summaryrefslogtreecommitdiff
path: root/training
diff options
context:
space:
mode:
authorPatrick Simianer <simianer@cl.uni-heidelberg.de>2012-07-08 14:26:51 +0200
committerPatrick Simianer <simianer@cl.uni-heidelberg.de>2012-07-08 14:26:51 +0200
commitc139ce495861bb341e1b86a85ad4559f9ad53c14 (patch)
tree1071839ee458f21f169ce06fc536fefe07e4c65d /training
parent3a94ac22e5c60aa205f2b3dadf81b0666500e0c3 (diff)
parentd01e5b66d3010d61b9b56301fd7f302dd4ea5bc8 (diff)
Merge branch 'master' of github.com:pks/cdec-dtrain
Diffstat (limited to 'training')
-rw-r--r--training/Makefile.am1
-rw-r--r--training/candidate_set.cc9
-rw-r--r--training/entropy.cc41
-rw-r--r--training/entropy.h22
4 files changed, 69 insertions, 4 deletions
diff --git a/training/Makefile.am b/training/Makefile.am
index 68ebfab4..4cef0d5b 100644
--- a/training/Makefile.am
+++ b/training/Makefile.am
@@ -26,6 +26,7 @@ TESTS = lbfgs_test optimize_test
noinst_LIBRARIES = libtraining.a
libtraining_a_SOURCES = \
candidate_set.cc \
+ entropy.cc \
optimize.cc \
online_optimizer.cc \
risk.cc
diff --git a/training/candidate_set.cc b/training/candidate_set.cc
index 8c086ece..087efec3 100644
--- a/training/candidate_set.cc
+++ b/training/candidate_set.cc
@@ -4,6 +4,7 @@
#include <boost/functional/hash.hpp>
+#include "verbose.h"
#include "ns.h"
#include "filelib.h"
#include "wordid.h"
@@ -118,7 +119,7 @@ void CandidateSet::WriteToFile(const string& file) const {
}
void CandidateSet::ReadFromFile(const string& file) {
- cerr << "Reading candidates from " << file << endl;
+ if(!SILENT) cerr << "Reading candidates from " << file << endl;
ReadFile rf(file);
istream& in = *rf.stream();
string cand;
@@ -133,11 +134,11 @@ void CandidateSet::ReadFromFile(const string& file) {
ParseSparseVector(feats, 0, &cs.back().fmap);
cs.back().eval_feats = SufficientStats(ss);
}
- cerr << " read " << cs.size() << " candidates\n";
+ if(!SILENT) cerr << " read " << cs.size() << " candidates\n";
}
void CandidateSet::Dedup() {
- cerr << "Dedup in=" << cs.size();
+ if(!SILENT) cerr << "Dedup in=" << cs.size();
tr1::unordered_set<Candidate, CandidateHasher, CandidateCompare> u;
while(cs.size() > 0) {
u.insert(cs.back());
@@ -148,7 +149,7 @@ void CandidateSet::Dedup() {
cs.push_back(*it);
it = u.erase(it);
}
- cerr << " out=" << cs.size() << endl;
+ if(!SILENT) cerr << " out=" << cs.size() << endl;
}
void CandidateSet::AddKBestCandidates(const Hypergraph& hg, size_t kbest_size, const SegmentEvaluator* scorer) {
diff --git a/training/entropy.cc b/training/entropy.cc
new file mode 100644
index 00000000..4fdbe2be
--- /dev/null
+++ b/training/entropy.cc
@@ -0,0 +1,41 @@
+#include "entropy.h"
+
+#include "prob.h"
+#include "candidate_set.h"
+
+using namespace std;
+
+namespace training {
+
+// see Mann and McCallum "Efficient Computation of Entropy Gradient ..." for
+// a mostly clear derivation of:
+// g = E[ F(x,y) * log p(y|x) ] + H(y | x) * E[ F(x,y) ]
+double CandidateSetEntropy::operator()(const vector<double>& params,
+ SparseVector<double>* g) const {
+ prob_t z;
+ vector<double> dps(cands_.size());
+ for (unsigned i = 0; i < cands_.size(); ++i) {
+ dps[i] = cands_[i].fmap.dot(params);
+ const prob_t u(dps[i], init_lnx());
+ z += u;
+ }
+ const double log_z = log(z);
+
+ SparseVector<double> exp_feats;
+ double entropy = 0;
+ for (unsigned i = 0; i < cands_.size(); ++i) {
+ const double log_prob = cands_[i].fmap.dot(params) - log_z;
+ const double prob = exp(log_prob);
+ const double e_logprob = prob * log_prob;
+ entropy -= e_logprob;
+ if (g) {
+ (*g) += cands_[i].fmap * e_logprob;
+ exp_feats += cands_[i].fmap * prob;
+ }
+ }
+ if (g) (*g) += exp_feats * entropy;
+ return entropy;
+}
+
+}
+
diff --git a/training/entropy.h b/training/entropy.h
new file mode 100644
index 00000000..796589ca
--- /dev/null
+++ b/training/entropy.h
@@ -0,0 +1,22 @@
+#ifndef _CSENTROPY_H_
+#define _CSENTROPY_H_
+
+#include <vector>
+#include "sparse_vector.h"
+
+namespace training {
+ class CandidateSet;
+
+ class CandidateSetEntropy {
+ public:
+ explicit CandidateSetEntropy(const CandidateSet& cs) : cands_(cs) {}
+ // compute the entropy (expected log likelihood) of a CandidateSet
+ // (optional) the gradient of the entropy with respect to params
+ double operator()(const std::vector<double>& params,
+ SparseVector<double>* g = NULL) const;
+ private:
+ const CandidateSet& cands_;
+ };
+};
+
+#endif