From 8fb12f413a0974ee0e6fe0c04b3f760463cf9e30 Mon Sep 17 00:00:00 2001 From: Patrick Simianer Date: Mon, 21 May 2012 11:42:10 +0200 Subject: merging --- dtrain/README.md | 7 +++++++ 1 file changed, 7 insertions(+) (limited to 'dtrain/README.md') diff --git a/dtrain/README.md b/dtrain/README.md index 9580df6d..45f21ad5 100644 --- a/dtrain/README.md +++ b/dtrain/README.md @@ -48,6 +48,13 @@ Next + make svm doable; no subgradient? + reranking while sgd? + try PRO, mira emulations ++ sampling (MBR) ++ forest (on train)? ++ best BLEU transl (Sokolov)? ++ entire reg. path ++ resharding [nfold cross val.] ++ bigger LM, feats (target side Ng., word alignments etc.) ++ merge kbest lists Legal ----- -- cgit v1.2.3 From aadabfdf37dfd451485277cb77fad02f77b361c6 Mon Sep 17 00:00:00 2001 From: Patrick Simianer Date: Wed, 23 May 2012 09:46:27 +0200 Subject: README++ --- dtrain/README.md | 2 ++ 1 file changed, 2 insertions(+) (limited to 'dtrain/README.md') diff --git a/dtrain/README.md b/dtrain/README.md index 45f21ad5..350c7423 100644 --- a/dtrain/README.md +++ b/dtrain/README.md @@ -41,6 +41,8 @@ DTRAIN_LOCAL. Next ---- ++ approx. Bleu? ++ turn off inclusion + (dtrain|decoder) more meta-parameters testing + feature selection directly in dtrain + feature template: target side rule ngrams -- cgit v1.2.3 From 62c805c90c5347b844f92574e240db5c65578e12 Mon Sep 17 00:00:00 2001 From: Patrick Simianer Date: Thu, 31 May 2012 14:33:59 +0200 Subject: new scorer, stuff --- dtrain/README.md | 3 ++- dtrain/dtrain.cc | 75 +++++++++++++++++++++++++++------------------------ dtrain/pairsampling.h | 6 ++--- dtrain/score.cc | 35 +++++++++++++++++++++++- dtrain/score.h | 49 ++++++++++++++++++++++++++++++--- 5 files changed, 125 insertions(+), 43 deletions(-) (limited to 'dtrain/README.md') diff --git a/dtrain/README.md b/dtrain/README.md index 350c7423..7aefcc55 100644 --- a/dtrain/README.md +++ b/dtrain/README.md @@ -41,7 +41,7 @@ DTRAIN_LOCAL. Next ---- -+ approx. Bleu? ++ approx. Bleu? proper lc_bleu (init with X) + turn off inclusion + (dtrain|decoder) more meta-parameters testing + feature selection directly in dtrain @@ -57,6 +57,7 @@ Next + resharding [nfold cross val.] + bigger LM, feats (target side Ng., word alignments etc.) + merge kbest lists ++ proper eval, pairwise ranking, forced transl Legal ----- diff --git a/dtrain/dtrain.cc b/dtrain/dtrain.cc index 717d47a2..88413a1d 100644 --- a/dtrain/dtrain.cc +++ b/dtrain/dtrain.cc @@ -6,38 +6,39 @@ dtrain_init(int argc, char** argv, po::variables_map* cfg) { po::options_description ini("Configuration File Options"); ini.add_options() - ("input", po::value()->default_value("-"), "input file") - ("output", po::value()->default_value("-"), "output weights file, '-' for STDOUT") - ("input_weights", po::value(), "input weights file (e.g. from previous iteration)") - ("decoder_config", po::value(), "configuration file for cdec") - ("print_weights", po::value(), "weights to print on each iteration") - ("stop_after", po::value()->default_value(0), "stop after X input sentences") - ("tmp", po::value()->default_value("/tmp"), "temp dir to use") - ("keep", po::value()->zero_tokens(), "keep weights files for each iteration") - ("hstreaming", po::value(), "run in hadoop streaming mode, arg is a task id") - ("epochs", po::value()->default_value(10), "# of iterations T (per shard)") - ("k", po::value()->default_value(100), "how many translations to sample") - ("sample_from", po::value()->default_value("kbest"), "where to sample translations from: 'kbest', 'forest'") - ("filter", po::value()->default_value("uniq"), "filter kbest list: 'not', 'uniq'") - ("pair_sampling", po::value()->default_value("XYX"), "how to sample pairs: 'all', 'XYX' or 'PRO'") - ("hi_lo", po::value()->default_value(0.1), "hi and lo (X) for XYX (default 0.1), <= 0.5") - ("pair_threshold", po::value()->default_value(0.), "bleu [0,1] threshold to filter pairs") - ("N", po::value()->default_value(4), "N for Ngrams (BLEU)") - ("scorer", po::value()->default_value("stupid_bleu"), "scoring: bleu, stupid_, smooth_, approx_") - ("learning_rate", po::value()->default_value(0.0001), "learning rate") - ("gamma", po::value()->default_value(0.), "gamma for SVM (0 for perceptron)") - ("select_weights", po::value()->default_value("last"), "output best, last, avg weights ('VOID' to throw away)") - ("rescale", po::value()->zero_tokens(), "rescale weight vector after each input") - ("l1_reg", po::value()->default_value("none"), "apply l1 regularization as in 'Tsuroka et al' (2010)") - ("l1_reg_strength", po::value(), "l1 regularization strength") - ("fselect", po::value()->default_value(-1), "TODO select top x percent (or by threshold) of features after each epoch") - ("approx_bleu_d", po::value()->default_value(0.9), "discount for approx. BLEU") - ("scale_bleu_diff", po::value()->zero_tokens(), "learning rate <- bleu diff of a misranked pair") - ("loss_margin", po::value()->default_value(0.), "update if no error in pref pair but model scores this near") + ("input", po::value()->default_value("-"), "input file") + ("output", po::value()->default_value("-"), "output weights file, '-' for STDOUT") + ("input_weights", po::value(), "input weights file (e.g. from previous iteration)") + ("decoder_config", po::value(), "configuration file for cdec") + ("print_weights", po::value(), "weights to print on each iteration") + ("stop_after", po::value()->default_value(0), "stop after X input sentences") + ("tmp", po::value()->default_value("/tmp"), "temp dir to use") + ("keep", po::value()->zero_tokens(), "keep weights files for each iteration") + ("hstreaming", po::value(), "run in hadoop streaming mode, arg is a task id") + ("epochs", po::value()->default_value(10), "# of iterations T (per shard)") + ("k", po::value()->default_value(100), "how many translations to sample") + ("sample_from", po::value()->default_value("kbest"), "where to sample translations from: 'kbest', 'forest'") + ("filter", po::value()->default_value("uniq"), "filter kbest list: 'not', 'uniq'") + ("pair_sampling", po::value()->default_value("XYX"), "how to sample pairs: 'all', 'XYX' or 'PRO'") + ("hi_lo", po::value()->default_value(0.1), "hi and lo (X) for XYX (default 0.1), <= 0.5") + ("pair_threshold", po::value()->default_value(0.), "bleu [0,1] threshold to filter pairs") + ("N", po::value()->default_value(4), "N for Ngrams (BLEU)") + ("scorer", po::value()->default_value("stupid_bleu"), "scoring: bleu, stupid_, smooth_, approx_, lc_") + ("learning_rate", po::value()->default_value(0.0001), "learning rate") + ("gamma", po::value()->default_value(0.), "gamma for SVM (0 for perceptron)") + ("select_weights", po::value()->default_value("last"), "output best, last, avg weights ('VOID' to throw away)") + ("rescale", po::value()->zero_tokens(), "rescale weight vector after each input") + ("l1_reg", po::value()->default_value("none"), "apply l1 regularization as in 'Tsuroka et al' (2010)") + ("l1_reg_strength", po::value(), "l1 regularization strength") + ("fselect", po::value()->default_value(-1), "select top x percent (or by threshold) of features after each epoch NOT IMPL") // TODO + ("approx_bleu_d", po::value()->default_value(0.9), "discount for approx. BLEU") + ("scale_bleu_diff", po::value()->zero_tokens(), "learning rate <- bleu diff of a misranked pair") + ("loss_margin", po::value()->default_value(0.), "update if no error in pref pair but model scores this near") + ("max_pairs", po::value()->default_value(std::numeric_limits::max()), "max. # of pairs per Sent.") #ifdef DTRAIN_LOCAL - ("refs,r", po::value(), "references in local mode") + ("refs,r", po::value(), "references in local mode") #endif - ("noup", po::value()->zero_tokens(), "do not update weights"); + ("noup", po::value()->zero_tokens(), "do not update weights"); po::options_description cl("Command Line Options"); cl.add_options() ("config,c", po::value(), "dtrain config file") @@ -135,6 +136,7 @@ main(int argc, char** argv) const string select_weights = cfg["select_weights"].as(); const float hi_lo = cfg["hi_lo"].as(); const score_t approx_bleu_d = cfg["approx_bleu_d"].as(); + const unsigned max_pairs = cfg["max_pairs"].as(); weight_t loss_margin = cfg["loss_margin"].as(); if (loss_margin > 9998.) loss_margin = std::numeric_limits::max(); bool scale_bleu_diff = false; @@ -167,6 +169,8 @@ main(int argc, char** argv) scorer = dynamic_cast(new SmoothSingleBleuScorer); } else if (scorer_str == "approx_bleu") { scorer = dynamic_cast(new ApproxBleuScorer(N, approx_bleu_d)); + } else if (scorer_str == "lc_bleu") { + scorer = dynamic_cast(new LinearBleuScorer(N)); } else { cerr << "Don't know scoring metric: '" << scorer_str << "', exiting." << endl; exit(1); @@ -257,6 +261,7 @@ main(int argc, char** argv) cerr << setw(25) << "l1 reg " << l1_reg << " '" << cfg["l1_reg"].as() << "'" << endl; if (rescale) cerr << setw(25) << "rescale " << rescale << endl; + cerr << "max pairs " << max_pairs << endl; cerr << setw(25) << "cdec cfg " << "'" << cfg["decoder_config"].as() << "'" << endl; cerr << setw(25) << "input " << "'" << input_fn << "'" << endl; #ifdef DTRAIN_LOCAL @@ -421,17 +426,17 @@ main(int argc, char** argv) // get pairs vector > pairs; if (pair_sampling == "all") - all_pairs(samples, pairs, pair_threshold); + all_pairs(samples, pairs, pair_threshold, max_pairs); if (pair_sampling == "XYX") - partXYX(samples, pairs, pair_threshold, hi_lo); + partXYX(samples, pairs, pair_threshold, max_pairs, hi_lo); if (pair_sampling == "PRO") - PROsampling(samples, pairs, pair_threshold); + PROsampling(samples, pairs, pair_threshold, max_pairs); npairs += pairs.size(); for (vector >::iterator it = pairs.begin(); it != pairs.end(); it++) { #ifdef DTRAIN_FASTER_PERCEPTRON - bool rank_error = true; // pair filtering already did this for us + bool rank_error = true; // pair sampling already did this for us rank_errors++; score_t margin = std::numeric_limits::max(); #else @@ -498,7 +503,7 @@ main(int argc, char** argv) if (average) w_average += lambdas; - if (scorer_str == "approx_bleu") scorer->Reset(); + if (scorer_str == "approx_bleu" || scorer_str == "lc_bleu") scorer->Reset(); if (t == 0) { in_sz = ii; // remember size of input (# lines) diff --git a/dtrain/pairsampling.h b/dtrain/pairsampling.h index 32006a41..71c8ae59 100644 --- a/dtrain/pairsampling.h +++ b/dtrain/pairsampling.h @@ -19,7 +19,7 @@ cmp_hyp_by_score_d(ScoredHyp a, ScoredHyp b) } inline void -all_pairs(vector* s, vector >& training, score_t threshold, float _unused=1) +all_pairs(vector* s, vector >& training, score_t threshold, unsigned max, float _unused=1) { sort(s->begin(), s->end(), cmp_hyp_by_score_d); unsigned sz = s->size(); @@ -44,7 +44,7 @@ all_pairs(vector* s, vector >& training, sc */ inline void -partXYX(vector* s, vector >& training, score_t threshold, float hi_lo) +partXYX(vector* s, vector >& training, score_t threshold, unsigned max, float hi_lo) { unsigned sz = s->size(); if (sz < 2) return; @@ -104,7 +104,7 @@ _PRO_cmp_pair_by_diff_d(pair a, pair b return (fabs(a.first.score - a.second.score)) > (fabs(b.first.score - b.second.score)); } inline void -PROsampling(vector* s, vector >& training, score_t threshold, float _unused=1) +PROsampling(vector* s, vector >& training, score_t threshold, unsigned max, float _unused=1) { unsigned max_count = 5000, count = 0, sz = s->size(); bool b = false; diff --git a/dtrain/score.cc b/dtrain/score.cc index b331dc4f..5c356c0f 100644 --- a/dtrain/score.cc +++ b/dtrain/score.cc @@ -122,12 +122,13 @@ SmoothSingleBleuScorer::Score(vector& hyp, vector& ref, unsigned j = 1; for (unsigned i = 0; i < M; i++) { if (counts.sum_[i] == 0 || counts.clipped_[i] == 0) break; - sum += ((score_t)counts.clipped_[i]/counts.sum_[i])/pow(2.0, N_-j+1); + sum += ((score_t)counts.clipped_[i]/counts.sum_[i])/pow(2., N_-j+1); j++; } return brevity_penalty(hyp_len, ref_len) * sum; } + /* * approx. bleu * @@ -160,6 +161,38 @@ ApproxBleuScorer::Score(vector& hyp, vector& ref, return (score_t)glob_src_len_ * score; } +/* + * Linear (Corpus) Bleu + * + * as in "Lattice Minimum Bayes-Risk Decoding + * for Statistical Machine Translation" + * (Tromble et al. '08) + * + */ +score_t +LinearBleuScorer::Score(vector& hyp, vector& ref, + const unsigned rank, const unsigned /*src_len*/) +{ + unsigned hyp_len = hyp.size(), ref_len = ref.size(); + if (ref_len == 0) return 0.; + unsigned M = N_; + if (ref_len < N_) M = ref_len; + NgramCounts counts(M); + if (hyp_len > 0) + counts = make_ngram_counts(hyp, ref, M); + score_t ret = 0.; + for (unsigned i = 0; i < M; i++) { + if (counts.sum_[i] == 0 || onebest_counts_.sum_[i] == 0) break; + ret += counts.sum_[i]/onebest_counts_.sum_[i]; + } + ret = -(hyp_len/(score_t)onebest_len_) + (1./M) * ret; + if (rank == 0) { + onebest_len_ += hyp_len; + onebest_counts_ += counts; + } + return ret; +} + } // namespace diff --git a/dtrain/score.h b/dtrain/score.h index d4fba22c..c5be2829 100644 --- a/dtrain/score.h +++ b/dtrain/score.h @@ -20,7 +20,7 @@ struct NgramCounts inline void operator+=(const NgramCounts& rhs) { - assert(N_ == rhs.N_); + if (rhs.N_ > N_) Resize(rhs.N_); for (unsigned i = 0; i < N_; i++) { this->clipped_[i] += rhs.clipped_.find(i)->second; this->sum_[i] += rhs.sum_.find(i)->second; @@ -59,13 +59,21 @@ struct NgramCounts inline void Zero() { - unsigned i; - for (i = 0; i < N_; i++) { + for (unsigned i = 0; i < N_; i++) { clipped_[i] = 0.; sum_[i] = 0.; } } + inline void + One() + { + for (unsigned i = 0; i < N_; i++) { + clipped_[i] = 1.; + sum_[i] = 1.; + } + } + inline void Print() { @@ -74,6 +82,23 @@ struct NgramCounts cout << i+1 << "grams:\t\t\t" << sum_[i] << endl; } } + + inline void Resize(unsigned N) + { + if (N == N_) return; + else if (N > N_) { + for (unsigned i = N_; i < N; i++) { + clipped_[i] = 0.; + sum_[i] = 0.; + } + } else { // N < N_ + for (unsigned i = N_-1; i > N-1; i--) { + clipped_.erase(i); + sum_.erase(i); + } + } + N_ = N; + } }; typedef map, unsigned> Ngrams; @@ -152,6 +177,24 @@ struct ApproxBleuScorer : public BleuScorer score_t Score(vector& hyp, vector& ref, const unsigned rank, const unsigned src_len); }; +struct LinearBleuScorer : public BleuScorer +{ + unsigned onebest_len_; + NgramCounts onebest_counts_; + + LinearBleuScorer(unsigned N) : onebest_len_(1), onebest_counts_(N) + { + onebest_counts_.One(); + } + + score_t Score(vector& hyp, vector& ref, const unsigned rank, const unsigned /*src_len*/); + + inline void Reset() { + onebest_len_ = 1; + onebest_counts_.One(); + } +}; + } // namespace -- cgit v1.2.3 From 1d1172a3f85ab7423ae9537cf3c73afdfe7dc693 Mon Sep 17 00:00:00 2001 From: Patrick Simianer Date: Thu, 14 Jun 2012 20:11:57 +0200 Subject: 'sum bleu' --- dtrain/README.md | 2 ++ dtrain/dtrain.cc | 8 ++++++-- dtrain/dtrain.h | 2 +- dtrain/score.cc | 62 ++++++++++++++++++++++++++++++++++++++++++++++++++++---- dtrain/score.h | 12 ++++++++++- 5 files changed, 78 insertions(+), 8 deletions(-) (limited to 'dtrain/README.md') diff --git a/dtrain/README.md b/dtrain/README.md index 7aefcc55..843874fa 100644 --- a/dtrain/README.md +++ b/dtrain/README.md @@ -58,6 +58,8 @@ Next + bigger LM, feats (target side Ng., word alignments etc.) + merge kbest lists + proper eval, pairwise ranking, forced transl ++ smmoth bleu variants X ++ MMERT exp Legal ----- diff --git a/dtrain/dtrain.cc b/dtrain/dtrain.cc index 7f44d4cf..eea58393 100644 --- a/dtrain/dtrain.cc +++ b/dtrain/dtrain.cc @@ -165,8 +165,12 @@ main(int argc, char** argv) scorer = dynamic_cast(new StupidBleuScorer); } else if (scorer_str == "smooth_bleu") { scorer = dynamic_cast(new SmoothBleuScorer); - } else if (scorer_str == "smooth_single_bleu") { - scorer = dynamic_cast(new SmoothSingleBleuScorer); + } else if (scorer_str == "sum_bleu") { + scorer = dynamic_cast(new SumBleuScorer); + } else if (scorer_str == "sumexp_bleu") { + scorer = dynamic_cast(new SumExpBleuScorer); + } else if (scorer_str == "sumwhatever_bleu") { + scorer = dynamic_cast(new SumWhateverBleuScorer); } else if (scorer_str == "approx_bleu") { scorer = dynamic_cast(new ApproxBleuScorer(N, approx_bleu_d)); } else if (scorer_str == "lc_bleu") { diff --git a/dtrain/dtrain.h b/dtrain/dtrain.h index d8dc14b6..7e084a79 100644 --- a/dtrain/dtrain.h +++ b/dtrain/dtrain.h @@ -3,7 +3,7 @@ #undef DTRAIN_FASTER_PERCEPTRON // only look at misranked pairs // DO NOT USE WITH SVM! -#undef DTRAIN_LOCAL +#define DTRAIN_LOCAL #define DTRAIN_DOTS 10 // after how many inputs to display a '.' #define DTRAIN_GRAMMAR_DELIM "########EOS########" #define DTRAIN_SCALE 100000 diff --git a/dtrain/score.cc b/dtrain/score.cc index 5bb0bcaa..4a7cac6e 100644 --- a/dtrain/score.cc +++ b/dtrain/score.cc @@ -80,7 +80,7 @@ StupidBleuScorer::Score(vector& hyp, vector& ref, * to Machine Translation" * (Liang et al. '06) * - * NOTE: max is 0.9375 + * NOTE: max is 0.9375 (with N=4) */ score_t SmoothBleuScorer::Score(vector& hyp, vector& ref, @@ -108,9 +108,13 @@ SmoothBleuScorer::Score(vector& hyp, vector& ref, return brevity_penalty(hyp_len, ref_len) * sum; } -// variant of smooth_bleu; i-Bleu scores only single 'i' +/* + * 'sum' bleu + * + * sum up Ngram precisions + */ score_t -SmoothSingleBleuScorer::Score(vector& hyp, vector& ref, +SumBleuScorer::Score(vector& hyp, vector& ref, const unsigned /*rank*/, const unsigned /*src_len*/) { unsigned hyp_len = hyp.size(), ref_len = ref.size(); @@ -128,7 +132,57 @@ SmoothSingleBleuScorer::Score(vector& hyp, vector& ref, return brevity_penalty(hyp_len, ref_len) * sum; } -// TODO single variants! +/* + * 'sum' (exp) bleu + * + * sum up exp(Ngram precisions) + */ +score_t +SumExpBleuScorer::Score(vector& hyp, vector& ref, + const unsigned /*rank*/, const unsigned /*src_len*/) +{ + unsigned hyp_len = hyp.size(), ref_len = ref.size(); + if (hyp_len == 0 || ref_len == 0) return 0.; + NgramCounts counts = make_ngram_counts(hyp, ref, N_); + unsigned M = N_; + if (ref_len < N_) M = ref_len; + score_t sum = 0.; + unsigned j = 1; + for (unsigned i = 0; i < M; i++) { + if (counts.sum_[i] == 0 || counts.clipped_[i] == 0) break; + sum += exp(((score_t)counts.clipped_[i]/counts.sum_[i]))/pow(2., N_-j+1); + j++; + } + return brevity_penalty(hyp_len, ref_len) * sum; +} + +/* + * 'sum' (whatever) bleu + * + * sum up exp(weight * log(Ngram precisions)) + */ +score_t +SumWhateverBleuScorer::Score(vector& hyp, vector& ref, + const unsigned /*rank*/, const unsigned /*src_len*/) +{ + unsigned hyp_len = hyp.size(), ref_len = ref.size(); + if (hyp_len == 0 || ref_len == 0) return 0.; + NgramCounts counts = make_ngram_counts(hyp, ref, N_); + unsigned M = N_; + vector v = w_; + if (ref_len < N_) { + M = ref_len; + for (unsigned i = 0; i < M; i++) v[i] = 1/((score_t)M); + } + score_t sum = 0.; + unsigned j = 1; + for (unsigned i = 0; i < M; i++) { + if (counts.sum_[i] == 0 || counts.clipped_[i] == 0) break; + sum += exp(v[i] * log(((score_t)counts.clipped_[i]/counts.sum_[i])))/pow(2., N_-j+1); + j++; + } + return brevity_penalty(hyp_len, ref_len) * sum; +} /* * approx. bleu diff --git a/dtrain/score.h b/dtrain/score.h index c5be2829..f317c903 100644 --- a/dtrain/score.h +++ b/dtrain/score.h @@ -153,7 +153,17 @@ struct SmoothBleuScorer : public LocalScorer score_t Score(vector& hyp, vector& ref, const unsigned /*rank*/, const unsigned /*src_len*/); }; -struct SmoothSingleBleuScorer : public LocalScorer +struct SumBleuScorer : public LocalScorer +{ + score_t Score(vector& hyp, vector& ref, const unsigned /*rank*/, const unsigned /*src_len*/); +}; + +struct SumExpBleuScorer : public LocalScorer +{ + score_t Score(vector& hyp, vector& ref, const unsigned /*rank*/, const unsigned /*src_len*/); +}; + +struct SumWhateverBleuScorer : public LocalScorer { score_t Score(vector& hyp, vector& ref, const unsigned /*rank*/, const unsigned /*src_len*/); }; -- cgit v1.2.3 From ee1520c5095ea8648617a3658b20eedfd4dd2007 Mon Sep 17 00:00:00 2001 From: Patrick Simianer Date: Mon, 18 Jun 2012 17:26:33 +0200 Subject: extract_rules cdec param --- .gitignore | 1 + decoder/decoder.cc | 22 +++++++++++++++------- decoder/viterbi.cc | 12 +++++++++++- decoder/viterbi.h | 5 ++++- dtrain/README.md | 22 ---------------------- 5 files changed, 31 insertions(+), 31 deletions(-) (limited to 'dtrain/README.md') diff --git a/.gitignore b/.gitignore index 27c6a739..943e6dc5 100644 --- a/.gitignore +++ b/.gitignore @@ -128,6 +128,7 @@ decoder/rule_lexer.cc training/atools training/collapse_weights training/lbfgs_test +training/libtraining.a training/mr_optimize_reduce training/mr_em_adapted_reduce training/mr_em_map_adapter diff --git a/decoder/decoder.cc b/decoder/decoder.cc index cbb97a0d..333f0fb6 100644 --- a/decoder/decoder.cc +++ b/decoder/decoder.cc @@ -3,6 +3,7 @@ #include #include #include +#include #include "program_options.h" #include "stringlib.h" @@ -187,8 +188,8 @@ struct DecoderImpl { } void SetId(int next_sent_id) { sent_id = next_sent_id - 1; } - void forest_stats(Hypergraph &forest,string name,bool show_tree,bool show_deriv=false) { - cerr << viterbi_stats(forest,name,true,show_tree,show_deriv); + void forest_stats(Hypergraph &forest,string name,bool show_tree,bool show_deriv=false, bool extract_rules=false, boost::shared_ptr extract_file = boost::make_shared()) { + cerr << viterbi_stats(forest,name,true,show_tree,show_deriv,extract_rules, extract_file); cerr << endl; } @@ -424,7 +425,7 @@ DecoderImpl::DecoderImpl(po::variables_map& conf, int argc, char** argv, istream ("tagger_tagset,t", po::value(), "(Tagger) file containing tag set") ("csplit_output_plf", "(Compound splitter) Output lattice in PLF format") ("csplit_preserve_full_word", "(Compound splitter) Always include the unsegmented form in the output lattice") - ("extract_rules", po::value(), "Extract the rules used in translation (de-duped) to this file") + ("extract_rules", po::value(), "Extract the rules used in translation (not de-duped!) to a file in this directory") ("show_derivations", po::value(), "Directory to print the derivation structures to") ("graphviz","Show (constrained) translation forest in GraphViz format") ("max_translation_beam,x", po::value(), "Beam approximation to get max translation from the chart") @@ -570,6 +571,11 @@ DecoderImpl::DecoderImpl(po::variables_map& conf, int argc, char** argv, istream // cube pruning pop-limit: we may want to configure this on a per-pass basis pop_limit = conf["cubepruning_pop_limit"].as(); + if (conf.count("extract_rules")) { + if (!DirectoryExists(conf["extract_rules"].as())) + MkDirP(conf["extract_rules"].as()); + } + // determine the number of rescoring/pruning/weighting passes configured const int MAX_PASSES = 3; for (int pass = 0; pass < MAX_PASSES; ++pass) { @@ -712,9 +718,11 @@ DecoderImpl::DecoderImpl(po::variables_map& conf, int argc, char** argv, istream cfg_options.Validate(); #endif - if (conf.count("extract_rules")) - extract_file.reset(new WriteFile(str("extract_rules",conf))); - + if (conf.count("extract_rules")) { + stringstream ss; + ss << sent_id; + extract_file.reset(new WriteFile(str("extract_rules",conf)+"/"+ss.str())); + } combine_size = conf["combine_size"].as(); if (combine_size < 1) combine_size = 1; sent_id = -1; @@ -851,7 +859,7 @@ bool DecoderImpl::Decode(const string& input, DecoderObserver* o) { #endif forest.swap(rescored_forest); forest.Reweight(cur_weights); - if (!SILENT) forest_stats(forest," " + passtr +" forest",show_tree_structure,oracle.show_derivation); + if (!SILENT) forest_stats(forest," " + passtr +" forest",show_tree_structure,oracle.show_derivation, conf.count("extract_rules"), extract_file); } if (conf.count("show_partition")) { diff --git a/decoder/viterbi.cc b/decoder/viterbi.cc index 9d19914b..1b9c6665 100644 --- a/decoder/viterbi.cc +++ b/decoder/viterbi.cc @@ -5,11 +5,12 @@ #include #include "hg.h" + //#define DEBUG_VITERBI_SORT using namespace std; -std::string viterbi_stats(Hypergraph const& hg, std::string const& name, bool estring, bool etree,bool show_derivation) +std::string viterbi_stats(Hypergraph const& hg, std::string const& name, bool estring, bool etree,bool show_derivation, bool extract_rules, boost::shared_ptr extract_file) { ostringstream o; o << hg.stats(name); @@ -22,6 +23,9 @@ std::string viterbi_stats(Hypergraph const& hg, std::string const& name, bool es if (etree) { o<stream()); + } if (show_derivation) { o< edges; + Viterbi(hg, &edges); + for (unsigned i = 0; i < edges.size(); i++) + (*o) << edges[i]->rule_->AsString(true) << endl; +} string ViterbiETree(const Hypergraph& hg) { vector tmp; diff --git a/decoder/viterbi.h b/decoder/viterbi.h index 3092f6da..03e961a2 100644 --- a/decoder/viterbi.h +++ b/decoder/viterbi.h @@ -5,8 +5,10 @@ #include "prob.h" #include "hg.h" #include "tdict.h" +#include "filelib.h" +#include -std::string viterbi_stats(Hypergraph const& hg, std::string const& name="forest", bool estring=true, bool etree=false, bool derivation_tree=false); +std::string viterbi_stats(Hypergraph const& hg, std::string const& name="forest", bool estring=true, bool etree=false, bool derivation_tree=false, bool extract_rules=false, boost::shared_ptr extract_file = boost::make_shared()); /// computes for each hg node the best (according to WeightType/WeightFunction) derivation, and some homomorphism (bottom up expression tree applied through Traversal) of it. T is the "return type" of Traversal, which is called only once for the best edge for a node's result (i.e. result will start default constructed) //TODO: make T a typename inside Traversal and WeightType a typename inside WeightFunction? @@ -201,6 +203,7 @@ struct FeatureVectorTraversal { std::string JoshuaVisualizationString(const Hypergraph& hg); prob_t ViterbiESentence(const Hypergraph& hg, std::vector* result); std::string ViterbiETree(const Hypergraph& hg); +void ViterbiRules(const Hypergraph& hg, std::ostream* s); prob_t ViterbiFSentence(const Hypergraph& hg, std::vector* result); std::string ViterbiFTree(const Hypergraph& hg); int ViterbiELength(const Hypergraph& hg); diff --git a/dtrain/README.md b/dtrain/README.md index 843874fa..7edabbf1 100644 --- a/dtrain/README.md +++ b/dtrain/README.md @@ -39,28 +39,6 @@ For an example of local usage (with the 'distributed' format) the see test/example/ . This expects dtrain to be built without DTRAIN_LOCAL. -Next ----- -+ approx. Bleu? proper lc_bleu (init with X) -+ turn off inclusion -+ (dtrain|decoder) more meta-parameters testing -+ feature selection directly in dtrain -+ feature template: target side rule ngrams -+ sa-extract -> leave-one-out for grammar of training set? -+ make svm doable; no subgradient? -+ reranking while sgd? -+ try PRO, mira emulations -+ sampling (MBR) -+ forest (on train)? -+ best BLEU transl (Sokolov)? -+ entire reg. path -+ resharding [nfold cross val.] -+ bigger LM, feats (target side Ng., word alignments etc.) -+ merge kbest lists -+ proper eval, pairwise ranking, forced transl -+ smmoth bleu variants X -+ MMERT exp - Legal ----- Copyright (c) 2012 by Patrick Simianer -- cgit v1.2.3