mosesdecoder/moses/FeatureVector.h

/*
 Moses - factored phrase-based language decoder
 Copyright (C) 2010 University of Edinburgh

 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
 the Free Software Foundation; either version 2 of the License, or
 (at your option) any later version.

 This program is distributed in the hope that it will be useful,
 but WITHOUT ANY WARRANTY; without even the implied warranty of
 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 GNU General Public License for more details.

 You should have received a copy of the GNU General Public License along
 with this program; if not, write to the Free Software Foundation, Inc.,
 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.

 */
#pragma once

#ifndef FEATUREVECTOR_H
#define FEATUREVECTOR_H

#include <iostream>
#include <map>
#include <sstream>
#include <string>
#include <valarray>
#include <vector>

#include <boost/functional/hash.hpp>
#include <boost/unordered_map.hpp>

#ifdef MPI_ENABLE
#include <boost/serialization/access.hpp>
#include <boost/serialization/split_member.hpp>
#include <boost/serialization/string.hpp>
#include <boost/serialization/vector.hpp>
#include <boost/serialization/valarray.hpp>
#endif

#ifdef WITH_THREADS
#include <boost/thread/shared_mutex.hpp>
#endif

#include "util/check.hh"
#include "util/string_piece.hh"

namespace Moses
{

typedef float FValue;

/**
 * Feature name
 **/
struct FName {

  static const std::string SEP;

  typedef boost::unordered_map<std::string,size_t> Name2Id;
  typedef boost::unordered_map<size_t,size_t> Id2Count;
  //typedef std::map<std::string, size_t> Name2Id;
  static Name2Id name2id;
  static std::vector<std::string> id2name;
  static Id2Count id2hopeCount;
  static Id2Count id2fearCount;

  //A feature name can either be initialised as a pair of strings,
  //which will be concatenated with a SEP between them, or as
  //a single string, which will be used as-is.
  FName(const StringPiece &root, const StringPiece &name) {
    std::string assembled(root.data(), root.size());
    assembled += SEP;
    assembled.append(name.data(), name.size());
    init(assembled);
  }
  explicit FName(const StringPiece &name) {
    init(name);
  }

  const std::string& name() const;
  //const std::string& root() const {return m_root;}

  size_t hash() const;

  bool operator==(const FName& rhs) const ;
  bool operator!=(const FName& rhs) const ;

  static size_t getId(const std::string& name);
  static size_t getHopeIdCount(const std::string& name);
  static size_t getFearIdCount(const std::string& name);
  static void incrementHopeId(const std::string& name);
  static void incrementFearId(const std::string& name);
  static void eraseId(size_t id);

private:
  void init(const StringPiece& name);
  size_t m_id;
#ifdef WITH_THREADS
  //reader-writer lock
  static boost::shared_mutex m_idLock;
#endif
};

std::ostream& operator<<(std::ostream& out,const FName& name);

struct FNameEquals {
  inline bool operator() (const FName& lhs, const FName& rhs) const {
    return (lhs == rhs);
  }
};

struct FNameHash
    : std::unary_function<FName, std::size_t> {
  std::size_t operator()(const FName& x) const {
    return x.hash();
  }
};

class ProxyFVector;

/**
 * A sparse feature (or weight) vector.
 **/
class FVector
{
public:
  /** Empty feature vector */
  FVector(size_t coreFeatures = 0);

  FVector& operator=( const FVector& rhs ) {
    m_features = rhs.m_features;
    m_coreFeatures = rhs.m_coreFeatures;
    return *this;
  }

  /*
   * Change the number of core features
  **/
  void resize(size_t newsize);

  typedef boost::unordered_map<FName,FValue,FNameHash, FNameEquals> FNVmap;
  /** Iterators */
  typedef FNVmap::iterator iterator;
  typedef FNVmap::const_iterator const_iterator;
  iterator begin() {
    return m_features.begin();
  }
  iterator end() {
    return m_features.end();
  }
  const_iterator cbegin() const {
    return m_features.cbegin();
  }
  const_iterator cend() const {
    return m_features.cend();
  }

  bool hasNonDefaultValue(FName name) const {
    return m_features.find(name) != m_features.end();
  }
  void clear();


  /** Load from file - each line should be 'root[_name] value' */
  bool load(const std::string& filename);
  void save(const std::string& filename) const;
  void write(std::ostream& out) const ;

  /** Element access */
  ProxyFVector operator[](const FName& name);
  FValue& operator[](size_t index);
  FValue operator[](const FName& name) const;
  FValue operator[](size_t index) const;

  /** Size */
  size_t size() const {
    return m_features.size() + m_coreFeatures.size();
  }

  size_t coreSize() const {
    return m_coreFeatures.size();
  }

  const std::valarray<FValue> &getCoreFeatures() const {
    return m_coreFeatures;
  }

  /** Equality */
  bool operator== (const FVector& rhs) const;
  bool operator!= (const FVector& rhs) const;

  FValue inner_product(const FVector& rhs) const;

  friend class ProxyFVector;

  /**arithmetic */
  //Element-wise
  //If one side has fewer core features, take the missing ones to be 0.
  FVector& operator+= (const FVector& rhs);
  FVector& operator-= (const FVector& rhs);
  FVector& operator*= (const FVector& rhs);
  FVector& operator/= (const FVector& rhs);
  //Scalar
  FVector& operator*= (const FValue& rhs);
  FVector& operator/= (const FValue& rhs);

  FVector& multiplyEqualsBackoff(const FVector& rhs, float backoff);
  FVector& multiplyEquals(float core_r0, float sparse_r0);

  FVector& max_equals(const FVector& rhs);

  /** norms and sums */
  FValue l1norm() const;
  FValue l1norm_coreFeatures() const;
  FValue l2norm() const;
  FValue linfnorm() const;
  size_t l1regularize(float lambda);
  void l2regularize(float lambda);
  size_t sparseL1regularize(float lambda);
  void sparseL2regularize(float lambda);
  FValue sum() const;

  /** pretty printing */
  std::ostream& print(std::ostream& out) const;

  /** additional */
  void printCoreFeatures();
  //scale so that abs. value is less than maxvalue
  void thresholdScale(float maxValue );

  void capMax(FValue maxValue);
  void capMin(FValue minValue);

  void sparsePlusEquals(const FVector& rhs);
  void coreAssign(const FVector& rhs);

  void incrementSparseHopeFeatures();
  void incrementSparseFearFeatures();
  void printSparseHopeFeatureCounts(std::ofstream& out);
  void printSparseFearFeatureCounts(std::ofstream& out);
  void printSparseHopeFeatureCounts();
  void printSparseFearFeatureCounts();
  size_t pruneSparseFeatures(size_t threshold);
  size_t pruneZeroWeightFeatures();
  void updateConfidenceCounts(const FVector& weightUpdate, bool signedCounts);
  void updateLearningRates(float decay_core, float decay_sparse, const FVector& confidence_counts, float core_r0, float sparse_r0);

  // vector which, for each element of the original vector, reflects whether an element is zero or non-zero
  void setToBinaryOf(const FVector& rhs);

  // divide only core features by scalar
  FVector& coreDivideEquals(float scalar);

  // divide each element by the number given in the rhs vector
  FVector& divideEquals(const FVector& rhs);

  void merge(const FVector &other);

#ifdef MPI_ENABLE
  friend class boost::serialization::access;
#endif

private:

  /** Internal get and set. */
  const FValue& get(const FName& name) const;
  FValue getBackoff(const FName& name, float backoff) const;
  void set(const FName& name, const FValue& value);

  FNVmap m_features;
  std::valarray<FValue> m_coreFeatures;

#ifdef MPI_ENABLE
  //serialization
  template<class Archive>
  void save(Archive &ar, const unsigned int version) const {
    std::vector<std::string> names;
    std::vector<FValue> values;
    for (const_iterator i = cbegin(); i != cend(); ++i) {
      std::ostringstream ostr;
      ostr << i->first;
      names.push_back(ostr.str());
      values.push_back(i->second);
    }
    ar << names;
    ar << values;
    ar << m_coreFeatures;
  }

  template<class Archive>
  void load(Archive &ar, const unsigned int version) {
    clear();
    std::vector<std::string> names;
    std::vector<FValue> values;
    ar >> names;
    ar >> values;
    ar >> m_coreFeatures;
    CHECK(names.size() == values.size());
    for (size_t i = 0; i < names.size(); ++i) {
      set(FName(names[i]), values[i]);
    }
  }

  BOOST_SERIALIZATION_SPLIT_MEMBER()

#endif

};

std::ostream& operator<<( std::ostream& out, const FVector& fv);
//Element-wise operations
const FVector operator+(const FVector& lhs, const FVector& rhs);
const FVector operator-(const FVector& lhs, const FVector& rhs);
const FVector operator*(const FVector& lhs, const FVector& rhs);
const FVector operator/(const FVector& lhs, const FVector& rhs);

//Scalar operations
const FVector operator*(const FVector& lhs, const FValue& rhs);
const FVector operator/(const FVector& lhs, const FValue& rhs);

const FVector fvmax(const FVector& lhs, const FVector& rhs);

FValue inner_product(const FVector& lhs, const FVector& rhs);

struct FVectorPlus {
  FVector operator()(const FVector& lhs, const FVector& rhs) const {
    return lhs + rhs;
  }
};

/**
 * Used to help with subscript operator overloading.
 * See http://stackoverflow.com/questions/1386075/overloading-operator-for-a-sparse-vector
 **/
class ProxyFVector
{
public:
  ProxyFVector(FVector *fv, const FName& name ) : m_fv(fv), m_name(name) {}
  ProxyFVector &operator=(const FValue& value) {
    // If we get here, we know that operator[] was called to perform a write access,
    // so we can insert an item in the vector if needed
    //std::cerr << "Inserting " << value << " into " << m_name << std::endl;
    m_fv->set(m_name,value);
    return *this;

  }

  operator FValue() {
    // If we get here, we know that operator[] was called to perform a read access,
    // so we can simply return the value from the vector
    return m_fv->get(m_name);
  }

  /*operator FValue&() {
   return m_fv->m_features[m_name];
   }*/

  FValue operator++() {
    return ++m_fv->m_features[m_name];
  }

  FValue operator +=(FValue lhs) {
    return (m_fv->m_features[m_name] += lhs);
  }

  FValue operator -=(FValue lhs) {
    return (m_fv->m_features[m_name] -= lhs);
  }

private:
  FValue m_tmp;

private:
  FVector* m_fv;
  const FName& m_name;

};

}

#endif