summaryrefslogtreecommitdiff
path: root/training/utils/entropy.cc
diff options
context:
space:
mode:
authorAvneesh Saluja <asaluja@gmail.com>2013-03-28 18:28:16 -0700
committerAvneesh Saluja <asaluja@gmail.com>2013-03-28 18:28:16 -0700
commit3d8d656fa7911524e0e6885647173474524e0784 (patch)
tree81b1ee2fcb67980376d03f0aa48e42e53abff222 /training/utils/entropy.cc
parentbe7f57fdd484e063775d7abf083b9fa4c403b610 (diff)
parent96fedabebafe7a38a6d5928be8fff767e411d705 (diff)
fixed conflicts
Diffstat (limited to 'training/utils/entropy.cc')
-rw-r--r--training/utils/entropy.cc41
1 files changed, 41 insertions, 0 deletions
diff --git a/training/utils/entropy.cc b/training/utils/entropy.cc
new file mode 100644
index 00000000..4fdbe2be
--- /dev/null
+++ b/training/utils/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;
+}
+
+}
+