summaryrefslogtreecommitdiff
path: root/training/dtrain
diff options
context:
space:
mode:
Diffstat (limited to 'training/dtrain')
-rw-r--r--training/dtrain/dtrain.cc28
-rw-r--r--training/dtrain/dtrain.h74
-rw-r--r--training/dtrain/examples/standard/dtrain.ini24
-rw-r--r--training/dtrain/examples/standard/expected-output84
-rw-r--r--training/dtrain/kbestget.h66
-rw-r--r--training/dtrain/ksampler.h5
-rw-r--r--training/dtrain/score.h17
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
generated by cgit v1.2.3 (git 2.39.1) at 2024-11-25 20:09:05 +0000