From fd577d7a65cab923b9102d61873a032654d573a1 Mon Sep 17 00:00:00 2001
From: Colin Cherry <colin.a.cherry@gmail.com>
Date: Tue, 29 May 2012 13:38:57 -0400
Subject: Batch k-best MIRA is written and integrated into mert-moses.pl
 Regression tests all check out, and kbmira seems to work fine on a Hansard
 French->English task.

HypPackEnumerator class may be of interest to pro.cpp and future
optimizers, as it abstracts a lot of the boilerplate involved in
enumerating multiple k-best lists.

MiraWeightVector is not really mira-specific - just a weight vector
that enables efficient averaging. Could be useful to a perceptron
as well. Same goes for MiraFeatureVector.

Interaction with sparse features is written, but untested.
---
 mert/MiraWeightVector.h | 106 ++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 106 insertions(+)
 create mode 100644 mert/MiraWeightVector.h

(limited to 'mert/MiraWeightVector.h')

diff --git a/mert/MiraWeightVector.h b/mert/MiraWeightVector.h
new file mode 100644
index 000000000..375858634
--- /dev/null
+++ b/mert/MiraWeightVector.h
@@ -0,0 +1,106 @@
+/*
+ * MiraWeightVector.h
+ * kbmira - k-best Batch MIRA
+ *
+ * A self-averaging weight-vector. Good for
+ * perceptron learning as well.
+ * 
+ */
+
+#ifndef MERT_MIRA_WEIGHT_VECTOR_H
+#define MERT_MIRA_WEIGHT_VECTOR_H
+
+#include <vector>
+
+#include "MiraFeatureVector.h"
+
+using namespace std;
+
+class AvgWeightVector;
+
+class MiraWeightVector {
+public:
+  /**
+   * Constructor, initializes to the zero vector
+   */
+  MiraWeightVector();
+
+  /**
+   * Constructor with provided initial vector
+   * \param init Initial feature values
+   */
+  MiraWeightVector(const vector<ValType>& init); 
+
+  /**
+   * Update a the model
+   * \param fv  Feature vector to be added to the weights
+   * \param tau FV will be scaled by this value before update
+   */
+  void update(const MiraFeatureVector& fv, float tau);
+
+  /**
+   * Perform an empty update (affects averaging)
+   */
+  void tick();
+
+  /**
+   * Score a feature vector according to the model
+   * \param fv Feature vector to be scored
+   */
+  ValType score(const MiraFeatureVector& fv) const;
+
+  /**
+   * Squared norm of the weight vector
+   */
+  ValType sqrNorm() const;
+
+  /**
+   * Return an averaged view of this weight vector
+   */
+  AvgWeightVector avg();
+
+  friend class AvgWeightVector;
+  
+private:
+  /**
+   * Updates a weight and lazily updates its total
+   */
+  void update(size_t index, ValType delta);
+
+  /**
+   * Make sure everyone's total is up-to-date
+   */
+  void fixTotals();
+
+  /**
+   * Helper to handle out-of-range weights
+   */
+  ValType weight(size_t index) const;
+  
+  vector<ValType> m_weights;
+  vector<ValType> m_totals;
+  vector<size_t>  m_lastUpdated;
+  size_t          m_numUpdates;
+};
+
+/**
+ * Averaged view of a weight vector
+ */
+class AvgWeightVector {
+public:
+  AvgWeightVector(const MiraWeightVector& wv);
+  ValType score(const MiraFeatureVector& fv) const;
+  ValType weight(size_t index) const;
+  size_t size() const;
+private:
+  const MiraWeightVector& m_wv;
+};
+
+
+#endif // MERT_WEIGHT_VECTOR_H
+
+// --Emacs trickery--
+// Local Variables:
+// mode:c++
+// c-basic-offset:2
+// End:
-- 
cgit v1.2.3