summaryrefslogtreecommitdiff
path: root/dtrain/test
diff options
context:
space:
mode:
Diffstat (limited to 'dtrain/test')
-rw-r--r--dtrain/test/example/cdec.ini24
-rw-r--r--dtrain/test/example/dtrain.ini20
-rw-r--r--dtrain/test/example/nc-wmt11.1k.gzbin0 -> 21185883 bytes
-rw-r--r--dtrain/test/example/nc-wmt11.en.srilm.gzbin0 -> 16017291 bytes
-rw-r--r--dtrain/test/mtm11/logreg_cd/bin_class.cc4
-rw-r--r--dtrain/test/mtm11/logreg_cd/bin_class.h22
-rw-r--r--dtrain/test/mtm11/logreg_cd/log_reg.cc39
-rw-r--r--dtrain/test/mtm11/logreg_cd/log_reg.h14
-rw-r--r--dtrain/test/mtm11/mira_update/Hildreth.cpp187
-rw-r--r--dtrain/test/mtm11/mira_update/Hildreth.h10
-rw-r--r--dtrain/test/mtm11/mira_update/dtrain.cc532
-rw-r--r--dtrain/test/mtm11/mira_update/sample.h101
-rw-r--r--dtrain/test/toy/cdec.ini2
-rw-r--r--dtrain/test/toy/dtrain.ini12
-rw-r--r--dtrain/test/toy/input2
15 files changed, 969 insertions, 0 deletions
diff --git a/dtrain/test/example/cdec.ini b/dtrain/test/example/cdec.ini
new file mode 100644
index 00000000..fe5ca759
--- /dev/null
+++ b/dtrain/test/example/cdec.ini
@@ -0,0 +1,24 @@
+formalism=scfg
+add_pass_through_rules=true
+scfg_max_span_limit=15
+intersection_strategy=cube_pruning
+cubepruning_pop_limit=30
+feature_function=WordPenalty
+feature_function=KLanguageModel test/example/nc-wmt11.en.srilm.gz
+# all currently working feature function for translation:
+#feature_function=ArityPenalty
+#feature_function=CMR2008ReorderingFeatures
+#feature_function=Dwarf
+#feature_function=InputIndicator
+#feature_function=LexNullJump
+#feature_function=NewJump
+#feature_function=NgramFeatures
+#feature_function=NonLatinCount
+#feature_function=OutputIndicator
+feature_function=RuleIdentityFeatures
+feature_function=RuleNgramFeatures
+feature_function=RuleShape
+#feature_function=SourceSpanSizeFeatures
+#feature_function=SourceWordPenalty
+#feature_function=SpanFeatures
+# ^^^ features active that were used in the ACL paper
diff --git a/dtrain/test/example/dtrain.ini b/dtrain/test/example/dtrain.ini
new file mode 100644
index 00000000..68173e11
--- /dev/null
+++ b/dtrain/test/example/dtrain.ini
@@ -0,0 +1,20 @@
+input=test/example/nc-wmt11.1k.gz # use '-' for stdin
+output=- # a weights file or stdout
+decoder_config=test/example/cdec.ini # ini for cdec
+# these will be printed on each iteration
+print_weights=Glue WordPenalty LanguageModel LanguageModel_OOV PhraseModel_0 PhraseModel_1 PhraseModel_2 PhraseModel_3 PhraseModel_4 PhraseModel_5 PhraseModel_6 PassThrough
+tmp=/tmp
+stop_after=10 # stop iteration after 10 inputs
+
+# interesting stuff
+epochs=3 # run over input 3 times
+k=200 # use 100best lists
+N=4 # optimize (approx) BLEU4
+learning_rate=0.0001 # learning rate
+gamma=0.00001 # use SVM reg
+scorer=stupid_bleu # use stupid BLEU+1 approx.
+sample_from=kbest # use kbest lists (as opposed to forest)
+filter=uniq # only uniq entries in kbest
+pair_sampling=108010 # 10 vs 80 vs 10 and 80 vs 10
+pair_threshold=0 # minimum distance in BLEU
+select_weights=last # just output last weights
diff --git a/dtrain/test/example/nc-wmt11.1k.gz b/dtrain/test/example/nc-wmt11.1k.gz
new file mode 100644
index 00000000..45496cd8
--- /dev/null
+++ b/dtrain/test/example/nc-wmt11.1k.gz
Binary files differ
diff --git a/dtrain/test/example/nc-wmt11.en.srilm.gz b/dtrain/test/example/nc-wmt11.en.srilm.gz
new file mode 100644
index 00000000..7ce81057
--- /dev/null
+++ b/dtrain/test/example/nc-wmt11.en.srilm.gz
Binary files differ
diff --git a/dtrain/test/mtm11/logreg_cd/bin_class.cc b/dtrain/test/mtm11/logreg_cd/bin_class.cc
new file mode 100644
index 00000000..19bcde25
--- /dev/null
+++ b/dtrain/test/mtm11/logreg_cd/bin_class.cc
@@ -0,0 +1,4 @@
+#include "bin_class.h"
+
+Objective::~Objective() {}
+
diff --git a/dtrain/test/mtm11/logreg_cd/bin_class.h b/dtrain/test/mtm11/logreg_cd/bin_class.h
new file mode 100644
index 00000000..3466109a
--- /dev/null
+++ b/dtrain/test/mtm11/logreg_cd/bin_class.h
@@ -0,0 +1,22 @@
+#ifndef _BIN_CLASS_H_
+#define _BIN_CLASS_H_
+
+#include <vector>
+#include "sparse_vector.h"
+
+struct TrainingInstance {
+ // TODO add other info? loss for MIRA-type updates?
+ SparseVector<double> x_feature_map;
+ bool y;
+};
+
+struct Objective {
+ virtual ~Objective();
+
+ // returns f(x) and f'(x)
+ virtual double ObjectiveAndGradient(const SparseVector<double>& x,
+ const std::vector<TrainingInstance>& training_instances,
+ SparseVector<double>* g) const = 0;
+};
+
+#endif
diff --git a/dtrain/test/mtm11/logreg_cd/log_reg.cc b/dtrain/test/mtm11/logreg_cd/log_reg.cc
new file mode 100644
index 00000000..ec2331fe
--- /dev/null
+++ b/dtrain/test/mtm11/logreg_cd/log_reg.cc
@@ -0,0 +1,39 @@
+#include "log_reg.h"
+
+#include <vector>
+#include <cmath>
+
+#include "sparse_vector.h"
+
+using namespace std;
+
+double LogisticRegression::ObjectiveAndGradient(const SparseVector<double>& x,
+ const vector<TrainingInstance>& training_instances,
+ SparseVector<double>* g) const {
+ double cll = 0;
+ for (int i = 0; i < training_instances.size(); ++i) {
+ const double dotprod = training_instances[i].x_feature_map.dot(x); // TODO no bias, if bias, add x[0]
+ double lp_false = dotprod;
+ double lp_true = -dotprod;
+ if (0 < lp_true) {
+ lp_true += log1p(exp(-lp_true));
+ lp_false = log1p(exp(lp_false));
+ } else {
+ lp_true = log1p(exp(lp_true));
+ lp_false += log1p(exp(-lp_false));
+ }
+ lp_true *= -1;
+ lp_false *= -1;
+ if (training_instances[i].y) { // true label
+ cll -= lp_true;
+ (*g) -= training_instances[i].x_feature_map * exp(lp_false);
+ // (*g)[0] -= exp(lp_false); // bias
+ } else { // false label
+ cll -= lp_false;
+ (*g) += training_instances[i].x_feature_map * exp(lp_true);
+ // g += corpus[i].second * exp(lp_true);
+ }
+ }
+ return cll;
+}
+
diff --git a/dtrain/test/mtm11/logreg_cd/log_reg.h b/dtrain/test/mtm11/logreg_cd/log_reg.h
new file mode 100644
index 00000000..ecc560b8
--- /dev/null
+++ b/dtrain/test/mtm11/logreg_cd/log_reg.h
@@ -0,0 +1,14 @@
+#ifndef _LOG_REG_H_
+#define _LOG_REG_H_
+
+#include <vector>
+#include "sparse_vector.h"
+#include "bin_class.h"
+
+struct LogisticRegression : public Objective {
+ double ObjectiveAndGradient(const SparseVector<double>& x,
+ const std::vector<TrainingInstance>& training_instances,
+ SparseVector<double>* g) const;
+};
+
+#endif
diff --git a/dtrain/test/mtm11/mira_update/Hildreth.cpp b/dtrain/test/mtm11/mira_update/Hildreth.cpp
new file mode 100644
index 00000000..0e67eb15
--- /dev/null
+++ b/dtrain/test/mtm11/mira_update/Hildreth.cpp
@@ -0,0 +1,187 @@
+#include "Hildreth.h"
+#include "sparse_vector.h"
+
+using namespace std;
+
+namespace Mira {
+ vector<double> Hildreth::optimise (vector< SparseVector<double> >& a, vector<double>& b) {
+
+ size_t i;
+ int max_iter = 10000;
+ double eps = 0.00000001;
+ double zero = 0.000000000001;
+
+ vector<double> alpha ( b.size() );
+ vector<double> F ( b.size() );
+ vector<double> kkt ( b.size() );
+
+ double max_kkt = -1e100;
+
+ size_t K = b.size();
+
+ double A[K][K];
+ bool is_computed[K];
+ for ( i = 0; i < K; i++ )
+ {
+ A[i][i] = a[i].dot(a[i]);
+ is_computed[i] = false;
+ }
+
+ int max_kkt_i = -1;
+
+
+ for ( i = 0; i < b.size(); i++ )
+ {
+ F[i] = b[i];
+ kkt[i] = F[i];
+ if ( kkt[i] > max_kkt )
+ {
+ max_kkt = kkt[i];
+ max_kkt_i = i;
+ }
+ }
+
+ int iter = 0;
+ double diff_alpha;
+ double try_alpha;
+ double add_alpha;
+
+ while ( max_kkt >= eps && iter < max_iter )
+ {
+
+ diff_alpha = A[max_kkt_i][max_kkt_i] <= zero ? 0.0 : F[max_kkt_i]/A[max_kkt_i][max_kkt_i];
+ try_alpha = alpha[max_kkt_i] + diff_alpha;
+ add_alpha = 0.0;
+
+ if ( try_alpha < 0.0 )
+ add_alpha = -1.0 * alpha[max_kkt_i];
+ else
+ add_alpha = diff_alpha;
+
+ alpha[max_kkt_i] = alpha[max_kkt_i] + add_alpha;
+
+ if ( !is_computed[max_kkt_i] )
+ {
+ for ( i = 0; i < K; i++ )
+ {
+ A[i][max_kkt_i] = a[i].dot(a[max_kkt_i] ); // for version 1
+ //A[i][max_kkt_i] = 0; // for version 1
+ is_computed[max_kkt_i] = true;
+ }
+ }
+
+ for ( i = 0; i < F.size(); i++ )
+ {
+ F[i] -= add_alpha * A[i][max_kkt_i];
+ kkt[i] = F[i];
+ if ( alpha[i] > zero )
+ kkt[i] = abs ( F[i] );
+ }
+ max_kkt = -1e100;
+ max_kkt_i = -1;
+ for ( i = 0; i < F.size(); i++ )
+ if ( kkt[i] > max_kkt )
+ {
+ max_kkt = kkt[i];
+ max_kkt_i = i;
+ }
+
+ iter++;
+ }
+
+ return alpha;
+ }
+
+ vector<double> Hildreth::optimise (vector< SparseVector<double> >& a, vector<double>& b, double C) {
+
+ size_t i;
+ int max_iter = 10000;
+ double eps = 0.00000001;
+ double zero = 0.000000000001;
+
+ vector<double> alpha ( b.size() );
+ vector<double> F ( b.size() );
+ vector<double> kkt ( b.size() );
+
+ double max_kkt = -1e100;
+
+ size_t K = b.size();
+
+ double A[K][K];
+ bool is_computed[K];
+ for ( i = 0; i < K; i++ )
+ {
+ A[i][i] = a[i].dot(a[i]);
+ is_computed[i] = false;
+ }
+
+ int max_kkt_i = -1;
+
+
+ for ( i = 0; i < b.size(); i++ )
+ {
+ F[i] = b[i];
+ kkt[i] = F[i];
+ if ( kkt[i] > max_kkt )
+ {
+ max_kkt = kkt[i];
+ max_kkt_i = i;
+ }
+ }
+
+ int iter = 0;
+ double diff_alpha;
+ double try_alpha;
+ double add_alpha;
+
+ while ( max_kkt >= eps && iter < max_iter )
+ {
+
+ diff_alpha = A[max_kkt_i][max_kkt_i] <= zero ? 0.0 : F[max_kkt_i]/A[max_kkt_i][max_kkt_i];
+ try_alpha = alpha[max_kkt_i] + diff_alpha;
+ add_alpha = 0.0;
+
+ if ( try_alpha < 0.0 )
+ add_alpha = -1.0 * alpha[max_kkt_i];
+ else if (try_alpha > C)
+ add_alpha = C - alpha[max_kkt_i];
+ else
+ add_alpha = diff_alpha;
+
+ alpha[max_kkt_i] = alpha[max_kkt_i] + add_alpha;
+
+ if ( !is_computed[max_kkt_i] )
+ {
+ for ( i = 0; i < K; i++ )
+ {
+ A[i][max_kkt_i] = a[i].dot(a[max_kkt_i] ); // for version 1
+ //A[i][max_kkt_i] = 0; // for version 1
+ is_computed[max_kkt_i] = true;
+ }
+ }
+
+ for ( i = 0; i < F.size(); i++ )
+ {
+ F[i] -= add_alpha * A[i][max_kkt_i];
+ kkt[i] = F[i];
+ if (alpha[i] > C - zero)
+ kkt[i]=-kkt[i];
+ else if (alpha[i] > zero)
+ kkt[i] = abs(F[i]);
+
+ }
+ max_kkt = -1e100;
+ max_kkt_i = -1;
+ for ( i = 0; i < F.size(); i++ )
+ if ( kkt[i] > max_kkt )
+ {
+ max_kkt = kkt[i];
+ max_kkt_i = i;
+ }
+
+ iter++;
+ }
+
+ return alpha;
+ }
+}
diff --git a/dtrain/test/mtm11/mira_update/Hildreth.h b/dtrain/test/mtm11/mira_update/Hildreth.h
new file mode 100644
index 00000000..8d791085
--- /dev/null
+++ b/dtrain/test/mtm11/mira_update/Hildreth.h
@@ -0,0 +1,10 @@
+#include "sparse_vector.h"
+
+namespace Mira {
+ class Hildreth {
+ public :
+ static std::vector<double> optimise(std::vector< SparseVector<double> >& a, std::vector<double>& b);
+ static std::vector<double> optimise(std::vector< SparseVector<double> >& a, std::vector<double>& b, double C);
+ };
+}
+
diff --git a/dtrain/test/mtm11/mira_update/dtrain.cc b/dtrain/test/mtm11/mira_update/dtrain.cc
new file mode 100644
index 00000000..933417a4
--- /dev/null
+++ b/dtrain/test/mtm11/mira_update/dtrain.cc
@@ -0,0 +1,532 @@
+#include "common.h"
+#include "kbestget.h"
+#include "util.h"
+#include "sample.h"
+#include "Hildreth.h"
+
+#include "ksampler.h"
+
+// boost compression
+#include <boost/iostreams/device/file.hpp>
+#include <boost/iostreams/filtering_stream.hpp>
+#include <boost/iostreams/filter/gzip.hpp>
+//#include <boost/iostreams/filter/zlib.hpp>
+//#include <boost/iostreams/filter/bzip2.hpp>
+using namespace boost::iostreams;
+
+
+#ifdef DTRAIN_DEBUG
+#include "tests.h"
+#endif
+
+
+/*
+ * init
+ *
+ */
+bool
+init(int argc, char** argv, po::variables_map* cfg)
+{
+ po::options_description conff( "Configuration File Options" );
+ size_t k, N, T, stop, n_pairs;
+ string s, f, update_type;
+ conff.add_options()
+ ( "decoder_config", po::value<string>(), "configuration file for cdec" )
+ ( "kbest", po::value<size_t>(&k)->default_value(DTRAIN_DEFAULT_K), "k for kbest" )
+ ( "ngrams", po::value<size_t>(&N)->default_value(DTRAIN_DEFAULT_N), "N for Ngrams" )
+ ( "filter", po::value<string>(&f)->default_value("unique"), "filter kbest list" )
+ ( "epochs", po::value<size_t>(&T)->default_value(DTRAIN_DEFAULT_T), "# of iterations T" )
+ ( "input", po::value<string>(), "input file" )
+ ( "scorer", po::value<string>(&s)->default_value(DTRAIN_DEFAULT_SCORER), "scoring metric" )
+ ( "output", po::value<string>(), "output weights file" )
+ ( "stop_after", po::value<size_t>(&stop)->default_value(0), "stop after X input sentences" )
+ ( "weights_file", po::value<string>(), "input weights file (e.g. from previous iteration)" )
+ ( "wprint", po::value<string>(), "weights to print on each iteration" )
+ ( "noup", po::value<bool>()->zero_tokens(), "do not update weights" );
+
+ po::options_description clo("Command Line Options");
+ clo.add_options()
+ ( "config,c", po::value<string>(), "dtrain config file" )
+ ( "quiet,q", po::value<bool>()->zero_tokens(), "be quiet" )
+ ( "update-type", po::value<string>(&update_type)->default_value("mira"), "perceptron or mira" )
+ ( "n-pairs", po::value<size_t>(&n_pairs)->default_value(10), "number of pairs used to compute update" )
+ ( "verbose,v", po::value<bool>()->zero_tokens(), "be verbose" )
+#ifndef DTRAIN_DEBUG
+ ;
+#else
+ ( "test", "run tests and exit");
+#endif
+ po::options_description config_options, cmdline_options;
+
+ config_options.add(conff);
+ cmdline_options.add(clo);
+ cmdline_options.add(conff);
+
+ po::store( parse_command_line(argc, argv, cmdline_options), *cfg );
+ if ( cfg->count("config") ) {
+ ifstream config( (*cfg)["config"].as<string>().c_str() );
+ po::store( po::parse_config_file(config, config_options), *cfg );
+ }
+ po::notify(*cfg);
+
+ if ( !cfg->count("decoder_config") || !cfg->count("input") ) {
+ cerr << cmdline_options << endl;
+ return false;
+ }
+ if ( cfg->count("noup") && cfg->count("decode") ) {
+ cerr << "You can't use 'noup' and 'decode' at once." << endl;
+ return false;
+ }
+ if ( cfg->count("filter") && (*cfg)["filter"].as<string>() != "unique"
+ && (*cfg)["filter"].as<string>() != "no" ) {
+ cerr << "Wrong 'filter' type: '" << (*cfg)["filter"].as<string>() << "'." << endl;
+ }
+ #ifdef DTRAIN_DEBUG
+ if ( !cfg->count("test") ) {
+ cerr << cmdline_options << endl;
+ return false;
+ }
+ #endif
+ return true;
+}
+
+
+// output formatting
+ostream& _nopos( ostream& out ) { return out << resetiosflags( ios::showpos ); }
+ostream& _pos( ostream& out ) { return out << setiosflags( ios::showpos ); }
+ostream& _prec2( ostream& out ) { return out << setprecision(2); }
+ostream& _prec5( ostream& out ) { return out << setprecision(5); }
+
+
+
+
+/*
+ * dtrain
+ *
+ */
+int
+main( int argc, char** argv )
+{
+ cout << setprecision( 5 );
+ // handle most parameters
+ po::variables_map cfg;
+ if ( ! init(argc, argv, &cfg) ) exit(1); // something is wrong
+#ifdef DTRAIN_DEBUG
+ if ( cfg.count("test") ) run_tests(); // run tests and exit
+#endif
+ bool quiet = false;
+ if ( cfg.count("quiet") ) quiet = true;
+ bool verbose = false;
+ if ( cfg.count("verbose") ) verbose = true;
+ bool noup = false;
+ if ( cfg.count("noup") ) noup = true;
+ const size_t k = cfg["kbest"].as<size_t>();
+ const size_t N = cfg["ngrams"].as<size_t>();
+ const size_t T = cfg["epochs"].as<size_t>();
+ const size_t stop_after = cfg["stop_after"].as<size_t>();
+ const string filter_type = cfg["filter"].as<string>();
+ const string update_type = cfg["update-type"].as<string>();
+ const size_t n_pairs = cfg["n-pairs"].as<size_t>();
+ const string output_file = cfg["output"].as<string>();
+ if ( !quiet ) {
+ cout << endl << "dtrain" << endl << "Parameters:" << endl;
+ cout << setw(25) << "k " << k << endl;
+ cout << setw(25) << "N " << N << endl;
+ cout << setw(25) << "T " << T << endl;
+ if ( cfg.count("stop-after") )
+ cout << setw(25) << "stop_after " << stop_after << endl;
+ if ( cfg.count("weights") )
+ cout << setw(25) << "weights " << cfg["weights"].as<string>() << endl;
+ cout << setw(25) << "input " << "'" << cfg["input"].as<string>() << "'" << endl;
+ cout << setw(25) << "filter " << "'" << filter_type << "'" << endl;
+ }
+
+ vector<string> wprint;
+ if ( cfg.count("wprint") ) {
+ boost::split( wprint, cfg["wprint"].as<string>(), boost::is_any_of(" ") );
+ }
+
+ // setup decoder, observer
+ register_feature_functions();
+ SetSilent(true);
+ ReadFile ini_rf( cfg["decoder_config"].as<string>() );
+ if ( !quiet )
+ cout << setw(25) << "cdec cfg " << "'" << cfg["decoder_config"].as<string>() << "'" << endl;
+ Decoder decoder( ini_rf.stream() );
+ //KBestGetter observer( k, filter_type );
+ MT19937 rng;
+ KSampler observer( k, &rng );
+
+ // scoring metric/scorer
+ string scorer_str = cfg["scorer"].as<string>();
+ double (*scorer)( NgramCounts&, const size_t, const size_t, size_t, vector<float> );
+ if ( scorer_str == "bleu" ) {
+ scorer = &bleu;
+ } else if ( scorer_str == "stupid_bleu" ) {
+ scorer = &stupid_bleu;
+ } else if ( scorer_str == "smooth_bleu" ) {
+ scorer = &smooth_bleu;
+ } else if ( scorer_str == "approx_bleu" ) {
+ scorer = &approx_bleu;
+ } else {
+ cerr << "Don't know scoring metric: '" << scorer_str << "', exiting." << endl;
+ exit(1);
+ }
+ // for approx_bleu
+ NgramCounts global_counts( N ); // counts for 1 best translations
+ size_t global_hyp_len = 0; // sum hypothesis lengths
+ size_t global_ref_len = 0; // sum reference lengths
+ // this is all BLEU implmentations
+ vector<float> bleu_weights; // we leave this empty -> 1/N; TODO?
+ if ( !quiet ) cout << setw(26) << "scorer '" << scorer_str << "'" << endl << endl;
+
+ // init weights
+ Weights weights;
+ if ( cfg.count("weights") ) weights.InitFromFile( cfg["weights"].as<string>() );
+ SparseVector<double> lambdas;
+ weights.InitSparseVector( &lambdas );
+ vector<double> dense_weights;
+
+ // input
+ if ( !quiet && !verbose )
+ cout << "(a dot represents " << DTRAIN_DOTS << " lines of input)" << endl;
+ string input_fn = cfg["input"].as<string>();
+ ifstream input;
+ if ( input_fn != "-" ) input.open( input_fn.c_str() );
+ string in;
+ vector<string> in_split; // input: src\tref\tpsg
+ vector<string> ref_tok; // tokenized reference
+ vector<WordID> ref_ids; // reference as vector of WordID
+ string grammar_str;
+
+ // buffer input for t > 0
+ vector<string> src_str_buf; // source strings, TODO? memory
+ vector<vector<WordID> > ref_ids_buf; // references as WordID vecs
+ filtering_ostream grammar_buf; // written to compressed file in /tmp
+ // this is for writing the grammar buffer file
+ grammar_buf.push( gzip_compressor() );
+ char grammar_buf_tmp_fn[] = DTRAIN_TMP_DIR"/dtrain-grammars-XXXXXX";
+ mkstemp( grammar_buf_tmp_fn );
+ grammar_buf.push( file_sink(grammar_buf_tmp_fn, ios::binary | ios::trunc) );
+
+ size_t sid = 0, in_sz = 99999999; // sentence id, input size
+ double acc_1best_score = 0., acc_1best_model = 0.;
+ vector<vector<double> > scores_per_iter;
+ double max_score = 0.;
+ size_t best_t = 0;
+ bool next = false, stop = false;
+ double score = 0.;
+ size_t cand_len = 0;
+ double overall_time = 0.;
+
+ // for the perceptron/SVM; TODO as params
+ double eta = 0.0005;
+ double gamma = 0.;//01; // -> SVM
+ lambdas.add_value( FD::Convert("__bias"), 0 );
+
+ // for random sampling
+ srand ( time(NULL) );
+
+
+ for ( size_t t = 0; t < T; t++ ) // T epochs
+ {
+
+ time_t start, end;
+ time( &start );
+
+ // actually, we need only need this if t > 0 FIXME
+ ifstream grammar_file( grammar_buf_tmp_fn, ios_base::in | ios_base::binary );
+ filtering_istream grammar_buf_in;
+ grammar_buf_in.push( gzip_decompressor() );
+ grammar_buf_in.push( grammar_file );
+
+ // reset average scores
+ acc_1best_score = acc_1best_model = 0.;
+
+ // reset sentence counter
+ sid = 0;
+
+ if ( !quiet ) cout << "Iteration #" << t+1 << " of " << T << "." << endl;
+
+ while( true )
+ {
+
+ // get input from stdin or file
+ in.clear();
+ next = stop = false; // next iteration, premature stop
+ if ( t == 0 ) {
+ if ( input_fn == "-" ) {
+ if ( !getline(cin, in) ) next = true;
+ } else {
+ if ( !getline(input, in) ) next = true;
+ }
+ } else {
+ if ( sid == in_sz ) next = true; // stop if we reach the end of our input
+ }
+ // stop after X sentences (but still iterate for those)
+ if ( stop_after > 0 && stop_after == sid && !next ) stop = true;
+
+ // produce some pretty output
+ if ( !quiet && !verbose ) {
+ if ( sid == 0 ) cout << " ";
+ if ( (sid+1) % (DTRAIN_DOTS) == 0 ) {
+ cout << ".";
+ cout.flush();
+ }
+ if ( (sid+1) % (20*DTRAIN_DOTS) == 0) {
+ cout << " " << sid+1 << endl;
+ if ( !next && !stop ) cout << " ";
+ }
+ if ( stop ) {
+ if ( sid % (20*DTRAIN_DOTS) != 0 ) cout << " " << sid << endl;
+ cout << "Stopping after " << stop_after << " input sentences." << endl;
+ } else {
+ if ( next ) {
+ if ( sid % (20*DTRAIN_DOTS) != 0 ) {
+ cout << " " << sid << endl;
+ }
+ }
+ }
+ }
+
+ // next iteration
+ if ( next || stop ) break;
+
+ // weights
+ dense_weights.clear();
+ weights.InitFromVector( lambdas );
+ weights.InitVector( &dense_weights );
+ decoder.SetWeights( dense_weights );
+
+ if ( t == 0 ) {
+ // handling input
+ in_split.clear();
+ boost::split( in_split, in, boost::is_any_of("\t") ); // in_split[0] is id
+ // getting reference
+ ref_tok.clear(); ref_ids.clear();
+ boost::split( ref_tok, in_split[2], boost::is_any_of(" ") );
+ register_and_convert( ref_tok, ref_ids );
+ ref_ids_buf.push_back( ref_ids );
+ // process and set grammar
+ bool broken_grammar = true;
+ for ( string::iterator ti = in_split[3].begin(); ti != in_split[3].end(); ti++ ) {
+ if ( !isspace(*ti) ) {
+ broken_grammar = false;
+ break;
+ }
+ }
+ if ( broken_grammar ) continue;
+ grammar_str = boost::replace_all_copy( in_split[3], " __NEXT__RULE__ ", "\n" ) + "\n"; // FIXME copy, __
+ grammar_buf << grammar_str << DTRAIN_GRAMMAR_DELIM << endl;
+ decoder.SetSentenceGrammarFromString( grammar_str );
+ // decode, kbest
+ src_str_buf.push_back( in_split[1] );
+ decoder.Decode( in_split[1], &observer );
+ } else {
+ // get buffered grammar
+ grammar_str.clear();
+ int i = 1;
+ while ( true ) {
+ string g;
+ getline( grammar_buf_in, g );
+ if ( g == DTRAIN_GRAMMAR_DELIM ) break;
+ grammar_str += g+"\n";
+ i += 1;
+ }
+ decoder.SetSentenceGrammarFromString( grammar_str );
+ // decode, kbest
+ decoder.Decode( src_str_buf[sid], &observer );
+ }
+
+ // get kbest list
+ KBestList* kb;
+ //if ( ) { // TODO get from forest
+ kb = observer.GetKBest();
+ //}
+
+ // scoring kbest
+ if ( t > 0 ) ref_ids = ref_ids_buf[sid];
+ for ( size_t i = 0; i < kb->GetSize(); i++ ) {
+ NgramCounts counts = make_ngram_counts( ref_ids, kb->sents[i], N );
+ // this is for approx bleu
+ if ( scorer_str == "approx_bleu" ) {
+ if ( i == 0 ) { // 'context of 1best translations'
+ global_counts += counts;
+ global_hyp_len += kb->sents[i].size();
+ global_ref_len += ref_ids.size();
+ counts.reset();
+ cand_len = 0;
+ } else {
+ 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 );
+ } else {
+ // other scorers
+ cand_len = kb->sents[i].size();
+ score = scorer( counts,
+ ref_ids.size(),
+ kb->sents[i].size(), N, bleu_weights );
+ }
+
+ kb->scores.push_back( score );
+
+ if ( i == 0 ) {
+ acc_1best_score += score;
+ acc_1best_model += kb->model_scores[i];
+ }
+
+ if ( verbose ) {
+ if ( i == 0 ) cout << "'" << TD::GetString( ref_ids ) << "' [ref]" << endl;
+ cout << _prec5 << _nopos << "[hyp " << i << "] " << "'" << TD::GetString( kb->sents[i] ) << "'";
+ cout << " [SCORE=" << score << ",model="<< kb->model_scores[i] << "]" << endl;
+ cout << kb->feats[i] << endl; // this is maybe too verbose
+ }
+ } // Nbest loop
+
+ if ( verbose ) cout << endl;
+
+
+ // UPDATE WEIGHTS
+ if ( !noup ) {
+
+ TrainingInstances pairs;
+ sample_all( kb, pairs, n_pairs );
+
+ vector< SparseVector<double> > featureValueDiffs;
+ vector<double> lossMinusModelScoreDiffs;
+ for ( TrainingInstances::iterator ti = pairs.begin();
+ ti != pairs.end(); ti++ ) {
+
+ SparseVector<double> dv;
+ if ( ti->first_score - ti->second_score < 0 ) {
+ dv = ti->second - ti->first;
+ dv.add_value( FD::Convert("__bias"), -1 );
+
+ featureValueDiffs.push_back(dv);
+ double lossMinusModelScoreDiff = ti->loss_diff - ti->model_score_diff;
+ lossMinusModelScoreDiffs.push_back(lossMinusModelScoreDiff);
+
+ if (update_type == "perceptron") {
+ lambdas += dv * eta;
+ cerr << "after perceptron update: " << lambdas << endl << endl;
+ }
+
+ if ( verbose ) {
+ cout << "{{ f("<< ti->first_rank <<") > f(" << ti->second_rank << ") but g(i)="<< ti->first_score <<" < g(j)="<< ti->second_score << " so update" << endl;
+ cout << " i " << TD::GetString(kb->sents[ti->first_rank]) << endl;
+ cout << " " << kb->feats[ti->first_rank] << endl;
+ cout << " j " << TD::GetString(kb->sents[ti->second_rank]) << endl;
+ cout << " " << kb->feats[ti->second_rank] << endl;
+ cout << " diff vec: " << dv << endl;
+ cout << " lambdas after update: " << lambdas << endl;
+ cout << "}}" << endl;
+ }
+ } else {
+ //SparseVector<double> reg;
+ //reg = lambdas * ( 2 * gamma );
+ //lambdas += reg * ( -eta );
+ }
+ }
+ cerr << "Collected " << featureValueDiffs.size() << " constraints." << endl;
+
+ double slack = 0.01;
+ if (update_type == "mira") {
+ if (featureValueDiffs.size() > 0) {
+ vector<double> alphas;
+ if (slack != 0) {
+ alphas = Mira::Hildreth::optimise(featureValueDiffs, lossMinusModelScoreDiffs, slack);
+ } else {
+ alphas = Mira::Hildreth::optimise(featureValueDiffs, lossMinusModelScoreDiffs);
+ }
+
+ for (size_t k = 0; k < featureValueDiffs.size(); ++k) {
+ lambdas += featureValueDiffs[k] * alphas[k];
+ }
+ // cerr << "after mira update: " << lambdas << endl << endl;
+ }
+ }
+ }
+
+ ++sid;
+
+ } // input loop
+
+ if ( t == 0 ) in_sz = sid; // remember size (lines) of input
+
+ // print some stats
+ double avg_1best_score = acc_1best_score/(double)in_sz;
+ double avg_1best_model = acc_1best_model/(double)in_sz;
+ double avg_1best_score_diff, avg_1best_model_diff;
+ if ( t > 0 ) {
+ avg_1best_score_diff = avg_1best_score - scores_per_iter[t-1][0];
+ avg_1best_model_diff = avg_1best_model - scores_per_iter[t-1][1];
+ } else {
+ avg_1best_score_diff = avg_1best_score;
+ avg_1best_model_diff = avg_1best_model;
+ }
+ cout << _prec5 << _pos << "WEIGHTS" << endl;
+ for (vector<string>::iterator it = wprint.begin(); it != wprint.end(); it++) {
+ cout << setw(16) << *it << " = " << dense_weights[FD::Convert( *it )] << endl;
+ }
+
+ cout << " ---" << endl;
+ cout << _nopos << " avg score: " << avg_1best_score;
+ cout << _pos << " (" << avg_1best_score_diff << ")" << endl;
+ cout << _nopos << "avg model score: " << avg_1best_model;
+ cout << _pos << " (" << avg_1best_model_diff << ")" << endl;
+ vector<double> remember_scores;
+ remember_scores.push_back( avg_1best_score );
+ remember_scores.push_back( avg_1best_model );
+ scores_per_iter.push_back( remember_scores );
+ if ( avg_1best_score > max_score ) {
+ max_score = avg_1best_score;
+ best_t = t;
+ }
+
+ // close open files
+ if ( input_fn != "-" ) input.close();
+ close( grammar_buf );
+ grammar_file.close();
+
+ time ( &end );
+ double time_dif = difftime( end, start );
+ overall_time += time_dif;
+ if ( !quiet ) {
+ cout << _prec2 << _nopos << "(time " << time_dif/60. << " min, ";
+ cout << time_dif/(double)in_sz<< " s/S)" << endl;
+ }
+
+ if ( t+1 != T ) cout << endl;
+
+ if ( noup ) break;
+
+ // write weights after every epoch
+ std::string s;
+ std::stringstream out;
+ out << t;
+ s = out.str();
+ string weights_file = output_file + "." + s;
+ weights.WriteToFile(weights_file, true );
+
+ } // outer loop
+
+ unlink( grammar_buf_tmp_fn );
+ if ( !noup ) {
+ if ( !quiet ) cout << endl << "writing weights file '" << cfg["output"].as<string>() << "' ...";
+ weights.WriteToFile( cfg["output"].as<string>(), true );
+ if ( !quiet ) cout << "done" << endl;
+ }
+
+ if ( !quiet ) {
+ cout << _prec5 << _nopos << endl << "---" << endl << "Best iteration: ";
+ cout << best_t+1 << " [SCORE '" << scorer_str << "'=" << max_score << "]." << endl;
+ cout << _prec2 << "This took " << overall_time/60. << " min." << endl;
+ }
+
+ return 0;
+}
+
diff --git a/dtrain/test/mtm11/mira_update/sample.h b/dtrain/test/mtm11/mira_update/sample.h
new file mode 100644
index 00000000..5c331bba
--- /dev/null
+++ b/dtrain/test/mtm11/mira_update/sample.h
@@ -0,0 +1,101 @@
+#ifndef _DTRAIN_SAMPLE_H_
+#define _DTRAIN_SAMPLE_H_
+
+
+#include "kbestget.h"
+
+
+namespace dtrain
+{
+
+
+struct TPair
+{
+ SparseVector<double> first, second;
+ size_t first_rank, second_rank;
+ double first_score, second_score;
+ double model_score_diff;
+ double loss_diff;
+};
+
+typedef vector<TPair> TrainingInstances;
+
+
+void
+ sample_all( KBestList* kb, TrainingInstances &training, size_t n_pairs )
+{
+ std::vector<double> loss_diffs;
+ TrainingInstances training_tmp;
+ 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];
+ p.first_rank = i;
+ p.second_rank = j;
+ p.first_score = kb->scores[i];
+ p.second_score = kb->scores[j];
+
+ bool conservative = 1;
+ if ( kb->scores[i] - kb->scores[j] < 0 ) {
+ // j=hope, i=fear
+ p.model_score_diff = kb->model_scores[j] - kb->model_scores[i];
+ p.loss_diff = kb->scores[j] - kb->scores[i];
+ training_tmp.push_back(p);
+ loss_diffs.push_back(p.loss_diff);
+ }
+ else if (!conservative) {
+ // i=hope, j=fear
+ p.model_score_diff = kb->model_scores[i] - kb->model_scores[j];
+ p.loss_diff = kb->scores[i] - kb->scores[j];
+ training_tmp.push_back(p);
+ loss_diffs.push_back(p.loss_diff);
+ }
+ }
+ }
+
+ if (training_tmp.size() > 0) {
+ double threshold;
+ std::sort(loss_diffs.begin(), loss_diffs.end());
+ std::reverse(loss_diffs.begin(), loss_diffs.end());
+ threshold = loss_diffs.size() >= n_pairs ? loss_diffs[n_pairs-1] : loss_diffs[loss_diffs.size()-1];
+ cerr << "threshold: " << threshold << endl;
+ size_t constraints = 0;
+ for (size_t i = 0; (i < training_tmp.size() && constraints < n_pairs); ++i) {
+ if (training_tmp[i].loss_diff >= threshold) {
+ training.push_back(training_tmp[i]);
+ constraints++;
+ }
+ }
+ }
+ else {
+ cerr << "No pairs selected." << endl;
+ }
+}
+
+void
+sample_rand( KBestList* kb, TrainingInstances &training )
+{
+ 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 ) {
+ TPair p;
+ p.first = kb->feats[i];
+ p.second = kb->feats[j];
+ p.first_rank = i;
+ p.second_rank = j;
+ p.first_score = kb->scores[i];
+ p.second_score = kb->scores[j];
+ training.push_back( p );
+ }
+ }
+ }
+}
+
+
+} // namespace
+
+
+#endif
+
diff --git a/dtrain/test/toy/cdec.ini b/dtrain/test/toy/cdec.ini
new file mode 100644
index 00000000..98b02d44
--- /dev/null
+++ b/dtrain/test/toy/cdec.ini
@@ -0,0 +1,2 @@
+formalism=scfg
+add_pass_through_rules=true
diff --git a/dtrain/test/toy/dtrain.ini b/dtrain/test/toy/dtrain.ini
new file mode 100644
index 00000000..abf22b94
--- /dev/null
+++ b/dtrain/test/toy/dtrain.ini
@@ -0,0 +1,12 @@
+decoder_config=test/toy/cdec.ini
+input=test/toy/input
+output=-
+print_weights=logp shell_rule house_rule small_rule little_rule PassThrough
+k=4
+N=4
+epochs=3
+scorer=stupid_bleu
+sample_from=kbest
+filter=uniq
+pair_sampling=all
+learning_rate=1
diff --git a/dtrain/test/toy/input b/dtrain/test/toy/input
new file mode 100644
index 00000000..4d10a9ea
--- /dev/null
+++ b/dtrain/test/toy/input
@@ -0,0 +1,2 @@
+0 ich sah ein kleines haus i saw a little house [S] ||| [NP,1] [VP,2] ||| [1] [2] ||| logp=0 [NP] ||| ich ||| i ||| logp=0 [NP] ||| ein [NN,1] ||| a [1] ||| logp=0 [NN] ||| [JJ,1] haus ||| [1] house ||| logp=0 house_rule=1 [NN] ||| [JJ,1] haus ||| [1] shell ||| logp=0 shell_rule=1 [JJ] ||| kleines ||| small ||| logp=0 small_rule=1 [JJ] ||| kleines ||| little ||| logp=0 little_rule=1 [JJ] ||| grosses ||| big ||| logp=0 [JJ] ||| grosses ||| large ||| logp=0 [VP] ||| [V,1] [NP,2] ||| [1] [2] ||| logp=0 [V] ||| sah ||| saw ||| logp=0 [V] ||| fand ||| found ||| logp=0
+1 ich fand ein kleines haus i found a little house [S] ||| [NP,1] [VP,2] ||| [1] [2] ||| logp=0 [NP] ||| ich ||| i ||| logp=0 [NP] ||| ein [NN,1] ||| a [1] ||| logp=0 [NN] ||| [JJ,1] haus ||| [1] house ||| logp=0 house_rule=1 [NN] ||| [JJ,1] haus ||| [1] shell ||| logp=0 shell_rule=1 [JJ] ||| kleines ||| small ||| logp=0 small_rule=1 [JJ] ||| kleines ||| little ||| logp=0 little_rule=1 [JJ] ||| grosses ||| big ||| logp=0 [JJ] ||| grosses ||| large ||| logp=0 [VP] ||| [V,1] [NP,2] ||| [1] [2] ||| logp=0 [V] ||| sah ||| saw ||| logp=0 [V] ||| fand ||| found ||| logp=0
generated by cgit v1.2.3 (git 2.39.1) at 2024-11-05 05:33:52 +0000