diff options
author | Patrick Simianer <p@simianer.de> | 2013-04-29 15:24:39 +0200 |
---|---|---|
committer | Patrick Simianer <p@simianer.de> | 2013-04-29 15:24:39 +0200 |
commit | b6754386f1109b960b05cdf2eabbc97bdd38e8df (patch) | |
tree | b0c180d31be4851bc23738f5efaa86d35b18e39a | |
parent | 4b942c98e9b86507efe426b5eb36066151e8566b (diff) |
fix, cleaned up headers
-rw-r--r-- | training/dtrain/dtrain.cc | 28 | ||||
-rw-r--r-- | training/dtrain/dtrain.h | 74 | ||||
-rw-r--r-- | training/dtrain/examples/standard/dtrain.ini | 24 | ||||
-rw-r--r-- | training/dtrain/examples/standard/expected-output | 84 | ||||
-rw-r--r-- | training/dtrain/kbestget.h | 66 | ||||
-rw-r--r-- | training/dtrain/ksampler.h | 5 | ||||
-rw-r--r-- | training/dtrain/score.h | 17 |
7 files changed, 144 insertions, 154 deletions
diff --git a/training/dtrain/dtrain.cc b/training/dtrain/dtrain.cc index 149f87d4..83e4e440 100644 --- a/training/dtrain/dtrain.cc +++ b/training/dtrain/dtrain.cc @@ -1,4 +1,10 @@ #include "dtrain.h" +#include "score.h" +#include "kbestget.h" +#include "ksampler.h" +#include "pairsampling.h" + +using namespace dtrain; bool @@ -138,23 +144,23 @@ main(int argc, char** argv) string scorer_str = cfg["scorer"].as<string>(); LocalScorer* scorer; if (scorer_str == "bleu") { - scorer = dynamic_cast<BleuScorer*>(new BleuScorer); + scorer = static_cast<BleuScorer*>(new BleuScorer); } else if (scorer_str == "stupid_bleu") { - scorer = dynamic_cast<StupidBleuScorer*>(new StupidBleuScorer); + scorer = static_cast<StupidBleuScorer*>(new StupidBleuScorer); } else if (scorer_str == "fixed_stupid_bleu") { - scorer = dynamic_cast<FixedStupidBleuScorer*>(new FixedStupidBleuScorer); + scorer = static_cast<FixedStupidBleuScorer*>(new FixedStupidBleuScorer); } else if (scorer_str == "smooth_bleu") { - scorer = dynamic_cast<SmoothBleuScorer*>(new SmoothBleuScorer); + scorer = static_cast<SmoothBleuScorer*>(new SmoothBleuScorer); } else if (scorer_str == "sum_bleu") { - scorer = dynamic_cast<SumBleuScorer*>(new SumBleuScorer); + scorer = static_cast<SumBleuScorer*>(new SumBleuScorer); } else if (scorer_str == "sumexp_bleu") { - scorer = dynamic_cast<SumExpBleuScorer*>(new SumExpBleuScorer); + scorer = static_cast<SumExpBleuScorer*>(new SumExpBleuScorer); } else if (scorer_str == "sumwhatever_bleu") { - scorer = dynamic_cast<SumWhateverBleuScorer*>(new SumWhateverBleuScorer); + scorer = static_cast<SumWhateverBleuScorer*>(new SumWhateverBleuScorer); } else if (scorer_str == "approx_bleu") { - scorer = dynamic_cast<ApproxBleuScorer*>(new ApproxBleuScorer(N, approx_bleu_d)); + scorer = static_cast<ApproxBleuScorer*>(new ApproxBleuScorer(N, approx_bleu_d)); } else if (scorer_str == "lc_bleu") { - scorer = dynamic_cast<LinearBleuScorer*>(new LinearBleuScorer(N)); + scorer = static_cast<LinearBleuScorer*>(new LinearBleuScorer(N)); } else { cerr << "Don't know scoring metric: '" << scorer_str << "', exiting." << endl; exit(1); @@ -166,9 +172,9 @@ main(int argc, char** argv) MT19937 rng; // random number generator, only for forest sampling HypSampler* observer; if (sample_from == "kbest") - observer = dynamic_cast<KBestGetter*>(new KBestGetter(k, filter_type)); + observer = static_cast<KBestGetter*>(new KBestGetter(k, filter_type)); else - observer = dynamic_cast<KSampler*>(new KSampler(k, &rng)); + observer = static_cast<KSampler*>(new KSampler(k, &rng)); observer->SetScorer(scorer); // init weights diff --git a/training/dtrain/dtrain.h b/training/dtrain/dtrain.h index eb0b9f17..3981fb39 100644 --- a/training/dtrain/dtrain.h +++ b/training/dtrain/dtrain.h @@ -11,16 +11,19 @@ #include <boost/algorithm/string.hpp> #include <boost/program_options.hpp> -#include "ksampler.h" -#include "pairsampling.h" - -#include "filelib.h" - +#include "decoder.h" +#include "ff_register.h" +#include "sentence_metadata.h" +#include "verbose.h" +#include "viterbi.h" using namespace std; -using namespace dtrain; namespace po = boost::program_options; +namespace dtrain +{ + + inline void register_and_convert(const vector<string>& strs, vector<WordID>& ids) { vector<string>::const_iterator it; @@ -42,17 +45,55 @@ inline string gettmpf(const string path, const string infix) return string(fn); } -inline void split_in(string& s, vector<string>& parts) +typedef double score_t; + +struct ScoredHyp { - unsigned f = 0; - for(unsigned i = 0; i < 3; i++) { - unsigned e = f; - f = s.find("\t", f+1); - if (e != 0) parts.push_back(s.substr(e+1, f-e-1)); - else parts.push_back(s.substr(0, f)); + vector<WordID> w; + SparseVector<double> f; + score_t model; + score_t score; + unsigned rank; +}; + +struct LocalScorer +{ + unsigned N_; + vector<score_t> w_; + + virtual score_t + Score(vector<WordID>& hyp, vector<WordID>& ref, const unsigned rank, const unsigned src_len)=0; + + virtual void Reset() {} // only for ApproxBleuScorer, LinearBleuScorer + + inline void + Init(unsigned N, vector<score_t> weights) + { + assert(N > 0); + N_ = N; + if (weights.empty()) for (unsigned i = 0; i < N_; i++) w_.push_back(1./N_); + else w_ = weights; } - s.erase(0, f+1); -} + + inline score_t + brevity_penalty(const unsigned hyp_len, const unsigned ref_len) + { + if (hyp_len > ref_len) return 1; + return exp(1 - (score_t)ref_len/hyp_len); + } +}; + +struct HypSampler : public DecoderObserver +{ + LocalScorer* scorer_; + vector<WordID>* ref_; + unsigned f_count_, sz_; + virtual vector<ScoredHyp>* GetSamples()=0; + inline void SetScorer(LocalScorer* scorer) { scorer_ = scorer; } + inline void SetRef(vector<WordID>& ref) { ref_ = &ref; } + inline unsigned get_f_count() { return f_count_; } + inline unsigned get_sz() { return sz_; } +}; struct HSReporter { @@ -88,5 +129,8 @@ inline T sign(T z) return z < 0 ? -1 : +1; } + +} // namespace + #endif diff --git a/training/dtrain/examples/standard/dtrain.ini b/training/dtrain/examples/standard/dtrain.ini index e1072d30..23e94285 100644 --- a/training/dtrain/examples/standard/dtrain.ini +++ b/training/dtrain/examples/standard/dtrain.ini @@ -10,15 +10,15 @@ print_weights=Glue WordPenalty LanguageModel LanguageModel_OOV PhraseModel_0 Phr stop_after=10 # stop epoch after 10 inputs # interesting stuff -epochs=2 # run over input 2 times -k=100 # use 100best lists -N=4 # optimize (approx) BLEU4 -scorer=stupid_bleu # use 'stupid' BLEU+1 -learning_rate=1.0 # learning rate, don't care if gamma=0 (perceptron) -gamma=0 # use SVM reg -sample_from=kbest # use kbest lists (as opposed to forest) -filter=uniq # only unique entries in kbest (surface form) -pair_sampling=XYX # -hi_lo=0.1 # 10 vs 80 vs 10 and 80 vs 10 here -pair_threshold=0 # minimum distance in BLEU (here: > 0) -loss_margin=0 # update if correctly ranked, but within this margin +epochs=2 # run over input 2 times +k=100 # use 100best lists +N=4 # optimize (approx) BLEU4 +scorer=fixed_stupid_bleu # use 'stupid' BLEU+1 +learning_rate=1.0 # learning rate, don't care if gamma=0 (perceptron) +gamma=0 # use SVM reg +sample_from=kbest # use kbest lists (as opposed to forest) +filter=uniq # only unique entries in kbest (surface form) +pair_sampling=XYX # +hi_lo=0.1 # 10 vs 80 vs 10 and 80 vs 10 here +pair_threshold=0 # minimum distance in BLEU (here: > 0) +loss_margin=0 # update if correctly ranked, but within this margin diff --git a/training/dtrain/examples/standard/expected-output b/training/dtrain/examples/standard/expected-output index 7cd09dbf..9a25062b 100644 --- a/training/dtrain/examples/standard/expected-output +++ b/training/dtrain/examples/standard/expected-output @@ -4,14 +4,14 @@ Reading ./nc-wmt11.en.srilm.gz ----5---10---15---20---25---30---35---40---45---50---55---60---65---70---75---80---85---90---95--100 **************************************************************************************************** Example feature: Shape_S00000_T00000 -Seeding random number sequence to 2679584485 +Seeding random number sequence to 1677737427 dtrain Parameters: k 100 N 4 T 2 - scorer 'stupid_bleu' + scorer 'fixed_stupid_bleu' sample from 'kbest' filter 'uniq' learning rate 1 @@ -34,58 +34,58 @@ Iteration #1 of 2. . 10 Stopping after 10 input sentences. WEIGHTS - Glue = -576 - WordPenalty = +417.79 - LanguageModel = +5117.5 - LanguageModel_OOV = -1307 - PhraseModel_0 = -1612 - PhraseModel_1 = -2159.6 - PhraseModel_2 = -677.36 - PhraseModel_3 = +2663.8 - PhraseModel_4 = -1025.9 - PhraseModel_5 = -8 - PhraseModel_6 = +70 - PassThrough = -1455 + Glue = -1155 + WordPenalty = -329.63 + LanguageModel = +3903 + LanguageModel_OOV = -1630 + PhraseModel_0 = +2746.9 + PhraseModel_1 = +1200.3 + PhraseModel_2 = -1004.1 + PhraseModel_3 = +2223.1 + PhraseModel_4 = +551.58 + PhraseModel_5 = +217 + PhraseModel_6 = +1816 + PassThrough = -1603 --- - 1best avg score: 0.27697 (+0.27697) - 1best avg model score: -47918 (-47918) - avg # pairs: 581.9 (meaningless) - avg # rank err: 581.9 + 1best avg score: 0.19344 (+0.19344) + 1best avg model score: 81387 (+81387) + avg # pairs: 616.3 (meaningless) + avg # rank err: 616.3 avg # margin viol: 0 - non0 feature count: 703 + non0 feature count: 673 avg list sz: 90.9 - avg f count: 100.09 -(time 0.25 min, 1.5 s/S) + avg f count: 104.26 +(time 0.38 min, 2.3 s/S) Iteration #2 of 2. . 10 WEIGHTS - Glue = -622 - WordPenalty = +898.56 - LanguageModel = +8066.2 - LanguageModel_OOV = -2590 - PhraseModel_0 = -4335.8 - PhraseModel_1 = -5864.4 - PhraseModel_2 = -1729.8 - PhraseModel_3 = +2831.9 - PhraseModel_4 = -5384.8 - PhraseModel_5 = +1449 - PhraseModel_6 = +480 - PassThrough = -2578 + Glue = -994 + WordPenalty = -778.69 + LanguageModel = +2348.9 + LanguageModel_OOV = -1967 + PhraseModel_0 = -412.72 + PhraseModel_1 = +1428.9 + PhraseModel_2 = +1967.4 + PhraseModel_3 = -944.99 + PhraseModel_4 = -239.7 + PhraseModel_5 = +708 + PhraseModel_6 = +645 + PassThrough = -1866 --- - 1best avg score: 0.37119 (+0.094226) - 1best avg model score: -1.3174e+05 (-83822) - avg # pairs: 584.1 (meaningless) - avg # rank err: 584.1 + 1best avg score: 0.22395 (+0.03051) + 1best avg model score: -31388 (-1.1278e+05) + avg # pairs: 702.3 (meaningless) + avg # rank err: 702.3 avg # margin viol: 0 - non0 feature count: 1115 + non0 feature count: 955 avg list sz: 91.3 - avg f count: 90.755 -(time 0.3 min, 1.8 s/S) + avg f count: 103.45 +(time 0.32 min, 1.9 s/S) Writing weights file to '-' ... done --- -Best iteration: 2 [SCORE 'stupid_bleu'=0.37119]. -This took 0.55 min. +Best iteration: 2 [SCORE 'fixed_stupid_bleu'=0.22395]. +This took 0.7 min. diff --git a/training/dtrain/kbestget.h b/training/dtrain/kbestget.h index dd8882e1..85252db3 100644 --- a/training/dtrain/kbestget.h +++ b/training/dtrain/kbestget.h @@ -1,76 +1,12 @@ #ifndef _DTRAIN_KBESTGET_H_ #define _DTRAIN_KBESTGET_H_ -#include "kbest.h" // cdec -#include "sentence_metadata.h" - -#include "verbose.h" -#include "viterbi.h" -#include "ff_register.h" -#include "decoder.h" -#include "weights.h" -#include "logval.h" - -using namespace std; +#include "kbest.h" namespace dtrain { -typedef double score_t; - -struct ScoredHyp -{ - vector<WordID> w; - SparseVector<double> f; - score_t model; - score_t score; - unsigned rank; -}; - -struct LocalScorer -{ - unsigned N_; - vector<score_t> w_; - - virtual score_t - Score(vector<WordID>& hyp, vector<WordID>& ref, const unsigned rank, const unsigned src_len)=0; - - void Reset() {} // only for approx bleu - - inline void - Init(unsigned N, vector<score_t> weights) - { - assert(N > 0); - N_ = N; - if (weights.empty()) for (unsigned i = 0; i < N_; i++) w_.push_back(1./N_); - else w_ = weights; - } - - inline score_t - brevity_penalty(const unsigned hyp_len, const unsigned ref_len) - { - if (hyp_len > ref_len) return 1; - return exp(1 - (score_t)ref_len/hyp_len); - } -}; - -struct HypSampler : public DecoderObserver -{ - LocalScorer* scorer_; - vector<WordID>* ref_; - unsigned f_count_, sz_; - virtual vector<ScoredHyp>* GetSamples()=0; - inline void SetScorer(LocalScorer* scorer) { scorer_ = scorer; } - inline void SetRef(vector<WordID>& ref) { ref_ = &ref; } - inline unsigned get_f_count() { return f_count_; } - inline unsigned get_sz() { return sz_; } -}; -//////////////////////////////////////////////////////////////////////////////// - - - - struct KBestGetter : public HypSampler { const unsigned k_; diff --git a/training/dtrain/ksampler.h b/training/dtrain/ksampler.h index bc2f56cd..29dab667 100644 --- a/training/dtrain/ksampler.h +++ b/training/dtrain/ksampler.h @@ -1,13 +1,12 @@ #ifndef _DTRAIN_KSAMPLER_H_ #define _DTRAIN_KSAMPLER_H_ -#include "hg_sampler.h" // cdec -#include "kbestget.h" -#include "score.h" +#include "hg_sampler.h" namespace dtrain { + bool cmp_hyp_by_model_d(ScoredHyp a, ScoredHyp b) { diff --git a/training/dtrain/score.h b/training/dtrain/score.h index bddaa071..53e970ba 100644 --- a/training/dtrain/score.h +++ b/training/dtrain/score.h @@ -1,9 +1,7 @@ #ifndef _DTRAIN_SCORE_H_ #define _DTRAIN_SCORE_H_ -#include "kbestget.h" - -using namespace std; +#include "dtrain.h" namespace dtrain { @@ -141,36 +139,43 @@ struct BleuScorer : public LocalScorer { score_t Bleu(NgramCounts& counts, const unsigned hyp_len, const unsigned ref_len); score_t Score(vector<WordID>& hyp, vector<WordID>& ref, const unsigned /*rank*/, const unsigned /*src_len*/); + void Reset() {} }; struct StupidBleuScorer : public LocalScorer { score_t Score(vector<WordID>& hyp, vector<WordID>& ref, const unsigned /*rank*/, const unsigned /*src_len*/); + void Reset() {} }; struct FixedStupidBleuScorer : public LocalScorer { score_t Score(vector<WordID>& hyp, vector<WordID>& ref, const unsigned /*rank*/, const unsigned /*src_len*/); + void Reset() {} }; struct SmoothBleuScorer : public LocalScorer { score_t Score(vector<WordID>& hyp, vector<WordID>& ref, const unsigned /*rank*/, const unsigned /*src_len*/); + void Reset() {} }; struct SumBleuScorer : public LocalScorer { - score_t Score(vector<WordID>& hyp, vector<WordID>& ref, const unsigned /*rank*/, const unsigned /*src_len*/); + score_t Score(vector<WordID>& hyp, vector<WordID>& ref, const unsigned /*rank*/, const unsigned /*src_len*/); + void Reset() {} }; struct SumExpBleuScorer : public LocalScorer { - score_t Score(vector<WordID>& hyp, vector<WordID>& ref, const unsigned /*rank*/, const unsigned /*src_len*/); + score_t Score(vector<WordID>& hyp, vector<WordID>& ref, const unsigned /*rank*/, const unsigned /*src_len*/); + void Reset() {} }; struct SumWhateverBleuScorer : public LocalScorer { - score_t Score(vector<WordID>& hyp, vector<WordID>& ref, const unsigned /*rank*/, const unsigned /*src_len*/); + score_t Score(vector<WordID>& hyp, vector<WordID>& ref, const unsigned /*rank*/, const unsigned /*src_len*/); + void Reset() {}; }; struct ApproxBleuScorer : public BleuScorer |