summaryrefslogtreecommitdiff
path: root/dtrain
diff options
context:
space:
mode:
authorPatrick Simianer <p@simianer.de>2011-09-23 22:02:45 +0200
committerPatrick Simianer <p@simianer.de>2011-09-23 22:02:45 +0200
commit0a47dda0e980d1fb6222a1f548649914427555b2 (patch)
treeed163959983dd503699c752964c5178208c61235 /dtrain
parentdc9fd7a3adc863510d79a718e919b6833a86729c (diff)
more renaming, random pair sampler uses boost rng
Diffstat (limited to 'dtrain')
-rw-r--r--dtrain/Makefile.am2
-rw-r--r--dtrain/dtrain.cc10
-rw-r--r--dtrain/hgsampler.cc (renamed from dtrain/sample_hg.cc)2
-rw-r--r--dtrain/hgsampler.h (renamed from dtrain/sample_hg.h)17
-rw-r--r--dtrain/ksampler.h2
-rw-r--r--dtrain/pairsampling.h35
-rw-r--r--dtrain/score.cc165
-rw-r--r--dtrain/score.h53
8 files changed, 133 insertions, 153 deletions
diff --git a/dtrain/Makefile.am b/dtrain/Makefile.am
index 9b5df8bf..12084a70 100644
--- a/dtrain/Makefile.am
+++ b/dtrain/Makefile.am
@@ -1,7 +1,7 @@
# TODO I'm sure I can leave something out.
bin_PROGRAMS = dtrain
-dtrain_SOURCES = dtrain.cc score.cc sample_hg.cc
+dtrain_SOURCES = dtrain.cc score.cc hgsampler.cc
dtrain_LDADD = $(top_srcdir)/decoder/libcdec.a $(top_srcdir)/mteval/libmteval.a $(top_srcdir)/utils/libutils.a ../klm/lm/libklm.a ../klm/util/libklm_util.a -lz -lboost_filesystem -lboost_iostreams
AM_CPPFLAGS = -W -Wall -Wno-sign-compare -I$(top_srcdir)/utils -I$(top_srcdir)/decoder -I$(top_srcdir)/mteval
diff --git a/dtrain/dtrain.cc b/dtrain/dtrain.cc
index 01821b30..01119997 100644
--- a/dtrain/dtrain.cc
+++ b/dtrain/dtrain.cc
@@ -347,10 +347,9 @@ main( int argc, char** argv )
cand_len = kb->sents[i].size();
}
NgramCounts counts_tmp = global_counts + counts;
- // TODO as param
- score = 0.9 * scorer( counts_tmp,
- global_ref_len,
- global_hyp_len + cand_len, N, bleu_weights );
+ score = .9*scorer( counts_tmp,
+ global_ref_len,
+ global_hyp_len + cand_len, N, bleu_weights );
} else {
// other scorers
cand_len = kb->sents[i].size();
@@ -381,7 +380,8 @@ main( int argc, char** argv )
if ( !noup ) {
TrainingInstances pairs;
- sample_all( kb, pairs );
+ sample_all_pairs(kb, pairs);
+ //sample_rand_pairs( kb, pairs, &rng );
for ( TrainingInstances::iterator ti = pairs.begin();
ti != pairs.end(); ti++ ) {
diff --git a/dtrain/sample_hg.cc b/dtrain/hgsampler.cc
index 33872fb8..7a00a3d3 100644
--- a/dtrain/sample_hg.cc
+++ b/dtrain/hgsampler.cc
@@ -1,4 +1,4 @@
-#include "sample_hg.h"
+#include "hgsampler.h"
#include <queue>
diff --git a/dtrain/sample_hg.h b/dtrain/hgsampler.h
index 932fd369..b840c07f 100644
--- a/dtrain/sample_hg.h
+++ b/dtrain/hgsampler.h
@@ -1,5 +1,6 @@
-#ifndef _SAMPLE_HG_H_
-#define _SAMPLE_HG_H_
+#ifndef _DTRAIN_HGSAMPLER_H_
+#define _DTRAIN_HGSAMPLER_H_
+
#include <vector>
#include "sparse_vector.h"
@@ -9,16 +10,20 @@
class Hypergraph;
struct HypergraphSampler {
+
struct Hypothesis {
std::vector<WordID> words;
SparseVector<double> fmap;
prob_t model_score;
};
- static void sample_hypotheses(const Hypergraph& hg,
- unsigned n,
- MT19937* rng,
- std::vector<Hypothesis>* hypos);
+ static void
+ sample_hypotheses(const Hypergraph& hg,
+ unsigned n,
+ MT19937* rng,
+ std::vector<Hypothesis>* hypos);
};
+
#endif
+
diff --git a/dtrain/ksampler.h b/dtrain/ksampler.h
index a28b69c9..914e9723 100644
--- a/dtrain/ksampler.h
+++ b/dtrain/ksampler.h
@@ -2,7 +2,7 @@
#define _DTRAIN_KSAMPLER_H_
#include "kbest.h"
-#include "sample_hg.h"
+#include "hgsampler.h"
#include "sampler.h"
namespace dtrain
diff --git a/dtrain/pairsampling.h b/dtrain/pairsampling.h
index 502901af..9774ba4a 100644
--- a/dtrain/pairsampling.h
+++ b/dtrain/pairsampling.h
@@ -1,9 +1,8 @@
-#ifndef _DTRAIN_SAMPLE_H_
-#define _DTRAIN_SAMPLE_H_
-
+#ifndef _DTRAIN_PAIRSAMPLING_H_
+#define _DTRAIN_PAIRSAMPLING_H_
#include "kbestget.h"
-
+#include "sampler.h" // cdec MT19937
namespace dtrain
{
@@ -11,19 +10,18 @@ namespace dtrain
struct TPair
{
- SparseVector<double> first, second;
- size_t first_rank, second_rank;
- double first_score, second_score;
+ SparseVector<double> first, second;
+ size_t first_rank, second_rank;
+ double first_score, second_score;
};
typedef vector<TPair> TrainingInstances;
-
void
-sample_all( KBestList* kb, TrainingInstances &training )
+sample_all_pairs(KBestList* kb, TrainingInstances &training)
{
- for ( size_t i = 0; i < kb->GetSize()-1; i++ ) {
- for ( size_t j = i+1; j < kb->GetSize(); j++ ) {
+ for (size_t i = 0; i < kb->GetSize()-1; i++) {
+ for (size_t j = i+1; j < kb->GetSize(); j++) {
TPair p;
p.first = kb->feats[i];
p.second = kb->feats[j];
@@ -31,18 +29,18 @@ sample_all( KBestList* kb, TrainingInstances &training )
p.second_rank = j;
p.first_score = kb->scores[i];
p.second_score = kb->scores[j];
- training.push_back( p );
+ training.push_back(p);
}
}
}
void
-sample_rand( KBestList* kb, TrainingInstances &training )
+sample_rand_pairs(KBestList* kb, TrainingInstances &training, MT19937* prng)
{
- srand( time(NULL) );
- for ( size_t i = 0; i < kb->GetSize()-1; i++ ) {
- for ( size_t j = i+1; j < kb->GetSize(); j++ ) {
- if ( rand() % 2 ) {
+ srand(time(NULL));
+ for (size_t i = 0; i < kb->GetSize()-1; i++) {
+ for (size_t j = i+1; j < kb->GetSize(); j++) {
+ if (prng->next() < .5) {
TPair p;
p.first = kb->feats[i];
p.second = kb->feats[j];
@@ -50,10 +48,11 @@ sample_rand( KBestList* kb, TrainingInstances &training )
p.second_rank = j;
p.first_score = kb->scores[i];
p.second_score = kb->scores[j];
- training.push_back( p );
+ training.push_back(p);
}
}
}
+ cout << training.size() << " sampled" << endl;
}
diff --git a/dtrain/score.cc b/dtrain/score.cc
index 1e98c11d..d08e87f3 100644
--- a/dtrain/score.cc
+++ b/dtrain/score.cc
@@ -1,166 +1,149 @@
#include "score.h"
-
namespace dtrain
{
-/******************************************************************************
- * NGRAMS
- *
- *
- * make_ngrams
- *
- */
-typedef map<vector<WordID>, size_t> Ngrams;
Ngrams
-make_ngrams( vector<WordID>& s, size_t N )
+make_ngrams(vector<WordID>& s, size_t N)
{
Ngrams ngrams;
vector<WordID> ng;
- for ( size_t i = 0; i < s.size(); i++ ) {
+ for (size_t i = 0; i < s.size(); i++) {
ng.clear();
- for ( size_t j = i; j < min( i+N, s.size() ); j++ ) {
- ng.push_back( s[j] );
+ for (size_t j = i; j < min(i+N, s.size()); j++) {
+ ng.push_back(s[j]);
ngrams[ng]++;
}
}
return ngrams;
}
-
-/*
- * ngram_matches
- *
- */
NgramCounts
-make_ngram_counts( vector<WordID> hyp, vector<WordID> ref, size_t N )
+make_ngram_counts(vector<WordID> hyp, vector<WordID> ref, size_t N)
{
- Ngrams hyp_ngrams = make_ngrams( hyp, N );
- Ngrams ref_ngrams = make_ngrams( ref, N );
- NgramCounts counts( N );
+ Ngrams hyp_ngrams = make_ngrams(hyp, N);
+ Ngrams ref_ngrams = make_ngrams(ref, N);
+ NgramCounts counts(N);
Ngrams::iterator it;
Ngrams::iterator ti;
- for ( it = hyp_ngrams.begin(); it != hyp_ngrams.end(); it++ ) {
- ti = ref_ngrams.find( it->first );
- if ( ti != ref_ngrams.end() ) {
- counts.add( it->second, ti->second, it->first.size() - 1 );
+ for (it = hyp_ngrams.begin(); it != hyp_ngrams.end(); it++) {
+ ti = ref_ngrams.find(it->first);
+ if (ti != ref_ngrams.end()) {
+ counts.add(it->second, ti->second, it->first.size() - 1);
} else {
- counts.add( it->second, 0, it->first.size() - 1 );
+ counts.add(it->second, 0, it->first.size() - 1);
}
}
return counts;
}
-
-/******************************************************************************
- * SCORERS
- *
+/*
+ * bleu
*
- * brevity_penaly
+ * as in "BLEU: a Method for Automatic Evaluation
+ * of Machine Translation"
+ * (Papineni et al. '02)
*
+ * NOTE: 0 if one n in {1..N} has 0 count
*/
double
-brevity_penaly( const size_t hyp_len, const size_t ref_len )
+brevity_penaly(const size_t hyp_len, const size_t ref_len)
{
- if ( hyp_len > ref_len ) return 1;
- return exp( 1 - (double)ref_len/(double)hyp_len );
+ if (hyp_len > ref_len) return 1;
+ return exp(1 - (double)ref_len/(double)hyp_len);
}
-
-
-/*
- * bleu
- * as in "BLEU: a Method for Automatic Evaluation of Machine Translation" (Papineni et al. '02)
- * page TODO
- * 0 if for N one of the counts = 0
- */
double
-bleu( NgramCounts& counts, const size_t hyp_len, const size_t ref_len,
- size_t N, vector<float> weights )
+bleu(NgramCounts& counts, const size_t hyp_len, const size_t ref_len,
+ size_t N, vector<float> weights )
{
- if ( hyp_len == 0 || ref_len == 0 ) return 0;
- if ( ref_len < N ) N = ref_len;
+ if (hyp_len == 0 || ref_len == 0) return 0;
+ if (ref_len < N) N = ref_len;
float N_ = (float)N;
- if ( weights.empty() )
+ if (weights.empty())
{
- for ( size_t i = 0; i < N; i++ ) weights.push_back( 1/N_ );
+ for (size_t i = 0; i < N; i++) weights.push_back(1/N_);
}
double sum = 0;
- for ( size_t i = 0; i < N; i++ ) {
- if ( counts.clipped[i] == 0 || counts.sum[i] == 0 ) return 0;
- sum += weights[i] * log( (double)counts.clipped[i] / (double)counts.sum[i] );
+ for (size_t i = 0; i < N; i++) {
+ if (counts.clipped[i] == 0 || counts.sum[i] == 0) return 0;
+ sum += weights[i] * log((double)counts.clipped[i] / (double)counts.sum[i]);
}
- return brevity_penaly( hyp_len, ref_len ) * exp( sum );
+ return brevity_penaly(hyp_len, ref_len) * exp(sum);
}
-
/*
- * stupid_bleu
- * as in "ORANGE: a Method for Evaluating Automatic Evaluation Metrics for Machine Translation (Lin & Och '04)
- * page TODO
- * 0 iff no 1gram match
+ * 'stupid' bleu
+ *
+ * as in "ORANGE: a Method for Evaluating
+ * Automatic Evaluation Metrics
+ * for Machine Translation"
+ * (Lin & Och '04)
+ *
+ * NOTE: 0 iff no 1gram match
*/
double
-stupid_bleu( NgramCounts& counts, const size_t hyp_len, const size_t ref_len,
- size_t N, vector<float> weights )
+stupid_bleu(NgramCounts& counts, const size_t hyp_len, const size_t ref_len,
+ size_t N, vector<float> weights )
{
- if ( hyp_len == 0 || ref_len == 0 ) return 0;
- if ( ref_len < N ) N = ref_len;
+ if (hyp_len == 0 || ref_len == 0) return 0;
+ if (ref_len < N) N = ref_len;
float N_ = (float)N;
- if ( weights.empty() )
+ if (weights.empty())
{
- for ( size_t i = 0; i < N; i++ ) weights.push_back( 1/N_ );
+ for (size_t i = 0; i < N; i++) weights.push_back(1/N_);
}
double sum = 0;
float add = 0;
- for ( size_t i = 0; i < N; i++ ) {
- if ( i == 1 ) add = 1;
- sum += weights[i] * log( ((double)counts.clipped[i] + add) / ((double)counts.sum[i] + add) );
+ for (size_t i = 0; i < N; i++) {
+ if (i == 1) add = 1;
+ sum += weights[i] * log(((double)counts.clipped[i] + add) / ((double)counts.sum[i] + add));
}
- return brevity_penaly( hyp_len, ref_len ) * exp( sum );
+ return brevity_penaly(hyp_len, ref_len) * exp(sum);
}
-
/*
- * smooth_bleu
- * as in "An End-to-End Discriminative Approach to Machine Translation" (Liang et al. '06)
- * page TODO
- * max. 0.9375
+ * smooth bleu
+ *
+ * as in "An End-to-End Discriminative Approach
+ * to Machine Translation"
+ * (Liang et al. '06)
+ *
+ * NOTE: max is 0.9375
*/
double
-smooth_bleu( NgramCounts& counts, const size_t hyp_len, const size_t ref_len,
- const size_t N, vector<float> weights )
+smooth_bleu(NgramCounts& counts, const size_t hyp_len, const size_t ref_len,
+ const size_t N, vector<float> weights )
{
- if ( hyp_len == 0 || ref_len == 0 ) return 0;
+ if (hyp_len == 0 || ref_len == 0) return 0;
float N_ = (float)N;
- if ( weights.empty() )
+ if (weights.empty())
{
- for ( size_t i = 0; i < N; i++ ) weights.push_back( 1/N_ );
+ for (size_t i = 0; i < N; i++) weights.push_back(1/N_);
}
double sum = 0;
float j = 1;
- for ( size_t i = 0; i < N; i++ ) {
- if ( counts.clipped[i] == 0 || counts.sum[i] == 0) continue;
- sum += exp((weights[i] * log((double)counts.clipped[i]/(double)counts.sum[i]))) / pow( 2, N_-j+1 );
+ for (size_t i = 0; i < N; i++) {
+ if (counts.clipped[i] == 0 || counts.sum[i] == 0) continue;
+ sum += exp((weights[i] * log((double)counts.clipped[i]/(double)counts.sum[i]))) / pow(2, N_-j+1);
j++;
}
- return brevity_penaly( hyp_len, ref_len ) * sum;
+ return brevity_penaly(hyp_len, ref_len) * sum;
}
-
/*
- * approx_bleu
- * as in "Online Large-Margin Training for Statistical Machine Translation" (Watanabe et al. '07)
- * CHIANG, RESNIK, synt struct features
- * .9*
- * page TODO
+ * approx. bleu
*
+ * as in "Online Large-Margin Training of Syntactic
+ * and Structural Translation Features"
+ * (Chiang et al. '08)
*/
double
-approx_bleu( NgramCounts& counts, const size_t hyp_len, const size_t ref_len,
- const size_t N, vector<float> weights )
+approx_bleu(NgramCounts& counts, const size_t hyp_len, const size_t ref_len,
+ const size_t N, vector<float> weights)
{
- return bleu( counts, hyp_len, ref_len, N, weights );
+ return brevity_penaly(hyp_len, ref_len)
+ * 0.9 * bleu(counts, hyp_len, ref_len, N, weights);
}
diff --git a/dtrain/score.h b/dtrain/score.h
index e88387c5..0afb6237 100644
--- a/dtrain/score.h
+++ b/dtrain/score.h
@@ -1,29 +1,23 @@
#ifndef _DTRAIN_SCORE_H_
#define _DTRAIN_SCORE_H_
-
#include <iostream>
#include <vector>
#include <map>
#include <cassert>
#include <cmath>
-#include "wordid.h"
+#include "wordid.h" // cdec
using namespace std;
-
namespace dtrain
{
-/*
- * NgramCounts
- *
- */
struct NgramCounts
{
- NgramCounts( const size_t N ) : N_( N ) {
+ NgramCounts(const size_t N) : N_(N) {
reset();
}
size_t N_;
@@ -31,17 +25,17 @@ struct NgramCounts
map<size_t, size_t> sum;
void
- operator+=( const NgramCounts& rhs )
+ operator+=(const NgramCounts& rhs)
{
- assert( N_ == rhs.N_ );
- for ( size_t i = 0; i < N_; i++ ) {
+ assert(N_ == rhs.N_);
+ for (size_t i = 0; i < N_; i++) {
this->clipped[i] += rhs.clipped.find(i)->second;
this->sum[i] += rhs.sum.find(i)->second;
}
}
const NgramCounts
- operator+( const NgramCounts &other ) const
+ operator+(const NgramCounts &other) const
{
NgramCounts result = *this;
result += other;
@@ -49,10 +43,10 @@ struct NgramCounts
}
void
- add( size_t count, size_t ref_count, size_t i )
+ add(size_t count, size_t ref_count, size_t i)
{
- assert( i < N_ );
- if ( count > ref_count ) {
+ assert(i < N_);
+ if (count > ref_count) {
clipped[i] += ref_count;
sum[i] += count;
} else {
@@ -65,7 +59,7 @@ struct NgramCounts
reset()
{
size_t i;
- for ( i = 0; i < N_; i++ ) {
+ for (i = 0; i < N_; i++) {
clipped[i] = 0;
sum[i] = 0;
}
@@ -74,27 +68,26 @@ struct NgramCounts
void
print()
{
- for ( size_t i = 0; i < N_; i++ ) {
+ for (size_t i = 0; i < N_; i++) {
cout << i+1 << "grams (clipped):\t" << clipped[i] << endl;
cout << i+1 << "grams:\t\t\t" << sum[i] << endl;
}
}
};
-
typedef map<vector<WordID>, size_t> Ngrams;
-Ngrams make_ngrams( vector<WordID>& s, size_t N );
-NgramCounts make_ngram_counts( vector<WordID> hyp, vector<WordID> ref, size_t N );
-
-double brevity_penaly( const size_t hyp_len, const size_t ref_len );
-double bleu( NgramCounts& counts, const size_t hyp_len, const size_t ref_len, const size_t N,
- vector<float> weights = vector<float>() );
-double stupid_bleu( NgramCounts& counts, const size_t hyp_len, const size_t ref_len, size_t N,
- vector<float> weights = vector<float>() );
-double smooth_bleu( NgramCounts& counts, const size_t hyp_len, const size_t ref_len, const size_t N,
- vector<float> weights = vector<float>() );
-double approx_bleu( NgramCounts& counts, const size_t hyp_len, const size_t ref_len, const size_t N,
- vector<float> weights = vector<float>() );
+Ngrams make_ngrams(vector<WordID>& s, size_t N);
+NgramCounts make_ngram_counts(vector<WordID> hyp, vector<WordID> ref, size_t N);
+
+double brevity_penaly(const size_t hyp_len, const size_t ref_len);
+double bleu(NgramCounts& counts, const size_t hyp_len, const size_t ref_len, const size_t N,
+ vector<float> weights = vector<float>());
+double stupid_bleu(NgramCounts& counts, const size_t hyp_len, const size_t ref_len, size_t N,
+ vector<float> weights = vector<float>());
+double smooth_bleu(NgramCounts& counts, const size_t hyp_len, const size_t ref_len, const size_t N,
+ vector<float> weights = vector<float>());
+double approx_bleu(NgramCounts& counts, const size_t hyp_len, const size_t ref_len, const size_t N,
+ vector<float> weights = vector<float>());
} // namespace