Hypergraph const& hg, std::string const& name, bool estring, bool etree,bool show_derivation) +std::string viterbi_stats(Hypergraph const& hg, std::string const& name, bool estring, bool etree,bool show_derivation, bool extract_rules, boost::shared_ptr extract_file) { ostringstream o; o << hg.stats(name); @@ -22,6 +23,9 @@ std::string viterbi_stats(Hypergraph const& hg, std::string const& name, bool es if (etree) { o<stream()); + } if (show_derivation) { o< edges; + Viterbi(hg, &edges); + for (unsigned i = 0; i < edges.size(); i++) + (*o) << edges[i]->rule_->AsString(true) << endl; +} string ViterbiETree(const Hypergraph& hg) { vector tmp; diff --git a/decoder/viterbi.h b/decoder/viterbi.h index 3092f6da..03e961a2 100644 --- a/decoder/viterbi.h +++ b/decoder/viterbi.h @@ -5,8 +5,10 @@ #include "prob.h" #include "hg.h" #include "tdict.h" +#include "filelib.h" +#include -std::string viterbi_stats(Hypergraph const& hg, std::string const& name="forest", bool estring=true, bool etree=false, bool derivation_tree=false); +std::string viterbi_stats(Hypergraph const& hg, std::string const& name="forest", bool estring=true, bool etree=false, bool derivation_tree=false, bool extract_rules=false, boost::shared_ptr extract_file = boost::make_shared()); /// computes for each hg node the best (according to WeightType/WeightFunction) derivation, and some homomorphism (bottom up expression tree applied through Traversal) of it. T is the "return type" of Traversal, which is called only once for the best edge for a node's result (i.e. result will start default constructed) //TODO: make T a typename inside Traversal and WeightType a typename inside WeightFunction? @@ -201,6 +203,7 @@ struct FeatureVectorTraversal { std::string JoshuaVisualizationString(const Hypergraph& hg); prob_t ViterbiESentence(const Hypergraph& hg, std::vector* result); std::string ViterbiETree(const Hypergraph& hg); +void ViterbiRules(const Hypergraph& hg, std::ostream* s); prob_t ViterbiFSentence(const Hypergraph& hg, std::vector* result); std::string ViterbiFTree(const Hypergraph& hg); int ViterbiELength(const Hypergraph& hg); diff --git a/dtrain/README.md b/dtrain/README.md index 843874fa..7edabbf1 100644 --- a/dtrain/README.md +++ b/dtrain/README.md @@ -39,28 +39,6 @@ For an example of local usage (with the 'distributed' format) the see test/example/ . This expects dtrain to be built without DTRAIN_LOCAL. -Next ----- -+ approx. Bleu? proper lc_bleu (init with X) -+ turn off inclusion -+ (dtrain|decoder) more meta-parameters testing -+ feature selection directly in dtrain -+ feature template: target side rule ngrams -+ sa-extract -> leave-one-out for grammar of training set? -+ make svm doable; no subgradient? -+ reranking while sgd? -+ try PRO, mira emulations -+ sampling (MBR) -+ forest (on train)? -+ best BLEU transl (Sokolov)? -+ entire reg. path -+ resharding [nfold cross val.] -+ bigger LM, feats (target side Ng., word alignments etc.) -+ merge kbest lists -+ proper eval, pairwise ranking, forced transl -+ smmoth bleu variants X -+ MMERT exp - Legal ----- Copyright (c) 2012 by Patrick Simianer -- cgit v1.2.3