From 1b8181bf0d6e9137e6b9ccdbe414aec37377a1a9 Mon Sep 17 00:00:00 2001
From: Chris Dyer <cdyer@allegro.clab.cs.cmu.edu>
Date: Sun, 18 Nov 2012 13:35:42 -0500
Subject: major restructure of the training code

---
 training/utils/entropy.cc | 41 +++++++++++++++++++++++++++++++++++++++++
 1 file changed, 41 insertions(+)
 create mode 100644 training/utils/entropy.cc

(limited to 'training/utils/entropy.cc')

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;
+}
+
+}
+
-- 
cgit v1.2.3