summaryrefslogtreecommitdiff
path: root/pro-train
diff options
context:
space:
mode:
authorPatrick Simianer <simianer@cl.uni-heidelberg.de>2012-05-31 13:57:24 +0200
committerPatrick Simianer <simianer@cl.uni-heidelberg.de>2012-05-31 13:57:24 +0200
commitf1ba05780db1705493d9afb562332498b93d26f1 (patch)
treefb429a657ba97f33e8140742de9bc74d9fc88e75 /pro-train
parentaadabfdf37dfd451485277cb77fad02f77b361c6 (diff)
parent317d650f6cb1e24ac6f3be6f7bf9d4246a59e0e5 (diff)
Merge remote-tracking branch 'upstream/master'
Diffstat (limited to 'pro-train')
-rw-r--r--pro-train/Makefile.am8
-rw-r--r--pro-train/mr_pro_map.cc181
-rw-r--r--pro-train/mr_pro_reduce.cc10
3 files changed, 22 insertions, 177 deletions
diff --git a/pro-train/Makefile.am b/pro-train/Makefile.am
index 11d26211..1e9d46b0 100644
--- a/pro-train/Makefile.am
+++ b/pro-train/Makefile.am
@@ -2,12 +2,10 @@ bin_PROGRAMS = \
mr_pro_map \
mr_pro_reduce
-TESTS = lo_test
-
mr_pro_map_SOURCES = mr_pro_map.cc
-mr_pro_map_LDADD = $(top_srcdir)/decoder/libcdec.a $(top_srcdir)/mteval/libmteval.a $(top_srcdir)/utils/libutils.a -lz
+mr_pro_map_LDADD = $(top_srcdir)/training/libtraining.a $(top_srcdir)/decoder/libcdec.a $(top_srcdir)/mteval/libmteval.a $(top_srcdir)/utils/libutils.a -lz
mr_pro_reduce_SOURCES = mr_pro_reduce.cc
-mr_pro_reduce_LDADD = $(top_srcdir)/decoder/libcdec.a $(top_srcdir)/training/optimize.o $(top_srcdir)/training/liblbfgs/liblbfgs.a $(top_srcdir)/mteval/libmteval.a $(top_srcdir)/utils/libutils.a -lz
+mr_pro_reduce_LDADD = $(top_srcdir)/training/liblbfgs/liblbfgs.a $(top_srcdir)/utils/libutils.a -lz
-AM_CPPFLAGS = -W -Wall -Wno-sign-compare $(GTEST_CPPFLAGS) -I$(top_srcdir)/utils -I$(top_srcdir)/decoder -I$(top_srcdir)/mteval -I$(top_srcdir)/training
+AM_CPPFLAGS = -W -Wall -Wno-sign-compare -I$(top_srcdir)/utils -I$(top_srcdir)/decoder -I$(top_srcdir)/mteval -I$(top_srcdir)/training
diff --git a/pro-train/mr_pro_map.cc b/pro-train/mr_pro_map.cc
index 52b67f32..eef40b8a 100644
--- a/pro-train/mr_pro_map.cc
+++ b/pro-train/mr_pro_map.cc
@@ -9,14 +9,13 @@
#include <boost/program_options.hpp>
#include <boost/program_options/variables_map.hpp>
+#include "candidate_set.h"
#include "sampler.h"
#include "filelib.h"
#include "stringlib.h"
#include "weights.h"
#include "inside_outside.h"
#include "hg_io.h"
-#include "kbest.h"
-#include "viterbi.h"
#include "ns.h"
#include "ns_docscorer.h"
@@ -25,52 +24,6 @@
using namespace std;
namespace po = boost::program_options;
-struct ApproxVectorHasher {
- static const size_t MASK = 0xFFFFFFFFull;
- union UType {
- double f; // leave as double
- size_t i;
- };
- static inline double round(const double x) {
- UType t;
- t.f = x;
- size_t r = t.i & MASK;
- if ((r << 1) > MASK)
- t.i += MASK - r + 1;
- else
- t.i &= (1ull - MASK);
- return t.f;
- }
- size_t operator()(const SparseVector<weight_t>& x) const {
- size_t h = 0x573915839;
- for (SparseVector<weight_t>::const_iterator it = x.begin(); it != x.end(); ++it) {
- UType t;
- t.f = it->second;
- if (t.f) {
- size_t z = (t.i >> 32);
- boost::hash_combine(h, it->first);
- boost::hash_combine(h, z);
- }
- }
- return h;
- }
-};
-
-struct ApproxVectorEquals {
- bool operator()(const SparseVector<weight_t>& a, const SparseVector<weight_t>& b) const {
- SparseVector<weight_t>::const_iterator bit = b.begin();
- for (SparseVector<weight_t>::const_iterator ait = a.begin(); ait != a.end(); ++ait) {
- if (bit == b.end() ||
- ait->first != bit->first ||
- ApproxVectorHasher::round(ait->second) != ApproxVectorHasher::round(bit->second))
- return false;
- ++bit;
- }
- if (bit != b.end()) return false;
- return true;
- }
-};
-
boost::shared_ptr<MT19937> rng;
void InitCommandLine(int argc, char** argv, po::variables_map* conf) {
@@ -105,107 +58,6 @@ void InitCommandLine(int argc, char** argv, po::variables_map* conf) {
}
}
-struct HypInfo {
- HypInfo() : g_(-100.0f) {}
- HypInfo(const vector<WordID>& h, const SparseVector<weight_t>& feats) : hyp(h), g_(-100.0f), x(feats) {}
-
- // lazy evaluation
- double g(const SegmentEvaluator& scorer, const EvaluationMetric* metric) const {
- if (g_ == -100.0f) {
- SufficientStats ss;
- scorer.Evaluate(hyp, &ss);
- g_ = metric->ComputeScore(ss);
- }
- return g_;
- }
- vector<WordID> hyp;
- mutable float g_;
- SparseVector<weight_t> x;
-};
-
-struct HypInfoCompare {
- bool operator()(const HypInfo& a, const HypInfo& b) const {
- ApproxVectorEquals comp;
- return (a.hyp == b.hyp && comp(a.x,b.x));
- }
-};
-
-struct HypInfoHasher {
- size_t operator()(const HypInfo& x) const {
- boost::hash<vector<WordID> > hhasher;
- ApproxVectorHasher vhasher;
- size_t ha = hhasher(x.hyp);
- boost::hash_combine(ha, vhasher(x.x));
- return ha;
- }
-};
-
-void WriteKBest(const string& file, const vector<HypInfo>& kbest) {
- WriteFile wf(file);
- ostream& out = *wf.stream();
- out.precision(10);
- for (int i = 0; i < kbest.size(); ++i) {
- out << TD::GetString(kbest[i].hyp) << endl;
- out << kbest[i].x << endl;
- }
-}
-
-void ParseSparseVector(string& line, size_t cur, SparseVector<weight_t>* out) {
- SparseVector<weight_t>& x = *out;
- size_t last_start = cur;
- size_t last_comma = string::npos;
- while(cur <= line.size()) {
- if (line[cur] == ' ' || cur == line.size()) {
- if (!(cur > last_start && last_comma != string::npos && cur > last_comma)) {
- cerr << "[ERROR] " << line << endl << " position = " << cur << endl;
- exit(1);
- }
- const int fid = FD::Convert(line.substr(last_start, last_comma - last_start));
- if (cur < line.size()) line[cur] = 0;
- const double val = strtod(&line[last_comma + 1], NULL);
- x.set_value(fid, val);
-
- last_comma = string::npos;
- last_start = cur+1;
- } else {
- if (line[cur] == '=')
- last_comma = cur;
- }
- ++cur;
- }
-}
-
-void ReadKBest(const string& file, vector<HypInfo>* kbest) {
- cerr << "Reading from " << file << endl;
- ReadFile rf(file);
- istream& in = *rf.stream();
- string cand;
- string feats;
- while(getline(in, cand)) {
- getline(in, feats);
- assert(in);
- kbest->push_back(HypInfo());
- TD::ConvertSentence(cand, &kbest->back().hyp);
- ParseSparseVector(feats, 0, &kbest->back().x);
- }
- cerr << " read " << kbest->size() << " hypotheses\n";
-}
-
-void Dedup(vector<HypInfo>* h) {
- cerr << "Dedup in=" << h->size();
- tr1::unordered_set<HypInfo, HypInfoHasher, HypInfoCompare> u;
- while(h->size() > 0) {
- u.insert(h->back());
- h->pop_back();
- }
- tr1::unordered_set<HypInfo, HypInfoHasher, HypInfoCompare>::iterator it = u.begin();
- while (it != u.end()) {
- h->push_back(*it);
- it = u.erase(it);
- }
- cerr << " out=" << h->size() << endl;
-}
-
struct ThresholdAlpha {
explicit ThresholdAlpha(double t = 0.05) : threshold(t) {}
double operator()(double mag) const {
@@ -239,8 +91,7 @@ struct DiffOrder {
void Sample(const unsigned gamma,
const unsigned xi,
- const vector<HypInfo>& J_i,
- const SegmentEvaluator& scorer,
+ const training::CandidateSet& J_i,
const EvaluationMetric* metric,
vector<TrainingInstance>* pv) {
const bool invert_score = metric->IsErrorMetric();
@@ -250,17 +101,17 @@ void Sample(const unsigned gamma,
const size_t a = rng->inclusive(0, J_i.size() - 1)();
const size_t b = rng->inclusive(0, J_i.size() - 1)();
if (a == b) continue;
- float ga = J_i[a].g(scorer, metric);
- float gb = J_i[b].g(scorer, metric);
+ float ga = metric->ComputeScore(J_i[a].eval_feats);
+ float gb = metric->ComputeScore(J_i[b].eval_feats);
bool positive = gb < ga;
if (invert_score) positive = !positive;
const float gdiff = fabs(ga - gb);
if (!gdiff) continue;
avg_diff += gdiff;
- SparseVector<weight_t> xdiff = (J_i[a].x - J_i[b].x).erase_zeros();
+ SparseVector<weight_t> xdiff = (J_i[a].fmap - J_i[b].fmap).erase_zeros();
if (xdiff.empty()) {
- cerr << "Empty diff:\n " << TD::GetString(J_i[a].hyp) << endl << "x=" << J_i[a].x << endl;
- cerr << " " << TD::GetString(J_i[b].hyp) << endl << "x=" << J_i[b].x << endl;
+ cerr << "Empty diff:\n " << TD::GetString(J_i[a].ewords) << endl << "x=" << J_i[a].fmap << endl;
+ cerr << " " << TD::GetString(J_i[b].ewords) << endl << "x=" << J_i[b].fmap << endl;
continue;
}
v1.push_back(TrainingInstance(xdiff, positive, gdiff));
@@ -328,25 +179,17 @@ int main(int argc, char** argv) {
is >> file >> sent_id;
ReadFile rf(file);
ostringstream os;
- vector<HypInfo> J_i;
+ training::CandidateSet J_i;
os << kbest_repo << "/kbest." << sent_id << ".txt.gz";
const string kbest_file = os.str();
if (FileExists(kbest_file))
- ReadKBest(kbest_file, &J_i);
+ J_i.ReadFromFile(kbest_file);
HypergraphIO::ReadFromJSON(rf.stream(), &hg);
hg.Reweight(weights);
- KBest::KBestDerivations<vector<WordID>, ESentenceTraversal> kbest(hg, kbest_size);
-
- for (int i = 0; i < kbest_size; ++i) {
- const KBest::KBestDerivations<vector<WordID>, ESentenceTraversal>::Derivation* d =
- kbest.LazyKthBest(hg.nodes_.size() - 1, i);
- if (!d) break;
- J_i.push_back(HypInfo(d->yield, d->feature_values));
- }
- Dedup(&J_i);
- WriteKBest(kbest_file, J_i);
+ J_i.AddKBestCandidates(hg, kbest_size, ds[sent_id]);
+ J_i.WriteToFile(kbest_file);
- Sample(gamma, xi, J_i, *ds[sent_id], metric, &v);
+ Sample(gamma, xi, J_i, metric, &v);
for (unsigned i = 0; i < v.size(); ++i) {
const TrainingInstance& vi = v[i];
cout << vi.y << "\t" << vi.x << endl;
diff --git a/pro-train/mr_pro_reduce.cc b/pro-train/mr_pro_reduce.cc
index d3fb8026..5ef9b470 100644
--- a/pro-train/mr_pro_reduce.cc
+++ b/pro-train/mr_pro_reduce.cc
@@ -25,6 +25,7 @@ void InitCommandLine(int argc, char** argv, po::variables_map* conf) {
opts.add_options()
("weights,w", po::value<string>(), "Weights from previous iteration (used as initialization and interpolation")
("regularization_strength,C",po::value<double>()->default_value(500.0), "l2 regularization strength")
+ ("l1",po::value<double>()->default_value(0.0), "l1 regularization strength")
("regularize_to_weights,y",po::value<double>()->default_value(5000.0), "Differences in learned weights to previous weights are penalized with an l2 penalty with this strength; 0.0 = no effect")
("memory_buffers,m",po::value<unsigned>()->default_value(100), "Number of memory buffers (LBFGS)")
("min_reg,r",po::value<double>()->default_value(0.01), "When tuning (-T) regularization strength, minimum regularization strenght")
@@ -180,12 +181,14 @@ struct ProLoss {
double LearnParameters(const vector<pair<bool, SparseVector<weight_t> > >& training,
const vector<pair<bool, SparseVector<weight_t> > >& testing,
const double C,
+ const double C1,
const double T,
const unsigned memory_buffers,
const vector<weight_t>& prev_x,
vector<weight_t>* px) {
+ assert(px->size() == prev_x.size());
ProLoss loss(training, testing, C, T, prev_x);
- LBFGS<ProLoss> lbfgs(px, loss, 0.0, memory_buffers);
+ LBFGS<ProLoss> lbfgs(px, loss, memory_buffers, C1);
lbfgs.MinimizeFunction();
return loss.tppl;
}
@@ -203,6 +206,7 @@ int main(int argc, char** argv) {
const double min_reg = conf["min_reg"].as<double>();
const double max_reg = conf["max_reg"].as<double>();
double C = conf["regularization_strength"].as<double>(); // will be overridden if parameter is tuned
+ double C1 = conf["l1"].as<double>(); // will be overridden if parameter is tuned
const double T = conf["regularize_to_weights"].as<double>();
assert(C >= 0.0);
assert(min_reg >= 0.0);
@@ -239,7 +243,7 @@ int main(int argc, char** argv) {
cerr << "SWEEP FACTOR: " << sweep_factor << endl;
while(C < max_reg) {
cerr << "C=" << C << "\tT=" <<T << endl;
- tppl = LearnParameters(training, testing, C, T, conf["memory_buffers"].as<unsigned>(), prev_x, &x);
+ tppl = LearnParameters(training, testing, C, C1, T, conf["memory_buffers"].as<unsigned>(), prev_x, &x);
sp.push_back(make_pair(C, tppl));
C *= sweep_factor;
}
@@ -262,7 +266,7 @@ int main(int argc, char** argv) {
}
C = sp[best_i].first;
} // tune regularizer
- tppl = LearnParameters(training, testing, C, T, conf["memory_buffers"].as<unsigned>(), prev_x, &x);
+ tppl = LearnParameters(training, testing, C, C1, T, conf["memory_buffers"].as<unsigned>(), prev_x, &x);
if (conf.count("weights")) {
for (int i = 1; i < x.size(); ++i) {
x[i] = (x[i] * psi) + prev_x[i] * (1.0 - psi);