diff options
Diffstat (limited to 'gi/pf')
50 files changed, 0 insertions, 9460 deletions
diff --git a/gi/pf/Makefile.am b/gi/pf/Makefile.am deleted file mode 100644 index 86f8e07b..00000000 --- a/gi/pf/Makefile.am +++ /dev/null @@ -1,44 +0,0 @@ -bin_PROGRAMS = cbgi brat dpnaive pfbrat pfdist itg pfnaive condnaive align-lexonly-pyp learn_cfg pyp_lm nuisance_test align-tl pf_test bayes_lattice_score - -noinst_LIBRARIES = libpf.a - -libpf_a_SOURCES = base_distributions.cc reachability.cc cfg_wfst_composer.cc corpus.cc unigrams.cc ngram_base.cc transliterations.cc backward.cc hpyp_tm.cc pyp_tm.cc - -bayes_lattice_score_SOURCES = bayes_lattice_score.cc -bayes_lattice_score_LDADD = libpf.a $(top_srcdir)/decoder/libcdec.a $(top_srcdir)/mteval/libmteval.a $(top_srcdir)/utils/libutils.a $(top_srcdir)/klm/lm/libklm.a $(top_srcdir)/klm/util/libklm_util.a -lz - -pf_test_SOURCES = pf_test.cc -pf_test_LDADD = libpf.a $(top_srcdir)/decoder/libcdec.a $(top_srcdir)/mteval/libmteval.a $(top_srcdir)/utils/libutils.a $(top_srcdir)/klm/lm/libklm.a $(top_srcdir)/klm/util/libklm_util.a -lz - -nuisance_test_SOURCES = nuisance_test.cc -nuisance_test_LDADD = libpf.a $(top_srcdir)/decoder/libcdec.a $(top_srcdir)/mteval/libmteval.a $(top_srcdir)/utils/libutils.a $(top_srcdir)/klm/lm/libklm.a $(top_srcdir)/klm/util/libklm_util.a -lz - -align_lexonly_pyp_SOURCES = align-lexonly-pyp.cc -align_lexonly_pyp_LDADD = libpf.a $(top_srcdir)/decoder/libcdec.a $(top_srcdir)/mteval/libmteval.a $(top_srcdir)/utils/libutils.a $(top_srcdir)/klm/lm/libklm.a $(top_srcdir)/klm/util/libklm_util.a -lz - -align_tl_SOURCES = align-tl.cc -align_tl_LDADD = libpf.a $(top_srcdir)/decoder/libcdec.a $(top_srcdir)/mteval/libmteval.a $(top_srcdir)/utils/libutils.a $(top_srcdir)/klm/lm/libklm.a $(top_srcdir)/klm/util/libklm_util.a -lz - -itg_SOURCES = itg.cc - -pyp_lm_SOURCES = pyp_lm.cc - -learn_cfg_SOURCES = learn_cfg.cc - -condnaive_SOURCES = condnaive.cc - -dpnaive_SOURCES = dpnaive.cc - -pfdist_SOURCES = pfdist.cc - -pfnaive_SOURCES = pfnaive.cc - -cbgi_SOURCES = cbgi.cc - -brat_SOURCES = brat.cc - -pfbrat_SOURCES = pfbrat.cc - -AM_CPPFLAGS = -W -Wall -Wno-sign-compare -funroll-loops -I$(top_srcdir)/utils $(GTEST_CPPFLAGS) -I$(top_srcdir)/decoder -I$(top_srcdir)/klm - -AM_LDFLAGS = libpf.a $(top_srcdir)/decoder/libcdec.a $(top_srcdir)/klm/lm/libklm.a $(top_srcdir)/klm/util/libklm_util.a $(top_srcdir)/utils/libutils.a -lz diff --git a/gi/pf/README b/gi/pf/README deleted file mode 100644 index 62e47541..00000000 --- a/gi/pf/README +++ /dev/null @@ -1,2 +0,0 @@ -Experimental Bayesian alignment tools. Nothing to see here. - diff --git a/gi/pf/align-lexonly-pyp.cc b/gi/pf/align-lexonly-pyp.cc deleted file mode 100644 index e7509f57..00000000 --- a/gi/pf/align-lexonly-pyp.cc +++ /dev/null @@ -1,243 +0,0 @@ -#include <iostream> -#include <queue> - -#include <boost/program_options.hpp> -#include <boost/program_options/variables_map.hpp> - -#include "tdict.h" -#include "stringlib.h" -#include "filelib.h" -#include "array2d.h" -#include "sampler.h" -#include "corpus.h" -#include "pyp_tm.h" -#include "hpyp_tm.h" -#include "quasi_model2.h" - -using namespace std; -namespace po = boost::program_options; - -void InitCommandLine(int argc, char** argv, po::variables_map* conf) { - po::options_description opts("Configuration options"); - opts.add_options() - ("samples,s",po::value<unsigned>()->default_value(1000),"Number of samples") - ("infer_alignment_hyperparameters,I", "Infer alpha and p_null, otherwise fixed values will be assumed") - ("p_null,0", po::value<double>()->default_value(0.08), "probability of aligning to null") - ("align_alpha,a", po::value<double>()->default_value(4.0), "how 'tight' is the bias toward be along the diagonal?") - ("input,i",po::value<string>(),"Read parallel data from") - ("random_seed,S",po::value<uint32_t>(), "Random seed"); - po::options_description clo("Command line options"); - clo.add_options() - ("config", po::value<string>(), "Configuration file") - ("help,h", "Print this help message and exit"); - po::options_description dconfig_options, dcmdline_options; - dconfig_options.add(opts); - dcmdline_options.add(opts).add(clo); - - po::store(parse_command_line(argc, argv, dcmdline_options), *conf); - if (conf->count("config")) { - ifstream config((*conf)["config"].as<string>().c_str()); - po::store(po::parse_config_file(config, dconfig_options), *conf); - } - po::notify(*conf); - - if (conf->count("help") || (conf->count("input") == 0)) { - cerr << dcmdline_options << endl; - exit(1); - } -} - -MT19937* prng; - -struct LexicalAlignment { - unsigned char src_index; - bool is_transliteration; - vector<pair<short, short> > derivation; -}; - -struct AlignedSentencePair { - vector<WordID> src; - vector<WordID> trg; - vector<LexicalAlignment> a; - Array2D<short> posterior; -}; - -template <class LexicalTranslationModel> -struct Aligner { - Aligner(const vector<vector<WordID> >& lets, - int vocab_size, - int num_letters, - const po::variables_map& conf, - vector<AlignedSentencePair>* c) : - corpus(*c), - paj_model(conf["align_alpha"].as<double>(), conf["p_null"].as<double>()), - infer_paj(conf.count("infer_alignment_hyperparameters") > 0), - model(lets, vocab_size, num_letters), - kNULL(TD::Convert("NULL")) { - assert(lets[kNULL].size() == 0); - } - - vector<AlignedSentencePair>& corpus; - QuasiModel2 paj_model; - const bool infer_paj; - LexicalTranslationModel model; - const WordID kNULL; - - void ResampleHyperparameters() { - model.ResampleHyperparameters(prng); - if (infer_paj) paj_model.ResampleHyperparameters(prng); - } - - void InitializeRandom() { - cerr << "Initializing with random alignments ...\n"; - for (unsigned i = 0; i < corpus.size(); ++i) { - AlignedSentencePair& asp = corpus[i]; - asp.a.resize(asp.trg.size()); - for (unsigned j = 0; j < asp.trg.size(); ++j) { - unsigned char& a_j = asp.a[j].src_index; - a_j = prng->next() * (1 + asp.src.size()); - const WordID f_a_j = (a_j ? asp.src[a_j - 1] : kNULL); - model.Increment(f_a_j, asp.trg[j], &*prng); - paj_model.Increment(a_j, j, asp.src.size(), asp.trg.size()); - } - } - cerr << "Corpus intialized randomly." << endl; - cerr << "LLH = " << Likelihood() << " \t(Amodel=" << paj_model.Likelihood() - << " TModel=" << model.Likelihood() << ") contexts=" << model.UniqueConditioningContexts() << endl; - } - - void ResampleCorpus() { - for (unsigned i = 0; i < corpus.size(); ++i) { - AlignedSentencePair& asp = corpus[i]; - SampleSet<prob_t> ss; ss.resize(asp.src.size() + 1); - for (unsigned j = 0; j < asp.trg.size(); ++j) { - unsigned char& a_j = asp.a[j].src_index; - const WordID e_j = asp.trg[j]; - WordID f_a_j = (a_j ? asp.src[a_j - 1] : kNULL); - model.Decrement(f_a_j, e_j, prng); - paj_model.Decrement(a_j, j, asp.src.size(), asp.trg.size()); - - for (unsigned prop_a_j = 0; prop_a_j <= asp.src.size(); ++prop_a_j) { - const WordID prop_f = (prop_a_j ? asp.src[prop_a_j - 1] : kNULL); - ss[prop_a_j] = model.Prob(prop_f, e_j); - ss[prop_a_j] *= paj_model.Prob(prop_a_j, j, asp.src.size(), asp.trg.size()); - } - a_j = prng->SelectSample(ss); - f_a_j = (a_j ? asp.src[a_j - 1] : kNULL); - model.Increment(f_a_j, e_j, prng); - paj_model.Increment(a_j, j, asp.src.size(), asp.trg.size()); - } - } - } - - prob_t Likelihood() const { - return model.Likelihood() * paj_model.Likelihood(); - } -}; - -void ExtractLetters(const set<WordID>& v, vector<vector<WordID> >* l, set<WordID>* letset = NULL) { - for (set<WordID>::const_iterator it = v.begin(); it != v.end(); ++it) { - vector<WordID>& letters = (*l)[*it]; - if (letters.size()) continue; // if e and f have the same word - - const string& w = TD::Convert(*it); - - size_t cur = 0; - while (cur < w.size()) { - const size_t len = UTF8Len(w[cur]); - letters.push_back(TD::Convert(w.substr(cur, len))); - if (letset) letset->insert(letters.back()); - cur += len; - } - } -} - -void Debug(const AlignedSentencePair& asp) { - cerr << TD::GetString(asp.src) << endl << TD::GetString(asp.trg) << endl; - Array2D<bool> a(asp.src.size(), asp.trg.size()); - for (unsigned j = 0; j < asp.trg.size(); ++j) { - assert(asp.a[j].src_index <= asp.src.size()); - if (asp.a[j].src_index) a(asp.a[j].src_index - 1, j) = true; - } - cerr << a << endl; -} - -void AddSample(AlignedSentencePair* asp) { - for (unsigned j = 0; j < asp->trg.size(); ++j) - asp->posterior(asp->a[j].src_index, j)++; -} - -void WriteAlignments(const AlignedSentencePair& asp) { - bool first = true; - for (unsigned j = 0; j < asp.trg.size(); ++j) { - int src_index = -1; - int mc = -1; - for (unsigned i = 0; i <= asp.src.size(); ++i) { - if (asp.posterior(i, j) > mc) { - mc = asp.posterior(i, j); - src_index = i; - } - } - - if (src_index) { - if (first) first = false; else cout << ' '; - cout << (src_index - 1) << '-' << j; - } - } - cout << endl; -} - -int main(int argc, char** argv) { - po::variables_map conf; - InitCommandLine(argc, argv, &conf); - - if (conf.count("random_seed")) - prng = new MT19937(conf["random_seed"].as<uint32_t>()); - else - prng = new MT19937; - - vector<vector<int> > corpuse, corpusf; - set<int> vocabe, vocabf; - corpus::ReadParallelCorpus(conf["input"].as<string>(), &corpusf, &corpuse, &vocabf, &vocabe); - cerr << "f-Corpus size: " << corpusf.size() << " sentences\n"; - cerr << "f-Vocabulary size: " << vocabf.size() << " types\n"; - cerr << "f-Corpus size: " << corpuse.size() << " sentences\n"; - cerr << "f-Vocabulary size: " << vocabe.size() << " types\n"; - assert(corpusf.size() == corpuse.size()); - - vector<AlignedSentencePair> corpus(corpuse.size()); - for (unsigned i = 0; i < corpuse.size(); ++i) { - corpus[i].src.swap(corpusf[i]); - corpus[i].trg.swap(corpuse[i]); - corpus[i].posterior.resize(corpus[i].src.size() + 1, corpus[i].trg.size()); - } - corpusf.clear(); corpuse.clear(); - - vocabf.insert(TD::Convert("NULL")); - vector<vector<WordID> > letters(TD::NumWords()); - set<WordID> letset; - ExtractLetters(vocabe, &letters, &letset); - ExtractLetters(vocabf, &letters, NULL); - letters[TD::Convert("NULL")].clear(); - - //Aligner<PYPLexicalTranslation> aligner(letters, vocabe.size(), letset.size(), conf, &corpus); - Aligner<HPYPLexicalTranslation> aligner(letters, vocabe.size(), letset.size(), conf, &corpus); - aligner.InitializeRandom(); - - const unsigned samples = conf["samples"].as<unsigned>(); - for (int i = 0; i < samples; ++i) { - for (int j = 65; j < 67; ++j) Debug(corpus[j]); - if (i % 10 == 9) { - aligner.ResampleHyperparameters(); - cerr << "LLH = " << aligner.Likelihood() << " \t(Amodel=" << aligner.paj_model.Likelihood() - << " TModel=" << aligner.model.Likelihood() << ") contexts=" << aligner.model.UniqueConditioningContexts() << endl; - } - aligner.ResampleCorpus(); - if (i > (samples / 5) && (i % 6 == 5)) for (int j = 0; j < corpus.size(); ++j) AddSample(&corpus[j]); - } - for (unsigned i = 0; i < corpus.size(); ++i) - WriteAlignments(corpus[i]); - aligner.model.Summary(); - - return 0; -} diff --git a/gi/pf/align-tl.cc b/gi/pf/align-tl.cc deleted file mode 100644 index f6608f1d..00000000 --- a/gi/pf/align-tl.cc +++ /dev/null @@ -1,339 +0,0 @@ -#include <iostream> -#include <tr1/memory> -#include <queue> - -#include <boost/multi_array.hpp> -#include <boost/program_options.hpp> -#include <boost/program_options/variables_map.hpp> - -#include "backward.h" -#include "array2d.h" -#include "base_distributions.h" -#include "monotonic_pseg.h" -#include "conditional_pseg.h" -#include "trule.h" -#include "tdict.h" -#include "stringlib.h" -#include "filelib.h" -#include "dict.h" -#include "sampler.h" -#include "mfcr.h" -#include "corpus.h" -#include "ngram_base.h" -#include "transliterations.h" - -using namespace std; -using namespace tr1; -namespace po = boost::program_options; - -void InitCommandLine(int argc, char** argv, po::variables_map* conf) { - po::options_description opts("Configuration options"); - opts.add_options() - ("samples,s",po::value<unsigned>()->default_value(1000),"Number of samples") - ("input,i",po::value<string>(),"Read parallel data from") - ("s2t", po::value<string>(), "character level source-to-target prior transliteration probabilities") - ("t2s", po::value<string>(), "character level target-to-source prior transliteration probabilities") - ("max_src_chunk", po::value<unsigned>()->default_value(4), "Maximum size of translitered chunk in source") - ("max_trg_chunk", po::value<unsigned>()->default_value(4), "Maximum size of translitered chunk in target") - ("expected_src_to_trg_ratio", po::value<double>()->default_value(1.0), "If a word is transliterated, what is the expected length ratio from source to target?") - ("random_seed,S",po::value<uint32_t>(), "Random seed"); - po::options_description clo("Command line options"); - clo.add_options() - ("config", po::value<string>(), "Configuration file") - ("help,h", "Print this help message and exit"); - po::options_description dconfig_options, dcmdline_options; - dconfig_options.add(opts); - dcmdline_options.add(opts).add(clo); - - po::store(parse_command_line(argc, argv, dcmdline_options), *conf); - if (conf->count("config")) { - ifstream config((*conf)["config"].as<string>().c_str()); - po::store(po::parse_config_file(config, dconfig_options), *conf); - } - po::notify(*conf); - - if (conf->count("help") || (conf->count("input") == 0)) { - cerr << dcmdline_options << endl; - exit(1); - } -} - -boost::shared_ptr<MT19937> prng; - -struct LexicalAlignment { - unsigned char src_index; - bool is_transliteration; - vector<pair<short, short> > derivation; -}; - -struct AlignedSentencePair { - vector<WordID> src; - vector<WordID> trg; - vector<LexicalAlignment> a; - Array2D<short> posterior; -}; - -struct HierarchicalWordBase { - explicit HierarchicalWordBase(const unsigned vocab_e_size) : - base(prob_t::One()), r(1,1,1,1,0.66,50.0), u0(-log(vocab_e_size)), l(1,prob_t::One()), v(1, prob_t::Zero()) {} - - void ResampleHyperparameters(MT19937* rng) { - r.resample_hyperparameters(rng); - } - - inline double logp0(const vector<WordID>& s) const { - return Md::log_poisson(s.size(), 7.5) + s.size() * u0; - } - - // return p0 of rule.e_ - prob_t operator()(const TRule& rule) const { - v[0].logeq(logp0(rule.e_)); - return r.prob(rule.e_, v.begin(), l.begin()); - } - - void Increment(const TRule& rule) { - v[0].logeq(logp0(rule.e_)); - if (r.increment(rule.e_, v.begin(), l.begin(), &*prng).count) { - base *= v[0] * l[0]; - } - } - - void Decrement(const TRule& rule) { - if (r.decrement(rule.e_, &*prng).count) { - base /= prob_t(exp(logp0(rule.e_))); - } - } - - prob_t Likelihood() const { - prob_t p; p.logeq(r.log_crp_prob()); - p *= base; - return p; - } - - void Summary() const { - cerr << "NUMBER OF CUSTOMERS: " << r.num_customers() << " (d=" << r.discount() << ",s=" << r.strength() << ')' << endl; - for (MFCR<1,vector<WordID> >::const_iterator it = r.begin(); it != r.end(); ++it) - cerr << " " << it->second.total_dish_count_ << " (on " << it->second.table_counts_.size() << " tables) " << TD::GetString(it->first) << endl; - } - - prob_t base; - MFCR<1,vector<WordID> > r; - const double u0; - const vector<prob_t> l; - mutable vector<prob_t> v; -}; - -struct BasicLexicalAlignment { - explicit BasicLexicalAlignment(const vector<vector<WordID> >& lets, - const unsigned words_e, - const unsigned letters_e, - vector<AlignedSentencePair>* corp) : - letters(lets), - corpus(*corp), - //up0(words_e), - //up0("en.chars.1gram", letters_e), - //up0("en.words.1gram"), - up0(letters_e), - //up0("en.chars.2gram"), - tmodel(up0) { - } - - void InstantiateRule(const WordID src, - const WordID trg, - TRule* rule) const { - static const WordID kX = TD::Convert("X") * -1; - rule->lhs_ = kX; - rule->e_ = letters[trg]; - rule->f_ = letters[src]; - } - - void InitializeRandom() { - const WordID kNULL = TD::Convert("NULL"); - cerr << "Initializing with random alignments ...\n"; - for (unsigned i = 0; i < corpus.size(); ++i) { - AlignedSentencePair& asp = corpus[i]; - asp.a.resize(asp.trg.size()); - for (unsigned j = 0; j < asp.trg.size(); ++j) { - const unsigned char a_j = prng->next() * (1 + asp.src.size()); - const WordID f_a_j = (a_j ? asp.src[a_j - 1] : kNULL); - TRule r; - InstantiateRule(f_a_j, asp.trg[j], &r); - asp.a[j].is_transliteration = false; - asp.a[j].src_index = a_j; - if (tmodel.IncrementRule(r, &*prng)) - up0.Increment(r); - } - } - cerr << " LLH = " << Likelihood() << endl; - } - - prob_t Likelihood() const { - prob_t p = tmodel.Likelihood(); - p *= up0.Likelihood(); - return p; - } - - void ResampleHyperparemeters() { - tmodel.ResampleHyperparameters(&*prng); - up0.ResampleHyperparameters(&*prng); - cerr << " (base d=" << up0.r.discount() << ",s=" << up0.r.strength() << ")\n"; - } - - void ResampleCorpus(); - - const vector<vector<WordID> >& letters; // spelling dictionary - vector<AlignedSentencePair>& corpus; - //PhraseConditionalUninformativeBase up0; - //PhraseConditionalUninformativeUnigramBase up0; - //UnigramWordBase up0; - //HierarchicalUnigramBase up0; - HierarchicalWordBase up0; - //CompletelyUniformBase up0; - //FixedNgramBase up0; - //ConditionalTranslationModel<PhraseConditionalUninformativeBase> tmodel; - //ConditionalTranslationModel<PhraseConditionalUninformativeUnigramBase> tmodel; - //ConditionalTranslationModel<UnigramWordBase> tmodel; - //ConditionalTranslationModel<HierarchicalUnigramBase> tmodel; - MConditionalTranslationModel<HierarchicalWordBase> tmodel; - //ConditionalTranslationModel<FixedNgramBase> tmodel; - //ConditionalTranslationModel<CompletelyUniformBase> tmodel; -}; - -void BasicLexicalAlignment::ResampleCorpus() { - static const WordID kNULL = TD::Convert("NULL"); - for (unsigned i = 0; i < corpus.size(); ++i) { - AlignedSentencePair& asp = corpus[i]; - SampleSet<prob_t> ss; ss.resize(asp.src.size() + 1); - for (unsigned j = 0; j < asp.trg.size(); ++j) { - TRule r; - unsigned char& a_j = asp.a[j].src_index; - WordID f_a_j = (a_j ? asp.src[a_j - 1] : kNULL); - InstantiateRule(f_a_j, asp.trg[j], &r); - if (tmodel.DecrementRule(r, &*prng)) - up0.Decrement(r); - - for (unsigned prop_a_j = 0; prop_a_j <= asp.src.size(); ++prop_a_j) { - const WordID prop_f = (prop_a_j ? asp.src[prop_a_j - 1] : kNULL); - InstantiateRule(prop_f, asp.trg[j], &r); - ss[prop_a_j] = tmodel.RuleProbability(r); - } - a_j = prng->SelectSample(ss); - f_a_j = (a_j ? asp.src[a_j - 1] : kNULL); - InstantiateRule(f_a_j, asp.trg[j], &r); - if (tmodel.IncrementRule(r, &*prng)) - up0.Increment(r); - } - } - cerr << " LLH = " << Likelihood() << endl; -} - -void ExtractLetters(const set<WordID>& v, vector<vector<WordID> >* l, set<WordID>* letset = NULL) { - for (set<WordID>::const_iterator it = v.begin(); it != v.end(); ++it) { - vector<WordID>& letters = (*l)[*it]; - if (letters.size()) continue; // if e and f have the same word - - const string& w = TD::Convert(*it); - - size_t cur = 0; - while (cur < w.size()) { - const size_t len = UTF8Len(w[cur]); - letters.push_back(TD::Convert(w.substr(cur, len))); - if (letset) letset->insert(letters.back()); - cur += len; - } - } -} - -void Debug(const AlignedSentencePair& asp) { - cerr << TD::GetString(asp.src) << endl << TD::GetString(asp.trg) << endl; - Array2D<bool> a(asp.src.size(), asp.trg.size()); - for (unsigned j = 0; j < asp.trg.size(); ++j) - if (asp.a[j].src_index) a(asp.a[j].src_index - 1, j) = true; - cerr << a << endl; -} - -void AddSample(AlignedSentencePair* asp) { - for (unsigned j = 0; j < asp->trg.size(); ++j) - asp->posterior(asp->a[j].src_index, j)++; -} - -void WriteAlignments(const AlignedSentencePair& asp) { - bool first = true; - for (unsigned j = 0; j < asp.trg.size(); ++j) { - int src_index = -1; - int mc = -1; - for (unsigned i = 0; i <= asp.src.size(); ++i) { - if (asp.posterior(i, j) > mc) { - mc = asp.posterior(i, j); - src_index = i; - } - } - - if (src_index) { - if (first) first = false; else cout << ' '; - cout << (src_index - 1) << '-' << j; - } - } - cout << endl; -} - -int main(int argc, char** argv) { - po::variables_map conf; - InitCommandLine(argc, argv, &conf); - - if (conf.count("random_seed")) - prng.reset(new MT19937(conf["random_seed"].as<uint32_t>())); - else - prng.reset(new MT19937); -// MT19937& rng = *prng; - - vector<vector<int> > corpuse, corpusf; - set<int> vocabe, vocabf; - corpus::ReadParallelCorpus(conf["input"].as<string>(), &corpusf, &corpuse, &vocabf, &vocabe); - cerr << "f-Corpus size: " << corpusf.size() << " sentences\n"; - cerr << "f-Vocabulary size: " << vocabf.size() << " types\n"; - cerr << "f-Corpus size: " << corpuse.size() << " sentences\n"; - cerr << "f-Vocabulary size: " << vocabe.size() << " types\n"; - assert(corpusf.size() == corpuse.size()); - - vector<AlignedSentencePair> corpus(corpuse.size()); - for (unsigned i = 0; i < corpuse.size(); ++i) { - corpus[i].src.swap(corpusf[i]); - corpus[i].trg.swap(corpuse[i]); - corpus[i].posterior.resize(corpus[i].src.size() + 1, corpus[i].trg.size()); - } - corpusf.clear(); corpuse.clear(); - - vocabf.insert(TD::Convert("NULL")); - vector<vector<WordID> > letters(TD::NumWords() + 1); - set<WordID> letset; - ExtractLetters(vocabe, &letters, &letset); - ExtractLetters(vocabf, &letters, NULL); - letters[TD::Convert("NULL")].clear(); - - // TODO configure this - const int max_src_chunk = conf["max_src_chunk"].as<unsigned>(); - const int max_trg_chunk = conf["max_trg_chunk"].as<unsigned>(); - const double s2t_rat = conf["expected_src_to_trg_ratio"].as<double>(); - const BackwardEstimator be(conf["s2t"].as<string>(), conf["t2s"].as<string>()); - Transliterations tl(max_src_chunk, max_trg_chunk, s2t_rat, be); - - cerr << "Initializing transliteration graph structures ...\n"; - for (int i = 0; i < corpus.size(); ++i) { - const vector<int>& src = corpus[i].src; - const vector<int>& trg = corpus[i].trg; - for (int j = 0; j < src.size(); ++j) { - const vector<int>& src_let = letters[src[j]]; - for (int k = 0; k < trg.size(); ++k) { - const vector<int>& trg_let = letters[trg[k]]; - tl.Initialize(src[j], src_let, trg[k], trg_let); - //if (src_let.size() < min_trans_src) - // tl.Forbid(src[j], src_let, trg[k], trg_let); - } - } - } - cerr << endl; - tl.GraphSummary(); - - return 0; -} diff --git a/gi/pf/backward.cc b/gi/pf/backward.cc deleted file mode 100644 index b92629fd..00000000 --- a/gi/pf/backward.cc +++ /dev/null @@ -1,89 +0,0 @@ -#include "backward.h" - -#include <queue> -#include <utility> - -#include "array2d.h" -#include "reachability.h" -#include "base_distributions.h" - -using namespace std; - -BackwardEstimator::BackwardEstimator(const string& s2t, - const string& t2s) : m1(new Model1(s2t)), m1inv(new Model1(t2s)) {} - -BackwardEstimator::~BackwardEstimator() { - delete m1; m1 = NULL; - delete m1inv; m1inv = NULL; -} - -float BackwardEstimator::ComputeBackwardProb(const std::vector<WordID>& src, - const std::vector<WordID>& trg, - unsigned src_covered, - unsigned trg_covered, - double s2t_ratio) const { - if (src_covered == src.size() || trg_covered == trg.size()) { - assert(src_covered == src.size()); - assert(trg_covered == trg.size()); - return 0; - } - static const WordID kNULL = TD::Convert("<eps>"); - const prob_t uniform_alignment(1.0 / (src.size() - src_covered + 1)); - // TODO factor in expected length ratio - prob_t e; e.logeq(Md::log_poisson(trg.size() - trg_covered, (src.size() - src_covered) * s2t_ratio)); // p(trg len remaining | src len remaining) - for (unsigned j = trg_covered; j < trg.size(); ++j) { - prob_t p = (*m1)(kNULL, trg[j]) + prob_t(1e-12); - for (unsigned i = src_covered; i < src.size(); ++i) - p += (*m1)(src[i], trg[j]); - if (p.is_0()) { - cerr << "ERROR: p(" << TD::Convert(trg[j]) << " | " << TD::GetString(src) << ") = 0!\n"; - assert(!"failed"); - } - p *= uniform_alignment; - e *= p; - } - // TODO factor in expected length ratio - const prob_t inv_uniform(1.0 / (trg.size() - trg_covered + 1.0)); - prob_t inv; - inv.logeq(Md::log_poisson(src.size() - src_covered, (trg.size() - trg_covered) / s2t_ratio)); - for (unsigned i = src_covered; i < src.size(); ++i) { - prob_t p = (*m1inv)(kNULL, src[i]) + prob_t(1e-12); - for (unsigned j = trg_covered; j < trg.size(); ++j) - p += (*m1inv)(trg[j], src[i]); - if (p.is_0()) { - cerr << "ERROR: p_inv(" << TD::Convert(src[i]) << " | " << TD::GetString(trg) << ") = 0!\n"; - assert(!"failed"); - } - p *= inv_uniform; - inv *= p; - } - return (log(e) + log(inv)) / 2; -} - -void BackwardEstimator::InitializeGrid(const vector<WordID>& src, - const vector<WordID>& trg, - const Reachability& r, - double s2t_ratio, - float* grid) const { - queue<pair<int,int> > q; - q.push(make_pair(0,0)); - Array2D<bool> done(src.size()+1, trg.size()+1, false); - //cerr << TD::GetString(src) << " ||| " << TD::GetString(trg) << endl; - while(!q.empty()) { - const pair<int,int> n = q.front(); - q.pop(); - if (done(n.first,n.second)) continue; - done(n.first,n.second) = true; - - float lp = ComputeBackwardProb(src, trg, n.first, n.second, s2t_ratio); - if (n.first == 0 && n.second == 0) grid[0] = lp; - //cerr << " " << n.first << "," << n.second << "\t" << lp << endl; - - if (n.first == src.size() || n.second == trg.size()) continue; - const vector<pair<short,short> >& edges = r.valid_deltas[n.first][n.second]; - for (int i = 0; i < edges.size(); ++i) - q.push(make_pair(n.first + edges[i].first, n.second + edges[i].second)); - } - //static int cc = 0; ++cc; if (cc == 80) exit(1); -} - diff --git a/gi/pf/backward.h b/gi/pf/backward.h deleted file mode 100644 index e67eff0c..00000000 --- a/gi/pf/backward.h +++ /dev/null @@ -1,33 +0,0 @@ -#ifndef _BACKWARD_H_ -#define _BACKWARD_H_ - -#include <vector> -#include <string> -#include "wordid.h" - -struct Reachability; -struct Model1; - -struct BackwardEstimator { - BackwardEstimator(const std::string& s2t, - const std::string& t2s); - ~BackwardEstimator(); - - void InitializeGrid(const std::vector<WordID>& src, - const std::vector<WordID>& trg, - const Reachability& r, - double src2trg_ratio, - float* grid) const; - - private: - float ComputeBackwardProb(const std::vector<WordID>& src, - const std::vector<WordID>& trg, - unsigned src_covered, - unsigned trg_covered, - double src2trg_ratio) const; - - Model1* m1; - Model1* m1inv; -}; - -#endif diff --git a/gi/pf/base_distributions.cc b/gi/pf/base_distributions.cc deleted file mode 100644 index 57e0bbe1..00000000 --- a/gi/pf/base_distributions.cc +++ /dev/null @@ -1,241 +0,0 @@ -#include "base_distributions.h" - -#include <iostream> - -#include "filelib.h" - -using namespace std; - -TableLookupBase::TableLookupBase(const string& fname) { - cerr << "TableLookupBase reading from " << fname << " ..." << endl; - ReadFile rf(fname); - istream& in = *rf.stream(); - string line; - unsigned lc = 0; - const WordID kDIV = TD::Convert("|||"); - vector<WordID> tmp; - vector<int> le, lf; - TRule x; - x.lhs_ = -TD::Convert("X"); - bool flag = false; - while(getline(in, line)) { - ++lc; - if (lc % 1000000 == 0) { cerr << " [" << lc << ']' << endl; flag = false; } - else if (lc % 25000 == 0) { cerr << '.' << flush; flag = true; } - tmp.clear(); - TD::ConvertSentence(line, &tmp); - x.f_.clear(); - x.e_.clear(); - size_t pos = 0; - int cc = 0; - while(pos < tmp.size()) { - const WordID cur = tmp[pos++]; - if (cur == kDIV) { - ++cc; - } else if (cc == 0) { - x.f_.push_back(cur); - } else if (cc == 1) { - x.e_.push_back(cur); - } else if (cc == 2) { - table[x].logeq(atof(TD::Convert(cur).c_str())); - ++cc; - } else { - if (flag) cerr << endl; - cerr << "Bad format in " << lc << ": " << line << endl; abort(); - } - } - if (cc != 3) { - if (flag) cerr << endl; - cerr << "Bad format in " << lc << ": " << line << endl; abort(); - } - } - if (flag) cerr << endl; - cerr << " read " << lc << " entries\n"; -} - -prob_t PhraseConditionalUninformativeUnigramBase::p0(const vector<WordID>& vsrc, - const vector<WordID>& vtrg, - int start_src, int start_trg) const { - const int flen = vsrc.size() - start_src; - const int elen = vtrg.size() - start_trg; - prob_t p; - p.logeq(Md::log_poisson(elen, flen + 0.01)); // elen | flen ~Pois(flen + 0.01) - //p.logeq(log_poisson(elen, 1)); // elen | flen ~Pois(flen + 0.01) - for (int i = 0; i < elen; ++i) - p *= u(vtrg[i + start_trg]); // draw e_i ~Uniform - return p; -} - -prob_t PhraseConditionalUninformativeBase::p0(const vector<WordID>& vsrc, - const vector<WordID>& vtrg, - int start_src, int start_trg) const { - const int flen = vsrc.size() - start_src; - const int elen = vtrg.size() - start_trg; - prob_t p; - //p.logeq(log_poisson(elen, flen + 0.01)); // elen | flen ~Pois(flen + 0.01) - p.logeq(Md::log_poisson(elen, 1)); // elen | flen ~Pois(flen + 0.01) - for (int i = 0; i < elen; ++i) - p *= kUNIFORM_TARGET; // draw e_i ~Uniform - return p; -} - -void Model1::LoadModel1(const string& fname) { - cerr << "Loading Model 1 parameters from " << fname << " ..." << endl; - ReadFile rf(fname); - istream& in = *rf.stream(); - string line; - unsigned lc = 0; - while(getline(in, line)) { - ++lc; - int cur = 0; - int start = 0; - while(cur < line.size() && line[cur] != ' ') { ++cur; } - assert(cur != line.size()); - line[cur] = 0; - const WordID src = TD::Convert(&line[0]); - ++cur; - start = cur; - while(cur < line.size() && line[cur] != ' ') { ++cur; } - assert(cur != line.size()); - line[cur] = 0; - WordID trg = TD::Convert(&line[start]); - const double logprob = strtod(&line[cur + 1], NULL); - if (src >= ttable.size()) ttable.resize(src + 1); - ttable[src][trg].logeq(logprob); - } - cerr << " read " << lc << " parameters.\n"; -} - -prob_t PhraseConditionalBase::p0(const vector<WordID>& vsrc, - const vector<WordID>& vtrg, - int start_src, int start_trg) const { - const int flen = vsrc.size() - start_src; - const int elen = vtrg.size() - start_trg; - prob_t uniform_src_alignment; uniform_src_alignment.logeq(-log(flen + 1)); - prob_t p; - p.logeq(Md::log_poisson(elen, flen + 0.01)); // elen | flen ~Pois(flen + 0.01) - for (int i = 0; i < elen; ++i) { // for each position i in e-RHS - const WordID trg = vtrg[i + start_trg]; - prob_t tp = prob_t::Zero(); - for (int j = -1; j < flen; ++j) { - const WordID src = j < 0 ? 0 : vsrc[j + start_src]; - tp += kM1MIXTURE * model1(src, trg); - tp += kUNIFORM_MIXTURE * kUNIFORM_TARGET; - } - tp *= uniform_src_alignment; // draw a_i ~uniform - p *= tp; // draw e_i ~Model1(f_a_i) / uniform - } - if (p.is_0()) { - cerr << "Zero! " << vsrc << "\nTRG=" << vtrg << endl; - abort(); - } - return p; -} - -prob_t PhraseJointBase::p0(const vector<WordID>& vsrc, - const vector<WordID>& vtrg, - int start_src, int start_trg) const { - const int flen = vsrc.size() - start_src; - const int elen = vtrg.size() - start_trg; - prob_t uniform_src_alignment; uniform_src_alignment.logeq(-log(flen + 1)); - prob_t p; - p.logeq(Md::log_poisson(flen, 1.0)); // flen ~Pois(1) - // elen | flen ~Pois(flen + 0.01) - prob_t ptrglen; ptrglen.logeq(Md::log_poisson(elen, flen + 0.01)); - p *= ptrglen; - p *= kUNIFORM_SOURCE.pow(flen); // each f in F ~Uniform - for (int i = 0; i < elen; ++i) { // for each position i in E - const WordID trg = vtrg[i + start_trg]; - prob_t tp = prob_t::Zero(); - for (int j = -1; j < flen; ++j) { - const WordID src = j < 0 ? 0 : vsrc[j + start_src]; - tp += kM1MIXTURE * model1(src, trg); - tp += kUNIFORM_MIXTURE * kUNIFORM_TARGET; - } - tp *= uniform_src_alignment; // draw a_i ~uniform - p *= tp; // draw e_i ~Model1(f_a_i) / uniform - } - if (p.is_0()) { - cerr << "Zero! " << vsrc << "\nTRG=" << vtrg << endl; - abort(); - } - return p; -} - -prob_t PhraseJointBase_BiDir::p0(const vector<WordID>& vsrc, - const vector<WordID>& vtrg, - int start_src, int start_trg) const { - const int flen = vsrc.size() - start_src; - const int elen = vtrg.size() - start_trg; - prob_t uniform_src_alignment; uniform_src_alignment.logeq(-log(flen + 1)); - prob_t uniform_trg_alignment; uniform_trg_alignment.logeq(-log(elen + 1)); - - prob_t p1; - p1.logeq(Md::log_poisson(flen, 1.0)); // flen ~Pois(1) - // elen | flen ~Pois(flen + 0.01) - prob_t ptrglen; ptrglen.logeq(Md::log_poisson(elen, flen + 0.01)); - p1 *= ptrglen; - p1 *= kUNIFORM_SOURCE.pow(flen); // each f in F ~Uniform - for (int i = 0; i < elen; ++i) { // for each position i in E - const WordID trg = vtrg[i + start_trg]; - prob_t tp = prob_t::Zero(); - for (int j = -1; j < flen; ++j) { - const WordID src = j < 0 ? 0 : vsrc[j + start_src]; - tp += kM1MIXTURE * model1(src, trg); - tp += kUNIFORM_MIXTURE * kUNIFORM_TARGET; - } - tp *= uniform_src_alignment; // draw a_i ~uniform - p1 *= tp; // draw e_i ~Model1(f_a_i) / uniform - } - if (p1.is_0()) { - cerr << "Zero! " << vsrc << "\nTRG=" << vtrg << endl; - abort(); - } - - prob_t p2; - p2.logeq(Md::log_poisson(elen, 1.0)); // elen ~Pois(1) - // flen | elen ~Pois(flen + 0.01) - prob_t psrclen; psrclen.logeq(Md::log_poisson(flen, elen + 0.01)); - p2 *= psrclen; - p2 *= kUNIFORM_TARGET.pow(elen); // each f in F ~Uniform - for (int i = 0; i < flen; ++i) { // for each position i in E - const WordID src = vsrc[i + start_src]; - prob_t tp = prob_t::Zero(); - for (int j = -1; j < elen; ++j) { - const WordID trg = j < 0 ? 0 : vtrg[j + start_trg]; - tp += kM1MIXTURE * invmodel1(trg, src); - tp += kUNIFORM_MIXTURE * kUNIFORM_SOURCE; - } - tp *= uniform_trg_alignment; // draw a_i ~uniform - p2 *= tp; // draw e_i ~Model1(f_a_i) / uniform - } - if (p2.is_0()) { - cerr << "Zero! " << vsrc << "\nTRG=" << vtrg << endl; - abort(); - } - - static const prob_t kHALF(0.5); - return (p1 + p2) * kHALF; -} - -JumpBase::JumpBase() : p(200) { - for (unsigned src_len = 1; src_len < 200; ++src_len) { - map<int, prob_t>& cpd = p[src_len]; - int min_jump = 1 - src_len; - int max_jump = src_len; - prob_t z; - for (int j = min_jump; j <= max_jump; ++j) { - prob_t& cp = cpd[j]; - if (j < 0) - cp.logeq(Md::log_poisson(1.5-j, 1)); - else if (j > 0) - cp.logeq(Md::log_poisson(j, 1)); - cp.poweq(0.2); - z += cp; - } - for (int j = min_jump; j <= max_jump; ++j) { - cpd[j] /= z; - } - } -} - diff --git a/gi/pf/base_distributions.h b/gi/pf/base_distributions.h deleted file mode 100644 index 41b513f8..00000000 --- a/gi/pf/base_distributions.h +++ /dev/null @@ -1,238 +0,0 @@ -#ifndef _BASE_MEASURES_H_ -#define _BASE_MEASURES_H_ - -#include <vector> -#include <map> -#include <string> -#include <cmath> -#include <iostream> -#include <cassert> - -#include "unigrams.h" -#include "trule.h" -#include "prob.h" -#include "tdict.h" -#include "sampler.h" -#include "m.h" -#include "os_phrase.h" - -struct Model1 { - explicit Model1(const std::string& fname) : - kNULL(TD::Convert("<eps>")), - kZERO() { - LoadModel1(fname); - } - - void LoadModel1(const std::string& fname); - - // returns prob 0 if src or trg is not found - const prob_t& operator()(WordID src, WordID trg) const { - if (src == 0) src = kNULL; - if (src < ttable.size()) { - const std::map<WordID, prob_t>& cpd = ttable[src]; - const std::map<WordID, prob_t>::const_iterator it = cpd.find(trg); - if (it != cpd.end()) - return it->second; - } - return kZERO; - } - - const WordID kNULL; - const prob_t kZERO; - std::vector<std::map<WordID, prob_t> > ttable; -}; - -struct PoissonUniformUninformativeBase { - explicit PoissonUniformUninformativeBase(const unsigned ves) : kUNIFORM(1.0 / ves) {} - prob_t operator()(const TRule& r) const { - prob_t p; p.logeq(Md::log_poisson(r.e_.size(), 1.0)); - prob_t q = kUNIFORM; q.poweq(r.e_.size()); - p *= q; - return p; - } - void Summary() const {} - void ResampleHyperparameters(MT19937*) {} - void Increment(const TRule&) {} - void Decrement(const TRule&) {} - prob_t Likelihood() const { return prob_t::One(); } - const prob_t kUNIFORM; -}; - -struct CompletelyUniformBase { - explicit CompletelyUniformBase(const unsigned ves) : kUNIFORM(1.0 / ves) {} - prob_t operator()(const TRule&) const { - return kUNIFORM; - } - void Summary() const {} - void ResampleHyperparameters(MT19937*) {} - void Increment(const TRule&) {} - void Decrement(const TRule&) {} - prob_t Likelihood() const { return prob_t::One(); } - const prob_t kUNIFORM; -}; - -struct UnigramWordBase { - explicit UnigramWordBase(const std::string& fname) : un(fname) {} - prob_t operator()(const TRule& r) const { - return un(r.e_); - } - const UnigramWordModel un; -}; - -struct RuleHasher { - size_t operator()(const TRule& r) const { - return hash_value(r); - } -}; - -struct TableLookupBase { - TableLookupBase(const std::string& fname); - - prob_t operator()(const TRule& rule) const { - const std::tr1::unordered_map<TRule,prob_t,RuleHasher>::const_iterator it = table.find(rule); - if (it == table.end()) { - std::cerr << rule << " not found\n"; - abort(); - } - return it->second; - } - - void ResampleHyperparameters(MT19937*) {} - void Increment(const TRule&) {} - void Decrement(const TRule&) {} - prob_t Likelihood() const { return prob_t::One(); } - void Summary() const {} - - std::tr1::unordered_map<TRule,prob_t,RuleHasher> table; -}; - -struct PhraseConditionalUninformativeBase { - explicit PhraseConditionalUninformativeBase(const unsigned vocab_e_size) : - kUNIFORM_TARGET(1.0 / vocab_e_size) { - assert(vocab_e_size > 0); - } - - // return p0 of rule.e_ | rule.f_ - prob_t operator()(const TRule& rule) const { - return p0(rule.f_, rule.e_, 0, 0); - } - - prob_t p0(const std::vector<WordID>& vsrc, const std::vector<WordID>& vtrg, int start_src, int start_trg) const; - - void Summary() const {} - void ResampleHyperparameters(MT19937*) {} - void Increment(const TRule&) {} - void Decrement(const TRule&) {} - prob_t Likelihood() const { return prob_t::One(); } - const prob_t kUNIFORM_TARGET; -}; - -struct PhraseConditionalUninformativeUnigramBase { - explicit PhraseConditionalUninformativeUnigramBase(const std::string& file, const unsigned vocab_e_size) : u(file, vocab_e_size) {} - - // return p0 of rule.e_ | rule.f_ - prob_t operator()(const TRule& rule) const { - return p0(rule.f_, rule.e_, 0, 0); - } - - prob_t p0(const std::vector<WordID>& vsrc, const std::vector<WordID>& vtrg, int start_src, int start_trg) const; - - const UnigramModel u; -}; - -struct PhraseConditionalBase { - explicit PhraseConditionalBase(const Model1& m1, const double m1mixture, const unsigned vocab_e_size) : - model1(m1), - kM1MIXTURE(m1mixture), - kUNIFORM_MIXTURE(1.0 - m1mixture), - kUNIFORM_TARGET(1.0 / vocab_e_size) { - assert(m1mixture >= 0.0 && m1mixture <= 1.0); - assert(vocab_e_size > 0); - } - - // return p0 of rule.e_ | rule.f_ - prob_t operator()(const TRule& rule) const { - return p0(rule.f_, rule.e_, 0, 0); - } - - prob_t p0(const std::vector<WordID>& vsrc, const std::vector<WordID>& vtrg, int start_src, int start_trg) const; - - const Model1& model1; - const prob_t kM1MIXTURE; // Model 1 mixture component - const prob_t kUNIFORM_MIXTURE; // uniform mixture component - const prob_t kUNIFORM_TARGET; -}; - -struct PhraseJointBase { - explicit PhraseJointBase(const Model1& m1, const double m1mixture, const unsigned vocab_e_size, const unsigned vocab_f_size) : - model1(m1), - kM1MIXTURE(m1mixture), - kUNIFORM_MIXTURE(1.0 - m1mixture), - kUNIFORM_SOURCE(1.0 / vocab_f_size), - kUNIFORM_TARGET(1.0 / vocab_e_size) { - assert(m1mixture >= 0.0 && m1mixture <= 1.0); - assert(vocab_e_size > 0); - } - - // return p0 of rule.e_ , rule.f_ - prob_t operator()(const TRule& rule) const { - return p0(rule.f_, rule.e_, 0, 0); - } - - prob_t p0(const std::vector<WordID>& vsrc, const std::vector<WordID>& vtrg, int start_src, int start_trg) const; - - const Model1& model1; - const prob_t kM1MIXTURE; // Model 1 mixture component - const prob_t kUNIFORM_MIXTURE; // uniform mixture component - const prob_t kUNIFORM_SOURCE; - const prob_t kUNIFORM_TARGET; -}; - -struct PhraseJointBase_BiDir { - explicit PhraseJointBase_BiDir(const Model1& m1, - const Model1& im1, - const double m1mixture, - const unsigned vocab_e_size, - const unsigned vocab_f_size) : - model1(m1), - invmodel1(im1), - kM1MIXTURE(m1mixture), - kUNIFORM_MIXTURE(1.0 - m1mixture), - kUNIFORM_SOURCE(1.0 / vocab_f_size), - kUNIFORM_TARGET(1.0 / vocab_e_size) { - assert(m1mixture >= 0.0 && m1mixture <= 1.0); - assert(vocab_e_size > 0); - } - - // return p0 of rule.e_ , rule.f_ - prob_t operator()(const TRule& rule) const { - return p0(rule.f_, rule.e_, 0, 0); - } - - prob_t p0(const std::vector<WordID>& vsrc, const std::vector<WordID>& vtrg, int start_src, int start_trg) const; - - const Model1& model1; - const Model1& invmodel1; - const prob_t kM1MIXTURE; // Model 1 mixture component - const prob_t kUNIFORM_MIXTURE; // uniform mixture component - const prob_t kUNIFORM_SOURCE; - const prob_t kUNIFORM_TARGET; -}; - -// base distribution for jump size multinomials -// basically p(0) = 0 and then, p(1) is max, and then -// you drop as you move to the max jump distance -struct JumpBase { - JumpBase(); - - const prob_t& operator()(int jump, unsigned src_len) const { - assert(jump != 0); - const std::map<int, prob_t>::const_iterator it = p[src_len].find(jump); - assert(it != p[src_len].end()); - return it->second; - } - std::vector<std::map<int, prob_t> > p; -}; - - -#endif diff --git a/gi/pf/bayes_lattice_score.cc b/gi/pf/bayes_lattice_score.cc deleted file mode 100644 index 70cb8dc2..00000000 --- a/gi/pf/bayes_lattice_score.cc +++ /dev/null @@ -1,309 +0,0 @@ -#include <iostream> -#include <queue> - -#include <boost/functional.hpp> -#include <boost/program_options.hpp> -#include <boost/program_options/variables_map.hpp> - -#include "inside_outside.h" -#include "hg.h" -#include "hg_io.h" -#include "bottom_up_parser.h" -#include "fdict.h" -#include "grammar.h" -#include "m.h" -#include "trule.h" -#include "tdict.h" -#include "filelib.h" -#include "dict.h" -#include "sampler.h" -#include "ccrp.h" -#include "ccrp_onetable.h" - -using namespace std; -using namespace tr1; -namespace po = boost::program_options; - -boost::shared_ptr<MT19937> prng; - -void InitCommandLine(int argc, char** argv, po::variables_map* conf) { - po::options_description opts("Configuration options"); - opts.add_options() - ("samples,s",po::value<unsigned>()->default_value(1000),"Number of samples") - ("input,i",po::value<string>(),"Read parallel data from") - ("random_seed,S",po::value<uint32_t>(), "Random seed"); - po::options_description clo("Command line options"); - clo.add_options() - ("config", po::value<string>(), "Configuration file") - ("help", "Print this help message and exit"); - po::options_description dconfig_options, dcmdline_options; - dconfig_options.add(opts); - dcmdline_options.add(opts).add(clo); - - po::store(parse_command_line(argc, argv, dcmdline_options), *conf); - if (conf->count("config")) { - ifstream config((*conf)["config"].as<string>().c_str()); - po::store(po::parse_config_file(config, dconfig_options), *conf); - } - po::notify(*conf); - - if (conf->count("help") || (conf->count("input") == 0)) { - cerr << dcmdline_options << endl; - exit(1); - } -} - -unsigned ReadCorpus(const string& filename, - vector<Lattice>* e, - set<WordID>* vocab_e) { - e->clear(); - vocab_e->clear(); - ReadFile rf(filename); - istream* in = rf.stream(); - assert(*in); - string line; - unsigned toks = 0; - while(*in) { - getline(*in, line); - if (line.empty() && !*in) break; - e->push_back(Lattice()); - Lattice& le = e->back(); - LatticeTools::ConvertTextOrPLF(line, & le); - for (unsigned i = 0; i < le.size(); ++i) - for (unsigned j = 0; j < le[i].size(); ++j) - vocab_e->insert(le[i][j].label); - toks += le.size(); - } - return toks; -} - -struct BaseModel { - explicit BaseModel(unsigned tc) : - unif(1.0 / tc), p(prob_t::One()) {} - prob_t prob(const TRule& r) const { - return unif; - } - void increment(const TRule& r, MT19937* rng) { - p *= prob(r); - } - void decrement(const TRule& r, MT19937* rng) { - p /= prob(r); - } - prob_t Likelihood() const { - return p; - } - const prob_t unif; - prob_t p; -}; - -struct UnigramModel { - explicit UnigramModel(unsigned tc) : base(tc), crp(1,1,1,1), glue(1,1,1,1) {} - BaseModel base; - CCRP<TRule> crp; - CCRP<TRule> glue; - - prob_t Prob(const TRule& r) const { - if (r.Arity() != 0) { - return glue.prob(r, prob_t(0.5)); - } - return crp.prob(r, base.prob(r)); - } - - int Increment(const TRule& r, MT19937* rng) { - if (r.Arity() != 0) { - glue.increment(r, 0.5, rng); - return 0; - } else { - if (crp.increment(r, base.prob(r), rng)) { - base.increment(r, rng); - return 1; - } - return 0; - } - } - - int Decrement(const TRule& r, MT19937* rng) { - if (r.Arity() != 0) { - glue.decrement(r, rng); - return 0; - } else { - if (crp.decrement(r, rng)) { - base.decrement(r, rng); - return -1; - } - return 0; - } - } - - prob_t Likelihood() const { - prob_t p; - p.logeq(crp.log_crp_prob() + glue.log_crp_prob()); - p *= base.Likelihood(); - return p; - } - - void ResampleHyperparameters(MT19937* rng) { - crp.resample_hyperparameters(rng); - glue.resample_hyperparameters(rng); - cerr << " d=" << crp.discount() << ", s=" << crp.strength() << "\t STOP d=" << glue.discount() << ", s=" << glue.strength() << endl; - } -}; - -UnigramModel* plm; - -void SampleDerivation(const Hypergraph& hg, MT19937* rng, vector<unsigned>* sampled_deriv) { - vector<prob_t> node_probs; - Inside<prob_t, EdgeProb>(hg, &node_probs); - queue<unsigned> q; - q.push(hg.nodes_.size() - 2); - while(!q.empty()) { - unsigned cur_node_id = q.front(); -// cerr << "NODE=" << cur_node_id << endl; - q.pop(); - const Hypergraph::Node& node = hg.nodes_[cur_node_id]; - const unsigned num_in_edges = node.in_edges_.size(); - unsigned sampled_edge = 0; - if (num_in_edges == 1) { - sampled_edge = node.in_edges_[0]; - } else { - //prob_t z; - assert(num_in_edges > 1); - SampleSet<prob_t> ss; - for (unsigned j = 0; j < num_in_edges; ++j) { - const Hypergraph::Edge& edge = hg.edges_[node.in_edges_[j]]; - prob_t p = edge.edge_prob_; - for (unsigned k = 0; k < edge.tail_nodes_.size(); ++k) - p *= node_probs[edge.tail_nodes_[k]]; - ss.add(p); -// cerr << log(ss[j]) << " ||| " << edge.rule_->AsString() << endl; - //z += p; - } -// for (unsigned j = 0; j < num_in_edges; ++j) { -// const Hypergraph::Edge& edge = hg.edges_[node.in_edges_[j]]; -// cerr << exp(log(ss[j] / z)) << " ||| " << edge.rule_->AsString() << endl; -// } -// cerr << " --- \n"; - sampled_edge = node.in_edges_[rng->SelectSample(ss)]; - } - sampled_deriv->push_back(sampled_edge); - const Hypergraph::Edge& edge = hg.edges_[sampled_edge]; - for (unsigned j = 0; j < edge.tail_nodes_.size(); ++j) { - q.push(edge.tail_nodes_[j]); - } - } -// for (unsigned i = 0; i < sampled_deriv->size(); ++i) { -// cerr << *hg.edges_[(*sampled_deriv)[i]].rule_ << endl; -// } -} - -void IncrementDerivation(const Hypergraph& hg, const vector<unsigned>& d, UnigramModel* plm, MT19937* rng) { - for (unsigned i = 0; i < d.size(); ++i) - plm->Increment(*hg.edges_[d[i]].rule_, rng); -} - -void DecrementDerivation(const Hypergraph& hg, const vector<unsigned>& d, UnigramModel* plm, MT19937* rng) { - for (unsigned i = 0; i < d.size(); ++i) - plm->Decrement(*hg.edges_[d[i]].rule_, rng); -} - -prob_t TotalProb(const Hypergraph& hg) { - return Inside<prob_t, EdgeProb>(hg); -} - -void IncrementLatticePath(const Hypergraph& hg, const vector<unsigned>& d, Lattice* pl) { - Lattice& lat = *pl; - for (int i = 0; i < d.size(); ++i) { - const Hypergraph::Edge& edge = hg.edges_[d[i]]; - if (edge.rule_->Arity() != 0) continue; - WordID sym = edge.rule_->e_[0]; - vector<LatticeArc>& las = lat[edge.i_]; - int dist = edge.j_ - edge.i_; - assert(dist > 0); - for (int j = 0; j < las.size(); ++j) { - if (las[j].dist2next == dist && - las[j].label == sym) { - las[j].cost += 1; - } - } - } -} - -int main(int argc, char** argv) { - po::variables_map conf; - - InitCommandLine(argc, argv, &conf); - vector<GrammarPtr> grammars(2); - grammars[0].reset(new GlueGrammar("S","X")); - const unsigned samples = conf["samples"].as<unsigned>(); - - if (conf.count("random_seed")) - prng.reset(new MT19937(conf["random_seed"].as<uint32_t>())); - else - prng.reset(new MT19937); - MT19937& rng = *prng; - vector<Lattice> corpuse; - set<WordID> vocabe; - cerr << "Reading corpus...\n"; - const unsigned toks = ReadCorpus(conf["input"].as<string>(), &corpuse, &vocabe); - cerr << "E-corpus size: " << corpuse.size() << " lattices\t (" << vocabe.size() << " word types)\n"; - UnigramModel lm(vocabe.size()); - vector<Hypergraph> hgs(corpuse.size()); - vector<vector<unsigned> > derivs(corpuse.size()); - for (int i = 0; i < corpuse.size(); ++i) { - grammars[1].reset(new PassThroughGrammar(corpuse[i], "X")); - ExhaustiveBottomUpParser parser("S", grammars); - bool res = parser.Parse(corpuse[i], &hgs[i]); // exhaustive parse - assert(res); - } - - double csamples = 0; - for (int SS=0; SS < samples; ++SS) { - const bool is_last = ((samples - 1) == SS); - prob_t dlh = prob_t::One(); - bool record_sample = (SS > (samples * 1 / 3) && (SS % 5 == 3)); - if (record_sample) csamples++; - for (int ci = 0; ci < corpuse.size(); ++ci) { - Lattice& lat = corpuse[ci]; - Hypergraph& hg = hgs[ci]; - vector<unsigned>& d = derivs[ci]; - if (!is_last) DecrementDerivation(hg, d, &lm, &rng); - for (unsigned i = 0; i < hg.edges_.size(); ++i) { - TRule& r = *hg.edges_[i].rule_; - if (r.Arity() != 0) - hg.edges_[i].edge_prob_ = prob_t::One(); - else - hg.edges_[i].edge_prob_ = lm.Prob(r); - } - if (!is_last) { - d.clear(); - SampleDerivation(hg, &rng, &d); - IncrementDerivation(hg, derivs[ci], &lm, &rng); - } else { - prob_t p = TotalProb(hg); - dlh *= p; - cerr << " p(sentence) = " << log(p) << "\t" << log(dlh) << endl; - } - if (record_sample) IncrementLatticePath(hg, derivs[ci], &lat); - } - double llh = log(lm.Likelihood()); - cerr << "LLH=" << llh << "\tENTROPY=" << (-llh / log(2) / toks) << "\tPPL=" << pow(2, -llh / log(2) / toks) << endl; - if (SS % 10 == 9) lm.ResampleHyperparameters(&rng); - if (is_last) { - double z = log(dlh); - cerr << "TOTAL_PROB=" << z << "\tENTROPY=" << (-z / log(2) / toks) << "\tPPL=" << pow(2, -z / log(2) / toks) << endl; - } - } - cerr << lm.crp << endl; - cerr << lm.glue << endl; - for (int i = 0; i < corpuse.size(); ++i) { - for (int j = 0; j < corpuse[i].size(); ++j) - for (int k = 0; k < corpuse[i][j].size(); ++k) { - corpuse[i][j][k].cost /= csamples; - corpuse[i][j][k].cost += 1e-3; - corpuse[i][j][k].cost = log(corpuse[i][j][k].cost); - } - cout << HypergraphIO::AsPLF(corpuse[i]) << endl; - } - return 0; -} - diff --git a/gi/pf/brat.cc b/gi/pf/brat.cc deleted file mode 100644 index 832f22cf..00000000 --- a/gi/pf/brat.cc +++ /dev/null @@ -1,543 +0,0 @@ -#include <iostream> -#include <tr1/memory> -#include <queue> - -#include <boost/functional.hpp> -#include <boost/multi_array.hpp> -#include <boost/program_options.hpp> -#include <boost/program_options/variables_map.hpp> - -#include "viterbi.h" -#include "hg.h" -#include "trule.h" -#include "tdict.h" -#include "filelib.h" -#include "dict.h" -#include "sampler.h" -#include "ccrp_nt.h" -#include "cfg_wfst_composer.h" - -using namespace std; -using namespace tr1; -namespace po = boost::program_options; - -static unsigned kMAX_SRC_PHRASE; -static unsigned kMAX_TRG_PHRASE; -struct FSTState; - -double log_poisson(unsigned x, const double& lambda) { - assert(lambda > 0.0); - return log(lambda) * x - lgamma(x + 1) - lambda; -} - -struct ConditionalBase { - explicit ConditionalBase(const double m1mixture, const unsigned vocab_e_size, const string& model1fname) : - kM1MIXTURE(m1mixture), - kUNIFORM_MIXTURE(1.0 - m1mixture), - kUNIFORM_TARGET(1.0 / vocab_e_size), - kNULL(TD::Convert("<eps>")) { - assert(m1mixture >= 0.0 && m1mixture <= 1.0); - assert(vocab_e_size > 0); - LoadModel1(model1fname); - } - - void LoadModel1(const string& fname) { - cerr << "Loading Model 1 parameters from " << fname << " ..." << endl; - ReadFile rf(fname); - istream& in = *rf.stream(); - string line; - unsigned lc = 0; - while(getline(in, line)) { - ++lc; - int cur = 0; - int start = 0; - while(cur < line.size() && line[cur] != ' ') { ++cur; } - assert(cur != line.size()); - line[cur] = 0; - const WordID src = TD::Convert(&line[0]); - ++cur; - start = cur; - while(cur < line.size() && line[cur] != ' ') { ++cur; } - assert(cur != line.size()); - line[cur] = 0; - WordID trg = TD::Convert(&line[start]); - const double logprob = strtod(&line[cur + 1], NULL); - if (src >= ttable.size()) ttable.resize(src + 1); - ttable[src][trg].logeq(logprob); - } - cerr << " read " << lc << " parameters.\n"; - } - - // return logp0 of rule.e_ | rule.f_ - prob_t operator()(const TRule& rule) const { - const int flen = rule.f_.size(); - const int elen = rule.e_.size(); - prob_t uniform_src_alignment; uniform_src_alignment.logeq(-log(flen + 1)); - prob_t p; - p.logeq(log_poisson(elen, flen + 0.01)); // elen | flen ~Pois(flen + 0.01) - for (int i = 0; i < elen; ++i) { // for each position i in e-RHS - const WordID trg = rule.e_[i]; - prob_t tp = prob_t::Zero(); - for (int j = -1; j < flen; ++j) { - const WordID src = j < 0 ? kNULL : rule.f_[j]; - const map<WordID, prob_t>::const_iterator it = ttable[src].find(trg); - if (it != ttable[src].end()) { - tp += kM1MIXTURE * it->second; - } - tp += kUNIFORM_MIXTURE * kUNIFORM_TARGET; - } - tp *= uniform_src_alignment; // draw a_i ~uniform - p *= tp; // draw e_i ~Model1(f_a_i) / uniform - } - return p; - } - - const prob_t kM1MIXTURE; // Model 1 mixture component - const prob_t kUNIFORM_MIXTURE; // uniform mixture component - const prob_t kUNIFORM_TARGET; - const WordID kNULL; - vector<map<WordID, prob_t> > ttable; -}; - -void InitCommandLine(int argc, char** argv, po::variables_map* conf) { - po::options_description opts("Configuration options"); - opts.add_options() - ("samples,s",po::value<unsigned>()->default_value(1000),"Number of samples") - ("input,i",po::value<string>(),"Read parallel data from") - ("max_src_phrase",po::value<unsigned>()->default_value(3),"Maximum length of source language phrases") - ("max_trg_phrase",po::value<unsigned>()->default_value(3),"Maximum length of target language phrases") - ("model1,m",po::value<string>(),"Model 1 parameters (used in base distribution)") - ("model1_interpolation_weight",po::value<double>()->default_value(0.95),"Mixing proportion of model 1 with uniform target distribution") - ("random_seed,S",po::value<uint32_t>(), "Random seed"); - po::options_description clo("Command line options"); - clo.add_options() - ("config", po::value<string>(), "Configuration file") - ("help,h", "Print this help message and exit"); - po::options_description dconfig_options, dcmdline_options; - dconfig_options.add(opts); - dcmdline_options.add(opts).add(clo); - - po::store(parse_command_line(argc, argv, dcmdline_options), *conf); - if (conf->count("config")) { - ifstream config((*conf)["config"].as<string>().c_str()); - po::store(po::parse_config_file(config, dconfig_options), *conf); - } - po::notify(*conf); - - if (conf->count("help") || (conf->count("input") == 0)) { - cerr << dcmdline_options << endl; - exit(1); - } -} - -void ReadParallelCorpus(const string& filename, - vector<vector<WordID> >* f, - vector<vector<int> >* e, - set<int>* vocab_f, - set<int>* vocab_e) { - f->clear(); - e->clear(); - vocab_f->clear(); - vocab_e->clear(); - istream* in; - if (filename == "-") - in = &cin; - else - in = new ifstream(filename.c_str()); - assert(*in); - string line; - const WordID kDIV = TD::Convert("|||"); - vector<WordID> tmp; - while(*in) { - getline(*in, line); - if (line.empty() && !*in) break; - e->push_back(vector<int>()); - f->push_back(vector<int>()); - vector<int>& le = e->back(); - vector<int>& lf = f->back(); - tmp.clear(); - TD::ConvertSentence(line, &tmp); - bool isf = true; - for (unsigned i = 0; i < tmp.size(); ++i) { - const int cur = tmp[i]; - if (isf) { - if (kDIV == cur) { isf = false; } else { - lf.push_back(cur); - vocab_f->insert(cur); - } - } else { - assert(cur != kDIV); - le.push_back(cur); - vocab_e->insert(cur); - } - } - assert(isf == false); - } - if (in != &cin) delete in; -} - -struct UniphraseLM { - UniphraseLM(const vector<vector<int> >& corpus, - const set<int>& vocab, - const po::variables_map& conf) : - phrases_(1,1), - gen_(1,1), - corpus_(corpus), - uniform_word_(1.0 / vocab.size()), - gen_p0_(0.5), - p_end_(0.5), - use_poisson_(conf.count("poisson_length") > 0) {} - - void ResampleHyperparameters(MT19937* rng) { - phrases_.resample_hyperparameters(rng); - gen_.resample_hyperparameters(rng); - cerr << " " << phrases_.alpha(); - } - - CCRP_NoTable<vector<int> > phrases_; - CCRP_NoTable<bool> gen_; - vector<vector<bool> > z_; // z_[i] is there a phrase boundary after the ith word - const vector<vector<int> >& corpus_; - const double uniform_word_; - const double gen_p0_; - const double p_end_; // in base length distribution, p of the end of a phrase - const bool use_poisson_; -}; - -struct Reachability { - boost::multi_array<bool, 4> edges; // edges[src_covered][trg_covered][x][trg_delta] is this edge worth exploring? - boost::multi_array<short, 2> max_src_delta; // msd[src_covered][trg_covered] -- the largest src delta that's valid - - Reachability(int srclen, int trglen, int src_max_phrase_len, int trg_max_phrase_len) : - edges(boost::extents[srclen][trglen][src_max_phrase_len+1][trg_max_phrase_len+1]), - max_src_delta(boost::extents[srclen][trglen]) { - ComputeReachability(srclen, trglen, src_max_phrase_len, trg_max_phrase_len); - } - - private: - struct SState { - SState() : prev_src_covered(), prev_trg_covered() {} - SState(int i, int j) : prev_src_covered(i), prev_trg_covered(j) {} - int prev_src_covered; - int prev_trg_covered; - }; - - struct NState { - NState() : next_src_covered(), next_trg_covered() {} - NState(int i, int j) : next_src_covered(i), next_trg_covered(j) {} - int next_src_covered; - int next_trg_covered; - }; - - void ComputeReachability(int srclen, int trglen, int src_max_phrase_len, int trg_max_phrase_len) { - typedef boost::multi_array<vector<SState>, 2> array_type; - array_type a(boost::extents[srclen + 1][trglen + 1]); - a[0][0].push_back(SState()); - for (int i = 0; i < srclen; ++i) { - for (int j = 0; j < trglen; ++j) { - if (a[i][j].size() == 0) continue; - const SState prev(i,j); - for (int k = 1; k <= src_max_phrase_len; ++k) { - if ((i + k) > srclen) continue; - for (int l = 1; l <= trg_max_phrase_len; ++l) { - if ((j + l) > trglen) continue; - a[i + k][j + l].push_back(prev); - } - } - } - } - a[0][0].clear(); - cerr << "Final cell contains " << a[srclen][trglen].size() << " back pointers\n"; - assert(a[srclen][trglen].size() > 0); - - typedef boost::multi_array<bool, 2> rarray_type; - rarray_type r(boost::extents[srclen + 1][trglen + 1]); -// typedef boost::multi_array<vector<NState>, 2> narray_type; -// narray_type b(boost::extents[srclen + 1][trglen + 1]); - r[srclen][trglen] = true; - for (int i = srclen; i >= 0; --i) { - for (int j = trglen; j >= 0; --j) { - vector<SState>& prevs = a[i][j]; - if (!r[i][j]) { prevs.clear(); } -// const NState nstate(i,j); - for (int k = 0; k < prevs.size(); ++k) { - r[prevs[k].prev_src_covered][prevs[k].prev_trg_covered] = true; - int src_delta = i - prevs[k].prev_src_covered; - edges[prevs[k].prev_src_covered][prevs[k].prev_trg_covered][src_delta][j - prevs[k].prev_trg_covered] = true; - short &msd = max_src_delta[prevs[k].prev_src_covered][prevs[k].prev_trg_covered]; - if (src_delta > msd) msd = src_delta; -// b[prevs[k].prev_src_covered][prevs[k].prev_trg_covered].push_back(nstate); - } - } - } - assert(!edges[0][0][1][0]); - assert(!edges[0][0][0][1]); - assert(!edges[0][0][0][0]); - cerr << " MAX SRC DELTA[0][0] = " << max_src_delta[0][0] << endl; - assert(max_src_delta[0][0] > 0); - //cerr << "First cell contains " << b[0][0].size() << " forward pointers\n"; - //for (int i = 0; i < b[0][0].size(); ++i) { - // cerr << " -> (" << b[0][0][i].next_src_covered << "," << b[0][0][i].next_trg_covered << ")\n"; - //} - } -}; - -ostream& operator<<(ostream& os, const FSTState& q); -struct FSTState { - explicit FSTState(int src_size) : - trg_covered_(), - src_covered_(), - src_coverage_(src_size) {} - - FSTState(short trg_covered, short src_covered, const vector<bool>& src_coverage, const vector<short>& src_prefix) : - trg_covered_(trg_covered), - src_covered_(src_covered), - src_coverage_(src_coverage), - src_prefix_(src_prefix) { - if (src_coverage_.size() == src_covered) { - assert(src_prefix.size() == 0); - } - } - - // if we extend by the word at src_position, what are - // the next states that are reachable and lie on a valid - // path to the final state? - vector<FSTState> Extensions(int src_position, int src_len, int trg_len, const Reachability& r) const { - assert(src_position < src_coverage_.size()); - if (src_coverage_[src_position]) { - cerr << "Trying to extend " << *this << " with position " << src_position << endl; - abort(); - } - vector<bool> ncvg = src_coverage_; - ncvg[src_position] = true; - - vector<FSTState> res; - const int trg_remaining = trg_len - trg_covered_; - if (trg_remaining <= 0) { - cerr << "Target appears to have been covered: " << *this << " (trg_len=" << trg_len << ",trg_covered=" << trg_covered_ << ")" << endl; - abort(); - } - const int src_remaining = src_len - src_covered_; - if (src_remaining <= 0) { - cerr << "Source appears to have been covered: " << *this << endl; - abort(); - } - - for (int tc = 1; tc <= kMAX_TRG_PHRASE; ++tc) { - if (r.edges[src_covered_][trg_covered_][src_prefix_.size() + 1][tc]) { - int nc = src_prefix_.size() + 1 + src_covered_; - res.push_back(FSTState(trg_covered_ + tc, nc, ncvg, vector<short>())); - } - } - - if ((src_prefix_.size() + 1) < r.max_src_delta[src_covered_][trg_covered_]) { - vector<short> nsp = src_prefix_; - nsp.push_back(src_position); - res.push_back(FSTState(trg_covered_, src_covered_, ncvg, nsp)); - } - - if (res.size() == 0) { - cerr << *this << " can't be extended!\n"; - abort(); - } - return res; - } - - short trg_covered_, src_covered_; - vector<bool> src_coverage_; - vector<short> src_prefix_; -}; -bool operator<(const FSTState& q, const FSTState& r) { - if (q.trg_covered_ != r.trg_covered_) return q.trg_covered_ < r.trg_covered_; - if (q.src_covered_!= r.src_covered_) return q.src_covered_ < r.src_covered_; - if (q.src_coverage_ != r.src_coverage_) return q.src_coverage_ < r.src_coverage_; - return q.src_prefix_ < r.src_prefix_; -} - -ostream& operator<<(ostream& os, const FSTState& q) { - os << "[" << q.trg_covered_ << " : "; - for (int i = 0; i < q.src_coverage_.size(); ++i) - os << q.src_coverage_[i]; - os << " : <"; - for (int i = 0; i < q.src_prefix_.size(); ++i) { - if (i != 0) os << ' '; - os << q.src_prefix_[i]; - } - return os << ">]"; -} - -struct MyModel { - MyModel(ConditionalBase& rcp0) : rp0(rcp0) {} - typedef unordered_map<vector<WordID>, CCRP_NoTable<TRule>, boost::hash<vector<WordID> > > SrcToRuleCRPMap; - - void DecrementRule(const TRule& rule) { - SrcToRuleCRPMap::iterator it = rules.find(rule.f_); - assert(it != rules.end()); - it->second.decrement(rule); - if (it->second.num_customers() == 0) rules.erase(it); - } - - void IncrementRule(const TRule& rule) { - SrcToRuleCRPMap::iterator it = rules.find(rule.f_); - if (it == rules.end()) { - CCRP_NoTable<TRule> crp(1,1); - it = rules.insert(make_pair(rule.f_, crp)).first; - } - it->second.increment(rule); - } - - // conditioned on rule.f_ - prob_t RuleConditionalProbability(const TRule& rule) const { - const prob_t base = rp0(rule); - SrcToRuleCRPMap::const_iterator it = rules.find(rule.f_); - if (it == rules.end()) { - return base; - } else { - const double lp = it->second.logprob(rule, log(base)); - prob_t q; q.logeq(lp); - return q; - } - } - - const ConditionalBase& rp0; - SrcToRuleCRPMap rules; -}; - -struct MyFST : public WFST { - MyFST(const vector<WordID>& ssrc, const vector<WordID>& strg, MyModel* m) : - src(ssrc), trg(strg), - r(src.size(),trg.size(),kMAX_SRC_PHRASE, kMAX_TRG_PHRASE), - model(m) { - FSTState in(src.size()); - cerr << " INIT: " << in << endl; - init = GetNode(in); - for (int i = 0; i < in.src_coverage_.size(); ++i) in.src_coverage_[i] = true; - in.src_covered_ = src.size(); - in.trg_covered_ = trg.size(); - cerr << "FINAL: " << in << endl; - final = GetNode(in); - } - virtual const WFSTNode* Final() const; - virtual const WFSTNode* Initial() const; - - const WFSTNode* GetNode(const FSTState& q); - map<FSTState, boost::shared_ptr<WFSTNode> > m; - const vector<WordID>& src; - const vector<WordID>& trg; - Reachability r; - const WFSTNode* init; - const WFSTNode* final; - MyModel* model; -}; - -struct MyNode : public WFSTNode { - MyNode(const FSTState& q, MyFST* fst) : state(q), container(fst) {} - virtual vector<pair<const WFSTNode*, TRulePtr> > ExtendInput(unsigned srcindex) const; - const FSTState state; - mutable MyFST* container; -}; - -vector<pair<const WFSTNode*, TRulePtr> > MyNode::ExtendInput(unsigned srcindex) const { - cerr << "EXTEND " << state << " with " << srcindex << endl; - vector<FSTState> ext = state.Extensions(srcindex, container->src.size(), container->trg.size(), container->r); - vector<pair<const WFSTNode*,TRulePtr> > res(ext.size()); - for (unsigned i = 0; i < ext.size(); ++i) { - res[i].first = container->GetNode(ext[i]); - if (ext[i].src_prefix_.size() == 0) { - const unsigned trg_from = state.trg_covered_; - const unsigned trg_to = ext[i].trg_covered_; - const unsigned prev_prfx_size = state.src_prefix_.size(); - res[i].second.reset(new TRule); - res[i].second->lhs_ = -TD::Convert("X"); - vector<WordID>& src = res[i].second->f_; - vector<WordID>& trg = res[i].second->e_; - src.resize(prev_prfx_size + 1); - for (unsigned j = 0; j < prev_prfx_size; ++j) - src[j] = container->src[state.src_prefix_[j]]; - src[prev_prfx_size] = container->src[srcindex]; - for (unsigned j = trg_from; j < trg_to; ++j) - trg.push_back(container->trg[j]); - res[i].second->scores_.set_value(FD::Convert("Proposal"), log(container->model->RuleConditionalProbability(*res[i].second))); - } - } - return res; -} - -const WFSTNode* MyFST::GetNode(const FSTState& q) { - boost::shared_ptr<WFSTNode>& res = m[q]; - if (!res) { - res.reset(new MyNode(q, this)); - } - return &*res; -} - -const WFSTNode* MyFST::Final() const { - return final; -} - -const WFSTNode* MyFST::Initial() const { - return init; -} - -int main(int argc, char** argv) { - po::variables_map conf; - InitCommandLine(argc, argv, &conf); - kMAX_TRG_PHRASE = conf["max_trg_phrase"].as<unsigned>(); - kMAX_SRC_PHRASE = conf["max_src_phrase"].as<unsigned>(); - - if (!conf.count("model1")) { - cerr << argv[0] << "Please use --model1 to specify model 1 parameters\n"; - return 1; - } - boost::shared_ptr<MT19937> prng; - if (conf.count("random_seed")) - prng.reset(new MT19937(conf["random_seed"].as<uint32_t>())); - else - prng.reset(new MT19937); - MT19937& rng = *prng; - - vector<vector<int> > corpuse, corpusf; - set<int> vocabe, vocabf; - ReadParallelCorpus(conf["input"].as<string>(), &corpusf, &corpuse, &vocabf, &vocabe); - cerr << "f-Corpus size: " << corpusf.size() << " sentences\n"; - cerr << "f-Vocabulary size: " << vocabf.size() << " types\n"; - cerr << "f-Corpus size: " << corpuse.size() << " sentences\n"; - cerr << "f-Vocabulary size: " << vocabe.size() << " types\n"; - assert(corpusf.size() == corpuse.size()); - - ConditionalBase lp0(conf["model1_interpolation_weight"].as<double>(), - vocabe.size(), - conf["model1"].as<string>()); - MyModel m(lp0); - - TRule x("[X] ||| kAnwntR myN ||| at the convent ||| 0"); - m.IncrementRule(x); - TRule y("[X] ||| nY dyN ||| gave ||| 0"); - m.IncrementRule(y); - - - MyFST fst(corpusf[0], corpuse[0], &m); - ifstream in("./kimura.g"); - assert(in); - CFG_WFSTComposer comp(fst); - Hypergraph hg; - bool succeed = comp.Compose(&in, &hg); - hg.PrintGraphviz(); - if (succeed) { cerr << "SUCCESS.\n"; } else { cerr << "FAILURE REPORTED.\n"; } - -#if 0 - ifstream in2("./amnabooks.g"); - assert(in2); - MyFST fst2(corpusf[1], corpuse[1], &m); - CFG_WFSTComposer comp2(fst2); - Hypergraph hg2; - bool succeed2 = comp2.Compose(&in2, &hg2); - if (succeed2) { cerr << "SUCCESS.\n"; } else { cerr << "FAILURE REPORTED.\n"; } -#endif - - SparseVector<double> w; w.set_value(FD::Convert("Proposal"), 1.0); - hg.Reweight(w); - cerr << ViterbiFTree(hg) << endl; - return 0; -} - diff --git a/gi/pf/cbgi.cc b/gi/pf/cbgi.cc deleted file mode 100644 index 97f1ba34..00000000 --- a/gi/pf/cbgi.cc +++ /dev/null @@ -1,330 +0,0 @@ -#include <queue> -#include <sstream> -#include <iostream> - -#include <boost/unordered_map.hpp> -#include <boost/functional/hash.hpp> - -#include "sampler.h" -#include "filelib.h" -#include "hg_io.h" -#include "hg.h" -#include "ccrp_nt.h" -#include "trule.h" -#include "inside_outside.h" - -using namespace std; -using namespace std::tr1; - -double log_poisson(unsigned x, const double& lambda) { - assert(lambda > 0.0); - return log(lambda) * x - lgamma(x + 1) - lambda; -} - -double log_decay(unsigned x, const double& b) { - assert(b > 1.0); - assert(x > 0); - return log(b - 1) - x * log(b); -} - -struct SimpleBase { - SimpleBase(unsigned esize, unsigned fsize, unsigned ntsize = 144) : - uniform_e(-log(esize)), - uniform_f(-log(fsize)), - uniform_nt(-log(ntsize)) { - } - - // binomial coefficient - static double choose(unsigned n, unsigned k) { - return exp(lgamma(n + 1) - lgamma(k + 1) - lgamma(n - k + 1)); - } - - // count the number of patterns of terminals and NTs in the rule, given elen and flen - static double log_number_of_patterns(const unsigned flen, const unsigned elen) { - static vector<vector<double> > counts; - if (elen >= counts.size()) counts.resize(elen + 1); - if (flen >= counts[elen].size()) counts[elen].resize(flen + 1); - double& count = counts[elen][flen]; - if (count) return log(count); - const unsigned max_arity = min(elen, flen); - for (unsigned a = 0; a <= max_arity; ++a) - count += choose(elen, a) * choose(flen, a); - return log(count); - } - - // return logp0 of rule | LHS - double operator()(const TRule& rule) const { - const unsigned flen = rule.f_.size(); - const unsigned elen = rule.e_.size(); -#if 0 - double p = 0; - p += log_poisson(flen, 0.5); // flen ~Pois(0.5) - p += log_poisson(elen, flen); // elen | flen ~Pois(flen) - p -= log_number_of_patterns(flen, elen); // pattern | flen,elen ~Uniform - for (unsigned i = 0; i < flen; ++i) { // for each position in f-RHS - if (rule.f_[i] <= 0) // according to pattern - p += uniform_nt; // draw NT ~Uniform - else - p += uniform_f; // draw f terminal ~Uniform - } - p -= lgamma(rule.Arity() + 1); // draw permutation ~Uniform - for (unsigned i = 0; i < elen; ++i) { // for each position in e-RHS - if (rule.e_[i] > 0) // according to pattern - p += uniform_e; // draw e|f term ~Uniform - // TODO this should prob be model 1 - } -#else - double p = 0; - bool is_abstract = rule.f_[0] <= 0; - p += log(0.5); - if (is_abstract) { - if (flen == 2) p += log(0.99); else p += log(0.01); - } else { - p += log_decay(flen, 3); - } - - for (unsigned i = 0; i < flen; ++i) { // for each position in f-RHS - if (rule.f_[i] <= 0) // according to pattern - p += uniform_nt; // draw NT ~Uniform - else - p += uniform_f; // draw f terminal ~Uniform - } -#endif - return p; - } - const double uniform_e; - const double uniform_f; - const double uniform_nt; - vector<double> arities; -}; - -MT19937* rng = NULL; - -template <typename Base> -struct MHSamplerEdgeProb { - MHSamplerEdgeProb(const Hypergraph& hg, - const map<int, CCRP_NoTable<TRule> >& rdp, - const Base& logp0, - const bool exclude_multiword_terminals) : edge_probs(hg.edges_.size()) { - for (int i = 0; i < edge_probs.size(); ++i) { - const TRule& rule = *hg.edges_[i].rule_; - const map<int, CCRP_NoTable<TRule> >::const_iterator it = rdp.find(rule.lhs_); - assert(it != rdp.end()); - const CCRP_NoTable<TRule>& crp = it->second; - edge_probs[i].logeq(crp.logprob(rule, logp0(rule))); - if (exclude_multiword_terminals && rule.f_[0] > 0 && rule.f_.size() > 1) - edge_probs[i] = prob_t::Zero(); - } - } - inline prob_t operator()(const Hypergraph::Edge& e) const { - return edge_probs[e.id_]; - } - prob_t DerivationProb(const vector<int>& d) const { - prob_t p = prob_t::One(); - for (unsigned i = 0; i < d.size(); ++i) - p *= edge_probs[d[i]]; - return p; - } - vector<prob_t> edge_probs; -}; - -template <typename Base> -struct ModelAndData { - ModelAndData() : - base_lh(prob_t::One()), - logp0(10000, 10000), - mh_samples(), - mh_rejects() {} - - void SampleCorpus(const string& hgpath, int i); - void ResampleHyperparameters() { - for (map<int, CCRP_NoTable<TRule> >::iterator it = rules.begin(); it != rules.end(); ++it) - it->second.resample_hyperparameters(rng); - } - - CCRP_NoTable<TRule>& RuleCRP(int lhs) { - map<int, CCRP_NoTable<TRule> >::iterator it = rules.find(lhs); - if (it == rules.end()) { - rules.insert(make_pair(lhs, CCRP_NoTable<TRule>(1,1))); - it = rules.find(lhs); - } - return it->second; - } - - void IncrementRule(const TRule& rule) { - CCRP_NoTable<TRule>& crp = RuleCRP(rule.lhs_); - if (crp.increment(rule)) { - prob_t p; p.logeq(logp0(rule)); - base_lh *= p; - } - } - - void DecrementRule(const TRule& rule) { - CCRP_NoTable<TRule>& crp = RuleCRP(rule.lhs_); - if (crp.decrement(rule)) { - prob_t p; p.logeq(logp0(rule)); - base_lh /= p; - } - } - - void DecrementDerivation(const Hypergraph& hg, const vector<int>& d) { - for (unsigned i = 0; i < d.size(); ++i) { - const TRule& rule = *hg.edges_[d[i]].rule_; - DecrementRule(rule); - } - } - - void IncrementDerivation(const Hypergraph& hg, const vector<int>& d) { - for (unsigned i = 0; i < d.size(); ++i) { - const TRule& rule = *hg.edges_[d[i]].rule_; - IncrementRule(rule); - } - } - - prob_t Likelihood() const { - prob_t p = prob_t::One(); - for (map<int, CCRP_NoTable<TRule> >::const_iterator it = rules.begin(); it != rules.end(); ++it) { - prob_t q; q.logeq(it->second.log_crp_prob()); - p *= q; - } - p *= base_lh; - return p; - } - - void ResampleDerivation(const Hypergraph& hg, vector<int>* sampled_derivation); - - map<int, CCRP_NoTable<TRule> > rules; // [lhs] -> distribution over RHSs - prob_t base_lh; - SimpleBase logp0; - vector<vector<int> > samples; // sampled derivations - unsigned int mh_samples; - unsigned int mh_rejects; -}; - -template <typename Base> -void ModelAndData<Base>::SampleCorpus(const string& hgpath, int n) { - vector<Hypergraph> hgs(n); hgs.clear(); - boost::unordered_map<TRule, unsigned> acc; - map<int, unsigned> tot; - for (int i = 0; i < n; ++i) { - ostringstream os; - os << hgpath << '/' << i << ".json.gz"; - if (!FileExists(os.str())) continue; - hgs.push_back(Hypergraph()); - ReadFile rf(os.str()); - HypergraphIO::ReadFromJSON(rf.stream(), &hgs.back()); - } - cerr << "Read " << hgs.size() << " alignment hypergraphs.\n"; - samples.resize(hgs.size()); - const unsigned SAMPLES = 2000; - const unsigned burnin = 3 * SAMPLES / 4; - const unsigned every = 20; - for (unsigned s = 0; s < SAMPLES; ++s) { - if (s % 10 == 0) { - if (s > 0) { cerr << endl; ResampleHyperparameters(); } - cerr << "[" << s << " LLH=" << log(Likelihood()) << " REJECTS=" << ((double)mh_rejects / mh_samples) << " LHS's=" << rules.size() << " base=" << log(base_lh) << "] "; - } - cerr << '.'; - for (unsigned i = 0; i < hgs.size(); ++i) { - ResampleDerivation(hgs[i], &samples[i]); - if (s > burnin && s % every == 0) { - for (unsigned j = 0; j < samples[i].size(); ++j) { - const TRule& rule = *hgs[i].edges_[samples[i][j]].rule_; - ++acc[rule]; - ++tot[rule.lhs_]; - } - } - } - } - cerr << endl; - for (boost::unordered_map<TRule,unsigned>::iterator it = acc.begin(); it != acc.end(); ++it) { - cout << it->first << " MyProb=" << log(it->second)-log(tot[it->first.lhs_]) << endl; - } -} - -template <typename Base> -void ModelAndData<Base>::ResampleDerivation(const Hypergraph& hg, vector<int>* sampled_deriv) { - vector<int> cur; - cur.swap(*sampled_deriv); - - const prob_t p_cur = Likelihood(); - DecrementDerivation(hg, cur); - if (cur.empty()) { - // first iteration, create restaurants - for (int i = 0; i < hg.edges_.size(); ++i) - RuleCRP(hg.edges_[i].rule_->lhs_); - } - MHSamplerEdgeProb<SimpleBase> wf(hg, rules, logp0, cur.empty()); -// MHSamplerEdgeProb<SimpleBase> wf(hg, rules, logp0, false); - const prob_t q_cur = wf.DerivationProb(cur); - vector<prob_t> node_probs; - Inside<prob_t, MHSamplerEdgeProb<SimpleBase> >(hg, &node_probs, wf); - queue<unsigned> q; - q.push(hg.nodes_.size() - 3); - while(!q.empty()) { - unsigned cur_node_id = q.front(); -// cerr << "NODE=" << cur_node_id << endl; - q.pop(); - const Hypergraph::Node& node = hg.nodes_[cur_node_id]; - const unsigned num_in_edges = node.in_edges_.size(); - unsigned sampled_edge = 0; - if (num_in_edges == 1) { - sampled_edge = node.in_edges_[0]; - } else { - prob_t z; - assert(num_in_edges > 1); - SampleSet<prob_t> ss; - for (unsigned j = 0; j < num_in_edges; ++j) { - const Hypergraph::Edge& edge = hg.edges_[node.in_edges_[j]]; - prob_t p = wf.edge_probs[edge.id_]; // edge proposal prob - for (unsigned k = 0; k < edge.tail_nodes_.size(); ++k) - p *= node_probs[edge.tail_nodes_[k]]; - ss.add(p); -// cerr << log(ss[j]) << " ||| " << edge.rule_->AsString() << endl; - z += p; - } -// for (unsigned j = 0; j < num_in_edges; ++j) { -// const Hypergraph::Edge& edge = hg.edges_[node.in_edges_[j]]; -// cerr << exp(log(ss[j] / z)) << " ||| " << edge.rule_->AsString() << endl; -// } -// cerr << " --- \n"; - sampled_edge = node.in_edges_[rng->SelectSample(ss)]; - } - sampled_deriv->push_back(sampled_edge); - const Hypergraph::Edge& edge = hg.edges_[sampled_edge]; - for (unsigned j = 0; j < edge.tail_nodes_.size(); ++j) { - q.push(edge.tail_nodes_[j]); - } - } - IncrementDerivation(hg, *sampled_deriv); - -// cerr << "sampled derivation contains " << sampled_deriv->size() << " edges\n"; -// cerr << "DERIV:\n"; -// for (int i = 0; i < sampled_deriv->size(); ++i) { -// cerr << " " << hg.edges_[(*sampled_deriv)[i]].rule_->AsString() << endl; -// } - - if (cur.empty()) return; // accept first sample - - ++mh_samples; - // only need to do MH if proposal is different to current state - if (cur != *sampled_deriv) { - const prob_t q_prop = wf.DerivationProb(*sampled_deriv); - const prob_t p_prop = Likelihood(); - if (!rng->AcceptMetropolisHastings(p_prop, p_cur, q_prop, q_cur)) { - ++mh_rejects; - DecrementDerivation(hg, *sampled_deriv); - IncrementDerivation(hg, cur); - swap(cur, *sampled_deriv); - } - } -} - -int main(int argc, char** argv) { - rng = new MT19937; - ModelAndData<SimpleBase> m; - m.SampleCorpus("./hgs", 50); - // m.SampleCorpus("./btec/hgs", 5000); - return 0; -} - diff --git a/gi/pf/cfg_wfst_composer.cc b/gi/pf/cfg_wfst_composer.cc deleted file mode 100644 index 21d5ec5b..00000000 --- a/gi/pf/cfg_wfst_composer.cc +++ /dev/null @@ -1,731 +0,0 @@ -#include "cfg_wfst_composer.h" - -#include <iostream> -#include <fstream> -#include <map> -#include <queue> -#include <tr1/unordered_map> -#include <tr1/unordered_set> - -#include <boost/shared_ptr.hpp> -#include <boost/program_options.hpp> -#include <boost/program_options/variables_map.hpp> -#include "fast_lexical_cast.hpp" - -#include "phrasetable_fst.h" -#include "sparse_vector.h" -#include "tdict.h" -#include "hg.h" -#include "hg_remove_eps.h" - -namespace po = boost::program_options; -using namespace std; -using namespace std::tr1; - -WFSTNode::~WFSTNode() {} -WFST::~WFST() {} - -// Define the following macro if you want to see lots of debugging output -// when you run the chart parser -#undef DEBUG_CHART_PARSER - -// A few constants used by the chart parser /////////////// -static const int kMAX_NODES = 2000000; -static const string kPHRASE_STRING = "X"; -static bool constants_need_init = true; -static WordID kUNIQUE_START; -static WordID kPHRASE; -static TRulePtr kX1X2; -static TRulePtr kX1; -static WordID kEPS; -static TRulePtr kEPSRule; - -static void InitializeConstants() { - if (constants_need_init) { - kPHRASE = TD::Convert(kPHRASE_STRING) * -1; - kUNIQUE_START = TD::Convert("S") * -1; - kX1X2.reset(new TRule("[X] ||| [X,1] [X,2] ||| [X,1] [X,2]")); - kX1.reset(new TRule("[X] ||| [X,1] ||| [X,1]")); - kEPSRule.reset(new TRule("[X] ||| <eps> ||| <eps>")); - kEPS = TD::Convert("<eps>"); - constants_need_init = false; - } -} -//////////////////////////////////////////////////////////// - -class EGrammarNode { - friend bool CFG_WFSTComposer::Compose(const Hypergraph& src_forest, Hypergraph* trg_forest); - friend void AddGrammarRule(const string& r, map<WordID, EGrammarNode>* g); - public: -#ifdef DEBUG_CHART_PARSER - string hint; -#endif - EGrammarNode() : is_some_rule_complete(false), is_root(false) {} - const map<WordID, EGrammarNode>& GetTerminals() const { return tptr; } - const map<WordID, EGrammarNode>& GetNonTerminals() const { return ntptr; } - bool HasNonTerminals() const { return (!ntptr.empty()); } - bool HasTerminals() const { return (!tptr.empty()); } - bool RuleCompletes() const { - return (is_some_rule_complete || (ntptr.empty() && tptr.empty())); - } - bool GrammarContinues() const { - return !(ntptr.empty() && tptr.empty()); - } - bool IsRoot() const { - return is_root; - } - // these are the features associated with the rule from the start - // node up to this point. If you use these features, you must - // not Extend() this rule. - const SparseVector<double>& GetCFGProductionFeatures() const { - return input_features; - } - - const EGrammarNode* Extend(const WordID& t) const { - if (t < 0) { - map<WordID, EGrammarNode>::const_iterator it = ntptr.find(t); - if (it == ntptr.end()) return NULL; - return &it->second; - } else { - map<WordID, EGrammarNode>::const_iterator it = tptr.find(t); - if (it == tptr.end()) return NULL; - return &it->second; - } - } - - private: - map<WordID, EGrammarNode> tptr; - map<WordID, EGrammarNode> ntptr; - SparseVector<double> input_features; - bool is_some_rule_complete; - bool is_root; -}; -typedef map<WordID, EGrammarNode> EGrammar; // indexed by the rule LHS - -// edges are immutable once created -struct Edge { -#ifdef DEBUG_CHART_PARSER - static int id_count; - const int id; -#endif - const WordID cat; // lhs side of rule proved/being proved - const EGrammarNode* const dot; // dot position - const WFSTNode* const q; // start of span - const WFSTNode* const r; // end of span - const Edge* const active_parent; // back pointer, NULL for PREDICT items - const Edge* const passive_parent; // back pointer, NULL for SCAN and PREDICT items - TRulePtr tps; // translations - boost::shared_ptr<SparseVector<double> > features; // features from CFG rule - - bool IsPassive() const { - // when a rule is completed, this value will be set - return static_cast<bool>(features); - } - bool IsActive() const { return !IsPassive(); } - bool IsInitial() const { - return !(active_parent || passive_parent); - } - bool IsCreatedByScan() const { - return active_parent && !passive_parent && !dot->IsRoot(); - } - bool IsCreatedByPredict() const { - return dot->IsRoot(); - } - bool IsCreatedByComplete() const { - return active_parent && passive_parent; - } - - // constructor for PREDICT - Edge(WordID c, const EGrammarNode* d, const WFSTNode* q_and_r) : -#ifdef DEBUG_CHART_PARSER - id(++id_count), -#endif - cat(c), dot(d), q(q_and_r), r(q_and_r), active_parent(NULL), passive_parent(NULL), tps() {} - Edge(WordID c, const EGrammarNode* d, const WFSTNode* q_and_r, const Edge* act_parent) : -#ifdef DEBUG_CHART_PARSER - id(++id_count), -#endif - cat(c), dot(d), q(q_and_r), r(q_and_r), active_parent(act_parent), passive_parent(NULL), tps() {} - - // constructors for SCAN - Edge(WordID c, const EGrammarNode* d, const WFSTNode* i, const WFSTNode* j, - const Edge* act_par, const TRulePtr& translations) : -#ifdef DEBUG_CHART_PARSER - id(++id_count), -#endif - cat(c), dot(d), q(i), r(j), active_parent(act_par), passive_parent(NULL), tps(translations) {} - - Edge(WordID c, const EGrammarNode* d, const WFSTNode* i, const WFSTNode* j, - const Edge* act_par, const TRulePtr& translations, - const SparseVector<double>& feats) : -#ifdef DEBUG_CHART_PARSER - id(++id_count), -#endif - cat(c), dot(d), q(i), r(j), active_parent(act_par), passive_parent(NULL), tps(translations), - features(new SparseVector<double>(feats)) {} - - // constructors for COMPLETE - Edge(WordID c, const EGrammarNode* d, const WFSTNode* i, const WFSTNode* j, - const Edge* act_par, const Edge *pas_par) : -#ifdef DEBUG_CHART_PARSER - id(++id_count), -#endif - cat(c), dot(d), q(i), r(j), active_parent(act_par), passive_parent(pas_par), tps() { - assert(pas_par->IsPassive()); - assert(act_par->IsActive()); - } - - Edge(WordID c, const EGrammarNode* d, const WFSTNode* i, const WFSTNode* j, - const Edge* act_par, const Edge *pas_par, const SparseVector<double>& feats) : -#ifdef DEBUG_CHART_PARSER - id(++id_count), -#endif - cat(c), dot(d), q(i), r(j), active_parent(act_par), passive_parent(pas_par), tps(), - features(new SparseVector<double>(feats)) { - assert(pas_par->IsPassive()); - assert(act_par->IsActive()); - } - - // constructor for COMPLETE query - Edge(const WFSTNode* _r) : -#ifdef DEBUG_CHART_PARSER - id(0), -#endif - cat(0), dot(NULL), q(NULL), - r(_r), active_parent(NULL), passive_parent(NULL), tps() {} - // constructor for MERGE quere - Edge(const WFSTNode* _q, int) : -#ifdef DEBUG_CHART_PARSER - id(0), -#endif - cat(0), dot(NULL), q(_q), - r(NULL), active_parent(NULL), passive_parent(NULL), tps() {} -}; -#ifdef DEBUG_CHART_PARSER -int Edge::id_count = 0; -#endif - -ostream& operator<<(ostream& os, const Edge& e) { - string type = "PREDICT"; - if (e.IsCreatedByScan()) - type = "SCAN"; - else if (e.IsCreatedByComplete()) - type = "COMPLETE"; - os << "[" -#ifdef DEBUG_CHART_PARSER - << '(' << e.id << ") " -#else - << '(' << &e << ") " -#endif - << "q=" << e.q << ", r=" << e.r - << ", cat="<< TD::Convert(e.cat*-1) << ", dot=" - << e.dot -#ifdef DEBUG_CHART_PARSER - << e.dot->hint -#endif - << (e.IsActive() ? ", Active" : ", Passive") - << ", " << type; -#ifdef DEBUG_CHART_PARSER - if (e.active_parent) { os << ", act.parent=(" << e.active_parent->id << ')'; } - if (e.passive_parent) { os << ", psv.parent=(" << e.passive_parent->id << ')'; } -#endif - if (e.tps) { os << ", tps=" << e.tps->AsString(); } - return os << ']'; -} - -struct Traversal { - const Edge* const edge; // result from the active / passive combination - const Edge* const active; - const Edge* const passive; - Traversal(const Edge* me, const Edge* a, const Edge* p) : edge(me), active(a), passive(p) {} -}; - -struct UniqueTraversalHash { - size_t operator()(const Traversal* t) const { - size_t x = 5381; - x = ((x << 5) + x) ^ reinterpret_cast<size_t>(t->active); - x = ((x << 5) + x) ^ reinterpret_cast<size_t>(t->passive); - x = ((x << 5) + x) ^ t->edge->IsActive(); - return x; - } -}; - -struct UniqueTraversalEquals { - size_t operator()(const Traversal* a, const Traversal* b) const { - return (a->passive == b->passive && a->active == b->active && a->edge->IsActive() == b->edge->IsActive()); - } -}; - -struct UniqueEdgeHash { - size_t operator()(const Edge* e) const { - size_t x = 5381; - if (e->IsActive()) { - x = ((x << 5) + x) ^ reinterpret_cast<size_t>(e->dot); - x = ((x << 5) + x) ^ reinterpret_cast<size_t>(e->q); - x = ((x << 5) + x) ^ reinterpret_cast<size_t>(e->r); - x = ((x << 5) + x) ^ static_cast<size_t>(e->cat); - x += 13; - } else { // with passive edges, we don't care about the dot - x = ((x << 5) + x) ^ reinterpret_cast<size_t>(e->q); - x = ((x << 5) + x) ^ reinterpret_cast<size_t>(e->r); - x = ((x << 5) + x) ^ static_cast<size_t>(e->cat); - } - return x; - } -}; - -struct UniqueEdgeEquals { - bool operator()(const Edge* a, const Edge* b) const { - if (a->IsActive() != b->IsActive()) return false; - if (a->IsActive()) { - return (a->cat == b->cat) && (a->dot == b->dot) && (a->q == b->q) && (a->r == b->r); - } else { - return (a->cat == b->cat) && (a->q == b->q) && (a->r == b->r); - } - } -}; - -struct REdgeHash { - size_t operator()(const Edge* e) const { - size_t x = 5381; - x = ((x << 5) + x) ^ reinterpret_cast<size_t>(e->r); - return x; - } -}; - -struct REdgeEquals { - bool operator()(const Edge* a, const Edge* b) const { - return (a->r == b->r); - } -}; - -struct QEdgeHash { - size_t operator()(const Edge* e) const { - size_t x = 5381; - x = ((x << 5) + x) ^ reinterpret_cast<size_t>(e->q); - return x; - } -}; - -struct QEdgeEquals { - bool operator()(const Edge* a, const Edge* b) const { - return (a->q == b->q); - } -}; - -struct EdgeQueue { - queue<const Edge*> q; - EdgeQueue() {} - void clear() { while(!q.empty()) q.pop(); } - bool HasWork() const { return !q.empty(); } - const Edge* Next() { const Edge* res = q.front(); q.pop(); return res; } - void AddEdge(const Edge* s) { q.push(s); } -}; - -class CFG_WFSTComposerImpl { - public: - CFG_WFSTComposerImpl(WordID start_cat, - const WFSTNode* q_0, - const WFSTNode* q_final) : start_cat_(start_cat), q_0_(q_0), q_final_(q_final) {} - - // returns false if the intersection is empty - bool Compose(const EGrammar& g, Hypergraph* forest) { - goal_node = NULL; - EGrammar::const_iterator sit = g.find(start_cat_); - forest->ReserveNodes(kMAX_NODES); - assert(sit != g.end()); - Edge* init = new Edge(start_cat_, &sit->second, q_0_); - assert(IncorporateNewEdge(init)); - while (exp_agenda.HasWork() || agenda.HasWork()) { - while(exp_agenda.HasWork()) { - const Edge* edge = exp_agenda.Next(); - FinishEdge(edge, forest); - } - if (agenda.HasWork()) { - const Edge* edge = agenda.Next(); -#ifdef DEBUG_CHART_PARSER - cerr << "processing (" << edge->id << ')' << endl; -#endif - if (edge->IsActive()) { - if (edge->dot->HasTerminals()) - DoScan(edge); - if (edge->dot->HasNonTerminals()) { - DoMergeWithPassives(edge); - DoPredict(edge, g); - } - } else { - DoComplete(edge); - } - } - } - if (goal_node) { - forest->PruneUnreachable(goal_node->id_); - RemoveEpsilons(forest, kEPS); - } - FreeAll(); - return goal_node; - } - - void FreeAll() { - for (int i = 0; i < free_list_.size(); ++i) - delete free_list_[i]; - free_list_.clear(); - for (int i = 0; i < traversal_free_list_.size(); ++i) - delete traversal_free_list_[i]; - traversal_free_list_.clear(); - all_traversals.clear(); - exp_agenda.clear(); - agenda.clear(); - tps2node.clear(); - edge2node.clear(); - all_edges.clear(); - passive_edges.clear(); - active_edges.clear(); - } - - ~CFG_WFSTComposerImpl() { - FreeAll(); - } - - // returns the total number of edges created during composition - int EdgesCreated() const { - return free_list_.size(); - } - - private: - void DoScan(const Edge* edge) { - // here, we assume that the FST will potentially have many more outgoing - // edges than the grammar, which will be just a couple. If you want to - // efficiently handle the case where both are relatively large, this code - // will need to change how the intersection is done. The best general - // solution would probably be the Baeza-Yates double binary search. - - const EGrammarNode* dot = edge->dot; - const WFSTNode* r = edge->r; - const map<WordID, EGrammarNode>& terms = dot->GetTerminals(); - for (map<WordID, EGrammarNode>::const_iterator git = terms.begin(); - git != terms.end(); ++git) { - - if (!(TD::Convert(git->first)[0] >= '0' && TD::Convert(git->first)[0] <= '9')) { - std::cerr << "TERMINAL SYMBOL: " << TD::Convert(git->first) << endl; - abort(); - } - std::vector<std::pair<const WFSTNode*, TRulePtr> > extensions = r->ExtendInput(atoi(TD::Convert(git->first).c_str())); - for (unsigned nsi = 0; nsi < extensions.size(); ++nsi) { - const WFSTNode* next_r = extensions[nsi].first; - const EGrammarNode* next_dot = &git->second; - const bool grammar_continues = next_dot->GrammarContinues(); - const bool rule_completes = next_dot->RuleCompletes(); - if (extensions[nsi].second) - cerr << "!!! " << extensions[nsi].second->AsString() << endl; - // cerr << " rule completes: " << rule_completes << " after consuming " << TD::Convert(git->first) << endl; - assert(grammar_continues || rule_completes); - const SparseVector<double>& input_features = next_dot->GetCFGProductionFeatures(); - if (rule_completes) - IncorporateNewEdge(new Edge(edge->cat, next_dot, edge->q, next_r, edge, extensions[nsi].second, input_features)); - if (grammar_continues) - IncorporateNewEdge(new Edge(edge->cat, next_dot, edge->q, next_r, edge, extensions[nsi].second)); - } - } - } - - void DoPredict(const Edge* edge, const EGrammar& g) { - const EGrammarNode* dot = edge->dot; - const map<WordID, EGrammarNode>& non_terms = dot->GetNonTerminals(); - for (map<WordID, EGrammarNode>::const_iterator git = non_terms.begin(); - git != non_terms.end(); ++git) { - const WordID nt_to_predict = git->first; - //cerr << edge->id << " -- " << TD::Convert(nt_to_predict*-1) << endl; - EGrammar::const_iterator egi = g.find(nt_to_predict); - if (egi == g.end()) { - cerr << "[ERROR] Can't find any grammar rules with a LHS of type " - << TD::Convert(-1*nt_to_predict) << '!' << endl; - continue; - } - assert(edge->IsActive()); - const EGrammarNode* new_dot = &egi->second; - Edge* new_edge = new Edge(nt_to_predict, new_dot, edge->r, edge); - IncorporateNewEdge(new_edge); - } - } - - void DoComplete(const Edge* passive) { -#ifdef DEBUG_CHART_PARSER - cerr << " complete: " << *passive << endl; -#endif - const WordID completed_nt = passive->cat; - const WFSTNode* q = passive->q; - const WFSTNode* next_r = passive->r; - const Edge query(q); - const pair<unordered_multiset<const Edge*, REdgeHash, REdgeEquals>::iterator, - unordered_multiset<const Edge*, REdgeHash, REdgeEquals>::iterator > p = - active_edges.equal_range(&query); - for (unordered_multiset<const Edge*, REdgeHash, REdgeEquals>::iterator it = p.first; - it != p.second; ++it) { - const Edge* active = *it; -#ifdef DEBUG_CHART_PARSER - cerr << " pos: " << *active << endl; -#endif - const EGrammarNode* next_dot = active->dot->Extend(completed_nt); - if (!next_dot) continue; - const SparseVector<double>& input_features = next_dot->GetCFGProductionFeatures(); - // add up to 2 rules - if (next_dot->RuleCompletes()) - IncorporateNewEdge(new Edge(active->cat, next_dot, active->q, next_r, active, passive, input_features)); - if (next_dot->GrammarContinues()) - IncorporateNewEdge(new Edge(active->cat, next_dot, active->q, next_r, active, passive)); - } - } - - void DoMergeWithPassives(const Edge* active) { - // edge is active, has non-terminals, we need to find the passives that can extend it - assert(active->IsActive()); - assert(active->dot->HasNonTerminals()); -#ifdef DEBUG_CHART_PARSER - cerr << " merge active with passives: ACT=" << *active << endl; -#endif - const Edge query(active->r, 1); - const pair<unordered_multiset<const Edge*, QEdgeHash, QEdgeEquals>::iterator, - unordered_multiset<const Edge*, QEdgeHash, QEdgeEquals>::iterator > p = - passive_edges.equal_range(&query); - for (unordered_multiset<const Edge*, QEdgeHash, QEdgeEquals>::iterator it = p.first; - it != p.second; ++it) { - const Edge* passive = *it; - const EGrammarNode* next_dot = active->dot->Extend(passive->cat); - if (!next_dot) continue; - const WFSTNode* next_r = passive->r; - const SparseVector<double>& input_features = next_dot->GetCFGProductionFeatures(); - if (next_dot->RuleCompletes()) - IncorporateNewEdge(new Edge(active->cat, next_dot, active->q, next_r, active, passive, input_features)); - if (next_dot->GrammarContinues()) - IncorporateNewEdge(new Edge(active->cat, next_dot, active->q, next_r, active, passive)); - } - } - - // take ownership of edge memory, add to various indexes, etc - // returns true if this edge is new - bool IncorporateNewEdge(Edge* edge) { - free_list_.push_back(edge); - if (edge->passive_parent && edge->active_parent) { - Traversal* t = new Traversal(edge, edge->active_parent, edge->passive_parent); - traversal_free_list_.push_back(t); - if (all_traversals.find(t) != all_traversals.end()) { - return false; - } else { - all_traversals.insert(t); - } - } - exp_agenda.AddEdge(edge); - return true; - } - - bool FinishEdge(const Edge* edge, Hypergraph* hg) { - bool is_new = false; - if (all_edges.find(edge) == all_edges.end()) { -#ifdef DEBUG_CHART_PARSER - cerr << *edge << " is NEW\n"; -#endif - all_edges.insert(edge); - is_new = true; - if (edge->IsPassive()) passive_edges.insert(edge); - if (edge->IsActive()) active_edges.insert(edge); - agenda.AddEdge(edge); - } else { -#ifdef DEBUG_CHART_PARSER - cerr << *edge << " is NOT NEW.\n"; -#endif - } - AddEdgeToTranslationForest(edge, hg); - return is_new; - } - - // build the translation forest - void AddEdgeToTranslationForest(const Edge* edge, Hypergraph* hg) { - assert(hg->nodes_.size() < kMAX_NODES); - Hypergraph::Node* tps = NULL; - // first add any target language rules - if (edge->tps) { - Hypergraph::Node*& node = tps2node[(size_t)edge->tps.get()]; - if (!node) { - // cerr << "Creating phrases for " << edge->tps << endl; - const TRulePtr& rule = edge->tps; - node = hg->AddNode(kPHRASE); - Hypergraph::Edge* hg_edge = hg->AddEdge(rule, Hypergraph::TailNodeVector()); - hg_edge->feature_values_ += rule->GetFeatureValues(); - hg->ConnectEdgeToHeadNode(hg_edge, node); - } - tps = node; - } - Hypergraph::Node*& head_node = edge2node[edge]; - if (!head_node) - head_node = hg->AddNode(kPHRASE); - if (edge->cat == start_cat_ && edge->q == q_0_ && edge->r == q_final_ && edge->IsPassive()) { - assert(goal_node == NULL || goal_node == head_node); - goal_node = head_node; - } - Hypergraph::TailNodeVector tail; - SparseVector<double> extra; - if (edge->IsCreatedByPredict()) { - // extra.set_value(FD::Convert("predict"), 1); - } else if (edge->IsCreatedByScan()) { - tail.push_back(edge2node[edge->active_parent]->id_); - if (tps) { - tail.push_back(tps->id_); - } - //extra.set_value(FD::Convert("scan"), 1); - } else if (edge->IsCreatedByComplete()) { - tail.push_back(edge2node[edge->active_parent]->id_); - tail.push_back(edge2node[edge->passive_parent]->id_); - //extra.set_value(FD::Convert("complete"), 1); - } else { - assert(!"unexpected edge type!"); - } - //cerr << head_node->id_ << "<--" << *edge << endl; - -#ifdef DEBUG_CHART_PARSER - for (int i = 0; i < tail.size(); ++i) - if (tail[i] == head_node->id_) { - cerr << "ERROR: " << *edge << "\n i=" << i << endl; - if (i == 1) { cerr << "\tP: " << *edge->passive_parent << endl; } - if (i == 0) { cerr << "\tA: " << *edge->active_parent << endl; } - assert(!"self-loop found!"); - } -#endif - Hypergraph::Edge* hg_edge = NULL; - if (tail.size() == 0) { - hg_edge = hg->AddEdge(kEPSRule, tail); - } else if (tail.size() == 1) { - hg_edge = hg->AddEdge(kX1, tail); - } else if (tail.size() == 2) { - hg_edge = hg->AddEdge(kX1X2, tail); - } - if (edge->features) - hg_edge->feature_values_ += *edge->features; - hg_edge->feature_values_ += extra; - hg->ConnectEdgeToHeadNode(hg_edge, head_node); - } - - Hypergraph::Node* goal_node; - EdgeQueue exp_agenda; - EdgeQueue agenda; - unordered_map<size_t, Hypergraph::Node*> tps2node; - unordered_map<const Edge*, Hypergraph::Node*, UniqueEdgeHash, UniqueEdgeEquals> edge2node; - unordered_set<const Traversal*, UniqueTraversalHash, UniqueTraversalEquals> all_traversals; - unordered_set<const Edge*, UniqueEdgeHash, UniqueEdgeEquals> all_edges; - unordered_multiset<const Edge*, QEdgeHash, QEdgeEquals> passive_edges; - unordered_multiset<const Edge*, REdgeHash, REdgeEquals> active_edges; - vector<Edge*> free_list_; - vector<Traversal*> traversal_free_list_; - const WordID start_cat_; - const WFSTNode* const q_0_; - const WFSTNode* const q_final_; -}; - -#ifdef DEBUG_CHART_PARSER -static string TrimRule(const string& r) { - size_t start = r.find(" |||") + 5; - size_t end = r.rfind(" |||"); - return r.substr(start, end - start); -} -#endif - -void AddGrammarRule(const string& r, EGrammar* g) { - const size_t pos = r.find(" ||| "); - if (pos == string::npos || r[0] != '[') { - cerr << "Bad rule: " << r << endl; - return; - } - const size_t rpos = r.rfind(" ||| "); - string feats; - string rs = r; - if (rpos != pos) { - feats = r.substr(rpos + 5); - rs = r.substr(0, rpos); - } - string rhs = rs.substr(pos + 5); - string trule = rs + " ||| " + rhs + " ||| " + feats; - TRule tr(trule); - cerr << "X: " << tr.e_[0] << endl; -#ifdef DEBUG_CHART_PARSER - string hint_last_rule; -#endif - EGrammarNode* cur = &(*g)[tr.GetLHS()]; - cur->is_root = true; - for (int i = 0; i < tr.FLength(); ++i) { - WordID sym = tr.f()[i]; -#ifdef DEBUG_CHART_PARSER - hint_last_rule = TD::Convert(sym < 0 ? -sym : sym); - cur->hint += " <@@> (*" + hint_last_rule + ") " + TrimRule(tr.AsString()); -#endif - if (sym < 0) - cur = &cur->ntptr[sym]; - else - cur = &cur->tptr[sym]; - } -#ifdef DEBUG_CHART_PARSER - cur->hint += " <@@> (" + hint_last_rule + "*) " + TrimRule(tr.AsString()); -#endif - cur->is_some_rule_complete = true; - cur->input_features = tr.GetFeatureValues(); -} - -CFG_WFSTComposer::~CFG_WFSTComposer() { - delete pimpl_; -} - -CFG_WFSTComposer::CFG_WFSTComposer(const WFST& wfst) { - InitializeConstants(); - pimpl_ = new CFG_WFSTComposerImpl(kUNIQUE_START, wfst.Initial(), wfst.Final()); -} - -bool CFG_WFSTComposer::Compose(const Hypergraph& src_forest, Hypergraph* trg_forest) { - // first, convert the src forest into an EGrammar - EGrammar g; - const int nedges = src_forest.edges_.size(); - const int nnodes = src_forest.nodes_.size(); - vector<int> cats(nnodes); - bool assign_cats = false; - for (int i = 0; i < nnodes; ++i) - if (assign_cats) { - cats[i] = TD::Convert("CAT_" + boost::lexical_cast<string>(i)) * -1; - } else { - cats[i] = src_forest.nodes_[i].cat_; - } - // construct the grammar - for (int i = 0; i < nedges; ++i) { - const Hypergraph::Edge& edge = src_forest.edges_[i]; - const vector<WordID>& src = edge.rule_->f(); - EGrammarNode* cur = &g[cats[edge.head_node_]]; - cur->is_root = true; - int ntc = 0; - for (int j = 0; j < src.size(); ++j) { - WordID sym = src[j]; - if (sym <= 0) { - sym = cats[edge.tail_nodes_[ntc]]; - ++ntc; - cur = &cur->ntptr[sym]; - } else { - cur = &cur->tptr[sym]; - } - } - cur->is_some_rule_complete = true; - cur->input_features = edge.feature_values_; - } - EGrammarNode& goal_rule = g[kUNIQUE_START]; - assert((goal_rule.ntptr.size() == 1 && goal_rule.tptr.size() == 0) || - (goal_rule.ntptr.size() == 0 && goal_rule.tptr.size() == 1)); - - return pimpl_->Compose(g, trg_forest); -} - -bool CFG_WFSTComposer::Compose(istream* in, Hypergraph* trg_forest) { - EGrammar g; - while(*in) { - string line; - getline(*in, line); - if (line.empty()) continue; - AddGrammarRule(line, &g); - } - - return pimpl_->Compose(g, trg_forest); -} diff --git a/gi/pf/cfg_wfst_composer.h b/gi/pf/cfg_wfst_composer.h deleted file mode 100644 index cf47f459..00000000 --- a/gi/pf/cfg_wfst_composer.h +++ /dev/null @@ -1,46 +0,0 @@ -#ifndef _CFG_WFST_COMPOSER_H_ -#define _CFG_WFST_COMPOSER_H_ - -#include <iostream> -#include <vector> -#include <utility> - -#include "trule.h" -#include "wordid.h" - -class CFG_WFSTComposerImpl; -class Hypergraph; - -struct WFSTNode { - virtual ~WFSTNode(); - // returns the next states reachable by consuming srcindex (which identifies a word) - // paired with the output string generated by taking that transition. - virtual std::vector<std::pair<const WFSTNode*,TRulePtr> > ExtendInput(unsigned srcindex) const = 0; -}; - -struct WFST { - virtual ~WFST(); - virtual const WFSTNode* Final() const = 0; - virtual const WFSTNode* Initial() const = 0; -}; - -class CFG_WFSTComposer { - public: - ~CFG_WFSTComposer(); - explicit CFG_WFSTComposer(const WFST& wfst); - bool Compose(const Hypergraph& in_forest, Hypergraph* trg_forest); - - // reads the grammar from a file. There must be a single top-level - // S -> X rule. Anything else is possible. Format is: - // [S] ||| [SS,1] - // [SS] ||| [NP,1] [VP,2] ||| Feature1=0.2 Feature2=-2.3 - // [SS] ||| [VP,1] [NP,2] ||| Feature1=0.8 - // [NP] ||| [DET,1] [N,2] ||| Feature3=2 - // ... - bool Compose(std::istream* grammar_file, Hypergraph* trg_forest); - - private: - CFG_WFSTComposerImpl* pimpl_; -}; - -#endif diff --git a/gi/pf/conditional_pseg.h b/gi/pf/conditional_pseg.h deleted file mode 100644 index 81ddb206..00000000 --- a/gi/pf/conditional_pseg.h +++ /dev/null @@ -1,275 +0,0 @@ -#ifndef _CONDITIONAL_PSEG_H_ -#define _CONDITIONAL_PSEG_H_ - -#include <vector> -#include <tr1/unordered_map> -#include <boost/functional/hash.hpp> -#include <iostream> - -#include "m.h" -#include "prob.h" -#include "ccrp_nt.h" -#include "mfcr.h" -#include "trule.h" -#include "base_distributions.h" -#include "tdict.h" - -template <typename ConditionalBaseMeasure> -struct MConditionalTranslationModel { - explicit MConditionalTranslationModel(ConditionalBaseMeasure& rcp0) : - rp0(rcp0), d(0.5), strength(1.0), lambdas(1, prob_t::One()), p0s(1) {} - - void Summary() const { - std::cerr << "Number of conditioning contexts: " << r.size() << std::endl; - for (RuleModelHash::const_iterator it = r.begin(); it != r.end(); ++it) { - std::cerr << TD::GetString(it->first) << " \t(d=" << it->second.discount() << ",s=" << it->second.strength() << ") --------------------------" << std::endl; - for (MFCR<1,TRule>::const_iterator i2 = it->second.begin(); i2 != it->second.end(); ++i2) - std::cerr << " " << i2->second.total_dish_count_ << '\t' << i2->first << std::endl; - } - } - - double log_likelihood(const double& dd, const double& aa) const { - if (aa <= -dd) return -std::numeric_limits<double>::infinity(); - //double llh = Md::log_beta_density(dd, 10, 3) + Md::log_gamma_density(aa, 1, 1); - double llh = Md::log_beta_density(dd, 1, 1) + - Md::log_gamma_density(dd + aa, 1, 1); - typename std::tr1::unordered_map<std::vector<WordID>, MFCR<1,TRule>, boost::hash<std::vector<WordID> > >::const_iterator it; - for (it = r.begin(); it != r.end(); ++it) - llh += it->second.log_crp_prob(dd, aa); - return llh; - } - - struct DiscountResampler { - DiscountResampler(const MConditionalTranslationModel& m) : m_(m) {} - const MConditionalTranslationModel& m_; - double operator()(const double& proposed_discount) const { - return m_.log_likelihood(proposed_discount, m_.strength); - } - }; - - struct AlphaResampler { - AlphaResampler(const MConditionalTranslationModel& m) : m_(m) {} - const MConditionalTranslationModel& m_; - double operator()(const double& proposed_strength) const { - return m_.log_likelihood(m_.d, proposed_strength); - } - }; - - void ResampleHyperparameters(MT19937* rng) { - typename std::tr1::unordered_map<std::vector<WordID>, MFCR<1,TRule>, boost::hash<std::vector<WordID> > >::iterator it; -#if 1 - for (it = r.begin(); it != r.end(); ++it) { - it->second.resample_hyperparameters(rng); - } -#else - const unsigned nloop = 5; - const unsigned niterations = 10; - DiscountResampler dr(*this); - AlphaResampler ar(*this); - for (int iter = 0; iter < nloop; ++iter) { - strength = slice_sampler1d(ar, strength, *rng, -d + std::numeric_limits<double>::min(), - std::numeric_limits<double>::infinity(), 0.0, niterations, 100*niterations); - double min_discount = std::numeric_limits<double>::min(); - if (strength < 0.0) min_discount -= strength; - d = slice_sampler1d(dr, d, *rng, min_discount, - 1.0, 0.0, niterations, 100*niterations); - } - strength = slice_sampler1d(ar, strength, *rng, -d, - std::numeric_limits<double>::infinity(), 0.0, niterations, 100*niterations); - std::cerr << "MConditionalTranslationModel(d=" << d << ",s=" << strength << ") = " << log_likelihood(d, strength) << std::endl; - for (it = r.begin(); it != r.end(); ++it) { - it->second.set_discount(d); - it->second.set_strength(strength); - } -#endif - } - - int DecrementRule(const TRule& rule, MT19937* rng) { - RuleModelHash::iterator it = r.find(rule.f_); - assert(it != r.end()); - const TableCount delta = it->second.decrement(rule, rng); - if (delta.count) { - if (it->second.num_customers() == 0) r.erase(it); - } - return delta.count; - } - - int IncrementRule(const TRule& rule, MT19937* rng) { - RuleModelHash::iterator it = r.find(rule.f_); - if (it == r.end()) { - //it = r.insert(make_pair(rule.f_, MFCR<1,TRule>(d, strength))).first; - it = r.insert(make_pair(rule.f_, MFCR<1,TRule>(1,1,1,1,0.6, -0.12))).first; - } - p0s[0] = rp0(rule); - TableCount delta = it->second.increment(rule, p0s.begin(), lambdas.begin(), rng); - return delta.count; - } - - prob_t RuleProbability(const TRule& rule) const { - prob_t p; - RuleModelHash::const_iterator it = r.find(rule.f_); - if (it == r.end()) { - p = rp0(rule); - } else { - p0s[0] = rp0(rule); - p = it->second.prob(rule, p0s.begin(), lambdas.begin()); - } - return p; - } - - prob_t Likelihood() const { - prob_t p; p.logeq(log_likelihood(d, strength)); - return p; - } - - const ConditionalBaseMeasure& rp0; - typedef std::tr1::unordered_map<std::vector<WordID>, - MFCR<1, TRule>, - boost::hash<std::vector<WordID> > > RuleModelHash; - RuleModelHash r; - double d, strength; - std::vector<prob_t> lambdas; - mutable std::vector<prob_t> p0s; -}; - -template <typename ConditionalBaseMeasure> -struct ConditionalTranslationModel { - explicit ConditionalTranslationModel(ConditionalBaseMeasure& rcp0) : - rp0(rcp0) {} - - void Summary() const { - std::cerr << "Number of conditioning contexts: " << r.size() << std::endl; - for (RuleModelHash::const_iterator it = r.begin(); it != r.end(); ++it) { - std::cerr << TD::GetString(it->first) << " \t(\\alpha = " << it->second.alpha() << ") --------------------------" << std::endl; - for (CCRP_NoTable<TRule>::const_iterator i2 = it->second.begin(); i2 != it->second.end(); ++i2) - std::cerr << " " << i2->second << '\t' << i2->first << std::endl; - } - } - - void ResampleHyperparameters(MT19937* rng) { - for (RuleModelHash::iterator it = r.begin(); it != r.end(); ++it) - it->second.resample_hyperparameters(rng); - } - - int DecrementRule(const TRule& rule) { - RuleModelHash::iterator it = r.find(rule.f_); - assert(it != r.end()); - int count = it->second.decrement(rule); - if (count) { - if (it->second.num_customers() == 0) r.erase(it); - } - return count; - } - - int IncrementRule(const TRule& rule) { - RuleModelHash::iterator it = r.find(rule.f_); - if (it == r.end()) { - it = r.insert(make_pair(rule.f_, CCRP_NoTable<TRule>(1.0, 1.0, 8.0))).first; - } - int count = it->second.increment(rule); - return count; - } - - void IncrementRules(const std::vector<TRulePtr>& rules) { - for (int i = 0; i < rules.size(); ++i) - IncrementRule(*rules[i]); - } - - void DecrementRules(const std::vector<TRulePtr>& rules) { - for (int i = 0; i < rules.size(); ++i) - DecrementRule(*rules[i]); - } - - prob_t RuleProbability(const TRule& rule) const { - prob_t p; - RuleModelHash::const_iterator it = r.find(rule.f_); - if (it == r.end()) { - p.logeq(log(rp0(rule))); - } else { - p.logeq(it->second.logprob(rule, log(rp0(rule)))); - } - return p; - } - - prob_t Likelihood() const { - prob_t p = prob_t::One(); - for (RuleModelHash::const_iterator it = r.begin(); it != r.end(); ++it) { - prob_t q; q.logeq(it->second.log_crp_prob()); - p *= q; - for (CCRP_NoTable<TRule>::const_iterator i2 = it->second.begin(); i2 != it->second.end(); ++i2) - p *= rp0(i2->first); - } - return p; - } - - const ConditionalBaseMeasure& rp0; - typedef std::tr1::unordered_map<std::vector<WordID>, - CCRP_NoTable<TRule>, - boost::hash<std::vector<WordID> > > RuleModelHash; - RuleModelHash r; -}; - -template <typename ConditionalBaseMeasure> -struct ConditionalParallelSegementationModel { - explicit ConditionalParallelSegementationModel(ConditionalBaseMeasure& rcp0) : - tmodel(rcp0), base(prob_t::One()), aligns(1,1) {} - - ConditionalTranslationModel<ConditionalBaseMeasure> tmodel; - - void DecrementRule(const TRule& rule) { - tmodel.DecrementRule(rule); - } - - void IncrementRule(const TRule& rule) { - tmodel.IncrementRule(rule); - } - - void IncrementRulesAndAlignments(const std::vector<TRulePtr>& rules) { - tmodel.IncrementRules(rules); - for (int i = 0; i < rules.size(); ++i) { - IncrementAlign(rules[i]->f_.size()); - } - } - - void DecrementRulesAndAlignments(const std::vector<TRulePtr>& rules) { - tmodel.DecrementRules(rules); - for (int i = 0; i < rules.size(); ++i) { - DecrementAlign(rules[i]->f_.size()); - } - } - - prob_t RuleProbability(const TRule& rule) const { - return tmodel.RuleProbability(rule); - } - - void IncrementAlign(unsigned span) { - if (aligns.increment(span)) { - // TODO - } - } - - void DecrementAlign(unsigned span) { - if (aligns.decrement(span)) { - // TODO - } - } - - prob_t AlignProbability(unsigned span) const { - prob_t p; - p.logeq(aligns.logprob(span, Md::log_poisson(span, 1.0))); - return p; - } - - prob_t Likelihood() const { - prob_t p; p.logeq(aligns.log_crp_prob()); - p *= base; - p *= tmodel.Likelihood(); - return p; - } - - prob_t base; - CCRP_NoTable<unsigned> aligns; -}; - -#endif - diff --git a/gi/pf/condnaive.cc b/gi/pf/condnaive.cc deleted file mode 100644 index 419731ac..00000000 --- a/gi/pf/condnaive.cc +++ /dev/null @@ -1,298 +0,0 @@ -#include <iostream> -#include <tr1/memory> -#include <queue> - -#include <boost/multi_array.hpp> -#include <boost/program_options.hpp> -#include <boost/program_options/variables_map.hpp> - -#include "base_distributions.h" -#include "monotonic_pseg.h" -#include "conditional_pseg.h" -#include "trule.h" -#include "tdict.h" -#include "filelib.h" -#include "dict.h" -#include "sampler.h" -#include "ccrp_nt.h" -#include "corpus.h" - -using namespace std; -using namespace std::tr1; -namespace po = boost::program_options; - -static unsigned kMAX_SRC_PHRASE; -static unsigned kMAX_TRG_PHRASE; - -void InitCommandLine(int argc, char** argv, po::variables_map* conf) { - po::options_description opts("Configuration options"); - opts.add_options() - ("samples,s",po::value<unsigned>()->default_value(1000),"Number of samples") - ("input,i",po::value<string>(),"Read parallel data from") - ("max_src_phrase",po::value<unsigned>()->default_value(4),"Maximum length of source language phrases") - ("max_trg_phrase",po::value<unsigned>()->default_value(4),"Maximum length of target language phrases") - ("model1,m",po::value<string>(),"Model 1 parameters (used in base distribution)") - ("model1_interpolation_weight",po::value<double>()->default_value(0.95),"Mixing proportion of model 1 with uniform target distribution") - ("random_seed,S",po::value<uint32_t>(), "Random seed"); - po::options_description clo("Command line options"); - clo.add_options() - ("config", po::value<string>(), "Configuration file") - ("help,h", "Print this help message and exit"); - po::options_description dconfig_options, dcmdline_options; - dconfig_options.add(opts); - dcmdline_options.add(opts).add(clo); - - po::store(parse_command_line(argc, argv, dcmdline_options), *conf); - if (conf->count("config")) { - ifstream config((*conf)["config"].as<string>().c_str()); - po::store(po::parse_config_file(config, dconfig_options), *conf); - } - po::notify(*conf); - - if (conf->count("help") || (conf->count("input") == 0)) { - cerr << dcmdline_options << endl; - exit(1); - } -} - -boost::shared_ptr<MT19937> prng; - -struct ModelAndData { - explicit ModelAndData(ConditionalParallelSegementationModel<PhraseConditionalBase>& m, const vector<vector<int> >& ce, const vector<vector<int> >& cf, const set<int>& ve, const set<int>& vf) : - model(m), - rng(&*prng), - corpuse(ce), - corpusf(cf), - vocabe(ve), - vocabf(vf), - mh_samples(), - mh_rejects(), - kX(-TD::Convert("X")), - derivations(corpuse.size()) {} - - void ResampleHyperparameters() { - } - - void InstantiateRule(const pair<short,short>& from, - const pair<short,short>& to, - const vector<int>& sentf, - const vector<int>& sente, - TRule* rule) const { - rule->f_.clear(); - rule->e_.clear(); - rule->lhs_ = kX; - for (short i = from.first; i < to.first; ++i) - rule->f_.push_back(sentf[i]); - for (short i = from.second; i < to.second; ++i) - rule->e_.push_back(sente[i]); - } - - void DecrementDerivation(const vector<pair<short,short> >& d, const vector<int>& sentf, const vector<int>& sente) { - if (d.size() < 2) return; - TRule x; - for (int i = 1; i < d.size(); ++i) { - InstantiateRule(d[i], d[i-1], sentf, sente, &x); - model.DecrementRule(x); - model.DecrementAlign(x.f_.size()); - } - } - - void PrintDerivation(const vector<pair<short,short> >& d, const vector<int>& sentf, const vector<int>& sente) { - if (d.size() < 2) return; - TRule x; - for (int i = 1; i < d.size(); ++i) { - InstantiateRule(d[i], d[i-1], sentf, sente, &x); - cerr << i << '/' << (d.size() - 1) << ": " << x << endl; - } - } - - void IncrementDerivation(const vector<pair<short,short> >& d, const vector<int>& sentf, const vector<int>& sente) { - if (d.size() < 2) return; - TRule x; - for (int i = 1; i < d.size(); ++i) { - InstantiateRule(d[i], d[i-1], sentf, sente, &x); - model.IncrementRule(x); - model.IncrementAlign(x.f_.size()); - } - } - - prob_t Likelihood() const { - return model.Likelihood(); - } - - prob_t DerivationProposalProbability(const vector<pair<short,short> >& d, const vector<int>& sentf, const vector<int>& sente) const { - prob_t p = prob_t::One(); - TRule x; - for (int i = 1; i < d.size(); ++i) { - InstantiateRule(d[i], d[i-1], sentf, sente, &x); - p *= model.RuleProbability(x); - p *= model.AlignProbability(x.f_.size()); - } - return p; - } - - void Sample(); - - ConditionalParallelSegementationModel<PhraseConditionalBase>& model; - MT19937* rng; - const vector<vector<int> >& corpuse, corpusf; - const set<int>& vocabe, vocabf; - unsigned mh_samples, mh_rejects; - const int kX; - vector<vector<pair<short, short> > > derivations; -}; - -void ModelAndData::Sample() { - unsigned MAXK = kMAX_SRC_PHRASE; - unsigned MAXL = kMAX_TRG_PHRASE; - TRule x; - x.lhs_ = -TD::Convert("X"); - - for (int samples = 0; samples < 1000; ++samples) { - if (samples % 1 == 0 && samples > 0) { - //ResampleHyperparameters(); - cerr << " [" << samples << " LLH=" << log(Likelihood()) << " MH=" << ((double)mh_rejects / mh_samples) << "]\n"; - for (int i = 0; i < 10; ++i) { - cerr << "SENTENCE: " << TD::GetString(corpusf[i]) << " ||| " << TD::GetString(corpuse[i]) << endl; - PrintDerivation(derivations[i], corpusf[i], corpuse[i]); - } - static TRule xx("[X] ||| w n ||| s h ||| X=0"); - const CCRP_NoTable<TRule>& dcrp = model.tmodel.r.find(xx.f_)->second; - for (CCRP_NoTable<TRule>::const_iterator it = dcrp.begin(); it != dcrp.end(); ++it) { - cerr << "\t" << it->second << "\t" << it->first << endl; - } - } - cerr << '.' << flush; - for (int s = 0; s < corpuse.size(); ++s) { - const vector<int>& sentf = corpusf[s]; - const vector<int>& sente = corpuse[s]; -// cerr << " CUSTOMERS: " << rules.num_customers() << endl; -// cerr << "SENTENCE: " << TD::GetString(sentf) << " ||| " << TD::GetString(sente) << endl; - - vector<pair<short, short> >& deriv = derivations[s]; - const prob_t p_cur = Likelihood(); - DecrementDerivation(deriv, sentf, sente); - - boost::multi_array<prob_t, 2> a(boost::extents[sentf.size() + 1][sente.size() + 1]); - boost::multi_array<prob_t, 4> trans(boost::extents[sentf.size() + 1][sente.size() + 1][MAXK][MAXL]); - a[0][0] = prob_t::One(); - for (int i = 0; i < sentf.size(); ++i) { - for (int j = 0; j < sente.size(); ++j) { - const prob_t src_a = a[i][j]; - x.f_.clear(); - for (int k = 1; k <= MAXK; ++k) { - if (i + k > sentf.size()) break; - x.f_.push_back(sentf[i + k - 1]); - x.e_.clear(); - const prob_t p_span = model.AlignProbability(k); // prob of consuming this much source - for (int l = 1; l <= MAXL; ++l) { - if (j + l > sente.size()) break; - x.e_.push_back(sente[j + l - 1]); - trans[i][j][k - 1][l - 1] = model.RuleProbability(x) * p_span; - a[i + k][j + l] += src_a * trans[i][j][k - 1][l - 1]; - } - } - } - } -// cerr << "Inside: " << log(a[sentf.size()][sente.size()]) << endl; - const prob_t q_cur = DerivationProposalProbability(deriv, sentf, sente); - - vector<pair<short,short> > newderiv; - int cur_i = sentf.size(); - int cur_j = sente.size(); - while(cur_i > 0 && cur_j > 0) { - newderiv.push_back(pair<short,short>(cur_i, cur_j)); -// cerr << "NODE: (" << cur_i << "," << cur_j << ")\n"; - SampleSet<prob_t> ss; - vector<pair<short,short> > nexts; - for (int k = 1; k <= MAXK; ++k) { - const int hyp_i = cur_i - k; - if (hyp_i < 0) break; - for (int l = 1; l <= MAXL; ++l) { - const int hyp_j = cur_j - l; - if (hyp_j < 0) break; - const prob_t& inside = a[hyp_i][hyp_j]; - if (inside == prob_t::Zero()) continue; - const prob_t& transp = trans[hyp_i][hyp_j][k - 1][l - 1]; - if (transp == prob_t::Zero()) continue; - const prob_t p = inside * transp; - ss.add(p); - nexts.push_back(pair<short,short>(hyp_i, hyp_j)); -// cerr << " (" << hyp_i << "," << hyp_j << ") <--- " << log(p) << endl; - } - } -// cerr << " sample set has " << nexts.size() << " elements.\n"; - const int selected = rng->SelectSample(ss); - cur_i = nexts[selected].first; - cur_j = nexts[selected].second; - } - newderiv.push_back(pair<short,short>(0,0)); - const prob_t q_new = DerivationProposalProbability(newderiv, sentf, sente); - IncrementDerivation(newderiv, sentf, sente); -// cerr << "SANITY: " << q_new << " " <<log(DerivationProposalProbability(newderiv, sentf, sente)) << endl; - if (deriv.empty()) { deriv = newderiv; continue; } - ++mh_samples; - - if (deriv != newderiv) { - const prob_t p_new = Likelihood(); -// cerr << "p_cur=" << log(p_cur) << "\t p_new=" << log(p_new) << endl; -// cerr << "q_cur=" << log(q_cur) << "\t q_new=" << log(q_new) << endl; - if (!rng->AcceptMetropolisHastings(p_new, p_cur, q_new, q_cur)) { - ++mh_rejects; - DecrementDerivation(newderiv, sentf, sente); - IncrementDerivation(deriv, sentf, sente); - } else { -// cerr << " ACCEPT\n"; - deriv = newderiv; - } - } - } - } -} - -int main(int argc, char** argv) { - po::variables_map conf; - InitCommandLine(argc, argv, &conf); - kMAX_TRG_PHRASE = conf["max_trg_phrase"].as<unsigned>(); - kMAX_SRC_PHRASE = conf["max_src_phrase"].as<unsigned>(); - - if (!conf.count("model1")) { - cerr << argv[0] << "Please use --model1 to specify model 1 parameters\n"; - return 1; - } - if (conf.count("random_seed")) - prng.reset(new MT19937(conf["random_seed"].as<uint32_t>())); - else - prng.reset(new MT19937); -// MT19937& rng = *prng; - - vector<vector<int> > corpuse, corpusf; - set<int> vocabe, vocabf; - corpus::ReadParallelCorpus(conf["input"].as<string>(), &corpusf, &corpuse, &vocabf, &vocabe); - cerr << "f-Corpus size: " << corpusf.size() << " sentences\n"; - cerr << "f-Vocabulary size: " << vocabf.size() << " types\n"; - cerr << "f-Corpus size: " << corpuse.size() << " sentences\n"; - cerr << "f-Vocabulary size: " << vocabe.size() << " types\n"; - assert(corpusf.size() == corpuse.size()); - - Model1 m1(conf["model1"].as<string>()); - - PhraseConditionalBase pcb0(m1, conf["model1_interpolation_weight"].as<double>(), vocabe.size()); - ConditionalParallelSegementationModel<PhraseConditionalBase> x(pcb0); - - ModelAndData posterior(x, corpuse, corpusf, vocabe, vocabf); - posterior.Sample(); - - TRule r1("[X] ||| x ||| l e ||| X=0"); - TRule r2("[X] ||| A ||| a d ||| X=0"); - TRule r3("[X] ||| n ||| e r ||| X=0"); - TRule r4("[X] ||| x A n ||| b l a g ||| X=0"); - - PhraseConditionalUninformativeBase u0(vocabe.size()); - - cerr << (pcb0(r1)*pcb0(r2)*pcb0(r3)) << endl; - cerr << (u0(r4)) << endl; - - return 0; -} - diff --git a/gi/pf/corpus.cc b/gi/pf/corpus.cc deleted file mode 100644 index cb6e4ed7..00000000 --- a/gi/pf/corpus.cc +++ /dev/null @@ -1,62 +0,0 @@ -#include "corpus.h" - -#include <set> -#include <vector> -#include <string> - -#include "tdict.h" -#include "filelib.h" - -using namespace std; - -namespace corpus { - -void ReadParallelCorpus(const string& filename, - vector<vector<WordID> >* f, - vector<vector<WordID> >* e, - set<WordID>* vocab_f, - set<WordID>* vocab_e) { - f->clear(); - e->clear(); - vocab_f->clear(); - vocab_e->clear(); - ReadFile rf(filename); - istream* in = rf.stream(); - assert(*in); - string line; - unsigned lc = 0; - const WordID kDIV = TD::Convert("|||"); - vector<WordID> tmp; - while(getline(*in, line)) { - ++lc; - e->push_back(vector<int>()); - f->push_back(vector<int>()); - vector<int>& le = e->back(); - vector<int>& lf = f->back(); - tmp.clear(); - TD::ConvertSentence(line, &tmp); - bool isf = true; - for (unsigned i = 0; i < tmp.size(); ++i) { - const int cur = tmp[i]; - if (isf) { - if (kDIV == cur) { - isf = false; - } else { - lf.push_back(cur); - vocab_f->insert(cur); - } - } else { - if (cur == kDIV) { - cerr << "ERROR in " << lc << ": " << line << endl << endl; - abort(); - } - le.push_back(cur); - vocab_e->insert(cur); - } - } - assert(isf == false); - } -} - -} - diff --git a/gi/pf/corpus.h b/gi/pf/corpus.h deleted file mode 100644 index e7febdb7..00000000 --- a/gi/pf/corpus.h +++ /dev/null @@ -1,19 +0,0 @@ -#ifndef _CORPUS_H_ -#define _CORPUS_H_ - -#include <string> -#include <vector> -#include <set> -#include "wordid.h" - -namespace corpus { - -void ReadParallelCorpus(const std::string& filename, - std::vector<std::vector<WordID> >* f, - std::vector<std::vector<WordID> >* e, - std::set<WordID>* vocab_f, - std::set<WordID>* vocab_e); - -} - -#endif diff --git a/gi/pf/dpnaive.cc b/gi/pf/dpnaive.cc deleted file mode 100644 index 75ccad72..00000000 --- a/gi/pf/dpnaive.cc +++ /dev/null @@ -1,301 +0,0 @@ -#include <iostream> -#include <tr1/memory> -#include <queue> - -#include <boost/multi_array.hpp> -#include <boost/program_options.hpp> -#include <boost/program_options/variables_map.hpp> - -#include "base_distributions.h" -#include "monotonic_pseg.h" -#include "trule.h" -#include "tdict.h" -#include "filelib.h" -#include "dict.h" -#include "sampler.h" -#include "ccrp_nt.h" -#include "corpus.h" - -using namespace std; -using namespace std::tr1; -namespace po = boost::program_options; - -static unsigned kMAX_SRC_PHRASE; -static unsigned kMAX_TRG_PHRASE; - -void InitCommandLine(int argc, char** argv, po::variables_map* conf) { - po::options_description opts("Configuration options"); - opts.add_options() - ("samples,s",po::value<unsigned>()->default_value(1000),"Number of samples") - ("input,i",po::value<string>(),"Read parallel data from") - ("max_src_phrase",po::value<unsigned>()->default_value(4),"Maximum length of source language phrases") - ("max_trg_phrase",po::value<unsigned>()->default_value(4),"Maximum length of target language phrases") - ("model1,m",po::value<string>(),"Model 1 parameters (used in base distribution)") - ("inverse_model1,M",po::value<string>(),"Inverse Model 1 parameters (used in base distribution)") - ("model1_interpolation_weight",po::value<double>()->default_value(0.95),"Mixing proportion of model 1 with uniform target distribution") - ("random_seed,S",po::value<uint32_t>(), "Random seed"); - po::options_description clo("Command line options"); - clo.add_options() - ("config", po::value<string>(), "Configuration file") - ("help,h", "Print this help message and exit"); - po::options_description dconfig_options, dcmdline_options; - dconfig_options.add(opts); - dcmdline_options.add(opts).add(clo); - - po::store(parse_command_line(argc, argv, dcmdline_options), *conf); - if (conf->count("config")) { - ifstream config((*conf)["config"].as<string>().c_str()); - po::store(po::parse_config_file(config, dconfig_options), *conf); - } - po::notify(*conf); - - if (conf->count("help") || (conf->count("input") == 0)) { - cerr << dcmdline_options << endl; - exit(1); - } -} - -boost::shared_ptr<MT19937> prng; - -template <typename Base> -struct ModelAndData { - explicit ModelAndData(MonotonicParallelSegementationModel<PhraseJointBase_BiDir>& m, const Base& b, const vector<vector<int> >& ce, const vector<vector<int> >& cf, const set<int>& ve, const set<int>& vf) : - model(m), - rng(&*prng), - p0(b), - baseprob(prob_t::One()), - corpuse(ce), - corpusf(cf), - vocabe(ve), - vocabf(vf), - mh_samples(), - mh_rejects(), - kX(-TD::Convert("X")), - derivations(corpuse.size()) {} - - void ResampleHyperparameters() { - } - - void InstantiateRule(const pair<short,short>& from, - const pair<short,short>& to, - const vector<int>& sentf, - const vector<int>& sente, - TRule* rule) const { - rule->f_.clear(); - rule->e_.clear(); - rule->lhs_ = kX; - for (short i = from.first; i < to.first; ++i) - rule->f_.push_back(sentf[i]); - for (short i = from.second; i < to.second; ++i) - rule->e_.push_back(sente[i]); - } - - void DecrementDerivation(const vector<pair<short,short> >& d, const vector<int>& sentf, const vector<int>& sente) { - if (d.size() < 2) return; - TRule x; - for (int i = 1; i < d.size(); ++i) { - InstantiateRule(d[i], d[i-1], sentf, sente, &x); - model.DecrementRule(x); - model.DecrementContinue(); - } - model.DecrementStop(); - } - - void PrintDerivation(const vector<pair<short,short> >& d, const vector<int>& sentf, const vector<int>& sente) { - if (d.size() < 2) return; - TRule x; - for (int i = 1; i < d.size(); ++i) { - InstantiateRule(d[i], d[i-1], sentf, sente, &x); - cerr << i << '/' << (d.size() - 1) << ": " << x << endl; - } - } - - void IncrementDerivation(const vector<pair<short,short> >& d, const vector<int>& sentf, const vector<int>& sente) { - if (d.size() < 2) return; - TRule x; - for (int i = 1; i < d.size(); ++i) { - InstantiateRule(d[i], d[i-1], sentf, sente, &x); - model.IncrementRule(x); - model.IncrementContinue(); - } - model.IncrementStop(); - } - - prob_t Likelihood() const { - return model.Likelihood(); - } - - prob_t DerivationProposalProbability(const vector<pair<short,short> >& d, const vector<int>& sentf, const vector<int>& sente) const { - prob_t p = model.StopProbability(); - if (d.size() < 2) return p; - TRule x; - const prob_t p_cont = model.ContinueProbability(); - for (int i = 1; i < d.size(); ++i) { - InstantiateRule(d[i], d[i-1], sentf, sente, &x); - p *= p_cont; - p *= model.RuleProbability(x); - } - return p; - } - - void Sample(); - - MonotonicParallelSegementationModel<PhraseJointBase_BiDir>& model; - MT19937* rng; - const Base& p0; - prob_t baseprob; // cached value of generating the table table labels from p0 - // this can't be used if we go to a hierarchical prior! - const vector<vector<int> >& corpuse, corpusf; - const set<int>& vocabe, vocabf; - unsigned mh_samples, mh_rejects; - const int kX; - vector<vector<pair<short, short> > > derivations; -}; - -template <typename Base> -void ModelAndData<Base>::Sample() { - unsigned MAXK = kMAX_SRC_PHRASE; - unsigned MAXL = kMAX_TRG_PHRASE; - TRule x; - x.lhs_ = -TD::Convert("X"); - for (int samples = 0; samples < 1000; ++samples) { - if (samples % 1 == 0 && samples > 0) { - //ResampleHyperparameters(); - cerr << " [" << samples << " LLH=" << log(Likelihood()) << " MH=" << ((double)mh_rejects / mh_samples) << "]\n"; - for (int i = 0; i < 10; ++i) { - cerr << "SENTENCE: " << TD::GetString(corpusf[i]) << " ||| " << TD::GetString(corpuse[i]) << endl; - PrintDerivation(derivations[i], corpusf[i], corpuse[i]); - } - } - cerr << '.' << flush; - for (int s = 0; s < corpuse.size(); ++s) { - const vector<int>& sentf = corpusf[s]; - const vector<int>& sente = corpuse[s]; -// cerr << " CUSTOMERS: " << rules.num_customers() << endl; -// cerr << "SENTENCE: " << TD::GetString(sentf) << " ||| " << TD::GetString(sente) << endl; - - vector<pair<short, short> >& deriv = derivations[s]; - const prob_t p_cur = Likelihood(); - DecrementDerivation(deriv, sentf, sente); - - boost::multi_array<prob_t, 2> a(boost::extents[sentf.size() + 1][sente.size() + 1]); - boost::multi_array<prob_t, 4> trans(boost::extents[sentf.size() + 1][sente.size() + 1][MAXK][MAXL]); - a[0][0] = prob_t::One(); - const prob_t q_stop = model.StopProbability(); - const prob_t q_cont = model.ContinueProbability(); - for (int i = 0; i < sentf.size(); ++i) { - for (int j = 0; j < sente.size(); ++j) { - const prob_t src_a = a[i][j]; - x.f_.clear(); - for (int k = 1; k <= MAXK; ++k) { - if (i + k > sentf.size()) break; - x.f_.push_back(sentf[i + k - 1]); - x.e_.clear(); - for (int l = 1; l <= MAXL; ++l) { - if (j + l > sente.size()) break; - x.e_.push_back(sente[j + l - 1]); - const bool stop_now = ((j + l) == sente.size()) && ((i + k) == sentf.size()); - const prob_t& cp = stop_now ? q_stop : q_cont; - trans[i][j][k - 1][l - 1] = model.RuleProbability(x) * cp; - a[i + k][j + l] += src_a * trans[i][j][k - 1][l - 1]; - } - } - } - } -// cerr << "Inside: " << log(a[sentf.size()][sente.size()]) << endl; - const prob_t q_cur = DerivationProposalProbability(deriv, sentf, sente); - - vector<pair<short,short> > newderiv; - int cur_i = sentf.size(); - int cur_j = sente.size(); - while(cur_i > 0 && cur_j > 0) { - newderiv.push_back(pair<short,short>(cur_i, cur_j)); -// cerr << "NODE: (" << cur_i << "," << cur_j << ")\n"; - SampleSet<prob_t> ss; - vector<pair<short,short> > nexts; - for (int k = 1; k <= MAXK; ++k) { - const int hyp_i = cur_i - k; - if (hyp_i < 0) break; - for (int l = 1; l <= MAXL; ++l) { - const int hyp_j = cur_j - l; - if (hyp_j < 0) break; - const prob_t& inside = a[hyp_i][hyp_j]; - if (inside == prob_t::Zero()) continue; - const prob_t& transp = trans[hyp_i][hyp_j][k - 1][l - 1]; - if (transp == prob_t::Zero()) continue; - const prob_t p = inside * transp; - ss.add(p); - nexts.push_back(pair<short,short>(hyp_i, hyp_j)); -// cerr << " (" << hyp_i << "," << hyp_j << ") <--- " << log(p) << endl; - } - } -// cerr << " sample set has " << nexts.size() << " elements.\n"; - const int selected = rng->SelectSample(ss); - cur_i = nexts[selected].first; - cur_j = nexts[selected].second; - } - newderiv.push_back(pair<short,short>(0,0)); - const prob_t q_new = DerivationProposalProbability(newderiv, sentf, sente); - IncrementDerivation(newderiv, sentf, sente); -// cerr << "SANITY: " << q_new << " " <<log(DerivationProposalProbability(newderiv, sentf, sente)) << endl; - if (deriv.empty()) { deriv = newderiv; continue; } - ++mh_samples; - - if (deriv != newderiv) { - const prob_t p_new = Likelihood(); -// cerr << "p_cur=" << log(p_cur) << "\t p_new=" << log(p_new) << endl; -// cerr << "q_cur=" << log(q_cur) << "\t q_new=" << log(q_new) << endl; - if (!rng->AcceptMetropolisHastings(p_new, p_cur, q_new, q_cur)) { - ++mh_rejects; - DecrementDerivation(newderiv, sentf, sente); - IncrementDerivation(deriv, sentf, sente); - } else { -// cerr << " ACCEPT\n"; - deriv = newderiv; - } - } - } - } -} - -int main(int argc, char** argv) { - po::variables_map conf; - InitCommandLine(argc, argv, &conf); - kMAX_TRG_PHRASE = conf["max_trg_phrase"].as<unsigned>(); - kMAX_SRC_PHRASE = conf["max_src_phrase"].as<unsigned>(); - - if (!conf.count("model1")) { - cerr << argv[0] << "Please use --model1 to specify model 1 parameters\n"; - return 1; - } - if (!conf.count("inverse_model1")) { - cerr << argv[0] << "Please use --inverse_model1 to specify inverse model 1 parameters\n"; - return 1; - } - if (conf.count("random_seed")) - prng.reset(new MT19937(conf["random_seed"].as<uint32_t>())); - else - prng.reset(new MT19937); -// MT19937& rng = *prng; - - vector<vector<int> > corpuse, corpusf; - set<int> vocabe, vocabf; - corpus::ReadParallelCorpus(conf["input"].as<string>(), &corpusf, &corpuse, &vocabf, &vocabe); - cerr << "f-Corpus size: " << corpusf.size() << " sentences\n"; - cerr << "f-Vocabulary size: " << vocabf.size() << " types\n"; - cerr << "f-Corpus size: " << corpuse.size() << " sentences\n"; - cerr << "f-Vocabulary size: " << vocabe.size() << " types\n"; - assert(corpusf.size() == corpuse.size()); - - Model1 m1(conf["model1"].as<string>()); - Model1 invm1(conf["inverse_model1"].as<string>()); -// PhraseJointBase lp0(m1, conf["model1_interpolation_weight"].as<double>(), vocabe.size(), vocabf.size()); - PhraseJointBase_BiDir alp0(m1, invm1, conf["model1_interpolation_weight"].as<double>(), vocabe.size(), vocabf.size()); - MonotonicParallelSegementationModel<PhraseJointBase_BiDir> m(alp0); - - ModelAndData<PhraseJointBase_BiDir> posterior(m, alp0, corpuse, corpusf, vocabe, vocabf); - posterior.Sample(); - - return 0; -} - diff --git a/gi/pf/guess-translits.pl b/gi/pf/guess-translits.pl deleted file mode 100755 index d00c2168..00000000 --- a/gi/pf/guess-translits.pl +++ /dev/null @@ -1,72 +0,0 @@ -#!/usr/bin/perl -w -use strict; -use utf8; - -my $MIN_PMI = -3; - -my %fs; -my %es; -my %ef; - -die "Usage: $0 < input.utf8.txt\n" if scalar @ARGV > 0; - -binmode(STDIN,":utf8"); -binmode(STDOUT,":utf8"); -binmode(STDERR,":utf8"); - -my $tot = 0; -print STDERR "Reading alignments from STDIN ...\n"; -while(<STDIN>) { - chomp; - my ($fsent, $esent, $alsent) = split / \|\|\| /; - die "Format should be 'foreign sentence ||| english sentence ||| 0-0 1-1 ...'\n" unless defined $fsent && defined $esent && defined $alsent; - - my @fws = split /\s+/, $fsent; - my @ews = split /\s+/, $esent; - my @as = split /\s+/, $alsent; - my %a2b; - my %b2a; - for my $ap (@as) { - my ($a,$b) = split /-/, $ap; - die "BAD INPUT: $_\n" unless defined $a && defined $b; - $a2b{$a}->{$b} = 1; - $b2a{$b}->{$a} = 1; - } - for my $a (keys %a2b) { - my $bref = $a2b{$a}; - next unless scalar keys %$bref < 2; - my $b = (keys %$bref)[0]; - next unless scalar keys %{$b2a{$b}} < 2; - my $f = $fws[$a]; - next unless defined $f; - next unless length($f) > 3; - my $e = $ews[$b]; - next unless defined $e; - next unless length($e) > 3; - - $ef{$f}->{$e}++; - $es{$e}++; - $fs{$f}++; - $tot++; - } -} -my $ltot = log($tot); -my $num = 0; -print STDERR "Extracting pairs for PMI > $MIN_PMI ...\n"; -for my $f (keys %fs) { - my $logf = log($fs{$f}); - my $esref = $ef{$f}; - for my $e (keys %$esref) { - my $loge = log($es{$e}); - my $ef = $esref->{$e}; - my $logef = log($ef); - my $pmi = $logef - ($loge + $logf); - next if $pmi < $MIN_PMI; - my @flets = split //, $f; - my @elets = split //, $e; - print "@flets ||| @elets\n"; - $num++; - } -} -print STDERR "Extracted $num pairs.\n"; -print STDERR "Recommend running:\n ../../training/model1 -v -d -t -99999 output.txt\n"; diff --git a/gi/pf/hpyp_tm.cc b/gi/pf/hpyp_tm.cc deleted file mode 100644 index f362d3f8..00000000 --- a/gi/pf/hpyp_tm.cc +++ /dev/null @@ -1,133 +0,0 @@ -#include "hpyp_tm.h" - -#include <tr1/unordered_map> -#include <iostream> -#include <queue> - -#include "tdict.h" -#include "ccrp.h" -#include "pyp_word_model.h" -#include "tied_resampler.h" - -using namespace std; -using namespace std::tr1; - -struct FreqBinner { - FreqBinner(const std::string& fname) { fd_.Load(fname); } - unsigned NumberOfBins() const { return fd_.Max() + 1; } - unsigned Bin(const WordID& w) const { return fd_.LookUp(w); } - FreqDict<unsigned> fd_; -}; - -template <typename Base, class Binner = FreqBinner> -struct ConditionalPYPWordModel { - ConditionalPYPWordModel(Base* b, const Binner* bnr = NULL) : - base(*b), - binner(bnr), - btr(binner ? binner->NumberOfBins() + 1u : 2u) {} - - void Summary() const { - cerr << "Number of conditioning contexts: " << r.size() << endl; - for (RuleModelHash::const_iterator it = r.begin(); it != r.end(); ++it) { - cerr << TD::Convert(it->first) << " \tPYP(d=" << it->second.discount() << ",s=" << it->second.strength() << ") --------------------------" << endl; - for (CCRP<vector<WordID> >::const_iterator i2 = it->second.begin(); i2 != it->second.end(); ++i2) - cerr << " " << i2->second << endl; - } - } - - void ResampleHyperparameters(MT19937* rng) { - btr.ResampleHyperparameters(rng); - } - - prob_t Prob(const WordID src, const vector<WordID>& trglets) const { - RuleModelHash::const_iterator it = r.find(src); - if (it == r.end()) { - return base(trglets); - } else { - return it->second.prob(trglets, base(trglets)); - } - } - - void Increment(const WordID src, const vector<WordID>& trglets, MT19937* rng) { - RuleModelHash::iterator it = r.find(src); - if (it == r.end()) { - it = r.insert(make_pair(src, CCRP<vector<WordID> >(0.5,1.0))).first; - static const WordID kNULL = TD::Convert("NULL"); - unsigned bin = (src == kNULL ? 0 : 1); - if (binner && bin) { bin = binner->Bin(src) + 1; } - btr.Add(bin, &it->second); - } - if (it->second.increment(trglets, base(trglets), rng)) - base.Increment(trglets, rng); - } - - void Decrement(const WordID src, const vector<WordID>& trglets, MT19937* rng) { - RuleModelHash::iterator it = r.find(src); - assert(it != r.end()); - if (it->second.decrement(trglets, rng)) { - base.Decrement(trglets, rng); - } - } - - prob_t Likelihood() const { - prob_t p = prob_t::One(); - for (RuleModelHash::const_iterator it = r.begin(); it != r.end(); ++it) { - prob_t q; q.logeq(it->second.log_crp_prob()); - p *= q; - } - return p; - } - - unsigned UniqueConditioningContexts() const { - return r.size(); - } - - // TODO tie PYP hyperparameters based on source word frequency bins - Base& base; - const Binner* binner; - BinTiedResampler<CCRP<vector<WordID> > > btr; - typedef unordered_map<WordID, CCRP<vector<WordID> > > RuleModelHash; - RuleModelHash r; -}; - -HPYPLexicalTranslation::HPYPLexicalTranslation(const vector<vector<WordID> >& lets, - const unsigned vocab_size, - const unsigned num_letters) : - letters(lets), - base(vocab_size, num_letters, 5), - up0(new PYPWordModel<PoissonUniformWordModel>(&base)), - tmodel(new ConditionalPYPWordModel<PYPWordModel<PoissonUniformWordModel> >(up0, new FreqBinner("10k.freq"))), - kX(-TD::Convert("X")) {} - -void HPYPLexicalTranslation::Summary() const { - tmodel->Summary(); - up0->Summary(); -} - -prob_t HPYPLexicalTranslation::Likelihood() const { - prob_t p = up0->Likelihood(); - p *= tmodel->Likelihood(); - return p; -} - -void HPYPLexicalTranslation::ResampleHyperparameters(MT19937* rng) { - tmodel->ResampleHyperparameters(rng); - up0->ResampleHyperparameters(rng); -} - -unsigned HPYPLexicalTranslation::UniqueConditioningContexts() const { - return tmodel->UniqueConditioningContexts(); -} - -prob_t HPYPLexicalTranslation::Prob(WordID src, WordID trg) const { - return tmodel->Prob(src, letters[trg]); -} - -void HPYPLexicalTranslation::Increment(WordID src, WordID trg, MT19937* rng) { - tmodel->Increment(src, letters[trg], rng); -} - -void HPYPLexicalTranslation::Decrement(WordID src, WordID trg, MT19937* rng) { - tmodel->Decrement(src, letters[trg], rng); -} - diff --git a/gi/pf/hpyp_tm.h b/gi/pf/hpyp_tm.h deleted file mode 100644 index af3215ba..00000000 --- a/gi/pf/hpyp_tm.h +++ /dev/null @@ -1,38 +0,0 @@ -#ifndef HPYP_LEX_TRANS -#define HPYP_LEX_TRANS - -#include <vector> -#include "wordid.h" -#include "prob.h" -#include "sampler.h" -#include "freqdict.h" -#include "poisson_uniform_word_model.h" - -struct FreqBinner; -template <class B> struct PYPWordModel; -template <typename T, class B> struct ConditionalPYPWordModel; - -struct HPYPLexicalTranslation { - explicit HPYPLexicalTranslation(const std::vector<std::vector<WordID> >& lets, - const unsigned vocab_size, - const unsigned num_letters); - - prob_t Likelihood() const; - - void ResampleHyperparameters(MT19937* rng); - prob_t Prob(WordID src, WordID trg) const; // return p(trg | src) - void Summary() const; - void Increment(WordID src, WordID trg, MT19937* rng); - void Decrement(WordID src, WordID trg, MT19937* rng); - unsigned UniqueConditioningContexts() const; - - private: - const std::vector<std::vector<WordID> >& letters; // spelling dictionary - PoissonUniformWordModel base; // "generator" of English types - PYPWordModel<PoissonUniformWordModel>* up0; // model English lexicon - ConditionalPYPWordModel<PYPWordModel<PoissonUniformWordModel>, FreqBinner>* tmodel; // translation distributions - // (model English word | French word) - const WordID kX; -}; - -#endif diff --git a/gi/pf/itg.cc b/gi/pf/itg.cc deleted file mode 100644 index 29ec3860..00000000 --- a/gi/pf/itg.cc +++ /dev/null @@ -1,275 +0,0 @@ -#include <iostream> -#include <tr1/memory> -#include <queue> - -#include <boost/functional.hpp> -#include <boost/program_options.hpp> -#include <boost/program_options/variables_map.hpp> - -#include "viterbi.h" -#include "hg.h" -#include "trule.h" -#include "tdict.h" -#include "filelib.h" -#include "dict.h" -#include "sampler.h" -#include "ccrp_nt.h" -#include "ccrp_onetable.h" - -using namespace std; -using namespace tr1; -namespace po = boost::program_options; - -ostream& operator<<(ostream& os, const vector<WordID>& p) { - os << '['; - for (int i = 0; i < p.size(); ++i) - os << (i==0 ? "" : " ") << TD::Convert(p[i]); - return os << ']'; -} - -struct UnigramModel { - explicit UnigramModel(const string& fname, unsigned vocab_size, double p0null = 0.05) : - use_uniform_(fname.size() == 0), - p0null_(p0null), - uniform_((1.0 - p0null) / vocab_size), - probs_(TD::NumWords() + 1) { - if (fname.size() > 0) LoadUnigrams(fname); - probs_[0] = p0null_; - } - -// -// \data\ -// ngram 1=9295 -// -// \1-grams: -// -3.191193 " - - void LoadUnigrams(const string& fname) { - cerr << "Loading unigram probabilities from " << fname << " ..." << endl; - ReadFile rf(fname); - string line; - istream& in = *rf.stream(); - assert(in); - getline(in, line); - assert(line.empty()); - getline(in, line); - assert(line == "\\data\\"); - getline(in, line); - size_t pos = line.find("ngram 1="); - assert(pos == 0); - assert(line.size() > 8); - const size_t num_unigrams = atoi(&line[8]); - getline(in, line); - assert(line.empty()); - getline(in, line); - assert(line == "\\1-grams:"); - for (size_t i = 0; i < num_unigrams; ++i) { - getline(in, line); - assert(line.size() > 0); - pos = line.find('\t'); - assert(pos > 0); - assert(pos + 1 < line.size()); - const WordID w = TD::Convert(line.substr(pos + 1)); - line[pos] = 0; - float p = atof(&line[0]); - const prob_t pnon_null(1.0 - p0null_.as_float()); - if (w < probs_.size()) probs_[w].logeq(p * log(10) + log(pnon_null)); else abort(); - } - } - - const prob_t& operator()(const WordID& w) const { - if (!w) return p0null_; - if (use_uniform_) return uniform_; - return probs_[w]; - } - - const bool use_uniform_; - const prob_t p0null_; - const prob_t uniform_; - vector<prob_t> probs_; -}; - -struct Model1 { - explicit Model1(const string& fname) : - kNULL(TD::Convert("<eps>")), - kZERO() { - LoadModel1(fname); - } - - void LoadModel1(const string& fname) { - cerr << "Loading Model 1 parameters from " << fname << " ..." << endl; - ReadFile rf(fname); - istream& in = *rf.stream(); - string line; - unsigned lc = 0; - while(getline(in, line)) { - ++lc; - int cur = 0; - int start = 0; - while(cur < line.size() && line[cur] != ' ') { ++cur; } - assert(cur != line.size()); - line[cur] = 0; - const WordID src = TD::Convert(&line[0]); - ++cur; - start = cur; - while(cur < line.size() && line[cur] != ' ') { ++cur; } - assert(cur != line.size()); - line[cur] = 0; - WordID trg = TD::Convert(&line[start]); - const double logprob = strtod(&line[cur + 1], NULL); - if (src >= ttable.size()) ttable.resize(src + 1); - ttable[src][trg].logeq(logprob); - } - cerr << " read " << lc << " parameters.\n"; - } - - // returns prob 0 if src or trg is not found! - const prob_t& operator()(WordID src, WordID trg) const { - if (src == 0) src = kNULL; - if (src < ttable.size()) { - const map<WordID, prob_t>& cpd = ttable[src]; - const map<WordID, prob_t>::const_iterator it = cpd.find(trg); - if (it != cpd.end()) - return it->second; - } - return kZERO; - } - - const WordID kNULL; - const prob_t kZERO; - vector<map<WordID, prob_t> > ttable; -}; - -void InitCommandLine(int argc, char** argv, po::variables_map* conf) { - po::options_description opts("Configuration options"); - opts.add_options() - ("samples,s",po::value<unsigned>()->default_value(1000),"Number of samples") - ("particles,p",po::value<unsigned>()->default_value(25),"Number of particles") - ("input,i",po::value<string>(),"Read parallel data from") - ("model1,m",po::value<string>(),"Model 1 parameters (used in base distribution)") - ("inverse_model1,M",po::value<string>(),"Inverse Model 1 parameters (used in backward estimate)") - ("model1_interpolation_weight",po::value<double>()->default_value(0.95),"Mixing proportion of model 1 with uniform target distribution") - ("src_unigram,u",po::value<string>()->default_value(""),"Source unigram distribution; empty for uniform") - ("trg_unigram,U",po::value<string>()->default_value(""),"Target unigram distribution; empty for uniform") - ("random_seed,S",po::value<uint32_t>(), "Random seed"); - po::options_description clo("Command line options"); - clo.add_options() - ("config", po::value<string>(), "Configuration file") - ("help,h", "Print this help message and exit"); - po::options_description dconfig_options, dcmdline_options; - dconfig_options.add(opts); - dcmdline_options.add(opts).add(clo); - - po::store(parse_command_line(argc, argv, dcmdline_options), *conf); - if (conf->count("config")) { - ifstream config((*conf)["config"].as<string>().c_str()); - po::store(po::parse_config_file(config, dconfig_options), *conf); - } - po::notify(*conf); - - if (conf->count("help") || (conf->count("input") == 0)) { - cerr << dcmdline_options << endl; - exit(1); - } -} - -void ReadParallelCorpus(const string& filename, - vector<vector<WordID> >* f, - vector<vector<WordID> >* e, - set<WordID>* vocab_f, - set<WordID>* vocab_e) { - f->clear(); - e->clear(); - vocab_f->clear(); - vocab_e->clear(); - istream* in; - if (filename == "-") - in = &cin; - else - in = new ifstream(filename.c_str()); - assert(*in); - string line; - const WordID kDIV = TD::Convert("|||"); - vector<WordID> tmp; - while(*in) { - getline(*in, line); - if (line.empty() && !*in) break; - e->push_back(vector<int>()); - f->push_back(vector<int>()); - vector<int>& le = e->back(); - vector<int>& lf = f->back(); - tmp.clear(); - TD::ConvertSentence(line, &tmp); - bool isf = true; - for (unsigned i = 0; i < tmp.size(); ++i) { - const int cur = tmp[i]; - if (isf) { - if (kDIV == cur) { isf = false; } else { - lf.push_back(cur); - vocab_f->insert(cur); - } - } else { - assert(cur != kDIV); - le.push_back(cur); - vocab_e->insert(cur); - } - } - assert(isf == false); - } - if (in != &cin) delete in; -} - -int main(int argc, char** argv) { - po::variables_map conf; - InitCommandLine(argc, argv, &conf); - const unsigned particles = conf["particles"].as<unsigned>(); - const unsigned samples = conf["samples"].as<unsigned>(); - TD::Convert("<s>"); - TD::Convert("</s>"); - TD::Convert("<unk>"); - if (!conf.count("model1")) { - cerr << argv[0] << "Please use --model1 to specify model 1 parameters\n"; - return 1; - } - boost::shared_ptr<MT19937> prng; - if (conf.count("random_seed")) - prng.reset(new MT19937(conf["random_seed"].as<uint32_t>())); - else - prng.reset(new MT19937); - MT19937& rng = *prng; - - vector<vector<WordID> > corpuse, corpusf; - set<WordID> vocabe, vocabf; - cerr << "Reading corpus...\n"; - ReadParallelCorpus(conf["input"].as<string>(), &corpusf, &corpuse, &vocabf, &vocabe); - cerr << "F-corpus size: " << corpusf.size() << " sentences\t (" << vocabf.size() << " word types)\n"; - cerr << "E-corpus size: " << corpuse.size() << " sentences\t (" << vocabe.size() << " word types)\n"; - assert(corpusf.size() == corpuse.size()); - UnigramModel src_unigram(conf["src_unigram"].as<string>(), vocabf.size()); - UnigramModel trg_unigram(conf["trg_unigram"].as<string>(), vocabe.size()); - const prob_t kHALF(0.5); - - const string kEMPTY = "NULL"; - const int kLHS = -TD::Convert("X"); - Model1 m1(conf["model1"].as<string>()); - Model1 invm1(conf["inverse_model1"].as<string>()); - for (int si = 0; si < conf["samples"].as<unsigned>(); ++si) { - cerr << '.' << flush; - for (int ci = 0; ci < corpusf.size(); ++ci) { - const vector<WordID>& trg = corpuse[ci]; - const vector<WordID>& src = corpusf[ci]; - for (int i = 0; i <= trg.size(); ++i) { - const WordID e_i = i > 0 ? trg[i-1] : 0; - for (int j = 0; j <= src.size(); ++j) { - const WordID f_j = j > 0 ? src[j-1] : 0; - if (e_i == 0 && f_j == 0) continue; - prob_t je = kHALF * src_unigram(f_j) * m1(f_j,e_i) + kHALF * trg_unigram(e_i) * invm1(e_i,f_j); - cerr << "p( " << (e_i ? TD::Convert(e_i) : kEMPTY) << " , " << (f_j ? TD::Convert(f_j) : kEMPTY) << " ) = " << log(je) << endl; - if (e_i && f_j) - cout << "[X] ||| " << TD::Convert(f_j) << " ||| " << TD::Convert(e_i) << " ||| LogProb=" << log(je) << endl; - } - } - } - } -} - diff --git a/gi/pf/learn_cfg.cc b/gi/pf/learn_cfg.cc deleted file mode 100644 index 1d5126e4..00000000 --- a/gi/pf/learn_cfg.cc +++ /dev/null @@ -1,428 +0,0 @@ -#include <iostream> -#include <tr1/memory> -#include <queue> - -#include <boost/functional.hpp> -#include <boost/program_options.hpp> -#include <boost/program_options/variables_map.hpp> - -#include "inside_outside.h" -#include "hg.h" -#include "bottom_up_parser.h" -#include "fdict.h" -#include "grammar.h" -#include "m.h" -#include "trule.h" -#include "tdict.h" -#include "filelib.h" -#include "dict.h" -#include "sampler.h" -#include "ccrp.h" -#include "ccrp_onetable.h" - -using namespace std; -using namespace tr1; -namespace po = boost::program_options; - -boost::shared_ptr<MT19937> prng; -vector<int> nt_vocab; -vector<int> nt_id_to_index; -static unsigned kMAX_RULE_SIZE = 0; -static unsigned kMAX_ARITY = 0; -static bool kALLOW_MIXED = true; // allow rules with mixed terminals and NTs -static bool kHIERARCHICAL_PRIOR = false; - -void InitCommandLine(int argc, char** argv, po::variables_map* conf) { - po::options_description opts("Configuration options"); - opts.add_options() - ("samples,s",po::value<unsigned>()->default_value(1000),"Number of samples") - ("input,i",po::value<string>(),"Read parallel data from") - ("max_rule_size,m", po::value<unsigned>()->default_value(0), "Maximum rule size (0 for unlimited)") - ("max_arity,a", po::value<unsigned>()->default_value(0), "Maximum number of nonterminals in a rule (0 for unlimited)") - ("no_mixed_rules,M", "Do not mix terminals and nonterminals in a rule RHS") - ("nonterminals,n", po::value<unsigned>()->default_value(1), "Size of nonterminal vocabulary") - ("hierarchical_prior,h", "Use hierarchical prior") - ("random_seed,S",po::value<uint32_t>(), "Random seed"); - po::options_description clo("Command line options"); - clo.add_options() - ("config", po::value<string>(), "Configuration file") - ("help", "Print this help message and exit"); - po::options_description dconfig_options, dcmdline_options; - dconfig_options.add(opts); - dcmdline_options.add(opts).add(clo); - - po::store(parse_command_line(argc, argv, dcmdline_options), *conf); - if (conf->count("config")) { - ifstream config((*conf)["config"].as<string>().c_str()); - po::store(po::parse_config_file(config, dconfig_options), *conf); - } - po::notify(*conf); - - if (conf->count("help") || (conf->count("input") == 0)) { - cerr << dcmdline_options << endl; - exit(1); - } -} - -unsigned ReadCorpus(const string& filename, - vector<vector<WordID> >* e, - set<WordID>* vocab_e) { - e->clear(); - vocab_e->clear(); - istream* in; - if (filename == "-") - in = &cin; - else - in = new ifstream(filename.c_str()); - assert(*in); - string line; - unsigned toks = 0; - while(*in) { - getline(*in, line); - if (line.empty() && !*in) break; - e->push_back(vector<int>()); - vector<int>& le = e->back(); - TD::ConvertSentence(line, &le); - for (unsigned i = 0; i < le.size(); ++i) - vocab_e->insert(le[i]); - toks += le.size(); - } - if (in != &cin) delete in; - return toks; -} - -struct Grid { - // a b c d e - // 0 - 0 - - - vector<int> grid; -}; - -struct BaseRuleModel { - explicit BaseRuleModel(unsigned term_size, - unsigned nonterm_size = 1) : - unif_term(1.0 / term_size), - unif_nonterm(1.0 / nonterm_size) {} - prob_t operator()(const TRule& r) const { - prob_t p; p.logeq(Md::log_poisson(1.0, r.f_.size())); - const prob_t term_prob((2.0 + 0.01*r.f_.size()) / (r.f_.size() + 2)); - const prob_t nonterm_prob(1.0 - term_prob.as_float()); - for (unsigned i = 0; i < r.f_.size(); ++i) { - if (r.f_[i] <= 0) { // nonterminal - if (kALLOW_MIXED) p *= nonterm_prob; - p *= unif_nonterm; - } else { // terminal - if (kALLOW_MIXED) p *= term_prob; - p *= unif_term; - } - } - return p; - } - const prob_t unif_term, unif_nonterm; -}; - -struct HieroLMModel { - explicit HieroLMModel(unsigned vocab_size, unsigned num_nts = 1) : - base(vocab_size, num_nts), - q0(1,1,1,1), - nts(num_nts, CCRP<TRule>(1,1,1,1)) {} - - prob_t Prob(const TRule& r) const { - return nts[nt_id_to_index[-r.lhs_]].prob(r, p0(r)); - } - - inline prob_t p0(const TRule& r) const { - if (kHIERARCHICAL_PRIOR) - return q0.prob(r, base(r)); - else - return base(r); - } - - int Increment(const TRule& r, MT19937* rng) { - const int delta = nts[nt_id_to_index[-r.lhs_]].increment(r, p0(r), rng); - if (kHIERARCHICAL_PRIOR && delta) - q0.increment(r, base(r), rng); - return delta; - // return x.increment(r); - } - - int Decrement(const TRule& r, MT19937* rng) { - const int delta = nts[nt_id_to_index[-r.lhs_]].decrement(r, rng); - if (kHIERARCHICAL_PRIOR && delta) - q0.decrement(r, rng); - return delta; - //return x.decrement(r); - } - - prob_t Likelihood() const { - prob_t p = prob_t::One(); - for (unsigned i = 0; i < nts.size(); ++i) { - prob_t q; q.logeq(nts[i].log_crp_prob()); - p *= q; - for (CCRP<TRule>::const_iterator it = nts[i].begin(); it != nts[i].end(); ++it) { - prob_t tp = p0(it->first); - tp.poweq(it->second.num_tables()); - p *= tp; - } - } - if (kHIERARCHICAL_PRIOR) { - prob_t q; q.logeq(q0.log_crp_prob()); - p *= q; - for (CCRP<TRule>::const_iterator it = q0.begin(); it != q0.end(); ++it) { - prob_t tp = base(it->first); - tp.poweq(it->second.num_tables()); - p *= tp; - } - } - //for (CCRP_OneTable<TRule>::const_iterator it = x.begin(); it != x.end(); ++it) - // p *= base(it->first); - return p; - } - - void ResampleHyperparameters(MT19937* rng) { - for (unsigned i = 0; i < nts.size(); ++i) - nts[i].resample_hyperparameters(rng); - if (kHIERARCHICAL_PRIOR) { - q0.resample_hyperparameters(rng); - cerr << "[base d=" << q0.discount() << ", s=" << q0.strength() << "]"; - } - cerr << " d=" << nts[0].discount() << ", s=" << nts[0].strength() << endl; - } - - const BaseRuleModel base; - CCRP<TRule> q0; - vector<CCRP<TRule> > nts; - //CCRP_OneTable<TRule> x; -}; - -vector<GrammarIter* > tofreelist; - -HieroLMModel* plm; - -struct NPGrammarIter : public GrammarIter, public RuleBin { - NPGrammarIter() : arity() { tofreelist.push_back(this); } - NPGrammarIter(const TRulePtr& inr, const int a, int symbol) : arity(a) { - if (inr) { - r.reset(new TRule(*inr)); - } else { - r.reset(new TRule); - } - TRule& rr = *r; - rr.lhs_ = nt_vocab[0]; - rr.f_.push_back(symbol); - rr.e_.push_back(symbol < 0 ? (1-int(arity)) : symbol); - tofreelist.push_back(this); - } - inline static unsigned NextArity(int cur_a, int symbol) { - return cur_a + (symbol <= 0 ? 1 : 0); - } - virtual int GetNumRules() const { - if (r) return nt_vocab.size(); else return 0; - } - virtual TRulePtr GetIthRule(int i) const { - if (i == 0) return r; - TRulePtr nr(new TRule(*r)); - nr->lhs_ = nt_vocab[i]; - return nr; - } - virtual int Arity() const { - return arity; - } - virtual const RuleBin* GetRules() const { - if (!r) return NULL; else return this; - } - virtual const GrammarIter* Extend(int symbol) const { - const int next_arity = NextArity(arity, symbol); - if (kMAX_ARITY && next_arity > kMAX_ARITY) - return NULL; - if (!kALLOW_MIXED && r) { - bool t1 = r->f_.front() <= 0; - bool t2 = symbol <= 0; - if (t1 != t2) return NULL; - } - if (!kMAX_RULE_SIZE || !r || (r->f_.size() < kMAX_RULE_SIZE)) - return new NPGrammarIter(r, next_arity, symbol); - else - return NULL; - } - const unsigned char arity; - TRulePtr r; -}; - -struct NPGrammar : public Grammar { - virtual const GrammarIter* GetRoot() const { - return new NPGrammarIter; - } -}; - -prob_t TotalProb(const Hypergraph& hg) { - return Inside<prob_t, EdgeProb>(hg); -} - -void SampleDerivation(const Hypergraph& hg, MT19937* rng, vector<unsigned>* sampled_deriv) { - vector<prob_t> node_probs; - Inside<prob_t, EdgeProb>(hg, &node_probs); - queue<unsigned> q; - q.push(hg.nodes_.size() - 2); - while(!q.empty()) { - unsigned cur_node_id = q.front(); -// cerr << "NODE=" << cur_node_id << endl; - q.pop(); - const Hypergraph::Node& node = hg.nodes_[cur_node_id]; - const unsigned num_in_edges = node.in_edges_.size(); - unsigned sampled_edge = 0; - if (num_in_edges == 1) { - sampled_edge = node.in_edges_[0]; - } else { - //prob_t z; - assert(num_in_edges > 1); - SampleSet<prob_t> ss; - for (unsigned j = 0; j < num_in_edges; ++j) { - const Hypergraph::Edge& edge = hg.edges_[node.in_edges_[j]]; - prob_t p = edge.edge_prob_; - for (unsigned k = 0; k < edge.tail_nodes_.size(); ++k) - p *= node_probs[edge.tail_nodes_[k]]; - ss.add(p); -// cerr << log(ss[j]) << " ||| " << edge.rule_->AsString() << endl; - //z += p; - } -// for (unsigned j = 0; j < num_in_edges; ++j) { -// const Hypergraph::Edge& edge = hg.edges_[node.in_edges_[j]]; -// cerr << exp(log(ss[j] / z)) << " ||| " << edge.rule_->AsString() << endl; -// } -// cerr << " --- \n"; - sampled_edge = node.in_edges_[rng->SelectSample(ss)]; - } - sampled_deriv->push_back(sampled_edge); - const Hypergraph::Edge& edge = hg.edges_[sampled_edge]; - for (unsigned j = 0; j < edge.tail_nodes_.size(); ++j) { - q.push(edge.tail_nodes_[j]); - } - } - for (unsigned i = 0; i < sampled_deriv->size(); ++i) { - cerr << *hg.edges_[(*sampled_deriv)[i]].rule_ << endl; - } -} - -void IncrementDerivation(const Hypergraph& hg, const vector<unsigned>& d, HieroLMModel* plm, MT19937* rng) { - for (unsigned i = 0; i < d.size(); ++i) - plm->Increment(*hg.edges_[d[i]].rule_, rng); -} - -void DecrementDerivation(const Hypergraph& hg, const vector<unsigned>& d, HieroLMModel* plm, MT19937* rng) { - for (unsigned i = 0; i < d.size(); ++i) - plm->Decrement(*hg.edges_[d[i]].rule_, rng); -} - -int main(int argc, char** argv) { - po::variables_map conf; - - InitCommandLine(argc, argv, &conf); - nt_vocab.resize(conf["nonterminals"].as<unsigned>()); - assert(nt_vocab.size() > 0); - assert(nt_vocab.size() < 26); - { - string nt = "X"; - for (unsigned i = 0; i < nt_vocab.size(); ++i) { - if (nt_vocab.size() > 1) nt[0] = ('A' + i); - int pid = TD::Convert(nt); - nt_vocab[i] = -pid; - if (pid >= nt_id_to_index.size()) { - nt_id_to_index.resize(pid + 1, -1); - } - nt_id_to_index[pid] = i; - } - } - vector<GrammarPtr> grammars; - grammars.push_back(GrammarPtr(new NPGrammar)); - - const unsigned samples = conf["samples"].as<unsigned>(); - kMAX_RULE_SIZE = conf["max_rule_size"].as<unsigned>(); - if (kMAX_RULE_SIZE == 1) { - cerr << "Invalid maximum rule size: must be 0 or >1\n"; - return 1; - } - kMAX_ARITY = conf["max_arity"].as<unsigned>(); - if (kMAX_ARITY == 1) { - cerr << "Invalid maximum arity: must be 0 or >1\n"; - return 1; - } - kALLOW_MIXED = !conf.count("no_mixed_rules"); - - kHIERARCHICAL_PRIOR = conf.count("hierarchical_prior"); - - if (conf.count("random_seed")) - prng.reset(new MT19937(conf["random_seed"].as<uint32_t>())); - else - prng.reset(new MT19937); - MT19937& rng = *prng; - vector<vector<WordID> > corpuse; - set<WordID> vocabe; - cerr << "Reading corpus...\n"; - const unsigned toks = ReadCorpus(conf["input"].as<string>(), &corpuse, &vocabe); - cerr << "E-corpus size: " << corpuse.size() << " sentences\t (" << vocabe.size() << " word types)\n"; - HieroLMModel lm(vocabe.size(), nt_vocab.size()); - - plm = &lm; - ExhaustiveBottomUpParser parser(TD::Convert(-nt_vocab[0]), grammars); - - Hypergraph hg; - const int kGoal = -TD::Convert("Goal"); - const int kLP = FD::Convert("LogProb"); - SparseVector<double> v; v.set_value(kLP, 1.0); - vector<vector<unsigned> > derivs(corpuse.size()); - vector<Lattice> cl(corpuse.size()); - for (int ci = 0; ci < corpuse.size(); ++ci) { - vector<int>& src = corpuse[ci]; - Lattice& lat = cl[ci]; - lat.resize(src.size()); - for (unsigned i = 0; i < src.size(); ++i) - lat[i].push_back(LatticeArc(src[i], 0.0, 1)); - } - for (int SS=0; SS < samples; ++SS) { - const bool is_last = ((samples - 1) == SS); - prob_t dlh = prob_t::One(); - for (int ci = 0; ci < corpuse.size(); ++ci) { - const vector<int>& src = corpuse[ci]; - const Lattice& lat = cl[ci]; - cerr << TD::GetString(src) << endl; - hg.clear(); - parser.Parse(lat, &hg); // exhaustive parse - vector<unsigned>& d = derivs[ci]; - if (!is_last) DecrementDerivation(hg, d, &lm, &rng); - for (unsigned i = 0; i < hg.edges_.size(); ++i) { - TRule& r = *hg.edges_[i].rule_; - if (r.lhs_ == kGoal) - hg.edges_[i].edge_prob_ = prob_t::One(); - else - hg.edges_[i].edge_prob_ = lm.Prob(r); - } - if (!is_last) { - d.clear(); - SampleDerivation(hg, &rng, &d); - IncrementDerivation(hg, derivs[ci], &lm, &rng); - } else { - prob_t p = TotalProb(hg); - dlh *= p; - cerr << " p(sentence) = " << log(p) << "\t" << log(dlh) << endl; - } - if (tofreelist.size() > 200000) { - cerr << "Freeing ... "; - for (unsigned i = 0; i < tofreelist.size(); ++i) - delete tofreelist[i]; - tofreelist.clear(); - cerr << "Freed.\n"; - } - } - double llh = log(lm.Likelihood()); - cerr << "LLH=" << llh << "\tENTROPY=" << (-llh / log(2) / toks) << "\tPPL=" << pow(2, -llh / log(2) / toks) << endl; - if (SS % 10 == 9) lm.ResampleHyperparameters(&rng); - if (is_last) { - double z = log(dlh); - cerr << "TOTAL_PROB=" << z << "\tENTROPY=" << (-z / log(2) / toks) << "\tPPL=" << pow(2, -z / log(2) / toks) << endl; - } - } - for (unsigned i = 0; i < nt_vocab.size(); ++i) - cerr << lm.nts[i] << endl; - return 0; -} - diff --git a/gi/pf/make-freq-bins.pl b/gi/pf/make-freq-bins.pl deleted file mode 100755 index fdcd3555..00000000 --- a/gi/pf/make-freq-bins.pl +++ /dev/null @@ -1,26 +0,0 @@ -#!/usr/bin/perl -w -use strict; - -my $BASE = 6; -my $CUTOFF = 3; - -my %d; -my $num = 0; -while(<>){ - chomp; - my @words = split /\s+/; - for my $w (@words) {$d{$w}++; $num++;} -} - -my @vocab = sort {$d{$b} <=> $d{$a}} keys %d; - -for (my $i=0; $i<scalar @vocab; $i++) { - my $most = $d{$vocab[$i]}; - my $least = 1; - - my $nl = -int(log($most / $num) / log($BASE) + $CUTOFF); - if ($nl < 0) { $nl = 0; } - print "$vocab[$i] $nl\n" -} - - diff --git a/gi/pf/mh_test.cc b/gi/pf/mh_test.cc deleted file mode 100644 index 296e7285..00000000 --- a/gi/pf/mh_test.cc +++ /dev/null @@ -1,148 +0,0 @@ -#include "ccrp.h" - -#include <vector> -#include <iostream> - -#include "tdict.h" -#include "transliterations.h" - -using namespace std; - -MT19937 rng; - -static bool verbose = false; - -struct Model { - - Model() : bp(), base(0.2, 0.6) , ccrps(5, CCRP<int>(0.8, 0.5)) {} - - double p0(int x) const { - assert(x > 0); - assert(x < 5); - return 1.0/4.0; - } - - double llh() const { - double lh = bp + base.log_crp_prob(); - for (int ctx = 1; ctx < 5; ++ctx) - lh += ccrps[ctx].log_crp_prob(); - return lh; - } - - double prob(int ctx, int x) const { - assert(ctx > 0 && ctx < 5); - return ccrps[ctx].prob(x, base.prob(x, p0(x))); - } - - void increment(int ctx, int x) { - assert(ctx > 0 && ctx < 5); - if (ccrps[ctx].increment(x, base.prob(x, p0(x)), &rng)) { - if (base.increment(x, p0(x), &rng)) { - bp += log(1.0 / 4.0); - } - } - } - - // this is just a biased estimate - double est_base_prob(int x) { - return (x + 1) * x / 40.0; - } - - void increment_is(int ctx, int x) { - assert(ctx > 0 && ctx < 5); - SampleSet<double> ss; - const int PARTICLES = 25; - vector<CCRP<int> > s1s(PARTICLES, CCRP<int>(0.5,0.5)); - vector<CCRP<int> > sbs(PARTICLES, CCRP<int>(0.5,0.5)); - vector<double> sp0s(PARTICLES); - - CCRP<int> s1 = ccrps[ctx]; - CCRP<int> sb = base; - double sp0 = bp; - for (int pp = 0; pp < PARTICLES; ++pp) { - if (pp > 0) { - ccrps[ctx] = s1; - base = sb; - bp = sp0; - } - - double q = 1; - double gamma = 1; - double est_p = est_base_prob(x); - //base.prob(x, p0(x)) + rng.next() * 0.1; - if (ccrps[ctx].increment(x, est_p, &rng, &q)) { - gamma = q * base.prob(x, p0(x)); - q *= est_p; - if (verbose) cerr << "(DP-base draw) "; - double qq = -1; - if (base.increment(x, p0(x), &rng, &qq)) { - if (verbose) cerr << "(G0 draw) "; - bp += log(p0(x)); - qq *= p0(x); - } - } else { gamma = q; } - double w = gamma / q; - if (verbose) - cerr << "gamma=" << gamma << " q=" << q << "\tw=" << w << endl; - ss.add(w); - s1s[pp] = ccrps[ctx]; - sbs[pp] = base; - sp0s[pp] = bp; - } - int ps = rng.SelectSample(ss); - ccrps[ctx] = s1s[ps]; - base = sbs[ps]; - bp = sp0s[ps]; - if (verbose) { - cerr << "SELECTED: " << ps << endl; - static int cc = 0; cc++; if (cc ==10) exit(1); - } - } - - void decrement(int ctx, int x) { - assert(ctx > 0 && ctx < 5); - if (ccrps[ctx].decrement(x, &rng)) { - if (base.decrement(x, &rng)) { - bp -= log(p0(x)); - } - } - } - - double bp; - CCRP<int> base; - vector<CCRP<int> > ccrps; - -}; - -int main(int argc, char** argv) { - if (argc > 1) { verbose = true; } - vector<int> counts(15, 0); - vector<int> tcounts(15, 0); - int points[] = {1,2, 2,2, 3,2, 4,1, 3, 4, 3, 3, 2, 3, 4, 1, 4, 1, 3, 2, 1, 3, 1, 4, 0, 0}; - double tlh = 0; - double tt = 0; - for (int n = 0; n < 1000; ++n) { - if (n % 10 == 0) cerr << '.'; - if ((n+1) % 400 == 0) cerr << " [" << (n+1) << "]\n"; - Model m; - for (int *x = points; *x; x += 2) - m.increment(x[0], x[1]); - - for (int j = 0; j < 24; ++j) { - for (int *x = points; *x; x += 2) { - if (rng.next() < 0.8) { - m.decrement(x[0], x[1]); - m.increment_is(x[0], x[1]); - } - } - } - counts[m.base.num_customers()]++; - tcounts[m.base.num_tables()]++; - tlh += m.llh(); - tt += 1.0; - } - cerr << "mean LLH = " << (tlh / tt) << endl; - for (int i = 0; i < 15; ++i) - cerr << i << ": " << (counts[i] / tt) << "\t" << (tcounts[i] / tt) << endl; -} - diff --git a/gi/pf/monotonic_pseg.h b/gi/pf/monotonic_pseg.h deleted file mode 100644 index 10d171fe..00000000 --- a/gi/pf/monotonic_pseg.h +++ /dev/null @@ -1,89 +0,0 @@ -#ifndef _MONOTONIC_PSEG_H_ -#define _MONOTONIC_PSEG_H_ - -#include <vector> - -#include "prob.h" -#include "ccrp_nt.h" -#include "trule.h" -#include "base_distributions.h" - -template <typename BaseMeasure> -struct MonotonicParallelSegementationModel { - explicit MonotonicParallelSegementationModel(BaseMeasure& rcp0) : - rp0(rcp0), base(prob_t::One()), rules(1,1), stop(1.0) {} - - void DecrementRule(const TRule& rule) { - if (rules.decrement(rule)) - base /= rp0(rule); - } - - void IncrementRule(const TRule& rule) { - if (rules.increment(rule)) - base *= rp0(rule); - } - - void IncrementRulesAndStops(const std::vector<TRulePtr>& rules) { - for (int i = 0; i < rules.size(); ++i) - IncrementRule(*rules[i]); - if (rules.size()) IncrementContinue(rules.size() - 1); - IncrementStop(); - } - - void DecrementRulesAndStops(const std::vector<TRulePtr>& rules) { - for (int i = 0; i < rules.size(); ++i) - DecrementRule(*rules[i]); - if (rules.size()) { - DecrementContinue(rules.size() - 1); - DecrementStop(); - } - } - - prob_t RuleProbability(const TRule& rule) const { - prob_t p; p.logeq(rules.logprob(rule, log(rp0(rule)))); - return p; - } - - prob_t Likelihood() const { - prob_t p = base; - prob_t q; q.logeq(rules.log_crp_prob()); - p *= q; - q.logeq(stop.log_crp_prob()); - p *= q; - return p; - } - - void IncrementStop() { - stop.increment(true); - } - - void IncrementContinue(int n = 1) { - for (int i = 0; i < n; ++i) - stop.increment(false); - } - - void DecrementStop() { - stop.decrement(true); - } - - void DecrementContinue(int n = 1) { - for (int i = 0; i < n; ++i) - stop.decrement(false); - } - - prob_t StopProbability() const { - return prob_t(stop.prob(true, 0.5)); - } - - prob_t ContinueProbability() const { - return prob_t(stop.prob(false, 0.5)); - } - - const BaseMeasure& rp0; - prob_t base; - CCRP_NoTable<TRule> rules; - CCRP_NoTable<bool> stop; -}; - -#endif - diff --git a/gi/pf/ngram_base.cc b/gi/pf/ngram_base.cc deleted file mode 100644 index 1299f06f..00000000 --- a/gi/pf/ngram_base.cc +++ /dev/null @@ -1,69 +0,0 @@ -#include "ngram_base.h" - -#include "lm/model.hh" -#include "tdict.h" - -using namespace std; - -namespace { -struct GICSVMapper : public lm::EnumerateVocab { - GICSVMapper(vector<lm::WordIndex>* out) : out_(out), kLM_UNKNOWN_TOKEN(0) { out_->clear(); } - 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, kLM_UNKNOWN_TOKEN); - (*out_)[cdec_id] = index; - } - vector<lm::WordIndex>* out_; - const lm::WordIndex kLM_UNKNOWN_TOKEN; -}; -} - -struct FixedNgramBaseImpl { - FixedNgramBaseImpl(const string& param) { - GICSVMapper vm(&cdec2klm_map_); - lm::ngram::Config conf; - conf.enumerate_vocab = &vm; - cerr << "Reading character LM from " << param << endl; - model = new lm::ngram::ProbingModel(param.c_str(), conf); - order = model->Order(); - kEOS = MapWord(TD::Convert("</s>")); - assert(kEOS > 0); - } - - lm::WordIndex MapWord(const WordID w) const { - if (w < cdec2klm_map_.size()) return cdec2klm_map_[w]; - return 0; - } - - ~FixedNgramBaseImpl() { delete model; } - - prob_t StringProbability(const vector<WordID>& s) const { - lm::ngram::State state = model->BeginSentenceState(); - double prob = 0; - for (unsigned i = 0; i < s.size(); ++i) { - const lm::ngram::State scopy(state); - prob += model->Score(scopy, MapWord(s[i]), state); - } - const lm::ngram::State scopy(state); - prob += model->Score(scopy, kEOS, state); - prob_t p; p.logeq(prob * log(10)); - return p; - } - - lm::ngram::ProbingModel* model; - unsigned order; - vector<lm::WordIndex> cdec2klm_map_; - lm::WordIndex kEOS; -}; - -FixedNgramBase::~FixedNgramBase() { delete impl; } - -FixedNgramBase::FixedNgramBase(const string& lmfname) { - impl = new FixedNgramBaseImpl(lmfname); -} - -prob_t FixedNgramBase::StringProbability(const vector<WordID>& s) const { - return impl->StringProbability(s); -} - diff --git a/gi/pf/ngram_base.h b/gi/pf/ngram_base.h deleted file mode 100644 index 4ea999f3..00000000 --- a/gi/pf/ngram_base.h +++ /dev/null @@ -1,25 +0,0 @@ -#ifndef _NGRAM_BASE_H_ -#define _NGRAM_BASE_H_ - -#include <string> -#include <vector> -#include "trule.h" -#include "wordid.h" -#include "prob.h" - -struct FixedNgramBaseImpl; -struct FixedNgramBase { - FixedNgramBase(const std::string& lmfname); - ~FixedNgramBase(); - prob_t StringProbability(const std::vector<WordID>& s) const; - - prob_t operator()(const TRule& rule) const { - return StringProbability(rule.e_); - } - - private: - FixedNgramBaseImpl* impl; - -}; - -#endif diff --git a/gi/pf/nuisance_test.cc b/gi/pf/nuisance_test.cc deleted file mode 100644 index fc0af9cb..00000000 --- a/gi/pf/nuisance_test.cc +++ /dev/null @@ -1,161 +0,0 @@ -#include "ccrp.h" - -#include <vector> -#include <iostream> - -#include "tdict.h" -#include "transliterations.h" - -using namespace std; - -MT19937 rng; - -ostream& operator<<(ostream&os, const vector<int>& v) { - os << '[' << v[0]; - if (v.size() == 2) os << ' ' << v[1]; - return os << ']'; -} - -struct Base { - Base() : llh(), v(2), v1(1), v2(1), crp(0.25, 0.5) {} - inline double p0(const vector<int>& x) const { - double p = 0.75; - if (x.size() == 2) p = 0.25; - p *= 1.0 / 3.0; - if (x.size() == 2) p *= 1.0 / 3.0; - return p; - } - double est_deriv_prob(int a, int b, int seg) const { - assert(a > 0 && a < 4); // a \in {1,2,3} - assert(b > 0 && b < 4); // b \in {1,2,3} - assert(seg == 0 || seg == 1); // seg \in {0,1} - if (seg == 0) { - v[0] = a; - v[1] = b; - return crp.prob(v, p0(v)); - } else { - v1[0] = a; - v2[0] = b; - return crp.prob(v1, p0(v1)) * crp.prob(v2, p0(v2)); - } - } - double est_marginal_prob(int a, int b) const { - return est_deriv_prob(a,b,0) + est_deriv_prob(a,b,1); - } - int increment(int a, int b, double* pw = NULL) { - double p1 = est_deriv_prob(a, b, 0); - double p2 = est_deriv_prob(a, b, 1); - //p1 = 0.5; p2 = 0.5; - int seg = rng.SelectSample(p1,p2); - double tmp = 0; - if (!pw) pw = &tmp; - double& w = *pw; - if (seg == 0) { - v[0] = a; - v[1] = b; - w = crp.prob(v, p0(v)) / p1; - if (crp.increment(v, p0(v), &rng)) { - llh += log(p0(v)); - } - } else { - v1[0] = a; - w = crp.prob(v1, p0(v1)) / p2; - if (crp.increment(v1, p0(v1), &rng)) { - llh += log(p0(v1)); - } - v2[0] = b; - w *= crp.prob(v2, p0(v2)); - if (crp.increment(v2, p0(v2), &rng)) { - llh += log(p0(v2)); - } - } - return seg; - } - void increment(int a, int b, int seg) { - if (seg == 0) { - v[0] = a; - v[1] = b; - if (crp.increment(v, p0(v), &rng)) { - llh += log(p0(v)); - } - } else { - v1[0] = a; - if (crp.increment(v1, p0(v1), &rng)) { - llh += log(p0(v1)); - } - v2[0] = b; - if (crp.increment(v2, p0(v2), &rng)) { - llh += log(p0(v2)); - } - } - } - void decrement(int a, int b, int seg) { - if (seg == 0) { - v[0] = a; - v[1] = b; - if (crp.decrement(v, &rng)) { - llh -= log(p0(v)); - } - } else { - v1[0] = a; - if (crp.decrement(v1, &rng)) { - llh -= log(p0(v1)); - } - v2[0] = b; - if (crp.decrement(v2, &rng)) { - llh -= log(p0(v2)); - } - } - } - double log_likelihood() const { - return llh + crp.log_crp_prob(); - } - double llh; - mutable vector<int> v, v1, v2; - CCRP<vector<int> > crp; -}; - -int main(int argc, char** argv) { - double tl = 0; - const int ITERS = 1000; - const int PARTICLES = 20; - const int DATAPOINTS = 50; - WordID x = TD::Convert("souvenons"); - WordID y = TD::Convert("remember"); - vector<WordID> src; TD::ConvertSentence("s o u v e n o n s", &src); - vector<WordID> trg; TD::ConvertSentence("r e m e m b e r", &trg); -// Transliterations xx; -// xx.Initialize(x, src, y, trg); -// return 1; - - for (int j = 0; j < ITERS; ++j) { - Base b; - vector<int> segs(DATAPOINTS); - SampleSet<double> ss; - vector<int> sss; - for (int i = 0; i < DATAPOINTS; i++) { - ss.clear(); - sss.clear(); - int x = ((i / 10) % 3) + 1; - int y = (i % 3) + 1; - //double ep = b.est_marginal_prob(x,y); - //cerr << "est p(" << x << "," << y << ") = " << ep << endl; - for (int n = 0; n < PARTICLES; ++n) { - double w; - int seg = b.increment(x,y,&w); - //cerr << seg << " w=" << w << endl; - ss.add(w); - sss.push_back(seg); - b.decrement(x,y,seg); - } - int seg = sss[rng.SelectSample(ss)]; - b.increment(x, y, seg); - //cerr << "Selected: " << seg << endl; - //return 1; - segs[i] = seg; - } - tl += b.log_likelihood(); - } - cerr << "LLH=" << tl / ITERS << endl; -} - diff --git a/gi/pf/os_phrase.h b/gi/pf/os_phrase.h deleted file mode 100644 index dfe40cb1..00000000 --- a/gi/pf/os_phrase.h +++ /dev/null @@ -1,15 +0,0 @@ -#ifndef _OS_PHRASE_H_ -#define _OS_PHRASE_H_ - -#include <iostream> -#include <vector> -#include "tdict.h" - -inline std::ostream& operator<<(std::ostream& os, const std::vector<WordID>& p) { - os << '['; - for (int i = 0; i < p.size(); ++i) - os << (i==0 ? "" : " ") << TD::Convert(p[i]); - return os << ']'; -} - -#endif diff --git a/gi/pf/pf.h b/gi/pf/pf.h deleted file mode 100644 index ede7cda8..00000000 --- a/gi/pf/pf.h +++ /dev/null @@ -1,84 +0,0 @@ -#ifndef _PF_H_ -#define _PF_H_ - -#include <cassert> -#include <vector> -#include "sampler.h" -#include "prob.h" - -template <typename ParticleType> -struct ParticleRenormalizer { - void operator()(std::vector<ParticleType>* pv) const { - if (pv->empty()) return; - prob_t z = prob_t::Zero(); - for (unsigned i = 0; i < pv->size(); ++i) - z += (*pv)[i].weight; - assert(z > prob_t::Zero()); - for (unsigned i = 0; i < pv->size(); ++i) - (*pv)[i].weight /= z; - } -}; - -template <typename ParticleType> -struct MultinomialResampleFilter { - explicit MultinomialResampleFilter(MT19937* rng) : rng_(rng) {} - - void operator()(std::vector<ParticleType>* pv) { - if (pv->empty()) return; - std::vector<ParticleType>& ps = *pv; - SampleSet<prob_t> ss; - for (int i = 0; i < ps.size(); ++i) - ss.add(ps[i].weight); - std::vector<ParticleType> nps; nps.reserve(ps.size()); - const prob_t uniform_weight(1.0 / ps.size()); - for (int i = 0; i < ps.size(); ++i) { - nps.push_back(ps[rng_->SelectSample(ss)]); - nps[i].weight = uniform_weight; - } - nps.swap(ps); - } - - private: - MT19937* rng_; -}; - -template <typename ParticleType> -struct SystematicResampleFilter { - explicit SystematicResampleFilter(MT19937* rng) : rng_(rng), renorm_() {} - - void operator()(std::vector<ParticleType>* pv) { - if (pv->empty()) return; - renorm_(pv); - std::vector<ParticleType>& ps = *pv; - std::vector<ParticleType> nps; nps.reserve(ps.size()); - double lower = 0, upper = 0; - const double skip = 1.0 / ps.size(); - double u_j = rng_->next() * skip; - //std::cerr << "u_0: " << u_j << std::endl; - int j = 0; - for (unsigned i = 0; i < ps.size(); ++i) { - upper += ps[i].weight.as_float(); - //std::cerr << "lower: " << lower << " upper: " << upper << std::endl; - // how many children does ps[i] have? - while (u_j < lower) { u_j += skip; ++j; } - while (u_j >= lower && u_j <= upper) { - assert(j < ps.size()); - nps.push_back(ps[i]); - u_j += skip; - //std::cerr << " add u_j=" << u_j << std::endl; - ++j; - } - lower = upper; - } - //std::cerr << ps.size() << " " << nps.size() << "\n"; - assert(ps.size() == nps.size()); - //exit(1); - ps.swap(nps); - } - - private: - MT19937* rng_; - ParticleRenormalizer<ParticleType> renorm_; -}; - -#endif diff --git a/gi/pf/pf_test.cc b/gi/pf/pf_test.cc deleted file mode 100644 index 296e7285..00000000 --- a/gi/pf/pf_test.cc +++ /dev/null @@ -1,148 +0,0 @@ -#include "ccrp.h" - -#include <vector> -#include <iostream> - -#include "tdict.h" -#include "transliterations.h" - -using namespace std; - -MT19937 rng; - -static bool verbose = false; - -struct Model { - - Model() : bp(), base(0.2, 0.6) , ccrps(5, CCRP<int>(0.8, 0.5)) {} - - double p0(int x) const { - assert(x > 0); - assert(x < 5); - return 1.0/4.0; - } - - double llh() const { - double lh = bp + base.log_crp_prob(); - for (int ctx = 1; ctx < 5; ++ctx) - lh += ccrps[ctx].log_crp_prob(); - return lh; - } - - double prob(int ctx, int x) const { - assert(ctx > 0 && ctx < 5); - return ccrps[ctx].prob(x, base.prob(x, p0(x))); - } - - void increment(int ctx, int x) { - assert(ctx > 0 && ctx < 5); - if (ccrps[ctx].increment(x, base.prob(x, p0(x)), &rng)) { - if (base.increment(x, p0(x), &rng)) { - bp += log(1.0 / 4.0); - } - } - } - - // this is just a biased estimate - double est_base_prob(int x) { - return (x + 1) * x / 40.0; - } - - void increment_is(int ctx, int x) { - assert(ctx > 0 && ctx < 5); - SampleSet<double> ss; - const int PARTICLES = 25; - vector<CCRP<int> > s1s(PARTICLES, CCRP<int>(0.5,0.5)); - vector<CCRP<int> > sbs(PARTICLES, CCRP<int>(0.5,0.5)); - vector<double> sp0s(PARTICLES); - - CCRP<int> s1 = ccrps[ctx]; - CCRP<int> sb = base; - double sp0 = bp; - for (int pp = 0; pp < PARTICLES; ++pp) { - if (pp > 0) { - ccrps[ctx] = s1; - base = sb; - bp = sp0; - } - - double q = 1; - double gamma = 1; - double est_p = est_base_prob(x); - //base.prob(x, p0(x)) + rng.next() * 0.1; - if (ccrps[ctx].increment(x, est_p, &rng, &q)) { - gamma = q * base.prob(x, p0(x)); - q *= est_p; - if (verbose) cerr << "(DP-base draw) "; - double qq = -1; - if (base.increment(x, p0(x), &rng, &qq)) { - if (verbose) cerr << "(G0 draw) "; - bp += log(p0(x)); - qq *= p0(x); - } - } else { gamma = q; } - double w = gamma / q; - if (verbose) - cerr << "gamma=" << gamma << " q=" << q << "\tw=" << w << endl; - ss.add(w); - s1s[pp] = ccrps[ctx]; - sbs[pp] = base; - sp0s[pp] = bp; - } - int ps = rng.SelectSample(ss); - ccrps[ctx] = s1s[ps]; - base = sbs[ps]; - bp = sp0s[ps]; - if (verbose) { - cerr << "SELECTED: " << ps << endl; - static int cc = 0; cc++; if (cc ==10) exit(1); - } - } - - void decrement(int ctx, int x) { - assert(ctx > 0 && ctx < 5); - if (ccrps[ctx].decrement(x, &rng)) { - if (base.decrement(x, &rng)) { - bp -= log(p0(x)); - } - } - } - - double bp; - CCRP<int> base; - vector<CCRP<int> > ccrps; - -}; - -int main(int argc, char** argv) { - if (argc > 1) { verbose = true; } - vector<int> counts(15, 0); - vector<int> tcounts(15, 0); - int points[] = {1,2, 2,2, 3,2, 4,1, 3, 4, 3, 3, 2, 3, 4, 1, 4, 1, 3, 2, 1, 3, 1, 4, 0, 0}; - double tlh = 0; - double tt = 0; - for (int n = 0; n < 1000; ++n) { - if (n % 10 == 0) cerr << '.'; - if ((n+1) % 400 == 0) cerr << " [" << (n+1) << "]\n"; - Model m; - for (int *x = points; *x; x += 2) - m.increment(x[0], x[1]); - - for (int j = 0; j < 24; ++j) { - for (int *x = points; *x; x += 2) { - if (rng.next() < 0.8) { - m.decrement(x[0], x[1]); - m.increment_is(x[0], x[1]); - } - } - } - counts[m.base.num_customers()]++; - tcounts[m.base.num_tables()]++; - tlh += m.llh(); - tt += 1.0; - } - cerr << "mean LLH = " << (tlh / tt) << endl; - for (int i = 0; i < 15; ++i) - cerr << i << ": " << (counts[i] / tt) << "\t" << (tcounts[i] / tt) << endl; -} - diff --git a/gi/pf/pfbrat.cc b/gi/pf/pfbrat.cc deleted file mode 100644 index 832f22cf..00000000 --- a/gi/pf/pfbrat.cc +++ /dev/null @@ -1,543 +0,0 @@ -#include <iostream> -#include <tr1/memory> -#include <queue> - -#include <boost/functional.hpp> -#include <boost/multi_array.hpp> -#include <boost/program_options.hpp> -#include <boost/program_options/variables_map.hpp> - -#include "viterbi.h" -#include "hg.h" -#include "trule.h" -#include "tdict.h" -#include "filelib.h" -#include "dict.h" -#include "sampler.h" -#include "ccrp_nt.h" -#include "cfg_wfst_composer.h" - -using namespace std; -using namespace tr1; -namespace po = boost::program_options; - -static unsigned kMAX_SRC_PHRASE; -static unsigned kMAX_TRG_PHRASE; -struct FSTState; - -double log_poisson(unsigned x, const double& lambda) { - assert(lambda > 0.0); - return log(lambda) * x - lgamma(x + 1) - lambda; -} - -struct ConditionalBase { - explicit ConditionalBase(const double m1mixture, const unsigned vocab_e_size, const string& model1fname) : - kM1MIXTURE(m1mixture), - kUNIFORM_MIXTURE(1.0 - m1mixture), - kUNIFORM_TARGET(1.0 / vocab_e_size), - kNULL(TD::Convert("<eps>")) { - assert(m1mixture >= 0.0 && m1mixture <= 1.0); - assert(vocab_e_size > 0); - LoadModel1(model1fname); - } - - void LoadModel1(const string& fname) { - cerr << "Loading Model 1 parameters from " << fname << " ..." << endl; - ReadFile rf(fname); - istream& in = *rf.stream(); - string line; - unsigned lc = 0; - while(getline(in, line)) { - ++lc; - int cur = 0; - int start = 0; - while(cur < line.size() && line[cur] != ' ') { ++cur; } - assert(cur != line.size()); - line[cur] = 0; - const WordID src = TD::Convert(&line[0]); - ++cur; - start = cur; - while(cur < line.size() && line[cur] != ' ') { ++cur; } - assert(cur != line.size()); - line[cur] = 0; - WordID trg = TD::Convert(&line[start]); - const double logprob = strtod(&line[cur + 1], NULL); - if (src >= ttable.size()) ttable.resize(src + 1); - ttable[src][trg].logeq(logprob); - } - cerr << " read " << lc << " parameters.\n"; - } - - // return logp0 of rule.e_ | rule.f_ - prob_t operator()(const TRule& rule) const { - const int flen = rule.f_.size(); - const int elen = rule.e_.size(); - prob_t uniform_src_alignment; uniform_src_alignment.logeq(-log(flen + 1)); - prob_t p; - p.logeq(log_poisson(elen, flen + 0.01)); // elen | flen ~Pois(flen + 0.01) - for (int i = 0; i < elen; ++i) { // for each position i in e-RHS - const WordID trg = rule.e_[i]; - prob_t tp = prob_t::Zero(); - for (int j = -1; j < flen; ++j) { - const WordID src = j < 0 ? kNULL : rule.f_[j]; - const map<WordID, prob_t>::const_iterator it = ttable[src].find(trg); - if (it != ttable[src].end()) { - tp += kM1MIXTURE * it->second; - } - tp += kUNIFORM_MIXTURE * kUNIFORM_TARGET; - } - tp *= uniform_src_alignment; // draw a_i ~uniform - p *= tp; // draw e_i ~Model1(f_a_i) / uniform - } - return p; - } - - const prob_t kM1MIXTURE; // Model 1 mixture component - const prob_t kUNIFORM_MIXTURE; // uniform mixture component - const prob_t kUNIFORM_TARGET; - const WordID kNULL; - vector<map<WordID, prob_t> > ttable; -}; - -void InitCommandLine(int argc, char** argv, po::variables_map* conf) { - po::options_description opts("Configuration options"); - opts.add_options() - ("samples,s",po::value<unsigned>()->default_value(1000),"Number of samples") - ("input,i",po::value<string>(),"Read parallel data from") - ("max_src_phrase",po::value<unsigned>()->default_value(3),"Maximum length of source language phrases") - ("max_trg_phrase",po::value<unsigned>()->default_value(3),"Maximum length of target language phrases") - ("model1,m",po::value<string>(),"Model 1 parameters (used in base distribution)") - ("model1_interpolation_weight",po::value<double>()->default_value(0.95),"Mixing proportion of model 1 with uniform target distribution") - ("random_seed,S",po::value<uint32_t>(), "Random seed"); - po::options_description clo("Command line options"); - clo.add_options() - ("config", po::value<string>(), "Configuration file") - ("help,h", "Print this help message and exit"); - po::options_description dconfig_options, dcmdline_options; - dconfig_options.add(opts); - dcmdline_options.add(opts).add(clo); - - po::store(parse_command_line(argc, argv, dcmdline_options), *conf); - if (conf->count("config")) { - ifstream config((*conf)["config"].as<string>().c_str()); - po::store(po::parse_config_file(config, dconfig_options), *conf); - } - po::notify(*conf); - - if (conf->count("help") || (conf->count("input") == 0)) { - cerr << dcmdline_options << endl; - exit(1); - } -} - -void ReadParallelCorpus(const string& filename, - vector<vector<WordID> >* f, - vector<vector<int> >* e, - set<int>* vocab_f, - set<int>* vocab_e) { - f->clear(); - e->clear(); - vocab_f->clear(); - vocab_e->clear(); - istream* in; - if (filename == "-") - in = &cin; - else - in = new ifstream(filename.c_str()); - assert(*in); - string line; - const WordID kDIV = TD::Convert("|||"); - vector<WordID> tmp; - while(*in) { - getline(*in, line); - if (line.empty() && !*in) break; - e->push_back(vector<int>()); - f->push_back(vector<int>()); - vector<int>& le = e->back(); - vector<int>& lf = f->back(); - tmp.clear(); - TD::ConvertSentence(line, &tmp); - bool isf = true; - for (unsigned i = 0; i < tmp.size(); ++i) { - const int cur = tmp[i]; - if (isf) { - if (kDIV == cur) { isf = false; } else { - lf.push_back(cur); - vocab_f->insert(cur); - } - } else { - assert(cur != kDIV); - le.push_back(cur); - vocab_e->insert(cur); - } - } - assert(isf == false); - } - if (in != &cin) delete in; -} - -struct UniphraseLM { - UniphraseLM(const vector<vector<int> >& corpus, - const set<int>& vocab, - const po::variables_map& conf) : - phrases_(1,1), - gen_(1,1), - corpus_(corpus), - uniform_word_(1.0 / vocab.size()), - gen_p0_(0.5), - p_end_(0.5), - use_poisson_(conf.count("poisson_length") > 0) {} - - void ResampleHyperparameters(MT19937* rng) { - phrases_.resample_hyperparameters(rng); - gen_.resample_hyperparameters(rng); - cerr << " " << phrases_.alpha(); - } - - CCRP_NoTable<vector<int> > phrases_; - CCRP_NoTable<bool> gen_; - vector<vector<bool> > z_; // z_[i] is there a phrase boundary after the ith word - const vector<vector<int> >& corpus_; - const double uniform_word_; - const double gen_p0_; - const double p_end_; // in base length distribution, p of the end of a phrase - const bool use_poisson_; -}; - -struct Reachability { - boost::multi_array<bool, 4> edges; // edges[src_covered][trg_covered][x][trg_delta] is this edge worth exploring? - boost::multi_array<short, 2> max_src_delta; // msd[src_covered][trg_covered] -- the largest src delta that's valid - - Reachability(int srclen, int trglen, int src_max_phrase_len, int trg_max_phrase_len) : - edges(boost::extents[srclen][trglen][src_max_phrase_len+1][trg_max_phrase_len+1]), - max_src_delta(boost::extents[srclen][trglen]) { - ComputeReachability(srclen, trglen, src_max_phrase_len, trg_max_phrase_len); - } - - private: - struct SState { - SState() : prev_src_covered(), prev_trg_covered() {} - SState(int i, int j) : prev_src_covered(i), prev_trg_covered(j) {} - int prev_src_covered; - int prev_trg_covered; - }; - - struct NState { - NState() : next_src_covered(), next_trg_covered() {} - NState(int i, int j) : next_src_covered(i), next_trg_covered(j) {} - int next_src_covered; - int next_trg_covered; - }; - - void ComputeReachability(int srclen, int trglen, int src_max_phrase_len, int trg_max_phrase_len) { - typedef boost::multi_array<vector<SState>, 2> array_type; - array_type a(boost::extents[srclen + 1][trglen + 1]); - a[0][0].push_back(SState()); - for (int i = 0; i < srclen; ++i) { - for (int j = 0; j < trglen; ++j) { - if (a[i][j].size() == 0) continue; - const SState prev(i,j); - for (int k = 1; k <= src_max_phrase_len; ++k) { - if ((i + k) > srclen) continue; - for (int l = 1; l <= trg_max_phrase_len; ++l) { - if ((j + l) > trglen) continue; - a[i + k][j + l].push_back(prev); - } - } - } - } - a[0][0].clear(); - cerr << "Final cell contains " << a[srclen][trglen].size() << " back pointers\n"; - assert(a[srclen][trglen].size() > 0); - - typedef boost::multi_array<bool, 2> rarray_type; - rarray_type r(boost::extents[srclen + 1][trglen + 1]); -// typedef boost::multi_array<vector<NState>, 2> narray_type; -// narray_type b(boost::extents[srclen + 1][trglen + 1]); - r[srclen][trglen] = true; - for (int i = srclen; i >= 0; --i) { - for (int j = trglen; j >= 0; --j) { - vector<SState>& prevs = a[i][j]; - if (!r[i][j]) { prevs.clear(); } -// const NState nstate(i,j); - for (int k = 0; k < prevs.size(); ++k) { - r[prevs[k].prev_src_covered][prevs[k].prev_trg_covered] = true; - int src_delta = i - prevs[k].prev_src_covered; - edges[prevs[k].prev_src_covered][prevs[k].prev_trg_covered][src_delta][j - prevs[k].prev_trg_covered] = true; - short &msd = max_src_delta[prevs[k].prev_src_covered][prevs[k].prev_trg_covered]; - if (src_delta > msd) msd = src_delta; -// b[prevs[k].prev_src_covered][prevs[k].prev_trg_covered].push_back(nstate); - } - } - } - assert(!edges[0][0][1][0]); - assert(!edges[0][0][0][1]); - assert(!edges[0][0][0][0]); - cerr << " MAX SRC DELTA[0][0] = " << max_src_delta[0][0] << endl; - assert(max_src_delta[0][0] > 0); - //cerr << "First cell contains " << b[0][0].size() << " forward pointers\n"; - //for (int i = 0; i < b[0][0].size(); ++i) { - // cerr << " -> (" << b[0][0][i].next_src_covered << "," << b[0][0][i].next_trg_covered << ")\n"; - //} - } -}; - -ostream& operator<<(ostream& os, const FSTState& q); -struct FSTState { - explicit FSTState(int src_size) : - trg_covered_(), - src_covered_(), - src_coverage_(src_size) {} - - FSTState(short trg_covered, short src_covered, const vector<bool>& src_coverage, const vector<short>& src_prefix) : - trg_covered_(trg_covered), - src_covered_(src_covered), - src_coverage_(src_coverage), - src_prefix_(src_prefix) { - if (src_coverage_.size() == src_covered) { - assert(src_prefix.size() == 0); - } - } - - // if we extend by the word at src_position, what are - // the next states that are reachable and lie on a valid - // path to the final state? - vector<FSTState> Extensions(int src_position, int src_len, int trg_len, const Reachability& r) const { - assert(src_position < src_coverage_.size()); - if (src_coverage_[src_position]) { - cerr << "Trying to extend " << *this << " with position " << src_position << endl; - abort(); - } - vector<bool> ncvg = src_coverage_; - ncvg[src_position] = true; - - vector<FSTState> res; - const int trg_remaining = trg_len - trg_covered_; - if (trg_remaining <= 0) { - cerr << "Target appears to have been covered: " << *this << " (trg_len=" << trg_len << ",trg_covered=" << trg_covered_ << ")" << endl; - abort(); - } - const int src_remaining = src_len - src_covered_; - if (src_remaining <= 0) { - cerr << "Source appears to have been covered: " << *this << endl; - abort(); - } - - for (int tc = 1; tc <= kMAX_TRG_PHRASE; ++tc) { - if (r.edges[src_covered_][trg_covered_][src_prefix_.size() + 1][tc]) { - int nc = src_prefix_.size() + 1 + src_covered_; - res.push_back(FSTState(trg_covered_ + tc, nc, ncvg, vector<short>())); - } - } - - if ((src_prefix_.size() + 1) < r.max_src_delta[src_covered_][trg_covered_]) { - vector<short> nsp = src_prefix_; - nsp.push_back(src_position); - res.push_back(FSTState(trg_covered_, src_covered_, ncvg, nsp)); - } - - if (res.size() == 0) { - cerr << *this << " can't be extended!\n"; - abort(); - } - return res; - } - - short trg_covered_, src_covered_; - vector<bool> src_coverage_; - vector<short> src_prefix_; -}; -bool operator<(const FSTState& q, const FSTState& r) { - if (q.trg_covered_ != r.trg_covered_) return q.trg_covered_ < r.trg_covered_; - if (q.src_covered_!= r.src_covered_) return q.src_covered_ < r.src_covered_; - if (q.src_coverage_ != r.src_coverage_) return q.src_coverage_ < r.src_coverage_; - return q.src_prefix_ < r.src_prefix_; -} - -ostream& operator<<(ostream& os, const FSTState& q) { - os << "[" << q.trg_covered_ << " : "; - for (int i = 0; i < q.src_coverage_.size(); ++i) - os << q.src_coverage_[i]; - os << " : <"; - for (int i = 0; i < q.src_prefix_.size(); ++i) { - if (i != 0) os << ' '; - os << q.src_prefix_[i]; - } - return os << ">]"; -} - -struct MyModel { - MyModel(ConditionalBase& rcp0) : rp0(rcp0) {} - typedef unordered_map<vector<WordID>, CCRP_NoTable<TRule>, boost::hash<vector<WordID> > > SrcToRuleCRPMap; - - void DecrementRule(const TRule& rule) { - SrcToRuleCRPMap::iterator it = rules.find(rule.f_); - assert(it != rules.end()); - it->second.decrement(rule); - if (it->second.num_customers() == 0) rules.erase(it); - } - - void IncrementRule(const TRule& rule) { - SrcToRuleCRPMap::iterator it = rules.find(rule.f_); - if (it == rules.end()) { - CCRP_NoTable<TRule> crp(1,1); - it = rules.insert(make_pair(rule.f_, crp)).first; - } - it->second.increment(rule); - } - - // conditioned on rule.f_ - prob_t RuleConditionalProbability(const TRule& rule) const { - const prob_t base = rp0(rule); - SrcToRuleCRPMap::const_iterator it = rules.find(rule.f_); - if (it == rules.end()) { - return base; - } else { - const double lp = it->second.logprob(rule, log(base)); - prob_t q; q.logeq(lp); - return q; - } - } - - const ConditionalBase& rp0; - SrcToRuleCRPMap rules; -}; - -struct MyFST : public WFST { - MyFST(const vector<WordID>& ssrc, const vector<WordID>& strg, MyModel* m) : - src(ssrc), trg(strg), - r(src.size(),trg.size(),kMAX_SRC_PHRASE, kMAX_TRG_PHRASE), - model(m) { - FSTState in(src.size()); - cerr << " INIT: " << in << endl; - init = GetNode(in); - for (int i = 0; i < in.src_coverage_.size(); ++i) in.src_coverage_[i] = true; - in.src_covered_ = src.size(); - in.trg_covered_ = trg.size(); - cerr << "FINAL: " << in << endl; - final = GetNode(in); - } - virtual const WFSTNode* Final() const; - virtual const WFSTNode* Initial() const; - - const WFSTNode* GetNode(const FSTState& q); - map<FSTState, boost::shared_ptr<WFSTNode> > m; - const vector<WordID>& src; - const vector<WordID>& trg; - Reachability r; - const WFSTNode* init; - const WFSTNode* final; - MyModel* model; -}; - -struct MyNode : public WFSTNode { - MyNode(const FSTState& q, MyFST* fst) : state(q), container(fst) {} - virtual vector<pair<const WFSTNode*, TRulePtr> > ExtendInput(unsigned srcindex) const; - const FSTState state; - mutable MyFST* container; -}; - -vector<pair<const WFSTNode*, TRulePtr> > MyNode::ExtendInput(unsigned srcindex) const { - cerr << "EXTEND " << state << " with " << srcindex << endl; - vector<FSTState> ext = state.Extensions(srcindex, container->src.size(), container->trg.size(), container->r); - vector<pair<const WFSTNode*,TRulePtr> > res(ext.size()); - for (unsigned i = 0; i < ext.size(); ++i) { - res[i].first = container->GetNode(ext[i]); - if (ext[i].src_prefix_.size() == 0) { - const unsigned trg_from = state.trg_covered_; - const unsigned trg_to = ext[i].trg_covered_; - const unsigned prev_prfx_size = state.src_prefix_.size(); - res[i].second.reset(new TRule); - res[i].second->lhs_ = -TD::Convert("X"); - vector<WordID>& src = res[i].second->f_; - vector<WordID>& trg = res[i].second->e_; - src.resize(prev_prfx_size + 1); - for (unsigned j = 0; j < prev_prfx_size; ++j) - src[j] = container->src[state.src_prefix_[j]]; - src[prev_prfx_size] = container->src[srcindex]; - for (unsigned j = trg_from; j < trg_to; ++j) - trg.push_back(container->trg[j]); - res[i].second->scores_.set_value(FD::Convert("Proposal"), log(container->model->RuleConditionalProbability(*res[i].second))); - } - } - return res; -} - -const WFSTNode* MyFST::GetNode(const FSTState& q) { - boost::shared_ptr<WFSTNode>& res = m[q]; - if (!res) { - res.reset(new MyNode(q, this)); - } - return &*res; -} - -const WFSTNode* MyFST::Final() const { - return final; -} - -const WFSTNode* MyFST::Initial() const { - return init; -} - -int main(int argc, char** argv) { - po::variables_map conf; - InitCommandLine(argc, argv, &conf); - kMAX_TRG_PHRASE = conf["max_trg_phrase"].as<unsigned>(); - kMAX_SRC_PHRASE = conf["max_src_phrase"].as<unsigned>(); - - if (!conf.count("model1")) { - cerr << argv[0] << "Please use --model1 to specify model 1 parameters\n"; - return 1; - } - boost::shared_ptr<MT19937> prng; - if (conf.count("random_seed")) - prng.reset(new MT19937(conf["random_seed"].as<uint32_t>())); - else - prng.reset(new MT19937); - MT19937& rng = *prng; - - vector<vector<int> > corpuse, corpusf; - set<int> vocabe, vocabf; - ReadParallelCorpus(conf["input"].as<string>(), &corpusf, &corpuse, &vocabf, &vocabe); - cerr << "f-Corpus size: " << corpusf.size() << " sentences\n"; - cerr << "f-Vocabulary size: " << vocabf.size() << " types\n"; - cerr << "f-Corpus size: " << corpuse.size() << " sentences\n"; - cerr << "f-Vocabulary size: " << vocabe.size() << " types\n"; - assert(corpusf.size() == corpuse.size()); - - ConditionalBase lp0(conf["model1_interpolation_weight"].as<double>(), - vocabe.size(), - conf["model1"].as<string>()); - MyModel m(lp0); - - TRule x("[X] ||| kAnwntR myN ||| at the convent ||| 0"); - m.IncrementRule(x); - TRule y("[X] ||| nY dyN ||| gave ||| 0"); - m.IncrementRule(y); - - - MyFST fst(corpusf[0], corpuse[0], &m); - ifstream in("./kimura.g"); - assert(in); - CFG_WFSTComposer comp(fst); - Hypergraph hg; - bool succeed = comp.Compose(&in, &hg); - hg.PrintGraphviz(); - if (succeed) { cerr << "SUCCESS.\n"; } else { cerr << "FAILURE REPORTED.\n"; } - -#if 0 - ifstream in2("./amnabooks.g"); - assert(in2); - MyFST fst2(corpusf[1], corpuse[1], &m); - CFG_WFSTComposer comp2(fst2); - Hypergraph hg2; - bool succeed2 = comp2.Compose(&in2, &hg2); - if (succeed2) { cerr << "SUCCESS.\n"; } else { cerr << "FAILURE REPORTED.\n"; } -#endif - - SparseVector<double> w; w.set_value(FD::Convert("Proposal"), 1.0); - hg.Reweight(w); - cerr << ViterbiFTree(hg) << endl; - return 0; -} - diff --git a/gi/pf/pfdist.cc b/gi/pf/pfdist.cc deleted file mode 100644 index a3e46064..00000000 --- a/gi/pf/pfdist.cc +++ /dev/null @@ -1,598 +0,0 @@ -#include <iostream> -#include <tr1/memory> -#include <queue> - -#include <boost/functional.hpp> -#include <boost/program_options.hpp> -#include <boost/program_options/variables_map.hpp> - -#include "pf.h" -#include "base_distributions.h" -#include "reachability.h" -#include "viterbi.h" -#include "hg.h" -#include "trule.h" -#include "tdict.h" -#include "filelib.h" -#include "dict.h" -#include "sampler.h" -#include "ccrp_nt.h" -#include "ccrp_onetable.h" - -using namespace std; -using namespace tr1; -namespace po = boost::program_options; - -boost::shared_ptr<MT19937> prng; - -void InitCommandLine(int argc, char** argv, po::variables_map* conf) { - po::options_description opts("Configuration options"); - opts.add_options() - ("samples,s",po::value<unsigned>()->default_value(1000),"Number of samples") - ("particles,p",po::value<unsigned>()->default_value(30),"Number of particles") - ("filter_frequency,f",po::value<unsigned>()->default_value(5),"Number of time steps between filterings") - ("input,i",po::value<string>(),"Read parallel data from") - ("max_src_phrase",po::value<unsigned>()->default_value(5),"Maximum length of source language phrases") - ("max_trg_phrase",po::value<unsigned>()->default_value(5),"Maximum length of target language phrases") - ("model1,m",po::value<string>(),"Model 1 parameters (used in base distribution)") - ("inverse_model1,M",po::value<string>(),"Inverse Model 1 parameters (used in backward estimate)") - ("model1_interpolation_weight",po::value<double>()->default_value(0.95),"Mixing proportion of model 1 with uniform target distribution") - ("random_seed,S",po::value<uint32_t>(), "Random seed"); - po::options_description clo("Command line options"); - clo.add_options() - ("config", po::value<string>(), "Configuration file") - ("help,h", "Print this help message and exit"); - po::options_description dconfig_options, dcmdline_options; - dconfig_options.add(opts); - dcmdline_options.add(opts).add(clo); - - po::store(parse_command_line(argc, argv, dcmdline_options), *conf); - if (conf->count("config")) { - ifstream config((*conf)["config"].as<string>().c_str()); - po::store(po::parse_config_file(config, dconfig_options), *conf); - } - po::notify(*conf); - - if (conf->count("help") || (conf->count("input") == 0)) { - cerr << dcmdline_options << endl; - exit(1); - } -} - -void ReadParallelCorpus(const string& filename, - vector<vector<WordID> >* f, - vector<vector<WordID> >* e, - set<WordID>* vocab_f, - set<WordID>* vocab_e) { - f->clear(); - e->clear(); - vocab_f->clear(); - vocab_e->clear(); - istream* in; - if (filename == "-") - in = &cin; - else - in = new ifstream(filename.c_str()); - assert(*in); - string line; - const WordID kDIV = TD::Convert("|||"); - vector<WordID> tmp; - while(*in) { - getline(*in, line); - if (line.empty() && !*in) break; - e->push_back(vector<int>()); - f->push_back(vector<int>()); - vector<int>& le = e->back(); - vector<int>& lf = f->back(); - tmp.clear(); - TD::ConvertSentence(line, &tmp); - bool isf = true; - for (unsigned i = 0; i < tmp.size(); ++i) { - const int cur = tmp[i]; - if (isf) { - if (kDIV == cur) { isf = false; } else { - lf.push_back(cur); - vocab_f->insert(cur); - } - } else { - assert(cur != kDIV); - le.push_back(cur); - vocab_e->insert(cur); - } - } - assert(isf == false); - } - if (in != &cin) delete in; -} - -#if 0 -struct MyConditionalModel { - MyConditionalModel(PhraseConditionalBase& rcp0) : rp0(&rcp0), base(prob_t::One()), src_phrases(1,1), src_jumps(200, CCRP_NoTable<int>(1,1)) {} - - prob_t srcp0(const vector<WordID>& src) const { - prob_t p(1.0 / 3000.0); - p.poweq(src.size()); - prob_t lenp; lenp.logeq(log_poisson(src.size(), 1.0)); - p *= lenp; - return p; - } - - void DecrementRule(const TRule& rule) { - const RuleCRPMap::iterator it = rules.find(rule.f_); - assert(it != rules.end()); - if (it->second.decrement(rule)) { - base /= (*rp0)(rule); - if (it->second.num_customers() == 0) - rules.erase(it); - } - if (src_phrases.decrement(rule.f_)) - base /= srcp0(rule.f_); - } - - void IncrementRule(const TRule& rule) { - RuleCRPMap::iterator it = rules.find(rule.f_); - if (it == rules.end()) - it = rules.insert(make_pair(rule.f_, CCRP_NoTable<TRule>(1,1))).first; - if (it->second.increment(rule)) { - base *= (*rp0)(rule); - } - if (src_phrases.increment(rule.f_)) - base *= srcp0(rule.f_); - } - - void IncrementRules(const vector<TRulePtr>& rules) { - for (int i = 0; i < rules.size(); ++i) - IncrementRule(*rules[i]); - } - - void DecrementRules(const vector<TRulePtr>& rules) { - for (int i = 0; i < rules.size(); ++i) - DecrementRule(*rules[i]); - } - - void IncrementJump(int dist, unsigned src_len) { - assert(src_len > 0); - if (src_jumps[src_len].increment(dist)) - base *= jp0(dist, src_len); - } - - void DecrementJump(int dist, unsigned src_len) { - assert(src_len > 0); - if (src_jumps[src_len].decrement(dist)) - base /= jp0(dist, src_len); - } - - void IncrementJumps(const vector<int>& js, unsigned src_len) { - for (unsigned i = 0; i < js.size(); ++i) - IncrementJump(js[i], src_len); - } - - void DecrementJumps(const vector<int>& js, unsigned src_len) { - for (unsigned i = 0; i < js.size(); ++i) - DecrementJump(js[i], src_len); - } - - // p(jump = dist | src_len , z) - prob_t JumpProbability(int dist, unsigned src_len) { - const prob_t p0 = jp0(dist, src_len); - const double lp = src_jumps[src_len].logprob(dist, log(p0)); - prob_t q; q.logeq(lp); - return q; - } - - // p(rule.f_ | z) * p(rule.e_ | rule.f_ , z) - prob_t RuleProbability(const TRule& rule) const { - const prob_t p0 = (*rp0)(rule); - prob_t srcp; srcp.logeq(src_phrases.logprob(rule.f_, log(srcp0(rule.f_)))); - const RuleCRPMap::const_iterator it = rules.find(rule.f_); - if (it == rules.end()) return srcp * p0; - const double lp = it->second.logprob(rule, log(p0)); - prob_t q; q.logeq(lp); - return q * srcp; - } - - prob_t Likelihood() const { - prob_t p = base; - for (RuleCRPMap::const_iterator it = rules.begin(); - it != rules.end(); ++it) { - prob_t cl; cl.logeq(it->second.log_crp_prob()); - p *= cl; - } - for (unsigned l = 1; l < src_jumps.size(); ++l) { - if (src_jumps[l].num_customers() > 0) { - prob_t q; - q.logeq(src_jumps[l].log_crp_prob()); - p *= q; - } - } - return p; - } - - JumpBase jp0; - const PhraseConditionalBase* rp0; - prob_t base; - typedef unordered_map<vector<WordID>, CCRP_NoTable<TRule>, boost::hash<vector<WordID> > > RuleCRPMap; - RuleCRPMap rules; - CCRP_NoTable<vector<WordID> > src_phrases; - vector<CCRP_NoTable<int> > src_jumps; -}; - -#endif - -struct MyJointModel { - MyJointModel(PhraseJointBase& rcp0) : - rp0(rcp0), base(prob_t::One()), rules(1,1), src_jumps(200, CCRP_NoTable<int>(1,1)) {} - - void DecrementRule(const TRule& rule) { - if (rules.decrement(rule)) - base /= rp0(rule); - } - - void IncrementRule(const TRule& rule) { - if (rules.increment(rule)) - base *= rp0(rule); - } - - void IncrementRules(const vector<TRulePtr>& rules) { - for (int i = 0; i < rules.size(); ++i) - IncrementRule(*rules[i]); - } - - void DecrementRules(const vector<TRulePtr>& rules) { - for (int i = 0; i < rules.size(); ++i) - DecrementRule(*rules[i]); - } - - void IncrementJump(int dist, unsigned src_len) { - assert(src_len > 0); - if (src_jumps[src_len].increment(dist)) - base *= jp0(dist, src_len); - } - - void DecrementJump(int dist, unsigned src_len) { - assert(src_len > 0); - if (src_jumps[src_len].decrement(dist)) - base /= jp0(dist, src_len); - } - - void IncrementJumps(const vector<int>& js, unsigned src_len) { - for (unsigned i = 0; i < js.size(); ++i) - IncrementJump(js[i], src_len); - } - - void DecrementJumps(const vector<int>& js, unsigned src_len) { - for (unsigned i = 0; i < js.size(); ++i) - DecrementJump(js[i], src_len); - } - - // p(jump = dist | src_len , z) - prob_t JumpProbability(int dist, unsigned src_len) { - const prob_t p0 = jp0(dist, src_len); - const double lp = src_jumps[src_len].logprob(dist, log(p0)); - prob_t q; q.logeq(lp); - return q; - } - - // p(rule.f_ | z) * p(rule.e_ | rule.f_ , z) - prob_t RuleProbability(const TRule& rule) const { - prob_t p; p.logeq(rules.logprob(rule, log(rp0(rule)))); - return p; - } - - prob_t Likelihood() const { - prob_t p = base; - prob_t q; q.logeq(rules.log_crp_prob()); - p *= q; - for (unsigned l = 1; l < src_jumps.size(); ++l) { - if (src_jumps[l].num_customers() > 0) { - prob_t q; - q.logeq(src_jumps[l].log_crp_prob()); - p *= q; - } - } - return p; - } - - JumpBase jp0; - const PhraseJointBase& rp0; - prob_t base; - CCRP_NoTable<TRule> rules; - vector<CCRP_NoTable<int> > src_jumps; -}; - -struct BackwardEstimate { - BackwardEstimate(const Model1& m1, const vector<WordID>& src, const vector<WordID>& trg) : - model1_(m1), src_(src), trg_(trg) { - } - const prob_t& operator()(const vector<bool>& src_cov, unsigned trg_cov) const { - assert(src_.size() == src_cov.size()); - assert(trg_cov <= trg_.size()); - prob_t& e = cache_[src_cov][trg_cov]; - if (e.is_0()) { - if (trg_cov == trg_.size()) { e = prob_t::One(); return e; } - vector<WordID> r(src_.size() + 1); r.clear(); - r.push_back(0); // NULL word - for (int i = 0; i < src_cov.size(); ++i) - if (!src_cov[i]) r.push_back(src_[i]); - const prob_t uniform_alignment(1.0 / r.size()); - e.logeq(Md::log_poisson(trg_.size() - trg_cov, r.size() - 1)); // p(trg len remaining | src len remaining) - for (unsigned j = trg_cov; j < trg_.size(); ++j) { - prob_t p; - for (unsigned i = 0; i < r.size(); ++i) - p += model1_(r[i], trg_[j]); - if (p.is_0()) { - cerr << "ERROR: p(" << TD::Convert(trg_[j]) << " | " << TD::GetString(r) << ") = 0!\n"; - abort(); - } - p *= uniform_alignment; - e *= p; - } - } - return e; - } - const Model1& model1_; - const vector<WordID>& src_; - const vector<WordID>& trg_; - mutable unordered_map<vector<bool>, map<unsigned, prob_t>, boost::hash<vector<bool> > > cache_; -}; - -struct BackwardEstimateSym { - BackwardEstimateSym(const Model1& m1, - const Model1& invm1, const vector<WordID>& src, const vector<WordID>& trg) : - model1_(m1), invmodel1_(invm1), src_(src), trg_(trg) { - } - const prob_t& operator()(const vector<bool>& src_cov, unsigned trg_cov) const { - assert(src_.size() == src_cov.size()); - assert(trg_cov <= trg_.size()); - prob_t& e = cache_[src_cov][trg_cov]; - if (e.is_0()) { - if (trg_cov == trg_.size()) { e = prob_t::One(); return e; } - vector<WordID> r(src_.size() + 1); r.clear(); - for (int i = 0; i < src_cov.size(); ++i) - if (!src_cov[i]) r.push_back(src_[i]); - r.push_back(0); // NULL word - const prob_t uniform_alignment(1.0 / r.size()); - e.logeq(Md::log_poisson(trg_.size() - trg_cov, r.size() - 1)); // p(trg len remaining | src len remaining) - for (unsigned j = trg_cov; j < trg_.size(); ++j) { - prob_t p; - for (unsigned i = 0; i < r.size(); ++i) - p += model1_(r[i], trg_[j]); - if (p.is_0()) { - cerr << "ERROR: p(" << TD::Convert(trg_[j]) << " | " << TD::GetString(r) << ") = 0!\n"; - abort(); - } - p *= uniform_alignment; - e *= p; - } - r.pop_back(); - const prob_t inv_uniform(1.0 / (trg_.size() - trg_cov + 1.0)); - prob_t inv; - inv.logeq(Md::log_poisson(r.size(), trg_.size() - trg_cov)); - for (unsigned i = 0; i < r.size(); ++i) { - prob_t p; - for (unsigned j = trg_cov - 1; j < trg_.size(); ++j) - p += invmodel1_(j < trg_cov ? 0 : trg_[j], r[i]); - if (p.is_0()) { - cerr << "ERROR: p_inv(" << TD::Convert(r[i]) << " | " << TD::GetString(trg_) << ") = 0!\n"; - abort(); - } - p *= inv_uniform; - inv *= p; - } - prob_t x = pow(e * inv, 0.5); - e = x; - //cerr << "Forward: " << log(e) << "\tBackward: " << log(inv) << "\t prop: " << log(x) << endl; - } - return e; - } - const Model1& model1_; - const Model1& invmodel1_; - const vector<WordID>& src_; - const vector<WordID>& trg_; - mutable unordered_map<vector<bool>, map<unsigned, prob_t>, boost::hash<vector<bool> > > cache_; -}; - -struct Particle { - Particle() : weight(prob_t::One()), src_cov(), trg_cov(), prev_pos(-1) {} - prob_t weight; - prob_t gamma_last; - vector<int> src_jumps; - vector<TRulePtr> rules; - vector<bool> src_cv; - int src_cov; - int trg_cov; - int prev_pos; -}; - -ostream& operator<<(ostream& o, const vector<bool>& v) { - for (int i = 0; i < v.size(); ++i) - o << (v[i] ? '1' : '0'); - return o; -} -ostream& operator<<(ostream& o, const Particle& p) { - o << "[cv=" << p.src_cv << " src_cov=" << p.src_cov << " trg_cov=" << p.trg_cov << " last_pos=" << p.prev_pos << " num_rules=" << p.rules.size() << " w=" << log(p.weight) << ']'; - return o; -} - -int main(int argc, char** argv) { - po::variables_map conf; - InitCommandLine(argc, argv, &conf); - const unsigned kMAX_TRG_PHRASE = conf["max_trg_phrase"].as<unsigned>(); - const unsigned kMAX_SRC_PHRASE = conf["max_src_phrase"].as<unsigned>(); - const unsigned particles = conf["particles"].as<unsigned>(); - const unsigned samples = conf["samples"].as<unsigned>(); - const unsigned rejuv_freq = conf["filter_frequency"].as<unsigned>(); - - if (!conf.count("model1")) { - cerr << argv[0] << "Please use --model1 to specify model 1 parameters\n"; - return 1; - } - if (conf.count("random_seed")) - prng.reset(new MT19937(conf["random_seed"].as<uint32_t>())); - else - prng.reset(new MT19937); - MT19937& rng = *prng; - - vector<vector<WordID> > corpuse, corpusf; - set<WordID> vocabe, vocabf; - cerr << "Reading corpus...\n"; - ReadParallelCorpus(conf["input"].as<string>(), &corpusf, &corpuse, &vocabf, &vocabe); - cerr << "F-corpus size: " << corpusf.size() << " sentences\t (" << vocabf.size() << " word types)\n"; - cerr << "E-corpus size: " << corpuse.size() << " sentences\t (" << vocabe.size() << " word types)\n"; - assert(corpusf.size() == corpuse.size()); - - const int kLHS = -TD::Convert("X"); - Model1 m1(conf["model1"].as<string>()); - Model1 invm1(conf["inverse_model1"].as<string>()); - -#if 0 - PhraseConditionalBase lp0(m1, conf["model1_interpolation_weight"].as<double>(), vocabe.size()); - MyConditionalModel m(lp0); -#else - PhraseJointBase lp0(m1, conf["model1_interpolation_weight"].as<double>(), vocabe.size(), vocabf.size()); - MyJointModel m(lp0); -#endif - - MultinomialResampleFilter<Particle> filter(&rng); - cerr << "Initializing reachability limits...\n"; - vector<Particle> ps(corpusf.size()); - vector<Reachability> reaches; reaches.reserve(corpusf.size()); - for (int ci = 0; ci < corpusf.size(); ++ci) - reaches.push_back(Reachability(corpusf[ci].size(), - corpuse[ci].size(), - kMAX_SRC_PHRASE, - kMAX_TRG_PHRASE)); - cerr << "Sampling...\n"; - vector<Particle> tmp_p(10000); // work space - SampleSet<prob_t> pfss; - for (int SS=0; SS < samples; ++SS) { - for (int ci = 0; ci < corpusf.size(); ++ci) { - vector<int>& src = corpusf[ci]; - vector<int>& trg = corpuse[ci]; - m.DecrementRules(ps[ci].rules); - m.DecrementJumps(ps[ci].src_jumps, src.size()); - - //BackwardEstimate be(m1, src, trg); - BackwardEstimateSym be(m1, invm1, src, trg); - const Reachability& r = reaches[ci]; - vector<Particle> lps(particles); - - for (int pi = 0; pi < particles; ++pi) { - Particle& p = lps[pi]; - p.src_cv.resize(src.size(), false); - } - - bool all_complete = false; - while(!all_complete) { - SampleSet<prob_t> ss; - - // all particles have now been extended a bit, we will reweight them now - if (lps[0].trg_cov > 0) - filter(&lps); - - // loop over all particles and extend them - bool done_nothing = true; - for (int pi = 0; pi < particles; ++pi) { - Particle& p = lps[pi]; - int tic = 0; - while(p.trg_cov < trg.size() && tic < rejuv_freq) { - ++tic; - done_nothing = false; - ss.clear(); - TRule x; x.lhs_ = kLHS; - prob_t z; - int first_uncovered = src.size(); - int last_uncovered = -1; - for (int i = 0; i < src.size(); ++i) { - const bool is_uncovered = !p.src_cv[i]; - if (i < first_uncovered && is_uncovered) first_uncovered = i; - if (is_uncovered && i > last_uncovered) last_uncovered = i; - } - assert(last_uncovered > -1); - assert(first_uncovered < src.size()); - - for (int trg_len = 1; trg_len <= kMAX_TRG_PHRASE; ++trg_len) { - x.e_.push_back(trg[trg_len - 1 + p.trg_cov]); - for (int src_len = 1; src_len <= kMAX_SRC_PHRASE; ++src_len) { - if (!r.edges[p.src_cov][p.trg_cov][src_len][trg_len]) continue; - - const int last_possible_start = last_uncovered - src_len + 1; - assert(last_possible_start >= 0); - //cerr << src_len << "," << trg_len << " is allowed. E=" << TD::GetString(x.e_) << endl; - //cerr << " first_uncovered=" << first_uncovered << " last_possible_start=" << last_possible_start << endl; - for (int i = first_uncovered; i <= last_possible_start; ++i) { - if (p.src_cv[i]) continue; - assert(ss.size() < tmp_p.size()); // if fails increase tmp_p size - Particle& np = tmp_p[ss.size()]; - np = p; - x.f_.clear(); - int gap_add = 0; - bool bad = false; - prob_t jp = prob_t::One(); - int prev_pos = p.prev_pos; - for (int j = 0; j < src_len; ++j) { - if ((j + i + gap_add) == src.size()) { bad = true; break; } - while ((i+j+gap_add) < src.size() && p.src_cv[i + j + gap_add]) { ++gap_add; } - if ((j + i + gap_add) == src.size()) { bad = true; break; } - np.src_cv[i + j + gap_add] = true; - x.f_.push_back(src[i + j + gap_add]); - jp *= m.JumpProbability(i + j + gap_add - prev_pos, src.size()); - int jump = i + j + gap_add - prev_pos; - assert(jump != 0); - np.src_jumps.push_back(jump); - prev_pos = i + j + gap_add; - } - if (bad) continue; - np.prev_pos = prev_pos; - np.src_cov += x.f_.size(); - np.trg_cov += x.e_.size(); - if (x.f_.size() != src_len) continue; - prob_t rp = m.RuleProbability(x); - np.gamma_last = rp * jp; - const prob_t u = pow(np.gamma_last * be(np.src_cv, np.trg_cov), 0.2); - //cerr << "**rule=" << x << endl; - //cerr << " u=" << log(u) << " rule=" << rp << " jump=" << jp << endl; - ss.add(u); - np.rules.push_back(TRulePtr(new TRule(x))); - z += u; - - const bool completed = (p.trg_cov == trg.size()); - if (completed) { - int last_jump = src.size() - p.prev_pos; - assert(last_jump > 0); - p.src_jumps.push_back(last_jump); - p.weight *= m.JumpProbability(last_jump, src.size()); - } - } - } - } - cerr << "number of edges to consider: " << ss.size() << endl; - const int sampled = rng.SelectSample(ss); - prob_t q_n = ss[sampled] / z; - p = tmp_p[sampled]; - //m.IncrementRule(*p.rules.back()); - p.weight *= p.gamma_last / q_n; - cerr << "[w=" << log(p.weight) << "]\tsampled rule: " << p.rules.back()->AsString() << endl; - cerr << p << endl; - } - } // loop over particles (pi = 0 .. particles) - if (done_nothing) all_complete = true; - } - pfss.clear(); - for (int i = 0; i < lps.size(); ++i) - pfss.add(lps[i].weight); - const int sampled = rng.SelectSample(pfss); - ps[ci] = lps[sampled]; - m.IncrementRules(lps[sampled].rules); - m.IncrementJumps(lps[sampled].src_jumps, src.size()); - for (int i = 0; i < lps[sampled].rules.size(); ++i) { cerr << "S:\t" << lps[sampled].rules[i]->AsString() << "\n"; } - cerr << "tmp-LLH: " << log(m.Likelihood()) << endl; - } - cerr << "LLH: " << log(m.Likelihood()) << endl; - for (int sni = 0; sni < 5; ++sni) { - for (int i = 0; i < ps[sni].rules.size(); ++i) { cerr << "\t" << ps[sni].rules[i]->AsString() << endl; } - } - } - return 0; -} - diff --git a/gi/pf/pfdist.new.cc b/gi/pf/pfdist.new.cc deleted file mode 100644 index 3169eb75..00000000 --- a/gi/pf/pfdist.new.cc +++ /dev/null @@ -1,620 +0,0 @@ -#include <iostream> -#include <tr1/memory> -#include <queue> - -#include <boost/functional.hpp> -#include <boost/program_options.hpp> -#include <boost/program_options/variables_map.hpp> - -#include "base_measures.h" -#include "reachability.h" -#include "viterbi.h" -#include "hg.h" -#include "trule.h" -#include "tdict.h" -#include "filelib.h" -#include "dict.h" -#include "sampler.h" -#include "ccrp_nt.h" -#include "ccrp_onetable.h" - -using namespace std; -using namespace tr1; -namespace po = boost::program_options; - -shared_ptr<MT19937> prng; - -size_t hash_value(const TRule& r) { - size_t h = boost::hash_value(r.e_); - boost::hash_combine(h, -r.lhs_); - boost::hash_combine(h, boost::hash_value(r.f_)); - return h; -} - -bool operator==(const TRule& a, const TRule& b) { - return (a.lhs_ == b.lhs_ && a.e_ == b.e_ && a.f_ == b.f_); -} - -void InitCommandLine(int argc, char** argv, po::variables_map* conf) { - po::options_description opts("Configuration options"); - opts.add_options() - ("samples,s",po::value<unsigned>()->default_value(1000),"Number of samples") - ("particles,p",po::value<unsigned>()->default_value(25),"Number of particles") - ("input,i",po::value<string>(),"Read parallel data from") - ("max_src_phrase",po::value<unsigned>()->default_value(5),"Maximum length of source language phrases") - ("max_trg_phrase",po::value<unsigned>()->default_value(5),"Maximum length of target language phrases") - ("model1,m",po::value<string>(),"Model 1 parameters (used in base distribution)") - ("inverse_model1,M",po::value<string>(),"Inverse Model 1 parameters (used in backward estimate)") - ("model1_interpolation_weight",po::value<double>()->default_value(0.95),"Mixing proportion of model 1 with uniform target distribution") - ("random_seed,S",po::value<uint32_t>(), "Random seed"); - po::options_description clo("Command line options"); - clo.add_options() - ("config", po::value<string>(), "Configuration file") - ("help,h", "Print this help message and exit"); - po::options_description dconfig_options, dcmdline_options; - dconfig_options.add(opts); - dcmdline_options.add(opts).add(clo); - - po::store(parse_command_line(argc, argv, dcmdline_options), *conf); - if (conf->count("config")) { - ifstream config((*conf)["config"].as<string>().c_str()); - po::store(po::parse_config_file(config, dconfig_options), *conf); - } - po::notify(*conf); - - if (conf->count("help") || (conf->count("input") == 0)) { - cerr << dcmdline_options << endl; - exit(1); - } -} - -void ReadParallelCorpus(const string& filename, - vector<vector<WordID> >* f, - vector<vector<WordID> >* e, - set<WordID>* vocab_f, - set<WordID>* vocab_e) { - f->clear(); - e->clear(); - vocab_f->clear(); - vocab_e->clear(); - istream* in; - if (filename == "-") - in = &cin; - else - in = new ifstream(filename.c_str()); - assert(*in); - string line; - const WordID kDIV = TD::Convert("|||"); - vector<WordID> tmp; - while(*in) { - getline(*in, line); - if (line.empty() && !*in) break; - e->push_back(vector<int>()); - f->push_back(vector<int>()); - vector<int>& le = e->back(); - vector<int>& lf = f->back(); - tmp.clear(); - TD::ConvertSentence(line, &tmp); - bool isf = true; - for (unsigned i = 0; i < tmp.size(); ++i) { - const int cur = tmp[i]; - if (isf) { - if (kDIV == cur) { isf = false; } else { - lf.push_back(cur); - vocab_f->insert(cur); - } - } else { - assert(cur != kDIV); - le.push_back(cur); - vocab_e->insert(cur); - } - } - assert(isf == false); - } - if (in != &cin) delete in; -} - -#if 0 -struct MyConditionalModel { - MyConditionalModel(PhraseConditionalBase& rcp0) : rp0(&rcp0), base(prob_t::One()), src_phrases(1,1), src_jumps(200, CCRP_NoTable<int>(1,1)) {} - - prob_t srcp0(const vector<WordID>& src) const { - prob_t p(1.0 / 3000.0); - p.poweq(src.size()); - prob_t lenp; lenp.logeq(log_poisson(src.size(), 1.0)); - p *= lenp; - return p; - } - - void DecrementRule(const TRule& rule) { - const RuleCRPMap::iterator it = rules.find(rule.f_); - assert(it != rules.end()); - if (it->second.decrement(rule)) { - base /= (*rp0)(rule); - if (it->second.num_customers() == 0) - rules.erase(it); - } - if (src_phrases.decrement(rule.f_)) - base /= srcp0(rule.f_); - } - - void IncrementRule(const TRule& rule) { - RuleCRPMap::iterator it = rules.find(rule.f_); - if (it == rules.end()) - it = rules.insert(make_pair(rule.f_, CCRP_NoTable<TRule>(1,1))).first; - if (it->second.increment(rule)) { - base *= (*rp0)(rule); - } - if (src_phrases.increment(rule.f_)) - base *= srcp0(rule.f_); - } - - void IncrementRules(const vector<TRulePtr>& rules) { - for (int i = 0; i < rules.size(); ++i) - IncrementRule(*rules[i]); - } - - void DecrementRules(const vector<TRulePtr>& rules) { - for (int i = 0; i < rules.size(); ++i) - DecrementRule(*rules[i]); - } - - void IncrementJump(int dist, unsigned src_len) { - assert(src_len > 0); - if (src_jumps[src_len].increment(dist)) - base *= jp0(dist, src_len); - } - - void DecrementJump(int dist, unsigned src_len) { - assert(src_len > 0); - if (src_jumps[src_len].decrement(dist)) - base /= jp0(dist, src_len); - } - - void IncrementJumps(const vector<int>& js, unsigned src_len) { - for (unsigned i = 0; i < js.size(); ++i) - IncrementJump(js[i], src_len); - } - - void DecrementJumps(const vector<int>& js, unsigned src_len) { - for (unsigned i = 0; i < js.size(); ++i) - DecrementJump(js[i], src_len); - } - - // p(jump = dist | src_len , z) - prob_t JumpProbability(int dist, unsigned src_len) { - const prob_t p0 = jp0(dist, src_len); - const double lp = src_jumps[src_len].logprob(dist, log(p0)); - prob_t q; q.logeq(lp); - return q; - } - - // p(rule.f_ | z) * p(rule.e_ | rule.f_ , z) - prob_t RuleProbability(const TRule& rule) const { - const prob_t p0 = (*rp0)(rule); - prob_t srcp; srcp.logeq(src_phrases.logprob(rule.f_, log(srcp0(rule.f_)))); - const RuleCRPMap::const_iterator it = rules.find(rule.f_); - if (it == rules.end()) return srcp * p0; - const double lp = it->second.logprob(rule, log(p0)); - prob_t q; q.logeq(lp); - return q * srcp; - } - - prob_t Likelihood() const { - prob_t p = base; - for (RuleCRPMap::const_iterator it = rules.begin(); - it != rules.end(); ++it) { - prob_t cl; cl.logeq(it->second.log_crp_prob()); - p *= cl; - } - for (unsigned l = 1; l < src_jumps.size(); ++l) { - if (src_jumps[l].num_customers() > 0) { - prob_t q; - q.logeq(src_jumps[l].log_crp_prob()); - p *= q; - } - } - return p; - } - - JumpBase jp0; - const PhraseConditionalBase* rp0; - prob_t base; - typedef unordered_map<vector<WordID>, CCRP_NoTable<TRule>, boost::hash<vector<WordID> > > RuleCRPMap; - RuleCRPMap rules; - CCRP_NoTable<vector<WordID> > src_phrases; - vector<CCRP_NoTable<int> > src_jumps; -}; - -#endif - -struct MyJointModel { - MyJointModel(PhraseJointBase& rcp0) : - rp0(rcp0), base(prob_t::One()), rules(1,1), src_jumps(200, CCRP_NoTable<int>(1,1)) {} - - void DecrementRule(const TRule& rule) { - if (rules.decrement(rule)) - base /= rp0(rule); - } - - void IncrementRule(const TRule& rule) { - if (rules.increment(rule)) - base *= rp0(rule); - } - - void IncrementRules(const vector<TRulePtr>& rules) { - for (int i = 0; i < rules.size(); ++i) - IncrementRule(*rules[i]); - } - - void DecrementRules(const vector<TRulePtr>& rules) { - for (int i = 0; i < rules.size(); ++i) - DecrementRule(*rules[i]); - } - - void IncrementJump(int dist, unsigned src_len) { - assert(src_len > 0); - if (src_jumps[src_len].increment(dist)) - base *= jp0(dist, src_len); - } - - void DecrementJump(int dist, unsigned src_len) { - assert(src_len > 0); - if (src_jumps[src_len].decrement(dist)) - base /= jp0(dist, src_len); - } - - void IncrementJumps(const vector<int>& js, unsigned src_len) { - for (unsigned i = 0; i < js.size(); ++i) - IncrementJump(js[i], src_len); - } - - void DecrementJumps(const vector<int>& js, unsigned src_len) { - for (unsigned i = 0; i < js.size(); ++i) - DecrementJump(js[i], src_len); - } - - // p(jump = dist | src_len , z) - prob_t JumpProbability(int dist, unsigned src_len) { - const prob_t p0 = jp0(dist, src_len); - const double lp = src_jumps[src_len].logprob(dist, log(p0)); - prob_t q; q.logeq(lp); - return q; - } - - // p(rule.f_ | z) * p(rule.e_ | rule.f_ , z) - prob_t RuleProbability(const TRule& rule) const { - prob_t p; p.logeq(rules.logprob(rule, log(rp0(rule)))); - return p; - } - - prob_t Likelihood() const { - prob_t p = base; - prob_t q; q.logeq(rules.log_crp_prob()); - p *= q; - for (unsigned l = 1; l < src_jumps.size(); ++l) { - if (src_jumps[l].num_customers() > 0) { - prob_t q; - q.logeq(src_jumps[l].log_crp_prob()); - p *= q; - } - } - return p; - } - - JumpBase jp0; - const PhraseJointBase& rp0; - prob_t base; - CCRP_NoTable<TRule> rules; - vector<CCRP_NoTable<int> > src_jumps; -}; - -struct BackwardEstimate { - BackwardEstimate(const Model1& m1, const vector<WordID>& src, const vector<WordID>& trg) : - model1_(m1), src_(src), trg_(trg) { - } - const prob_t& operator()(const vector<bool>& src_cov, unsigned trg_cov) const { - assert(src_.size() == src_cov.size()); - assert(trg_cov <= trg_.size()); - prob_t& e = cache_[src_cov][trg_cov]; - if (e.is_0()) { - if (trg_cov == trg_.size()) { e = prob_t::One(); return e; } - vector<WordID> r(src_.size() + 1); r.clear(); - r.push_back(0); // NULL word - for (int i = 0; i < src_cov.size(); ++i) - if (!src_cov[i]) r.push_back(src_[i]); - const prob_t uniform_alignment(1.0 / r.size()); - e.logeq(log_poisson(trg_.size() - trg_cov, r.size() - 1)); // p(trg len remaining | src len remaining) - for (unsigned j = trg_cov; j < trg_.size(); ++j) { - prob_t p; - for (unsigned i = 0; i < r.size(); ++i) - p += model1_(r[i], trg_[j]); - if (p.is_0()) { - cerr << "ERROR: p(" << TD::Convert(trg_[j]) << " | " << TD::GetString(r) << ") = 0!\n"; - abort(); - } - p *= uniform_alignment; - e *= p; - } - } - return e; - } - const Model1& model1_; - const vector<WordID>& src_; - const vector<WordID>& trg_; - mutable unordered_map<vector<bool>, map<unsigned, prob_t>, boost::hash<vector<bool> > > cache_; -}; - -struct BackwardEstimateSym { - BackwardEstimateSym(const Model1& m1, - const Model1& invm1, const vector<WordID>& src, const vector<WordID>& trg) : - model1_(m1), invmodel1_(invm1), src_(src), trg_(trg) { - } - const prob_t& operator()(const vector<bool>& src_cov, unsigned trg_cov) const { - assert(src_.size() == src_cov.size()); - assert(trg_cov <= trg_.size()); - prob_t& e = cache_[src_cov][trg_cov]; - if (e.is_0()) { - if (trg_cov == trg_.size()) { e = prob_t::One(); return e; } - vector<WordID> r(src_.size() + 1); r.clear(); - for (int i = 0; i < src_cov.size(); ++i) - if (!src_cov[i]) r.push_back(src_[i]); - r.push_back(0); // NULL word - const prob_t uniform_alignment(1.0 / r.size()); - e.logeq(log_poisson(trg_.size() - trg_cov, r.size() - 1)); // p(trg len remaining | src len remaining) - for (unsigned j = trg_cov; j < trg_.size(); ++j) { - prob_t p; - for (unsigned i = 0; i < r.size(); ++i) - p += model1_(r[i], trg_[j]); - if (p.is_0()) { - cerr << "ERROR: p(" << TD::Convert(trg_[j]) << " | " << TD::GetString(r) << ") = 0!\n"; - abort(); - } - p *= uniform_alignment; - e *= p; - } - r.pop_back(); - const prob_t inv_uniform(1.0 / (trg_.size() - trg_cov + 1.0)); - prob_t inv; - inv.logeq(log_poisson(r.size(), trg_.size() - trg_cov)); - for (unsigned i = 0; i < r.size(); ++i) { - prob_t p; - for (unsigned j = trg_cov - 1; j < trg_.size(); ++j) - p += invmodel1_(j < trg_cov ? 0 : trg_[j], r[i]); - if (p.is_0()) { - cerr << "ERROR: p_inv(" << TD::Convert(r[i]) << " | " << TD::GetString(trg_) << ") = 0!\n"; - abort(); - } - p *= inv_uniform; - inv *= p; - } - prob_t x = pow(e * inv, 0.5); - e = x; - //cerr << "Forward: " << log(e) << "\tBackward: " << log(inv) << "\t prop: " << log(x) << endl; - } - return e; - } - const Model1& model1_; - const Model1& invmodel1_; - const vector<WordID>& src_; - const vector<WordID>& trg_; - mutable unordered_map<vector<bool>, map<unsigned, prob_t>, boost::hash<vector<bool> > > cache_; -}; - -struct Particle { - Particle() : weight(prob_t::One()), src_cov(), trg_cov(), prev_pos(-1) {} - prob_t weight; - prob_t gamma_last; - vector<int> src_jumps; - vector<TRulePtr> rules; - vector<bool> src_cv; - int src_cov; - int trg_cov; - int prev_pos; -}; - -ostream& operator<<(ostream& o, const vector<bool>& v) { - for (int i = 0; i < v.size(); ++i) - o << (v[i] ? '1' : '0'); - return o; -} -ostream& operator<<(ostream& o, const Particle& p) { - o << "[cv=" << p.src_cv << " src_cov=" << p.src_cov << " trg_cov=" << p.trg_cov << " last_pos=" << p.prev_pos << " num_rules=" << p.rules.size() << " w=" << log(p.weight) << ']'; - return o; -} - -void FilterCrapParticlesAndReweight(vector<Particle>* pps) { - vector<Particle>& ps = *pps; - SampleSet<prob_t> ss; - for (int i = 0; i < ps.size(); ++i) - ss.add(ps[i].weight); - vector<Particle> nps; nps.reserve(ps.size()); - const prob_t uniform_weight(1.0 / ps.size()); - for (int i = 0; i < ps.size(); ++i) { - nps.push_back(ps[prng->SelectSample(ss)]); - nps[i].weight = uniform_weight; - } - nps.swap(ps); -} - -int main(int argc, char** argv) { - po::variables_map conf; - InitCommandLine(argc, argv, &conf); - const unsigned kMAX_TRG_PHRASE = conf["max_trg_phrase"].as<unsigned>(); - const unsigned kMAX_SRC_PHRASE = conf["max_src_phrase"].as<unsigned>(); - const unsigned particles = conf["particles"].as<unsigned>(); - const unsigned samples = conf["samples"].as<unsigned>(); - - if (!conf.count("model1")) { - cerr << argv[0] << "Please use --model1 to specify model 1 parameters\n"; - return 1; - } - if (conf.count("random_seed")) - prng.reset(new MT19937(conf["random_seed"].as<uint32_t>())); - else - prng.reset(new MT19937); - MT19937& rng = *prng; - - vector<vector<WordID> > corpuse, corpusf; - set<WordID> vocabe, vocabf; - cerr << "Reading corpus...\n"; - ReadParallelCorpus(conf["input"].as<string>(), &corpusf, &corpuse, &vocabf, &vocabe); - cerr << "F-corpus size: " << corpusf.size() << " sentences\t (" << vocabf.size() << " word types)\n"; - cerr << "E-corpus size: " << corpuse.size() << " sentences\t (" << vocabe.size() << " word types)\n"; - assert(corpusf.size() == corpuse.size()); - - const int kLHS = -TD::Convert("X"); - Model1 m1(conf["model1"].as<string>()); - Model1 invm1(conf["inverse_model1"].as<string>()); - -#if 0 - PhraseConditionalBase lp0(m1, conf["model1_interpolation_weight"].as<double>(), vocabe.size()); - MyConditionalModel m(lp0); -#else - PhraseJointBase lp0(m1, conf["model1_interpolation_weight"].as<double>(), vocabe.size(), vocabf.size()); - MyJointModel m(lp0); -#endif - - cerr << "Initializing reachability limits...\n"; - vector<Particle> ps(corpusf.size()); - vector<Reachability> reaches; reaches.reserve(corpusf.size()); - for (int ci = 0; ci < corpusf.size(); ++ci) - reaches.push_back(Reachability(corpusf[ci].size(), - corpuse[ci].size(), - kMAX_SRC_PHRASE, - kMAX_TRG_PHRASE)); - cerr << "Sampling...\n"; - vector<Particle> tmp_p(10000); // work space - SampleSet<prob_t> pfss; - for (int SS=0; SS < samples; ++SS) { - for (int ci = 0; ci < corpusf.size(); ++ci) { - vector<int>& src = corpusf[ci]; - vector<int>& trg = corpuse[ci]; - m.DecrementRules(ps[ci].rules); - m.DecrementJumps(ps[ci].src_jumps, src.size()); - - //BackwardEstimate be(m1, src, trg); - BackwardEstimateSym be(m1, invm1, src, trg); - const Reachability& r = reaches[ci]; - vector<Particle> lps(particles); - - for (int pi = 0; pi < particles; ++pi) { - Particle& p = lps[pi]; - p.src_cv.resize(src.size(), false); - } - - bool all_complete = false; - while(!all_complete) { - SampleSet<prob_t> ss; - - // all particles have now been extended a bit, we will reweight them now - if (lps[0].trg_cov > 0) - FilterCrapParticlesAndReweight(&lps); - - // loop over all particles and extend them - bool done_nothing = true; - for (int pi = 0; pi < particles; ++pi) { - Particle& p = lps[pi]; - int tic = 0; - const int rejuv_freq = 1; - while(p.trg_cov < trg.size() && tic < rejuv_freq) { - ++tic; - done_nothing = false; - ss.clear(); - TRule x; x.lhs_ = kLHS; - prob_t z; - int first_uncovered = src.size(); - int last_uncovered = -1; - for (int i = 0; i < src.size(); ++i) { - const bool is_uncovered = !p.src_cv[i]; - if (i < first_uncovered && is_uncovered) first_uncovered = i; - if (is_uncovered && i > last_uncovered) last_uncovered = i; - } - assert(last_uncovered > -1); - assert(first_uncovered < src.size()); - - for (int trg_len = 1; trg_len <= kMAX_TRG_PHRASE; ++trg_len) { - x.e_.push_back(trg[trg_len - 1 + p.trg_cov]); - for (int src_len = 1; src_len <= kMAX_SRC_PHRASE; ++src_len) { - if (!r.edges[p.src_cov][p.trg_cov][src_len][trg_len]) continue; - - const int last_possible_start = last_uncovered - src_len + 1; - assert(last_possible_start >= 0); - //cerr << src_len << "," << trg_len << " is allowed. E=" << TD::GetString(x.e_) << endl; - //cerr << " first_uncovered=" << first_uncovered << " last_possible_start=" << last_possible_start << endl; - for (int i = first_uncovered; i <= last_possible_start; ++i) { - if (p.src_cv[i]) continue; - assert(ss.size() < tmp_p.size()); // if fails increase tmp_p size - Particle& np = tmp_p[ss.size()]; - np = p; - x.f_.clear(); - int gap_add = 0; - bool bad = false; - prob_t jp = prob_t::One(); - int prev_pos = p.prev_pos; - for (int j = 0; j < src_len; ++j) { - if ((j + i + gap_add) == src.size()) { bad = true; break; } - while ((i+j+gap_add) < src.size() && p.src_cv[i + j + gap_add]) { ++gap_add; } - if ((j + i + gap_add) == src.size()) { bad = true; break; } - np.src_cv[i + j + gap_add] = true; - x.f_.push_back(src[i + j + gap_add]); - jp *= m.JumpProbability(i + j + gap_add - prev_pos, src.size()); - int jump = i + j + gap_add - prev_pos; - assert(jump != 0); - np.src_jumps.push_back(jump); - prev_pos = i + j + gap_add; - } - if (bad) continue; - np.prev_pos = prev_pos; - np.src_cov += x.f_.size(); - np.trg_cov += x.e_.size(); - if (x.f_.size() != src_len) continue; - prob_t rp = m.RuleProbability(x); - np.gamma_last = rp * jp; - const prob_t u = pow(np.gamma_last * be(np.src_cv, np.trg_cov), 0.2); - //cerr << "**rule=" << x << endl; - //cerr << " u=" << log(u) << " rule=" << rp << " jump=" << jp << endl; - ss.add(u); - np.rules.push_back(TRulePtr(new TRule(x))); - z += u; - - const bool completed = (p.trg_cov == trg.size()); - if (completed) { - int last_jump = src.size() - p.prev_pos; - assert(last_jump > 0); - p.src_jumps.push_back(last_jump); - p.weight *= m.JumpProbability(last_jump, src.size()); - } - } - } - } - cerr << "number of edges to consider: " << ss.size() << endl; - const int sampled = rng.SelectSample(ss); - prob_t q_n = ss[sampled] / z; - p = tmp_p[sampled]; - //m.IncrementRule(*p.rules.back()); - p.weight *= p.gamma_last / q_n; - cerr << "[w=" << log(p.weight) << "]\tsampled rule: " << p.rules.back()->AsString() << endl; - cerr << p << endl; - } - } // loop over particles (pi = 0 .. particles) - if (done_nothing) all_complete = true; - } - pfss.clear(); - for (int i = 0; i < lps.size(); ++i) - pfss.add(lps[i].weight); - const int sampled = rng.SelectSample(pfss); - ps[ci] = lps[sampled]; - m.IncrementRules(lps[sampled].rules); - m.IncrementJumps(lps[sampled].src_jumps, src.size()); - for (int i = 0; i < lps[sampled].rules.size(); ++i) { cerr << "S:\t" << lps[sampled].rules[i]->AsString() << "\n"; } - cerr << "tmp-LLH: " << log(m.Likelihood()) << endl; - } - cerr << "LLH: " << log(m.Likelihood()) << endl; - for (int sni = 0; sni < 5; ++sni) { - for (int i = 0; i < ps[sni].rules.size(); ++i) { cerr << "\t" << ps[sni].rules[i]->AsString() << endl; } - } - } - return 0; -} - diff --git a/gi/pf/pfnaive.cc b/gi/pf/pfnaive.cc deleted file mode 100644 index 958ec4e2..00000000 --- a/gi/pf/pfnaive.cc +++ /dev/null @@ -1,284 +0,0 @@ -#include <iostream> -#include <tr1/memory> -#include <queue> - -#include <boost/functional.hpp> -#include <boost/program_options.hpp> -#include <boost/program_options/variables_map.hpp> - -#include "pf.h" -#include "base_distributions.h" -#include "monotonic_pseg.h" -#include "reachability.h" -#include "viterbi.h" -#include "hg.h" -#include "trule.h" -#include "tdict.h" -#include "filelib.h" -#include "dict.h" -#include "sampler.h" -#include "ccrp_nt.h" -#include "ccrp_onetable.h" -#include "corpus.h" - -using namespace std; -using namespace tr1; -namespace po = boost::program_options; - -boost::shared_ptr<MT19937> prng; - -void InitCommandLine(int argc, char** argv, po::variables_map* conf) { - po::options_description opts("Configuration options"); - opts.add_options() - ("samples,s",po::value<unsigned>()->default_value(1000),"Number of samples") - ("particles,p",po::value<unsigned>()->default_value(30),"Number of particles") - ("filter_frequency,f",po::value<unsigned>()->default_value(5),"Number of time steps between filterings") - ("input,i",po::value<string>(),"Read parallel data from") - ("max_src_phrase",po::value<unsigned>()->default_value(5),"Maximum length of source language phrases") - ("max_trg_phrase",po::value<unsigned>()->default_value(5),"Maximum length of target language phrases") - ("model1,m",po::value<string>(),"Model 1 parameters (used in base distribution)") - ("inverse_model1,M",po::value<string>(),"Inverse Model 1 parameters (used in backward estimate)") - ("model1_interpolation_weight",po::value<double>()->default_value(0.95),"Mixing proportion of model 1 with uniform target distribution") - ("random_seed,S",po::value<uint32_t>(), "Random seed"); - po::options_description clo("Command line options"); - clo.add_options() - ("config", po::value<string>(), "Configuration file") - ("help,h", "Print this help message and exit"); - po::options_description dconfig_options, dcmdline_options; - dconfig_options.add(opts); - dcmdline_options.add(opts).add(clo); - - po::store(parse_command_line(argc, argv, dcmdline_options), *conf); - if (conf->count("config")) { - ifstream config((*conf)["config"].as<string>().c_str()); - po::store(po::parse_config_file(config, dconfig_options), *conf); - } - po::notify(*conf); - - if (conf->count("help") || (conf->count("input") == 0)) { - cerr << dcmdline_options << endl; - exit(1); - } -} - -struct BackwardEstimateSym { - BackwardEstimateSym(const Model1& m1, - const Model1& invm1, const vector<WordID>& src, const vector<WordID>& trg) : - model1_(m1), invmodel1_(invm1), src_(src), trg_(trg) { - } - const prob_t& operator()(unsigned src_cov, unsigned trg_cov) const { - assert(src_cov <= src_.size()); - assert(trg_cov <= trg_.size()); - prob_t& e = cache_[src_cov][trg_cov]; - if (e.is_0()) { - if (trg_cov == trg_.size()) { e = prob_t::One(); return e; } - vector<WordID> r(src_.size() + 1); r.clear(); - for (int i = src_cov; i < src_.size(); ++i) - r.push_back(src_[i]); - r.push_back(0); // NULL word - const prob_t uniform_alignment(1.0 / r.size()); - e.logeq(Md::log_poisson(trg_.size() - trg_cov, r.size() - 1)); // p(trg len remaining | src len remaining) - for (unsigned j = trg_cov; j < trg_.size(); ++j) { - prob_t p; - for (unsigned i = 0; i < r.size(); ++i) - p += model1_(r[i], trg_[j]); - if (p.is_0()) { - cerr << "ERROR: p(" << TD::Convert(trg_[j]) << " | " << TD::GetString(r) << ") = 0!\n"; - abort(); - } - p *= uniform_alignment; - e *= p; - } - r.pop_back(); - const prob_t inv_uniform(1.0 / (trg_.size() - trg_cov + 1.0)); - prob_t inv; - inv.logeq(Md::log_poisson(r.size(), trg_.size() - trg_cov)); - for (unsigned i = 0; i < r.size(); ++i) { - prob_t p; - for (unsigned j = trg_cov - 1; j < trg_.size(); ++j) - p += invmodel1_(j < trg_cov ? 0 : trg_[j], r[i]); - if (p.is_0()) { - cerr << "ERROR: p_inv(" << TD::Convert(r[i]) << " | " << TD::GetString(trg_) << ") = 0!\n"; - abort(); - } - p *= inv_uniform; - inv *= p; - } - prob_t x = pow(e * inv, 0.5); - e = x; - //cerr << "Forward: " << log(e) << "\tBackward: " << log(inv) << "\t prop: " << log(x) << endl; - } - return e; - } - const Model1& model1_; - const Model1& invmodel1_; - const vector<WordID>& src_; - const vector<WordID>& trg_; - mutable unordered_map<unsigned, map<unsigned, prob_t> > cache_; -}; - -struct Particle { - Particle() : weight(prob_t::One()), src_cov(), trg_cov() {} - prob_t weight; - prob_t gamma_last; - vector<TRulePtr> rules; - int src_cov; - int trg_cov; -}; - -ostream& operator<<(ostream& o, const vector<bool>& v) { - for (int i = 0; i < v.size(); ++i) - o << (v[i] ? '1' : '0'); - return o; -} -ostream& operator<<(ostream& o, const Particle& p) { - o << "[src_cov=" << p.src_cov << " trg_cov=" << p.trg_cov << " num_rules=" << p.rules.size() << " w=" << log(p.weight) << ']'; - return o; -} - -int main(int argc, char** argv) { - po::variables_map conf; - InitCommandLine(argc, argv, &conf); - const unsigned kMAX_TRG_PHRASE = conf["max_trg_phrase"].as<unsigned>(); - const unsigned kMAX_SRC_PHRASE = conf["max_src_phrase"].as<unsigned>(); - const unsigned particles = conf["particles"].as<unsigned>(); - const unsigned samples = conf["samples"].as<unsigned>(); - const unsigned rejuv_freq = conf["filter_frequency"].as<unsigned>(); - - if (!conf.count("model1")) { - cerr << argv[0] << "Please use --model1 to specify model 1 parameters\n"; - return 1; - } - if (conf.count("random_seed")) - prng.reset(new MT19937(conf["random_seed"].as<uint32_t>())); - else - prng.reset(new MT19937); - MT19937& rng = *prng; - - vector<vector<WordID> > corpuse, corpusf; - set<WordID> vocabe, vocabf; - cerr << "Reading corpus...\n"; - corpus::ReadParallelCorpus(conf["input"].as<string>(), &corpusf, &corpuse, &vocabf, &vocabe); - cerr << "F-corpus size: " << corpusf.size() << " sentences\t (" << vocabf.size() << " word types)\n"; - cerr << "E-corpus size: " << corpuse.size() << " sentences\t (" << vocabe.size() << " word types)\n"; - assert(corpusf.size() == corpuse.size()); - - const int kLHS = -TD::Convert("X"); - Model1 m1(conf["model1"].as<string>()); - Model1 invm1(conf["inverse_model1"].as<string>()); - - PhraseJointBase lp0(m1, conf["model1_interpolation_weight"].as<double>(), vocabe.size(), vocabf.size()); - PhraseJointBase_BiDir alp0(m1, invm1, conf["model1_interpolation_weight"].as<double>(), vocabe.size(), vocabf.size()); - MonotonicParallelSegementationModel<PhraseJointBase_BiDir> m(alp0); - TRule xx("[X] ||| ms. kimura ||| MS. KIMURA ||| X=0"); - cerr << xx << endl << lp0(xx) << " " << alp0(xx) << endl; - TRule xx12("[X] ||| . ||| PHARMACY . ||| X=0"); - TRule xx21("[X] ||| pharmacy . ||| . ||| X=0"); -// TRule xx22("[X] ||| . ||| . ||| X=0"); - TRule xx22("[X] ||| . ||| THE . ||| X=0"); - cerr << xx12 << "\t" << lp0(xx12) << " " << alp0(xx12) << endl; - cerr << xx21 << "\t" << lp0(xx21) << " " << alp0(xx21) << endl; - cerr << xx22 << "\t" << lp0(xx22) << " " << alp0(xx22) << endl; - - cerr << "Initializing reachability limits...\n"; - vector<Particle> ps(corpusf.size()); - vector<Reachability> reaches; reaches.reserve(corpusf.size()); - for (int ci = 0; ci < corpusf.size(); ++ci) - reaches.push_back(Reachability(corpusf[ci].size(), - corpuse[ci].size(), - kMAX_SRC_PHRASE, - kMAX_TRG_PHRASE)); - cerr << "Sampling...\n"; - vector<Particle> tmp_p(10000); // work space - SampleSet<prob_t> pfss; - SystematicResampleFilter<Particle> filter(&rng); - // MultinomialResampleFilter<Particle> filter(&rng); - for (int SS=0; SS < samples; ++SS) { - for (int ci = 0; ci < corpusf.size(); ++ci) { - vector<int>& src = corpusf[ci]; - vector<int>& trg = corpuse[ci]; - m.DecrementRulesAndStops(ps[ci].rules); - const prob_t q_stop = m.StopProbability(); - const prob_t q_cont = m.ContinueProbability(); - cerr << "P(stop)=" << q_stop << "\tP(continue)=" <<q_cont << endl; - - BackwardEstimateSym be(m1, invm1, src, trg); - const Reachability& r = reaches[ci]; - vector<Particle> lps(particles); - - bool all_complete = false; - while(!all_complete) { - SampleSet<prob_t> ss; - - // all particles have now been extended a bit, we will reweight them now - if (lps[0].trg_cov > 0) - filter(&lps); - - // loop over all particles and extend them - bool done_nothing = true; - for (int pi = 0; pi < particles; ++pi) { - Particle& p = lps[pi]; - int tic = 0; - while(p.trg_cov < trg.size() && tic < rejuv_freq) { - ++tic; - done_nothing = false; - ss.clear(); - TRule x; x.lhs_ = kLHS; - prob_t z; - - for (int trg_len = 1; trg_len <= kMAX_TRG_PHRASE; ++trg_len) { - x.e_.push_back(trg[trg_len - 1 + p.trg_cov]); - for (int src_len = 1; src_len <= kMAX_SRC_PHRASE; ++src_len) { - if (!r.edges[p.src_cov][p.trg_cov][src_len][trg_len]) continue; - - int i = p.src_cov; - assert(ss.size() < tmp_p.size()); // if fails increase tmp_p size - Particle& np = tmp_p[ss.size()]; - np = p; - x.f_.clear(); - for (int j = 0; j < src_len; ++j) - x.f_.push_back(src[i + j]); - np.src_cov += x.f_.size(); - np.trg_cov += x.e_.size(); - const bool stop_now = (np.src_cov == src_len && np.trg_cov == trg_len); - prob_t rp = m.RuleProbability(x) * (stop_now ? q_stop : q_cont); - np.gamma_last = rp; - const prob_t u = pow(np.gamma_last * pow(be(np.src_cov, np.trg_cov), 1.2), 0.1); - //cerr << "**rule=" << x << endl; - //cerr << " u=" << log(u) << " rule=" << rp << endl; - ss.add(u); - np.rules.push_back(TRulePtr(new TRule(x))); - z += u; - } - } - //cerr << "number of edges to consider: " << ss.size() << endl; - const int sampled = rng.SelectSample(ss); - prob_t q_n = ss[sampled] / z; - p = tmp_p[sampled]; - //m.IncrementRule(*p.rules.back()); - p.weight *= p.gamma_last / q_n; - //cerr << "[w=" << log(p.weight) << "]\tsampled rule: " << p.rules.back()->AsString() << endl; - //cerr << p << endl; - } - } // loop over particles (pi = 0 .. particles) - if (done_nothing) all_complete = true; - prob_t wv = prob_t::Zero(); - for (int pp = 0; pp < lps.size(); ++pp) - wv += lps[pp].weight; - for (int pp = 0; pp < lps.size(); ++pp) - lps[pp].weight /= wv; - } - pfss.clear(); - for (int i = 0; i < lps.size(); ++i) - pfss.add(lps[i].weight); - const int sampled = rng.SelectSample(pfss); - ps[ci] = lps[sampled]; - m.IncrementRulesAndStops(lps[sampled].rules); - for (int i = 0; i < lps[sampled].rules.size(); ++i) { cerr << "S:\t" << lps[sampled].rules[i]->AsString() << "\n"; } - cerr << "tmp-LLH: " << log(m.Likelihood()) << endl; - } - cerr << "LLH: " << log(m.Likelihood()) << endl; - } - return 0; -} - diff --git a/gi/pf/poisson_uniform_word_model.h b/gi/pf/poisson_uniform_word_model.h deleted file mode 100644 index 76204a0e..00000000 --- a/gi/pf/poisson_uniform_word_model.h +++ /dev/null @@ -1,50 +0,0 @@ -#ifndef _POISSON_UNIFORM_WORD_MODEL_H_ -#define _POISSON_UNIFORM_WORD_MODEL_H_ - -#include <cmath> -#include <vector> -#include "prob.h" -#include "m.h" - -// len ~ Poisson(lambda) -// for (1..len) -// e_i ~ Uniform({Vocabulary}) -struct PoissonUniformWordModel { - explicit PoissonUniformWordModel(const unsigned vocab_size, - const unsigned alphabet_size, - const double mean_len = 5) : - lh(prob_t::One()), - v0(-std::log(vocab_size)), - u0(-std::log(alphabet_size)), - mean_length(mean_len) {} - - void ResampleHyperparameters(MT19937*) {} - - inline prob_t operator()(const std::vector<WordID>& s) const { - prob_t p; - p.logeq(Md::log_poisson(s.size(), mean_length) + s.size() * u0); - //p.logeq(v0); - return p; - } - - inline void Increment(const std::vector<WordID>& w, MT19937*) { - lh *= (*this)(w); - } - - inline void Decrement(const std::vector<WordID>& w, MT19937 *) { - lh /= (*this)(w); - } - - inline prob_t Likelihood() const { return lh; } - - void Summary() const {} - - private: - - prob_t lh; // keeps track of the draws from the base distribution - const double v0; // uniform log prob of generating a word - const double u0; // uniform log prob of generating a letter - const double mean_length; // mean length of a word in the base distribution -}; - -#endif diff --git a/gi/pf/pyp_lm.cc b/gi/pf/pyp_lm.cc deleted file mode 100644 index 605d8206..00000000 --- a/gi/pf/pyp_lm.cc +++ /dev/null @@ -1,273 +0,0 @@ -#include <iostream> -#include <tr1/memory> -#include <queue> - -#include <boost/functional.hpp> -#include <boost/program_options.hpp> -#include <boost/program_options/variables_map.hpp> - -#include "gamma_poisson.h" -#include "corpus_tools.h" -#include "m.h" -#include "tdict.h" -#include "sampler.h" -#include "ccrp.h" -#include "tied_resampler.h" - -// A not very memory-efficient implementation of an N-gram LM based on PYPs -// as described in Y.-W. Teh. (2006) A Hierarchical Bayesian Language Model -// based on Pitman-Yor Processes. In Proc. ACL. - -// I use templates to handle the recursive formalation of the prior, so -// the order of the model has to be specified here, at compile time: -#define kORDER 3 - -using namespace std; -using namespace tr1; -namespace po = boost::program_options; - -boost::shared_ptr<MT19937> prng; - -void InitCommandLine(int argc, char** argv, po::variables_map* conf) { - po::options_description opts("Configuration options"); - opts.add_options() - ("samples,n",po::value<unsigned>()->default_value(300),"Number of samples") - ("train,i",po::value<string>(),"Training data file") - ("test,T",po::value<string>(),"Test data file") - ("discount_prior_a,a",po::value<double>()->default_value(1.0), "discount ~ Beta(a,b): a=this") - ("discount_prior_b,b",po::value<double>()->default_value(1.0), "discount ~ Beta(a,b): b=this") - ("strength_prior_s,s",po::value<double>()->default_value(1.0), "strength ~ Gamma(s,r): s=this") - ("strength_prior_r,r",po::value<double>()->default_value(1.0), "strength ~ Gamma(s,r): r=this") - ("random_seed,S",po::value<uint32_t>(), "Random seed"); - po::options_description clo("Command line options"); - clo.add_options() - ("config", po::value<string>(), "Configuration file") - ("help", "Print this help message and exit"); - po::options_description dconfig_options, dcmdline_options; - dconfig_options.add(opts); - dcmdline_options.add(opts).add(clo); - - po::store(parse_command_line(argc, argv, dcmdline_options), *conf); - if (conf->count("config")) { - ifstream config((*conf)["config"].as<string>().c_str()); - po::store(po::parse_config_file(config, dconfig_options), *conf); - } - po::notify(*conf); - - if (conf->count("help") || (conf->count("train") == 0)) { - cerr << dcmdline_options << endl; - exit(1); - } -} - -// uniform distribution over a fixed vocabulary -struct UniformVocabulary { - UniformVocabulary(unsigned vs, double, double, double, double) : p0(1.0 / vs), draws() {} - void increment(WordID, const vector<WordID>&, MT19937*) { ++draws; } - void decrement(WordID, const vector<WordID>&, MT19937*) { --draws; assert(draws >= 0); } - double prob(WordID, const vector<WordID>&) const { return p0; } - void resample_hyperparameters(MT19937*) {} - double log_likelihood() const { return draws * log(p0); } - const double p0; - int draws; -}; - -// Lord Rothschild. 1986. THE DISTRIBUTION OF ENGLISH DICTIONARY WORD LENGTHS. -// Journal of Statistical Planning and Inference 14 (1986) 311-322 -struct PoissonLengthUniformCharWordModel { - explicit PoissonLengthUniformCharWordModel(unsigned vocab_size, double, double, double, double) : plen(5,5), uc(-log(95)), llh() {} - void increment(WordID w, const vector<WordID>& v, MT19937*) { - llh += log(prob(w, v)); // this isn't quite right - plen.increment(TD::Convert(w).size() - 1); - } - void decrement(WordID w, const vector<WordID>& v, MT19937*) { - plen.decrement(TD::Convert(w).size() - 1); - llh -= log(prob(w, v)); // this isn't quite right - } - double prob(WordID w, const vector<WordID>&) const { - const unsigned len = TD::Convert(w).size(); - return plen.prob(len - 1) * exp(uc * len); - } - double log_likelihood() const { return llh; } - void resample_hyperparameters(MT19937*) {} - GammaPoisson plen; - const double uc; - double llh; -}; - -struct PYPAdaptedPoissonLengthUniformCharWordModel { - explicit PYPAdaptedPoissonLengthUniformCharWordModel(unsigned vocab_size, double, double, double, double) : - base(vocab_size,1,1,1,1), - crp(1,1,1,1) {} - void increment(WordID w, const vector<WordID>& v, MT19937* rng) { - double p0 = base.prob(w, v); - if (crp.increment(w, p0, rng)) - base.increment(w, v, rng); - } - void decrement(WordID w, const vector<WordID>& v, MT19937* rng) { - if (crp.decrement(w, rng)) - base.decrement(w, v, rng); - } - double prob(WordID w, const vector<WordID>& v) const { - double p0 = base.prob(w, v); - return crp.prob(w, p0); - } - double log_likelihood() const { return crp.log_crp_prob() + base.log_likelihood(); } - void resample_hyperparameters(MT19937* rng) { crp.resample_hyperparameters(rng); } - PoissonLengthUniformCharWordModel base; - CCRP<WordID> crp; -}; - -template <unsigned N> struct PYPLM; - -#if 1 -template<> struct PYPLM<0> : public UniformVocabulary { - PYPLM(unsigned vs, double a, double b, double c, double d) : - UniformVocabulary(vs, a, b, c, d) {} -}; -#else -#if 0 -template<> struct PYPLM<0> : public PoissonLengthUniformCharWordModel { - PYPLM(unsigned vs, double a, double b, double c, double d) : - PoissonLengthUniformCharWordModel(vs, a, b, c, d) {} -}; -#else -template<> struct PYPLM<0> : public PYPAdaptedPoissonLengthUniformCharWordModel { - PYPLM(unsigned vs, double a, double b, double c, double d) : - PYPAdaptedPoissonLengthUniformCharWordModel(vs, a, b, c, d) {} -}; -#endif -#endif - -// represents an N-gram LM -template <unsigned N> struct PYPLM { - PYPLM(unsigned vs, double da, double db, double ss, double sr) : - backoff(vs, da, db, ss, sr), - tr(da, db, ss, sr, 0.8, 1.0), - lookup(N-1) {} - void increment(WordID w, const vector<WordID>& context, MT19937* rng) { - const double bo = backoff.prob(w, context); - for (unsigned i = 0; i < N-1; ++i) - lookup[i] = context[context.size() - 1 - i]; - typename unordered_map<vector<WordID>, CCRP<WordID>, boost::hash<vector<WordID> > >::iterator it = p.find(lookup); - if (it == p.end()) { - it = p.insert(make_pair(lookup, CCRP<WordID>(0.5,1))).first; - tr.Add(&it->second); // add to resampler - } - if (it->second.increment(w, bo, rng)) - backoff.increment(w, context, rng); - } - void decrement(WordID w, const vector<WordID>& context, MT19937* rng) { - for (unsigned i = 0; i < N-1; ++i) - lookup[i] = context[context.size() - 1 - i]; - typename unordered_map<vector<WordID>, CCRP<WordID>, boost::hash<vector<WordID> > >::iterator it = p.find(lookup); - assert(it != p.end()); - if (it->second.decrement(w, rng)) - backoff.decrement(w, context, rng); - } - double prob(WordID w, const vector<WordID>& context) const { - const double bo = backoff.prob(w, context); - for (unsigned i = 0; i < N-1; ++i) - lookup[i] = context[context.size() - 1 - i]; - typename unordered_map<vector<WordID>, CCRP<WordID>, boost::hash<vector<WordID> > >::const_iterator it = p.find(lookup); - if (it == p.end()) return bo; - return it->second.prob(w, bo); - } - - double log_likelihood() const { - double llh = backoff.log_likelihood(); - typename unordered_map<vector<WordID>, CCRP<WordID>, boost::hash<vector<WordID> > >::const_iterator it; - for (it = p.begin(); it != p.end(); ++it) - llh += it->second.log_crp_prob(); - llh += tr.LogLikelihood(); - return llh; - } - - void resample_hyperparameters(MT19937* rng) { - tr.ResampleHyperparameters(rng); - backoff.resample_hyperparameters(rng); - } - - PYPLM<N-1> backoff; - TiedResampler<CCRP<WordID> > tr; - double discount_a, discount_b, strength_s, strength_r; - double d, strength; - mutable vector<WordID> lookup; // thread-local - unordered_map<vector<WordID>, CCRP<WordID>, boost::hash<vector<WordID> > > p; -}; - -int main(int argc, char** argv) { - po::variables_map conf; - - InitCommandLine(argc, argv, &conf); - const unsigned samples = conf["samples"].as<unsigned>(); - if (conf.count("random_seed")) - prng.reset(new MT19937(conf["random_seed"].as<uint32_t>())); - else - prng.reset(new MT19937); - MT19937& rng = *prng; - vector<vector<WordID> > corpuse; - set<WordID> vocabe; - const WordID kEOS = TD::Convert("</s>"); - cerr << "Reading corpus...\n"; - CorpusTools::ReadFromFile(conf["train"].as<string>(), &corpuse, &vocabe); - cerr << "E-corpus size: " << corpuse.size() << " sentences\t (" << vocabe.size() << " word types)\n"; - vector<vector<WordID> > test; - if (conf.count("test")) - CorpusTools::ReadFromFile(conf["test"].as<string>(), &test); - else - test = corpuse; - PYPLM<kORDER> lm(vocabe.size(), - conf["discount_prior_a"].as<double>(), - conf["discount_prior_b"].as<double>(), - conf["strength_prior_s"].as<double>(), - conf["strength_prior_r"].as<double>()); - vector<WordID> ctx(kORDER - 1, TD::Convert("<s>")); - for (int SS=0; SS < samples; ++SS) { - for (int ci = 0; ci < corpuse.size(); ++ci) { - ctx.resize(kORDER - 1); - const vector<WordID>& s = corpuse[ci]; - for (int i = 0; i <= s.size(); ++i) { - WordID w = (i < s.size() ? s[i] : kEOS); - if (SS > 0) lm.decrement(w, ctx, &rng); - lm.increment(w, ctx, &rng); - ctx.push_back(w); - } - } - if (SS % 10 == 9) { - cerr << " [LLH=" << lm.log_likelihood() << "]" << endl; - if (SS % 30 == 29) lm.resample_hyperparameters(&rng); - } else { cerr << '.' << flush; } - } - double llh = 0; - unsigned cnt = 0; - unsigned oovs = 0; - for (int ci = 0; ci < test.size(); ++ci) { - ctx.resize(kORDER - 1); - const vector<WordID>& s = test[ci]; - for (int i = 0; i <= s.size(); ++i) { - WordID w = (i < s.size() ? s[i] : kEOS); - double lp = log(lm.prob(w, ctx)) / log(2); - if (i < s.size() && vocabe.count(w) == 0) { - cerr << "**OOV "; - ++oovs; - lp = 0; - } - cerr << "p(" << TD::Convert(w) << " |"; - for (int j = ctx.size() + 1 - kORDER; j < ctx.size(); ++j) - cerr << ' ' << TD::Convert(ctx[j]); - cerr << ") = " << lp << endl; - ctx.push_back(w); - llh -= lp; - cnt++; - } - } - cerr << " Log_10 prob: " << (-llh * log(2) / log(10)) << endl; - cerr << " Count: " << cnt << endl; - cerr << " OOVs: " << oovs << endl; - cerr << "Cross-entropy: " << (llh / cnt) << endl; - cerr << " Perplexity: " << pow(2, llh / cnt) << endl; - return 0; -} - - diff --git a/gi/pf/pyp_tm.cc b/gi/pf/pyp_tm.cc deleted file mode 100644 index 37b9a604..00000000 --- a/gi/pf/pyp_tm.cc +++ /dev/null @@ -1,128 +0,0 @@ -#include "pyp_tm.h" - -#include <tr1/unordered_map> -#include <iostream> -#include <queue> - -#include "tdict.h" -#include "ccrp.h" -#include "pyp_word_model.h" -#include "tied_resampler.h" - -using namespace std; -using namespace std::tr1; - -struct FreqBinner { - FreqBinner(const std::string& fname) { fd_.Load(fname); } - unsigned NumberOfBins() const { return fd_.Max() + 1; } - unsigned Bin(const WordID& w) const { return fd_.LookUp(w); } - FreqDict<unsigned> fd_; -}; - -template <typename Base, class Binner = FreqBinner> -struct ConditionalPYPWordModel { - ConditionalPYPWordModel(Base* b, const Binner* bnr = NULL) : - base(*b), - binner(bnr), - btr(binner ? binner->NumberOfBins() + 1u : 2u) {} - - void Summary() const { - cerr << "Number of conditioning contexts: " << r.size() << endl; - for (RuleModelHash::const_iterator it = r.begin(); it != r.end(); ++it) { - cerr << TD::Convert(it->first) << " \tPYP(d=" << it->second.discount() << ",s=" << it->second.strength() << ") --------------------------" << endl; - for (CCRP<vector<WordID> >::const_iterator i2 = it->second.begin(); i2 != it->second.end(); ++i2) - cerr << " " << i2->second << '\t' << TD::GetString(i2->first) << endl; - } - } - - void ResampleHyperparameters(MT19937* rng) { - btr.ResampleHyperparameters(rng); - } - - prob_t Prob(const WordID src, const vector<WordID>& trglets) const { - RuleModelHash::const_iterator it = r.find(src); - if (it == r.end()) { - return base(trglets); - } else { - return it->second.prob(trglets, base(trglets)); - } - } - - void Increment(const WordID src, const vector<WordID>& trglets, MT19937* rng) { - RuleModelHash::iterator it = r.find(src); - if (it == r.end()) { - it = r.insert(make_pair(src, CCRP<vector<WordID> >(0.5,1.0))).first; - static const WordID kNULL = TD::Convert("NULL"); - unsigned bin = (src == kNULL ? 0 : 1); - if (binner && bin) { bin = binner->Bin(src) + 1; } - btr.Add(bin, &it->second); - } - if (it->second.increment(trglets, base(trglets), rng)) - base.Increment(trglets, rng); - } - - void Decrement(const WordID src, const vector<WordID>& trglets, MT19937* rng) { - RuleModelHash::iterator it = r.find(src); - assert(it != r.end()); - if (it->second.decrement(trglets, rng)) { - base.Decrement(trglets, rng); - } - } - - prob_t Likelihood() const { - prob_t p = prob_t::One(); - for (RuleModelHash::const_iterator it = r.begin(); it != r.end(); ++it) { - prob_t q; q.logeq(it->second.log_crp_prob()); - p *= q; - } - return p; - } - - unsigned UniqueConditioningContexts() const { - return r.size(); - } - - // TODO tie PYP hyperparameters based on source word frequency bins - Base& base; - const Binner* binner; - BinTiedResampler<CCRP<vector<WordID> > > btr; - typedef unordered_map<WordID, CCRP<vector<WordID> > > RuleModelHash; - RuleModelHash r; -}; - -PYPLexicalTranslation::PYPLexicalTranslation(const vector<vector<WordID> >& lets, - const unsigned vocab_size, - const unsigned num_letters) : - letters(lets), - base(vocab_size, num_letters, 5), - tmodel(new ConditionalPYPWordModel<PoissonUniformWordModel>(&base, new FreqBinner("10k.freq"))), - kX(-TD::Convert("X")) {} - -void PYPLexicalTranslation::Summary() const { - tmodel->Summary(); -} - -prob_t PYPLexicalTranslation::Likelihood() const { - return tmodel->Likelihood() * base.Likelihood(); -} - -void PYPLexicalTranslation::ResampleHyperparameters(MT19937* rng) { - tmodel->ResampleHyperparameters(rng); -} - -unsigned PYPLexicalTranslation::UniqueConditioningContexts() const { - return tmodel->UniqueConditioningContexts(); -} - -prob_t PYPLexicalTranslation::Prob(WordID src, WordID trg) const { - return tmodel->Prob(src, letters[trg]); -} - -void PYPLexicalTranslation::Increment(WordID src, WordID trg, MT19937* rng) { - tmodel->Increment(src, letters[trg], rng); -} - -void PYPLexicalTranslation::Decrement(WordID src, WordID trg, MT19937* rng) { - tmodel->Decrement(src, letters[trg], rng); -} - diff --git a/gi/pf/pyp_tm.h b/gi/pf/pyp_tm.h deleted file mode 100644 index 2b076a25..00000000 --- a/gi/pf/pyp_tm.h +++ /dev/null @@ -1,36 +0,0 @@ -#ifndef PYP_LEX_TRANS -#define PYP_LEX_TRANS - -#include <vector> -#include "wordid.h" -#include "prob.h" -#include "sampler.h" -#include "freqdict.h" -#include "poisson_uniform_word_model.h" - -struct FreqBinner; -template <typename T, class B> struct ConditionalPYPWordModel; - -struct PYPLexicalTranslation { - explicit PYPLexicalTranslation(const std::vector<std::vector<WordID> >& lets, - const unsigned vocab_size, - const unsigned num_letters); - - prob_t Likelihood() const; - - void ResampleHyperparameters(MT19937* rng); - prob_t Prob(WordID src, WordID trg) const; // return p(trg | src) - void Summary() const; - void Increment(WordID src, WordID trg, MT19937* rng); - void Decrement(WordID src, WordID trg, MT19937* rng); - unsigned UniqueConditioningContexts() const; - - private: - const std::vector<std::vector<WordID> >& letters; // spelling dictionary - PoissonUniformWordModel base; // "generator" of English types - ConditionalPYPWordModel<PoissonUniformWordModel, FreqBinner>* tmodel; // translation distributions - // (model English word | French word) - const WordID kX; -}; - -#endif diff --git a/gi/pf/pyp_word_model.h b/gi/pf/pyp_word_model.h deleted file mode 100644 index 0bebb751..00000000 --- a/gi/pf/pyp_word_model.h +++ /dev/null @@ -1,61 +0,0 @@ -#ifndef _PYP_WORD_MODEL_H_ -#define _PYP_WORD_MODEL_H_ - -#include <iostream> -#include <cmath> -#include <vector> -#include "prob.h" -#include "ccrp.h" -#include "m.h" -#include "tdict.h" -#include "os_phrase.h" - -// PYP(d,s,poisson-uniform) represented as a CRP -template <class Base> -struct PYPWordModel { - explicit PYPWordModel(Base* b) : - base(*b), - r(1,1,1,1,0.66,50.0) - {} - - void ResampleHyperparameters(MT19937* rng) { - r.resample_hyperparameters(rng); - std::cerr << " PYPWordModel(d=" << r.discount() << ",s=" << r.strength() << ")\n"; - } - - inline prob_t operator()(const std::vector<WordID>& s) const { - return r.prob(s, base(s)); - } - - inline void Increment(const std::vector<WordID>& s, MT19937* rng) { - if (r.increment(s, base(s), rng)) - base.Increment(s, rng); - } - - inline void Decrement(const std::vector<WordID>& s, MT19937 *rng) { - if (r.decrement(s, rng)) - base.Decrement(s, rng); - } - - inline prob_t Likelihood() const { - prob_t p; p.logeq(r.log_crp_prob()); - p *= base.Likelihood(); - return p; - } - - void Summary() const { - std::cerr << "PYPWordModel: generations=" << r.num_customers() - << " PYP(d=" << r.discount() << ",s=" << r.strength() << ')' << std::endl; - for (typename CCRP<std::vector<WordID> >::const_iterator it = r.begin(); it != r.end(); ++it) { - std::cerr << " " << it->second - << TD::GetString(it->first) << std::endl; - } - } - - private: - - Base& base; // keeps track of the draws from the base distribution - CCRP<std::vector<WordID> > r; -}; - -#endif diff --git a/gi/pf/quasi_model2.h b/gi/pf/quasi_model2.h deleted file mode 100644 index 4075affe..00000000 --- a/gi/pf/quasi_model2.h +++ /dev/null @@ -1,177 +0,0 @@ -#ifndef _QUASI_MODEL2_H_ -#define _QUASI_MODEL2_H_ - -#include <vector> -#include <cmath> -#include <tr1/unordered_map> -#include "boost/functional.hpp" -#include "prob.h" -#include "array2d.h" -#include "slice_sampler.h" -#include "m.h" -#include "have_64_bits.h" - -struct AlignmentObservation { - AlignmentObservation() : src_len(), trg_len(), j(), a_j() {} - AlignmentObservation(unsigned sl, unsigned tl, unsigned tw, unsigned sw) : - src_len(sl), trg_len(tl), j(tw), a_j(sw) {} - unsigned short src_len; - unsigned short trg_len; - unsigned short j; - unsigned short a_j; -}; - -#ifdef HAVE_64_BITS -inline size_t hash_value(const AlignmentObservation& o) { - return reinterpret_cast<const size_t&>(o); -} -inline bool operator==(const AlignmentObservation& a, const AlignmentObservation& b) { - return hash_value(a) == hash_value(b); -} -#else -inline size_t hash_value(const AlignmentObservation& o) { - size_t h = 1; - boost::hash_combine(h, o.src_len); - boost::hash_combine(h, o.trg_len); - boost::hash_combine(h, o.j); - boost::hash_combine(h, o.a_j); - return h; -} -#endif - -struct QuasiModel2 { - explicit QuasiModel2(double alpha, double pnull = 0.1) : - alpha_(alpha), - pnull_(pnull), - pnotnull_(1 - pnull) {} - - // a_j = 0 => NULL; src_len does *not* include null - prob_t Prob(unsigned a_j, unsigned j, unsigned src_len, unsigned trg_len) const { - if (!a_j) return pnull_; - return pnotnull_ * - prob_t(UnnormalizedProb(a_j, j, src_len, trg_len, alpha_) / GetOrComputeZ(j, src_len, trg_len)); - } - - void Increment(unsigned a_j, unsigned j, unsigned src_len, unsigned trg_len) { - assert(a_j <= src_len); - assert(j < trg_len); - ++obs_[AlignmentObservation(src_len, trg_len, j, a_j)]; - } - - void Decrement(unsigned a_j, unsigned j, unsigned src_len, unsigned trg_len) { - const AlignmentObservation ao(src_len, trg_len, j, a_j); - int &cc = obs_[ao]; - assert(cc > 0); - --cc; - if (!cc) obs_.erase(ao); - } - - struct PNullResampler { - PNullResampler(const QuasiModel2& m) : m_(m) {} - const QuasiModel2& m_; - double operator()(const double& proposed_pnull) const { - return log(m_.Likelihood(m_.alpha_, proposed_pnull)); - } - }; - - struct AlphaResampler { - AlphaResampler(const QuasiModel2& m) : m_(m) {} - const QuasiModel2& m_; - double operator()(const double& proposed_alpha) const { - return log(m_.Likelihood(proposed_alpha, m_.pnull_.as_float())); - } - }; - - void ResampleHyperparameters(MT19937* rng, const unsigned nloop = 5, const unsigned niterations = 10) { - const PNullResampler dr(*this); - const AlphaResampler ar(*this); - for (unsigned i = 0; i < nloop; ++i) { - double pnull = slice_sampler1d(dr, pnull_.as_float(), *rng, 0.00000001, - 1.0, 0.0, niterations, 100*niterations); - pnull_ = prob_t(pnull); - alpha_ = slice_sampler1d(ar, alpha_, *rng, 0.00000001, - std::numeric_limits<double>::infinity(), 0.0, niterations, 100*niterations); - } - std::cerr << "QuasiModel2(alpha=" << alpha_ << ",p_null=" - << pnull_.as_float() << ") = " << Likelihood() << std::endl; - zcache_.clear(); - } - - prob_t Likelihood() const { - return Likelihood(alpha_, pnull_.as_float()); - } - - prob_t Likelihood(double alpha, double ppnull) const { - const prob_t pnull(ppnull); - const prob_t pnotnull(1 - ppnull); - - prob_t p; - p.logeq(Md::log_gamma_density(alpha, 0.1, 25)); // TODO configure - assert(!p.is_0()); - prob_t prob_of_ppnull; prob_of_ppnull.logeq(Md::log_beta_density(ppnull, 2, 10)); - assert(!prob_of_ppnull.is_0()); - p *= prob_of_ppnull; - for (ObsCount::const_iterator it = obs_.begin(); it != obs_.end(); ++it) { - const AlignmentObservation& ao = it->first; - if (ao.a_j) { - prob_t u = XUnnormalizedProb(ao.a_j, ao.j, ao.src_len, ao.trg_len, alpha); - prob_t z = XComputeZ(ao.j, ao.src_len, ao.trg_len, alpha); - prob_t pa(u / z); - pa *= pnotnull; - pa.poweq(it->second); - p *= pa; - } else { - p *= pnull.pow(it->second); - } - } - return p; - } - - private: - static prob_t XUnnormalizedProb(unsigned a_j, unsigned j, unsigned src_len, unsigned trg_len, double alpha) { - prob_t p; - p.logeq(-fabs(double(a_j - 1) / src_len - double(j) / trg_len) * alpha); - return p; - } - - static prob_t XComputeZ(unsigned j, unsigned src_len, unsigned trg_len, double alpha) { - prob_t z = prob_t::Zero(); - for (int a_j = 1; a_j <= src_len; ++a_j) - z += XUnnormalizedProb(a_j, j, src_len, trg_len, alpha); - return z; - } - - static double UnnormalizedProb(unsigned a_j, unsigned j, unsigned src_len, unsigned trg_len, double alpha) { - return exp(-fabs(double(a_j - 1) / src_len - double(j) / trg_len) * alpha); - } - - static double ComputeZ(unsigned j, unsigned src_len, unsigned trg_len, double alpha) { - double z = 0; - for (int a_j = 1; a_j <= src_len; ++a_j) - z += UnnormalizedProb(a_j, j, src_len, trg_len, alpha); - return z; - } - - const double& GetOrComputeZ(unsigned j, unsigned src_len, unsigned trg_len) const { - if (src_len >= zcache_.size()) - zcache_.resize(src_len + 1); - if (trg_len >= zcache_[src_len].size()) - zcache_[src_len].resize(trg_len + 1); - std::vector<double>& zv = zcache_[src_len][trg_len]; - if (zv.size() == 0) - zv.resize(trg_len); - double& z = zv[j]; - if (!z) - z = ComputeZ(j, src_len, trg_len, alpha_); - return z; - } - - double alpha_; - prob_t pnull_; - prob_t pnotnull_; - mutable std::vector<std::vector<std::vector<double> > > zcache_; - typedef std::tr1::unordered_map<AlignmentObservation, int, boost::hash<AlignmentObservation> > ObsCount; - ObsCount obs_; -}; - -#endif diff --git a/gi/pf/reachability.cc b/gi/pf/reachability.cc deleted file mode 100644 index 7d0d04ac..00000000 --- a/gi/pf/reachability.cc +++ /dev/null @@ -1,74 +0,0 @@ -#include "reachability.h" - -#include <vector> -#include <iostream> - -using namespace std; - -struct SState { - SState() : prev_src_covered(), prev_trg_covered() {} - SState(int i, int j) : prev_src_covered(i), prev_trg_covered(j) {} - int prev_src_covered; - int prev_trg_covered; -}; - -void Reachability::ComputeReachability(int srclen, int trglen, int src_max_phrase_len, int trg_max_phrase_len) { - typedef boost::multi_array<vector<SState>, 2> array_type; - array_type a(boost::extents[srclen + 1][trglen + 1]); - a[0][0].push_back(SState()); - for (int i = 0; i < srclen; ++i) { - for (int j = 0; j < trglen; ++j) { - if (a[i][j].size() == 0) continue; - const SState prev(i,j); - for (int k = 1; k <= src_max_phrase_len; ++k) { - if ((i + k) > srclen) continue; - for (int l = 1; l <= trg_max_phrase_len; ++l) { - if ((j + l) > trglen) continue; - a[i + k][j + l].push_back(prev); - } - } - } - } - a[0][0].clear(); - //cerr << srclen << "," << trglen << ": Final cell contains " << a[srclen][trglen].size() << " back pointers\n"; - if (a[srclen][trglen].empty()) { - cerr << "Sequence pair with lengths (" << srclen << ',' << trglen << ") violates reachability constraints\n"; - nodes = 0; - return; - } - - typedef boost::multi_array<bool, 2> rarray_type; - rarray_type r(boost::extents[srclen + 1][trglen + 1]); - r[srclen][trglen] = true; - nodes = 0; - for (int i = srclen; i >= 0; --i) { - for (int j = trglen; j >= 0; --j) { - vector<SState>& prevs = a[i][j]; - if (!r[i][j]) { prevs.clear(); } - for (int k = 0; k < prevs.size(); ++k) { - r[prevs[k].prev_src_covered][prevs[k].prev_trg_covered] = true; - int src_delta = i - prevs[k].prev_src_covered; - edges[prevs[k].prev_src_covered][prevs[k].prev_trg_covered][src_delta][j - prevs[k].prev_trg_covered] = true; - valid_deltas[prevs[k].prev_src_covered][prevs[k].prev_trg_covered].push_back(make_pair<short,short>(src_delta,j - prevs[k].prev_trg_covered)); - short &msd = max_src_delta[prevs[k].prev_src_covered][prevs[k].prev_trg_covered]; - if (src_delta > msd) msd = src_delta; - } - } - } - assert(!edges[0][0][1][0]); - assert(!edges[0][0][0][1]); - assert(!edges[0][0][0][0]); - assert(max_src_delta[0][0] > 0); - nodes = 0; - for (int i = 0; i < srclen; ++i) { - for (int j = 0; j < trglen; ++j) { - if (valid_deltas[i][j].size() > 0) { - node_addresses[i][j] = nodes++; - } else { - node_addresses[i][j] = -1; - } - } - } - cerr << "Sequence pair with lengths (" << srclen << ',' << trglen << ") has " << valid_deltas[0][0].size() << " out edges in its root node, " << nodes << " nodes in total, and outside estimate matrix will require " << sizeof(float)*nodes << " bytes\n"; - } - diff --git a/gi/pf/reachability.h b/gi/pf/reachability.h deleted file mode 100644 index 1e22c76a..00000000 --- a/gi/pf/reachability.h +++ /dev/null @@ -1,34 +0,0 @@ -#ifndef _REACHABILITY_H_ -#define _REACHABILITY_H_ - -#include "boost/multi_array.hpp" - -// determines minimum and maximum lengths of outgoing edges from all -// coverage positions such that the alignment path respects src and -// trg maximum phrase sizes -// -// runs in O(n^2 * src_max * trg_max) time but should be relatively fast -// -// currently forbids 0 -> n and n -> 0 alignments - -struct Reachability { - unsigned nodes; - boost::multi_array<bool, 4> edges; // edges[src_covered][trg_covered][src_delta][trg_delta] is this edge worth exploring? - boost::multi_array<short, 2> max_src_delta; // msd[src_covered][trg_covered] -- the largest src delta that's valid - boost::multi_array<short, 2> node_addresses; // na[src_covered][trg_covered] -- the index of the node in a one-dimensional array (of size "nodes") - boost::multi_array<std::vector<std::pair<short,short> >, 2> valid_deltas; // valid_deltas[src_covered][trg_covered] list of valid transitions leaving a particular node - - Reachability(int srclen, int trglen, int src_max_phrase_len, int trg_max_phrase_len) : - nodes(), - edges(boost::extents[srclen][trglen][src_max_phrase_len+1][trg_max_phrase_len+1]), - max_src_delta(boost::extents[srclen][trglen]), - node_addresses(boost::extents[srclen][trglen]), - valid_deltas(boost::extents[srclen][trglen]) { - ComputeReachability(srclen, trglen, src_max_phrase_len, trg_max_phrase_len); - } - - private: - void ComputeReachability(int srclen, int trglen, int src_max_phrase_len, int trg_max_phrase_len); -}; - -#endif diff --git a/gi/pf/tied_resampler.h b/gi/pf/tied_resampler.h deleted file mode 100644 index a4f4af36..00000000 --- a/gi/pf/tied_resampler.h +++ /dev/null @@ -1,122 +0,0 @@ -#ifndef _TIED_RESAMPLER_H_ -#define _TIED_RESAMPLER_H_ - -#include <set> -#include <vector> -#include "sampler.h" -#include "slice_sampler.h" -#include "m.h" - -template <class CRP> -struct TiedResampler { - explicit TiedResampler(double da, double db, double ss, double sr, double d=0.5, double s=1.0) : - d_alpha(da), - d_beta(db), - s_shape(ss), - s_rate(sr), - discount(d), - strength(s) {} - - void Add(CRP* crp) { - crps.insert(crp); - crp->set_discount(discount); - crp->set_strength(strength); - assert(!crp->has_discount_prior()); - assert(!crp->has_strength_prior()); - } - - void Remove(CRP* crp) { - crps.erase(crp); - } - - size_t size() const { - return crps.size(); - } - - double LogLikelihood(double d, double s) const { - if (s <= -d) return -std::numeric_limits<double>::infinity(); - double llh = Md::log_beta_density(d, d_alpha, d_beta) + - Md::log_gamma_density(d + s, s_shape, s_rate); - for (typename std::set<CRP*>::iterator it = crps.begin(); it != crps.end(); ++it) - llh += (*it)->log_crp_prob(d, s); - return llh; - } - - double LogLikelihood() const { - return LogLikelihood(discount, strength); - } - - struct DiscountResampler { - DiscountResampler(const TiedResampler& m) : m_(m) {} - const TiedResampler& m_; - double operator()(const double& proposed_discount) const { - return m_.LogLikelihood(proposed_discount, m_.strength); - } - }; - - struct AlphaResampler { - AlphaResampler(const TiedResampler& m) : m_(m) {} - const TiedResampler& m_; - double operator()(const double& proposed_strength) const { - return m_.LogLikelihood(m_.discount, proposed_strength); - } - }; - - void ResampleHyperparameters(MT19937* rng, const unsigned nloop = 5, const unsigned niterations = 10) { - if (size() == 0) { std::cerr << "EMPTY - not resampling\n"; return; } - const DiscountResampler dr(*this); - const AlphaResampler ar(*this); - for (int iter = 0; iter < nloop; ++iter) { - strength = slice_sampler1d(ar, strength, *rng, -discount + std::numeric_limits<double>::min(), - std::numeric_limits<double>::infinity(), 0.0, niterations, 100*niterations); - double min_discount = std::numeric_limits<double>::min(); - if (strength < 0.0) min_discount -= strength; - discount = slice_sampler1d(dr, discount, *rng, min_discount, - 1.0, 0.0, niterations, 100*niterations); - } - strength = slice_sampler1d(ar, strength, *rng, -discount + std::numeric_limits<double>::min(), - std::numeric_limits<double>::infinity(), 0.0, niterations, 100*niterations); - std::cerr << "TiedCRPs(d=" << discount << ",s=" - << strength << ") = " << LogLikelihood(discount, strength) << std::endl; - for (typename std::set<CRP*>::iterator it = crps.begin(); it != crps.end(); ++it) - (*it)->set_hyperparameters(discount, strength); - } - private: - std::set<CRP*> crps; - const double d_alpha, d_beta, s_shape, s_rate; - double discount, strength; -}; - -// split according to some criterion -template <class CRP> -struct BinTiedResampler { - explicit BinTiedResampler(unsigned nbins) : - resamplers(nbins, TiedResampler<CRP>(1,1,1,1)) {} - - void Add(unsigned bin, CRP* crp) { - resamplers[bin].Add(crp); - } - - void Remove(unsigned bin, CRP* crp) { - resamplers[bin].Remove(crp); - } - - void ResampleHyperparameters(MT19937* rng) { - for (unsigned i = 0; i < resamplers.size(); ++i) { - std::cerr << "BIN " << i << " (" << resamplers[i].size() << " CRPs): " << std::flush; - resamplers[i].ResampleHyperparameters(rng); - } - } - - double LogLikelihood() const { - double llh = 0; - for (unsigned i = 0; i < resamplers.size(); ++i) - llh += resamplers[i].LogLikelihood(); - return llh; - } - - private: - std::vector<TiedResampler<CRP> > resamplers; -}; - -#endif diff --git a/gi/pf/tpf.cc b/gi/pf/tpf.cc deleted file mode 100644 index 7348d21c..00000000 --- a/gi/pf/tpf.cc +++ /dev/null @@ -1,99 +0,0 @@ -#include <iostream> -#include <tr1/memory> -#include <queue> - -#include "sampler.h" - -using namespace std; -using namespace tr1; - -shared_ptr<MT19937> prng; - -struct Particle { - Particle() : weight(prob_t::One()) {} - vector<int> states; - prob_t weight; - prob_t gamma_last; -}; - -ostream& operator<<(ostream& os, const Particle& p) { - os << "["; - for (int i = 0; i < p.states.size(); ++i) os << p.states[i] << ' '; - os << "| w=" << log(p.weight) << ']'; - return os; -} - -void Rejuvenate(vector<Particle>& pps) { - SampleSet<prob_t> ss; - vector<Particle> nps(pps.size()); - for (int i = 0; i < pps.size(); ++i) { -// cerr << pps[i] << endl; - ss.add(pps[i].weight); - } -// cerr << "REJUVINATING...\n"; - for (int i = 0; i < pps.size(); ++i) { - nps[i] = pps[prng->SelectSample(ss)]; - nps[i].weight = prob_t(1.0 / pps.size()); -// cerr << nps[i] << endl; - } - nps.swap(pps); -// exit(1); -} - -int main(int argc, char** argv) { - const unsigned particles = 100; - prng.reset(new MT19937); - MT19937& rng = *prng; - - // q(a) = 0.8 - // q(b) = 0.8 - // q(c) = 0.4 - SampleSet<double> ssq; - ssq.add(0.4); - ssq.add(0.6); - ssq.add(0); - double qz = 1; - - // p(a) = 0.2 - // p(b) = 0.8 - vector<double> p(3); - p[0] = 0.2; - p[1] = 0.8; - p[2] = 0; - - vector<int> counts(3); - int tot = 0; - - vector<Particle> pps(particles); - SampleSet<prob_t> ppss; - int LEN = 12; - int PP = 1; - while (pps[0].states.size() < LEN) { - for (int pi = 0; pi < particles; ++pi) { - Particle& prt = pps[pi]; - - bool redo = true; - const Particle savedp = prt; - while (redo) { - redo = false; - for (int i = 0; i < PP; ++i) { - int s = rng.SelectSample(ssq); - double gamma_last = p[s]; - if (!gamma_last) { redo = true; break; } - double q = ssq[s] / qz; - prt.states.push_back(s); - prt.weight *= prob_t(gamma_last / q); - } - if (redo) { prt = savedp; continue; } - } - } - Rejuvenate(pps); - } - ppss.clear(); - for (int i = 0; i < particles; ++i) { ppss.add(pps[i].weight); } - int sp = rng.SelectSample(ppss); - cerr << pps[sp] << endl; - - return 0; -} - diff --git a/gi/pf/transliterations.cc b/gi/pf/transliterations.cc deleted file mode 100644 index b2996f65..00000000 --- a/gi/pf/transliterations.cc +++ /dev/null @@ -1,334 +0,0 @@ -#include "transliterations.h" - -#include <iostream> -#include <vector> - -#include "boost/shared_ptr.hpp" - -#include "backward.h" -#include "filelib.h" -#include "tdict.h" -#include "trule.h" -#include "filelib.h" -#include "ccrp_nt.h" -#include "m.h" -#include "reachability.h" - -using namespace std; -using namespace std::tr1; - -struct TruncatedConditionalLengthModel { - TruncatedConditionalLengthModel(unsigned max_src_size, unsigned max_trg_size, double expected_src_to_trg_ratio) : - plens(max_src_size+1, vector<prob_t>(max_trg_size+1, 0.0)) { - for (unsigned i = 1; i <= max_src_size; ++i) { - prob_t z = prob_t::Zero(); - for (unsigned j = 1; j <= max_trg_size; ++j) - z += (plens[i][j] = prob_t(0.01 + exp(Md::log_poisson(j, i * expected_src_to_trg_ratio)))); - for (unsigned j = 1; j <= max_trg_size; ++j) - plens[i][j] /= z; - //for (unsigned j = 1; j <= max_trg_size; ++j) - // cerr << "P(trg_len=" << j << " | src_len=" << i << ") = " << plens[i][j] << endl; - } - } - - // return p(tlen | slen) for *chunks* not full words - inline const prob_t& operator()(int slen, int tlen) const { - return plens[slen][tlen]; - } - - vector<vector<prob_t> > plens; -}; - -struct CondBaseDist { - CondBaseDist(unsigned max_src_size, unsigned max_trg_size, double expected_src_to_trg_ratio) : - tclm(max_src_size, max_trg_size, expected_src_to_trg_ratio) {} - - prob_t operator()(const vector<WordID>& src, unsigned sf, unsigned st, - const vector<WordID>& trg, unsigned tf, unsigned tt) const { - prob_t p = tclm(st - sf, tt - tf); // target len | source length ~ TCLM(source len) - assert(!"not impl"); - return p; - } - inline prob_t operator()(const vector<WordID>& src, const vector<WordID>& trg) const { - return (*this)(src, 0, src.size(), trg, 0, trg.size()); - } - TruncatedConditionalLengthModel tclm; -}; - -// represents transliteration phrase probabilities, e.g. -// p( a l - | A l ) , p( o | A w ) , ... -struct TransliterationChunkConditionalModel { - explicit TransliterationChunkConditionalModel(const CondBaseDist& pp0) : - d(0.0), - strength(1.0), - rp0(pp0) { - } - - void Summary() const { - std::cerr << "Number of conditioning contexts: " << r.size() << std::endl; - for (RuleModelHash::const_iterator it = r.begin(); it != r.end(); ++it) { - std::cerr << TD::GetString(it->first) << " \t(\\alpha = " << it->second.alpha() << ") --------------------------" << std::endl; - for (CCRP_NoTable<TRule>::const_iterator i2 = it->second.begin(); i2 != it->second.end(); ++i2) - std::cerr << " " << i2->second << '\t' << i2->first << std::endl; - } - } - - int DecrementRule(const TRule& rule) { - RuleModelHash::iterator it = r.find(rule.f_); - assert(it != r.end()); - int count = it->second.decrement(rule); - if (count) { - if (it->second.num_customers() == 0) r.erase(it); - } - return count; - } - - int IncrementRule(const TRule& rule) { - RuleModelHash::iterator it = r.find(rule.f_); - if (it == r.end()) { - it = r.insert(make_pair(rule.f_, CCRP_NoTable<TRule>(strength))).first; - } - int count = it->second.increment(rule); - return count; - } - - void IncrementRules(const std::vector<TRulePtr>& rules) { - for (int i = 0; i < rules.size(); ++i) - IncrementRule(*rules[i]); - } - - void DecrementRules(const std::vector<TRulePtr>& rules) { - for (int i = 0; i < rules.size(); ++i) - DecrementRule(*rules[i]); - } - - prob_t RuleProbability(const TRule& rule) const { - prob_t p; - RuleModelHash::const_iterator it = r.find(rule.f_); - if (it == r.end()) { - p = rp0(rule.f_, rule.e_); - } else { - p = it->second.prob(rule, rp0(rule.f_, rule.e_)); - } - return p; - } - - double LogLikelihood(const double& dd, const double& aa) const { - if (aa <= -dd) return -std::numeric_limits<double>::infinity(); - //double llh = Md::log_beta_density(dd, 10, 3) + Md::log_gamma_density(aa, 1, 1); - double llh = //Md::log_beta_density(dd, 1, 1) + - Md::log_gamma_density(dd + aa, 1, 1); - std::tr1::unordered_map<std::vector<WordID>, CCRP_NoTable<TRule>, boost::hash<std::vector<WordID> > >::const_iterator it; - for (it = r.begin(); it != r.end(); ++it) - llh += it->second.log_crp_prob(aa); - return llh; - } - - struct AlphaResampler { - AlphaResampler(const TransliterationChunkConditionalModel& m) : m_(m) {} - const TransliterationChunkConditionalModel& m_; - double operator()(const double& proposed_strength) const { - return m_.LogLikelihood(m_.d, proposed_strength); - } - }; - - void ResampleHyperparameters(MT19937* rng) { - std::tr1::unordered_map<std::vector<WordID>, CCRP_NoTable<TRule>, boost::hash<std::vector<WordID> > >::iterator it; - //const unsigned nloop = 5; - const unsigned niterations = 10; - //DiscountResampler dr(*this); - AlphaResampler ar(*this); -#if 0 - for (int iter = 0; iter < nloop; ++iter) { - strength = slice_sampler1d(ar, strength, *rng, -d + std::numeric_limits<double>::min(), - std::numeric_limits<double>::infinity(), 0.0, niterations, 100*niterations); - double min_discount = std::numeric_limits<double>::min(); - if (strength < 0.0) min_discount -= strength; - d = slice_sampler1d(dr, d, *rng, min_discount, - 1.0, 0.0, niterations, 100*niterations); - } -#endif - strength = slice_sampler1d(ar, strength, *rng, -d, - std::numeric_limits<double>::infinity(), 0.0, niterations, 100*niterations); - std::cerr << "CTMModel(alpha=" << strength << ") = " << LogLikelihood(d, strength) << std::endl; - for (it = r.begin(); it != r.end(); ++it) { -#if 0 - it->second.set_discount(d); -#endif - it->second.set_alpha(strength); - } - } - - prob_t Likelihood() const { - prob_t p; p.logeq(LogLikelihood(d, strength)); - return p; - } - - const CondBaseDist& rp0; - typedef std::tr1::unordered_map<std::vector<WordID>, - CCRP_NoTable<TRule>, - boost::hash<std::vector<WordID> > > RuleModelHash; - RuleModelHash r; - double d, strength; -}; - -struct GraphStructure { - GraphStructure() : r() {} - // leak memory - these are basically static - const Reachability* r; - bool IsReachable() const { return r->nodes > 0; } -}; - -struct ProbabilityEstimates { - ProbabilityEstimates() : gs(), backward() {} - explicit ProbabilityEstimates(const GraphStructure& g) : - gs(&g), backward() { - if (g.r->nodes > 0) - backward = new float[g.r->nodes]; - } - // leak memory, these are static - - // returns an estimate of the marginal probability - double MarginalEstimate() const { - if (!backward) return 0; - return backward[0]; - } - - // returns an backward estimate - double Backward(int src_covered, int trg_covered) const { - if (!backward) return 0; - int ind = gs->r->node_addresses[src_covered][trg_covered]; - if (ind < 0) return 0; - return backward[ind]; - } - - prob_t estp; - float* backward; - private: - const GraphStructure* gs; -}; - -struct TransliterationsImpl { - TransliterationsImpl(int max_src, int max_trg, double sr, const BackwardEstimator& b) : - cp0(max_src, max_trg, sr), - tccm(cp0), - be(b), - kMAX_SRC_CHUNK(max_src), - kMAX_TRG_CHUNK(max_trg), - kS2T_RATIO(sr), - tot_pairs(), tot_mem() { - } - const CondBaseDist cp0; - TransliterationChunkConditionalModel tccm; - const BackwardEstimator& be; - - void Initialize(WordID src, const vector<WordID>& src_lets, WordID trg, const vector<WordID>& trg_lets) { - const size_t src_len = src_lets.size(); - const size_t trg_len = trg_lets.size(); - - // init graph structure - if (src_len >= graphs.size()) graphs.resize(src_len + 1); - if (trg_len >= graphs[src_len].size()) graphs[src_len].resize(trg_len + 1); - GraphStructure& gs = graphs[src_len][trg_len]; - if (!gs.r) { - double rat = exp(fabs(log(trg_len / (src_len * kS2T_RATIO)))); - if (rat > 1.5 || (rat > 2.4 && src_len < 6)) { - cerr << " ** Forbidding transliterations of size " << src_len << "," << trg_len << ": " << rat << endl; - gs.r = new Reachability(src_len, trg_len, 0, 0); - } else { - gs.r = new Reachability(src_len, trg_len, kMAX_SRC_CHUNK, kMAX_TRG_CHUNK); - } - } - - const Reachability& r = *gs.r; - - // init backward estimates - if (src >= ests.size()) ests.resize(src + 1); - unordered_map<WordID, ProbabilityEstimates>::iterator it = ests[src].find(trg); - if (it != ests[src].end()) return; // already initialized - - it = ests[src].insert(make_pair(trg, ProbabilityEstimates(gs))).first; - ProbabilityEstimates& est = it->second; - if (!gs.r->nodes) return; // not derivable subject to length constraints - - be.InitializeGrid(src_lets, trg_lets, r, kS2T_RATIO, est.backward); - cerr << TD::GetString(src_lets) << " ||| " << TD::GetString(trg_lets) << " ||| " << (est.backward[0] / trg_lets.size()) << endl; - tot_pairs++; - tot_mem += sizeof(float) * gs.r->nodes; - } - - void Forbid(WordID src, const vector<WordID>& src_lets, WordID trg, const vector<WordID>& trg_lets) { - const size_t src_len = src_lets.size(); - const size_t trg_len = trg_lets.size(); - // TODO - } - - prob_t EstimateProbability(WordID s, const vector<WordID>& src, WordID t, const vector<WordID>& trg) const { - assert(src.size() < graphs.size()); - const vector<GraphStructure>& tv = graphs[src.size()]; - assert(trg.size() < tv.size()); - const GraphStructure& gs = tv[trg.size()]; - if (gs.r->nodes == 0) - return prob_t::Zero(); - const unordered_map<WordID, ProbabilityEstimates>::const_iterator it = ests[s].find(t); - assert(it != ests[s].end()); - return it->second.estp; - } - - void GraphSummary() const { - double to = 0; - double tn = 0; - double tt = 0; - for (int i = 0; i < graphs.size(); ++i) { - const vector<GraphStructure>& vt = graphs[i]; - for (int j = 0; j < vt.size(); ++j) { - const GraphStructure& gs = vt[j]; - if (!gs.r) continue; - tt++; - for (int k = 0; k < i; ++k) { - for (int l = 0; l < j; ++l) { - size_t c = gs.r->valid_deltas[k][l].size(); - if (c) { - tn += 1; - to += c; - } - } - } - } - } - cerr << " Average nodes = " << (tn / tt) << endl; - cerr << "Average out-degree = " << (to / tn) << endl; - cerr << " Unique structures = " << tt << endl; - cerr << " Unique pairs = " << tot_pairs << endl; - cerr << " BEs size = " << (tot_mem / (1024.0*1024.0)) << " MB" << endl; - } - - const int kMAX_SRC_CHUNK; - const int kMAX_TRG_CHUNK; - const double kS2T_RATIO; - unsigned tot_pairs; - size_t tot_mem; - vector<vector<GraphStructure> > graphs; // graphs[src_len][trg_len] - vector<unordered_map<WordID, ProbabilityEstimates> > ests; // ests[src][trg] -}; - -Transliterations::Transliterations(int max_src, int max_trg, double sr, const BackwardEstimator& be) : - pimpl_(new TransliterationsImpl(max_src, max_trg, sr, be)) {} -Transliterations::~Transliterations() { delete pimpl_; } - -void Transliterations::Initialize(WordID src, const vector<WordID>& src_lets, WordID trg, const vector<WordID>& trg_lets) { - pimpl_->Initialize(src, src_lets, trg, trg_lets); -} - -prob_t Transliterations::EstimateProbability(WordID s, const vector<WordID>& src, WordID t, const vector<WordID>& trg) const { - return pimpl_->EstimateProbability(s, src,t, trg); -} - -void Transliterations::Forbid(WordID src, const vector<WordID>& src_lets, WordID trg, const vector<WordID>& trg_lets) { - pimpl_->Forbid(src, src_lets, trg, trg_lets); -} - -void Transliterations::GraphSummary() const { - pimpl_->GraphSummary(); -} - diff --git a/gi/pf/transliterations.h b/gi/pf/transliterations.h deleted file mode 100644 index 49d14684..00000000 --- a/gi/pf/transliterations.h +++ /dev/null @@ -1,24 +0,0 @@ -#ifndef _TRANSLITERATIONS_H_ -#define _TRANSLITERATIONS_H_ - -#include <vector> -#include "wordid.h" -#include "prob.h" - -struct BackwardEstimator; -struct TransliterationsImpl; -struct Transliterations { - // max_src and max_trg indicate how big the transliteration phrases can be - // see reachability.h for information about filter_ratio - explicit Transliterations(int max_src, int max_trg, double s2t_rat, const BackwardEstimator& be); - ~Transliterations(); - void Initialize(WordID src, const std::vector<WordID>& src_lets, WordID trg, const std::vector<WordID>& trg_lets); - void Forbid(WordID src, const std::vector<WordID>& src_lets, WordID trg, const std::vector<WordID>& trg_lets); - void GraphSummary() const; - prob_t EstimateProbability(WordID s, const std::vector<WordID>& src, WordID t, const std::vector<WordID>& trg) const; - private: - TransliterationsImpl* pimpl_; -}; - -#endif - diff --git a/gi/pf/unigrams.cc b/gi/pf/unigrams.cc deleted file mode 100644 index 40829775..00000000 --- a/gi/pf/unigrams.cc +++ /dev/null @@ -1,80 +0,0 @@ -#include "unigrams.h" - -#include <string> -#include <cmath> - -#include "stringlib.h" -#include "filelib.h" - -using namespace std; - -void UnigramModel::LoadUnigrams(const string& fname) { - cerr << "Loading unigram probabilities from " << fname << " ..." << endl; - ReadFile rf(fname); - string line; - istream& in = *rf.stream(); - assert(in); - getline(in, line); - assert(line.empty()); - getline(in, line); - assert(line == "\\data\\"); - getline(in, line); - size_t pos = line.find("ngram 1="); - assert(pos == 0); - assert(line.size() > 8); - const size_t num_unigrams = atoi(&line[8]); - getline(in, line); - assert(line.empty()); - getline(in, line); - assert(line == "\\1-grams:"); - for (size_t i = 0; i < num_unigrams; ++i) { - getline(in, line); - assert(line.size() > 0); - pos = line.find('\t'); - assert(pos > 0); - assert(pos + 1 < line.size()); - const WordID w = TD::Convert(line.substr(pos + 1)); - line[pos] = 0; - float p = atof(&line[0]); - if (w < probs_.size()) probs_[w].logeq(p * log(10)); else cerr << "WARNING: don't know about '" << TD::Convert(w) << "'\n"; - } -} - -void UnigramWordModel::LoadUnigrams(const string& fname) { - cerr << "Loading unigram probabilities from " << fname << " ..." << endl; - ReadFile rf(fname); - string line; - istream& in = *rf.stream(); - assert(in); - getline(in, line); - assert(line.empty()); - getline(in, line); - assert(line == "\\data\\"); - getline(in, line); - size_t pos = line.find("ngram 1="); - assert(pos == 0); - assert(line.size() > 8); - const size_t num_unigrams = atoi(&line[8]); - getline(in, line); - assert(line.empty()); - getline(in, line); - assert(line == "\\1-grams:"); - for (size_t i = 0; i < num_unigrams; ++i) { - getline(in, line); - assert(line.size() > 0); - pos = line.find('\t'); - assert(pos > 0); - assert(pos + 1 < line.size()); - size_t cur = pos + 1; - vector<WordID> w; - while (cur < line.size()) { - const size_t len = UTF8Len(line[cur]); - w.push_back(TD::Convert(line.substr(cur, len))); - cur += len; - } - line[pos] = 0; - float p = atof(&line[0]); - probs_[w].logeq(p * log(10.0)); - } -} - diff --git a/gi/pf/unigrams.h b/gi/pf/unigrams.h deleted file mode 100644 index 1660d1ed..00000000 --- a/gi/pf/unigrams.h +++ /dev/null @@ -1,69 +0,0 @@ -#ifndef _UNIGRAMS_H_ -#define _UNIGRAMS_H_ - -#include <vector> -#include <string> -#include <tr1/unordered_map> -#include <boost/functional.hpp> - -#include "wordid.h" -#include "prob.h" -#include "tdict.h" - -struct UnigramModel { - explicit UnigramModel(const std::string& fname, unsigned vocab_size) : - use_uniform_(fname.size() == 0), - uniform_(1.0 / vocab_size), - probs_() { - if (fname.size() > 0) { - probs_.resize(TD::NumWords() + 1); - LoadUnigrams(fname); - } - } - - const prob_t& operator()(const WordID& w) const { - assert(w); - if (use_uniform_) return uniform_; - return probs_[w]; - } - - private: - void LoadUnigrams(const std::string& fname); - - const bool use_uniform_; - const prob_t uniform_; - std::vector<prob_t> probs_; -}; - - -// reads an ARPA unigram file and converts words like 'cat' into a string 'c a t' -struct UnigramWordModel { - explicit UnigramWordModel(const std::string& fname) : - use_uniform_(false), - uniform_(1.0), - probs_() { - LoadUnigrams(fname); - } - - explicit UnigramWordModel(const unsigned vocab_size) : - use_uniform_(true), - uniform_(1.0 / vocab_size), - probs_() {} - - const prob_t& operator()(const std::vector<WordID>& s) const { - if (use_uniform_) return uniform_; - const VectorProbHash::const_iterator it = probs_.find(s); - assert(it != probs_.end()); - return it->second; - } - - private: - void LoadUnigrams(const std::string& fname); - - const bool use_uniform_; - const prob_t uniform_; - typedef std::tr1::unordered_map<std::vector<WordID>, prob_t, boost::hash<std::vector<WordID> > > VectorProbHash; - VectorProbHash probs_; -}; - -#endif |